# Landing-Ellipse Contraction as a Technology Learning Curve: Quantifying the Precision-Landing Improvement Rate Across Mars Missions

**A Design-Stage Doctoral Dissertation**

| | |
|---|---|
| **Candidate** | JPL_AUTONOMY_EDL_06 |
| **Program** | COLLEGIUM 1st Battalion, PhD-candidate cohort |
| **Category** | NORTH STAR / JPL, Entry, Descent, and Landing Systems |
| **Hall-of-Shoulders anchors** | Joel Mokyr (economic history of technology); Simon Kuznets (cliometric measurement) |
| **Date** | 2026-06-15 |



## Abstract

The improvement of Mars landing precision across half a century of robotic exploration is the work of many hands, a record held in trust by successive generations of mission engineers and the institutions that sustain them. This dissertation seeks to honor that inheritance by measuring it honestly, in the service of the missions still to come. The precision with which spacecraft can be delivered to the Martian surface has improved by more than two orders of magnitude across the robotic exploration era. The Viking landers of 1976 were targeted to landing ellipses on the order of hundreds of kilometers in their major dimension; the Mars 2020 Perseverance rover was delivered inside an effective targeting region of a few kilometers. This dissertation asks whether that improvement follows a measurable learning curve, the empirical regularity from the economics of technology in which a performance metric improves at a roughly constant proportional rate as a function of cumulative experience or successive technology generations. The contribution is a single falsifiable proposition: the three-sigma landing-ellipse semi-major axis, equivalently the ellipse area, for Mars surface missions contracts along an exponential learning rate driven primarily by the insertion of onboard entry-descent-and-landing (EDL) guidance technologies, specifically guided (lifting) entry, the range-to-go parachute trigger, and terrain-relative navigation (TRN), rather than by improvements in launch-vehicle injection accuracy. The null hypothesis is that ellipse contraction is unrelated to onboard EDL-guidance technology generation.

The method is a log-linear learning-curve regression of landing-ellipse area on mission sequence and a set of EDL-guidance technology covariates, estimated across the United States-led Mars surface missions from Viking 1 through Mars 2020, with supporting reconstruction data drawn from the NASA Technical Reports Server, the Planetary Data System, and TechPort. Following Kuznets, the work treats the landing-ellipse series as a constructed measurement whose definitional boundary, valuation convention, and decomposition must be stated before any inference is drawn. Following Mokyr, it frames the technology covariates as discrete additions to the prescriptive knowledge base that make the landing technique self-correcting and extensible. Identification does not rest on asymptotic significance, which a sample of nine to eleven missions cannot support, but on two design features the historical record happens to contain: the InSight counterfactual, a late mission that deliberately flew an unguided ballistic entry into a large ellipse and so decouples technology generation from calendar date, and the approach-accuracy control, which holds constant the part of delivery accuracy attributable to launch and interplanetary navigation.

This document is a design-stage dissertation. The model, identification strategy, and pre-registered analysis plan are complete, and the data sources are named and accessible, but the regression has not yet been executed on the assembled dataset. Every numerical result shown is illustrative and labeled as such; no estimate is reported as an empirical finding. The strongest claim the executed design can support is a signed, order-of-magnitude, counterfactual-surviving attribution rather than a precise point estimate of a learning rate, and the design is built to be informative whichever way the central coefficient falls. The work matters to NASA and JPL because a defensible learning rate, attributed to specific technologies, converts qualitative claims about EDL maturity into a quantitative basis for setting landing-accuracy requirements and for valuing candidate guidance investments in the human-Mars and Mars Sample Return architectures now under study.



## Table of Contents

- **Abstract**
- **Chapter 1: Introduction**
  - 1.1 The chapter thesis
  - 1.2 The problem in full
  - 1.3 Institutional and historical context
  - 1.4 The research questions, broken out
  - 1.5 The falsifiable contribution stated as H0 and H1
  - 1.6 Significance for NASA, JPL, and named stakeholders
  - 1.7 Scope and delimitations
  - 1.8 Definitions of key terms
  - 1.9 The argument spine: why the design follows from the problem
  - 1.10 Roadmap of the dissertation
- **Chapter 2: Theoretical Framework**
  - 2.1 The chapter's answer, and the problem it solves
  - 2.2 The learning curve as an empirical regularity, and its limits
  - 2.3 Mokyr in depth: propositional knowledge, prescriptive technique, and the three levers
  - 2.4 Kuznets in depth: boundary, valuation, netting, and the transient-secular distinction
  - 2.5 From firm learning curves to a generation-indexed aerospace series
  - 2.6 The conceptual model the empirical work will test
  - 2.7 Confidence, and what would move it
- **Chapter 3: Literature Review**
  - 3.1 The chapter thesis and the shape of the gap
  - 3.2 Guided lifting entry and the entry-guidance algorithms
  - 3.3 The range-to-go parachute trigger and footprint reduction
  - 3.4 Terrain-relative navigation and the Lander Vision System
  - 3.5 Hazard detection and avoidance and precision-landing capability programs
  - 3.6 Supersonic retropropulsion and high-mass, human-class EDL
  - 3.7 Mission-context studies, instruments, and the public record
  - 3.8 Synthesis: levers to error sources, and what each literature contributes
  - 3.9 The gap and the propositions that follow
- **Chapter 4: Data and Measurement**
  - 4.1 The chapter thesis
  - 4.2 The three named datasets in depth
  - 4.3 Unit of analysis and the operationalization of every variable
  - 4.4 Data quality, validation, and coverage limitations
  - 4.5 Ethics, access, and reproducibility
  - 4.6 What this chapter establishes for the argument
- **Chapter 5: Research Design and Identification**
  - 5.1 The chapter thesis
  - 5.2 The problem this chapter addresses
  - 5.3 The estimator and why it is chosen
  - 5.4 The specifications written out
  - 5.5 The nested-specification hierarchy
  - 5.6 Identification strategy
  - 5.7 Threats to validity
  - 5.8 The robustness battery
  - 5.9 Power and minimum-detectable-effect analysis
  - 5.10 The pre-registration commitment
  - 5.11 The computational and software plan
  - 5.12 The dispersion-analysis precedent for constructing the series
  - 5.13 Chapter summary and how it advances the argument
- **Chapter 6: Analysis Plan and Expected Results**
  - 6.1 The chapter thesis
  - 6.2 The pre-registered estimation procedure
  - 6.3 The fixed decision rule
  - 6.4 Expected signs and their mechanisms
  - 6.5 The design of the illustrative simulation
  - 6.6 The event-study and profile interpretation of the InSight residual
  - 6.7 The robustness battery in detail
  - 6.8 The Kuznetsian transient-versus-secular reading
  - 6.9 Reporting discipline and provenance
  - 6.10 How the analysis plan advances the argument
  - 6.11 Summary of the chapter's commitments
- **Chapter 7: Discussion**
  - 7.1 The chapter thesis
  - 7.2 The problem this chapter addresses
  - 7.3 Implications if H1 holds
  - 7.4 Implications if H0 holds
  - 7.5 The theoretical contribution back to each anchor framework
  - 7.6 Policy and mission implications for NASA, JPL, and stakeholders
  - 7.7 Full engagement with rival explanations
  - 7.8 External-validity statement
  - 7.9 Confidence summary and the evidence that would move it
- **Chapter 8: Conclusion**
  - 8.1 The chapter thesis
  - 8.2 Restatement of the contribution
  - 8.3 What holds even if the hypothesis is not confirmed
  - 8.4 Honest limitations
  - 8.5 A concrete future-research program
  - 8.6 How the dissertation's argument stands at its conclusion
  - 8.7 Closing
- **References**
- **Appendix A: Variable and Data Dictionary**
- **Appendix B: Derivations**
- **Appendix C: Instrument and Query Details**
- **Appendix D: Supplementary Tables**



## List of Tables and Figures

**Tables**

- Table 3.1. Technology levers, the error source each attacks, and the primary literature (Chapter 3).
- Table 3.2. Literature themes, contribution, and gap (Chapter 3).
- Table 4.1. Measurement table: operationalization of every variable (Chapter 4, Section 4.3.3).
- Table 5.1. Nested-specification template, illustrative (Chapter 5, Section 5.5).
- Table 6.1. Illustrative coefficient table, nested hierarchy (Chapter 6, Section 6.5).
- Table D.1. Mission-by-mission technology-indicator coding sheet (Appendix D).
- Table D.2. Nested-specification results template, unpopulated by design (Appendix D).

**Figures**

- Figure (primary, specified): log ellipse area against mission generation, with the fitted line and the InSight residual marked. Specified in Chapters 5 and 6; unpopulated by design at the design stage.

The dissertation is a design-stage document: every results table is specified but, by design, unpopulated, and every illustrative value is labeled as such.



# Chapter 1: Introduction

## 1.1 The chapter thesis

The shrinking of the Mars landing ellipse is, before it is a research question, a record of decades of disciplined service by the engineers and institutions who carried each mission to its surface. This dissertation takes up that record in a spirit of stewardship, seeking to measure faithfully what others built so that the capability may be passed forward to the missions now under preparation. It converts a qualitative, mission-by-mission narrative of shrinking Mars landing ellipses into a single falsifiable, technology-attributed measurement claim. The claim is that the three-sigma landing-ellipse semi-major axis (equivalently, ellipse area) for Mars surface missions has contracted along an exponential, roughly log-linear learning curve whose discrete steps align with identifiable onboard entry-descent-and-landing (EDL) guidance insertions, specifically guided lifting entry, the range-to-go parachute trigger, and terrain-relative navigation (TRN), rather than with launch-vehicle and interplanetary injection accuracy once the standard approach-navigation corrections are accounted for. Stated as a competing pair, this is hypothesis H1 against the null H0 that ellipse contraction is unrelated to onboard EDL-guidance technology generation. The contribution is neither the observation that landings have grown more precise, which is well documented [\[1\]](#ref-1), [\[5\]](#ref-5), [\[8\]](#ref-8), nor a new guidance algorithm. It is the construction of the comparison the existing literature has never set up: one cross-mission ellipse series, treated as a constructed measurement in the cliometric sense, fitted with a learning-rate model, and used to arbitrate between the onboard-guidance explanation and its principal rival. Everything that follows in this chapter develops, defends, and bounds that thesis. The remaining chapters specify the framework that generates it (Chapter 2), the literature that establishes its gap (Chapter 3), the data that operationalize it (Chapter 4), the design that identifies it (Chapter 5), the pre-registered plan that will test it (Chapter 6), the implications under each outcome (Chapter 7), and the synthesis that survives whether or not the hypothesis is confirmed (Chapter 8).

A word on register, and on what kind of dissertation this is. It is a design-stage dissertation. The model, the identification strategy, the named data sources, and the analysis plan are complete and the data are accessible, but the regression has not been executed on the assembled dataset. Wherever a number appears in service of the argument below, it is drawn from the published record of a specific mission and cited, or it is labeled illustrative. No estimate is reported as an empirical finding, because none has yet been produced. This honesty is not a hedge; it is the point. The strongest defensible product of the work, once executed, is a signed, order-of-magnitude, counterfactual-surviving attribution, and the design is built to deliver exactly that and no more.

## 1.2 The problem in full

### 1.2.1 The ellipse as the most consequential number in mission design
Every Mars surface mission is committed, years before launch, to a landing ellipse: a probabilistic region on the surface, conventionally reported at three standard deviations, inside which the vehicle is expected to touch down. Calling this the most consequential number in mission design is not rhetorical inflation. The size of the ellipse propagates into nearly every downstream decision a project makes. It determines which candidate sites can even be considered, because a site is admissible only if the entire ellipse footprint placed over it is certifiably safe and trafficable. It determines how much of the scientifically interesting terrain must be sacrificed to that safety constraint, because the richest geology, a delta front, a crater rim, a layered outcrop, an ancient shoreline, is also the most hazardous and the most heterogeneous, and a large ellipse cannot be guaranteed to avoid the hazards it contains. It determines the mass and complexity budget allocated to the EDL system itself, because tightening the ellipse is bought with guidance hardware, software, and test campaigns that compete for resources against the science payload. It determines the residual risk the project carries to its review boards, because the ellipse is the formal statement of how confidently the agency believes it can place a vehicle where it intends.

The history of Mars surface exploration is, in large part, the history of shrinking that ellipse, and the magnitude of the shrinkage is the first fact the dissertation must establish, because the entire enterprise is unmotivated if the contraction is small or ambiguous. It is neither. The Viking landers of 1976 were targeted to regions whose major dimension was on the order of hundreds of kilometers, a consequence of ballistic entry in which a capsule, once released to the atmosphere, simply follows wherever its delivered state and the atmosphere carry it. The Mars Exploration Rovers in 2004, still flying an essentially ballistic, Pathfinder-heritage entry, landed inside ellipses on the order of a hundred kilometers in the major dimension. The Mars Science Laboratory in 2012 introduced guided lifting entry and a range-triggered parachute deploy and cut the targeting region to roughly twenty kilometers [\[5\]](#ref-5). The Mars 2020 mission added terrain-relative navigation and an autonomous divert, effectively delivering Perseverance to a targeting region of a few kilometers inside the hazardous Jezero crater site, a site that earlier systems could not have attempted at all [\[8\]](#ref-8), [\[5\]](#ref-5). Across the era, the targeting capability improved by more than two orders of magnitude. That is the current state: a large, real, well-attested contraction, narrated repeatedly in the EDL literature mission by mission.

### 1.2.2 Current state, desired state, gap, consequence

The problem this dissertation addresses is not the existence of the contraction but its causal ambiguity. It is useful to frame that problem precisely as current state, desired state, gap, and consequence, because the four together define the contribution.

The **current state** is that the cross-mission ellipse contraction is documented qualitatively, one mission at a time, with no joint time-series model and no formal attribution of the contraction to specific technologies against rival causes [\[1\]](#ref-1), [\[5\]](#ref-5), [\[8\]](#ref-8). Each engineering paper does its job within its scope, reporting the design ellipse and the reconstructed performance for its own mission. None was ever designed to arbitrate causes across missions, because that was not its purpose.

The **desired state** is a single constructed ellipse series, fitted with a learning-rate model, that attributes the contraction to identifiable technology insertions while controlling for the alternative explanation that the vehicles are simply being delivered to the top of the atmosphere more accurately. In the desired state, a project manager facing a future architecture can read off the literature an answer to a concrete question: how much of the achievable landing precision is bought by onboard guidance and how much by approach navigation, and therefore where the marginal investment dollar buys the most precision.

The **gap** is that no published study joins the EDL-engineering literature to the technology-economics learning-curve apparatus [\[18\]](#ref-18), [\[19\]](#ref-19), and none arbitrates among the rival causes of the contraction. The two literatures, the rich and specific EDL engineering canon and the mature learning-curve and cliometric-measurement tradition, have never been brought into contact. The contraction sits in one, and the apparatus that could quantify and attribute it sits in the other.

The **consequence of inaction** is concrete and forward-looking. Future architectures, the Mars Sample Return retrieval and the eventual human landing, must each specify a landing-accuracy requirement years in advance, and they currently do so by qualitative negotiation rather than by reading a defensible rate off an attributed curve [\[112\]](#ref-112), [\[113\]](#ref-113), [\[122\]](#ref-122). They also cannot value a guidance investment against an approach-navigation investment on a common quantitative basis, because no such basis exists. The cost of leaving the gap open is therefore not academic; it is paid in less-informed requirements and less-disciplined investment decisions on the most expensive and most scrutinized landings NASA and JPL will ever attempt.

### 1.2.3 Why the qualitative story is ambiguous about cause

The reason the quantitative story matters is that the qualitative one cannot settle cause, and the mechanism of that ambiguity is worth stating because it is the seed of the whole research design. Three things changed together, and monotonically, across the robotic era. The vehicles acquired more capable onboard guidance, generation by generation. The agency learned to model the entry environment, the atmosphere, the aerodynamics, the parachute, with steadily greater fidelity. And the approach navigation that delivers the vehicle to the atmospheric interface improved with better tracking and better delivery. Any one of these could, in principle, explain a shrinking ellipse. Because the engineering papers each examine a single mission, they hold none of the others fixed, and so they cannot, even in aggregate, separate the three. A decision-maker who reads the entire EDL literature still cannot answer whether buying terrain-relative navigation for a future mission will shrink its ellipse, or whether the same money spent on approach navigation would do as well. The driver here is a structural feature of how the literature is organized; the mechanism is the absence of any cross-mission model that holds rival causes constant; the observable effect is a body of work that narrates the contraction without attributing it; the operational consequence is qualitative requirement-setting; and the strategic implication is that landing precision remains an un-valued design variable. Setting up the comparison the literature has not, in a form where the data can settle it, is the contribution.

## 1.3 Institutional and historical context

### 1.3.1 The robotic-era arc, Viking through Mars 2020

The population this dissertation studies is not an abstraction; it is the concrete sequence of United States-led Mars surface missions that successfully entered, descended, and landed, and a brief institutional history of that sequence locates the technology insertions the work treats as covariates. Viking 1 and Viking 2 (1976) established the ballistic-entry, large-ellipse baseline. Mars Pathfinder (1997) revived surface landing after a long hiatus with an airbag-cushioned, essentially ballistic architecture, and the Mars Exploration Rovers Spirit and Opportunity (2004) inherited and refined that architecture, still without closed-loop entry guidance. Phoenix (2008) flew a Viking-heritage, ballistic, instrumented entry to a high-latitude plain, again accepting a large ellipse because its science did not demand precision [\[6\]](#ref-6). The Mars Science Laboratory, Curiosity (2012), was the inflection: it introduced the first guided lifting entry at Mars together with a range-triggered parachute deploy and the sky-crane terminal architecture, and it cut the ellipse by roughly an order of magnitude relative to its ballistic predecessors [\[1\]](#ref-1), [\[2\]](#ref-2), [\[5\]](#ref-5). InSight (2018) is the decisive counter-current in this arc: a modern, late-era lander that deliberately reverted to a Phoenix-heritage ballistic entry into a large, flat ellipse in Elysium Planitia, because its seismology mission did not require precision and its budget did not justify the guidance suite [\[15\]](#ref-15), [\[14\]](#ref-14). Mars 2020, Perseverance (2021), completed the arc by adding terrain-relative navigation, the Lander Vision System, and Safe Target Selection to the MSL architecture, enabling an autonomous divert that placed the rover inside a few-kilometer region in a site that the prior generation's ellipse would have ruled inadmissible [\[8\]](#ref-8), [\[121\]](#ref-121). Tianwen-1 (2021) landed in the same year under a different agency and is held out as an external-validity reference rather than treated as in-sample.

This arc matters institutionally because the technology insertions did not arrive smoothly; they arrived as discrete program decisions, each tied to a specific mission and a specific propositional base that had matured enough to support it. Guided entry rested on decades of entry-aerodynamics and atmosphere-reconstruction knowledge before it was trusted to fly [\[3\]](#ref-3), [\[4\]](#ref-4). The range trigger was a comparatively modest refinement of parachute-deploy logic [\[5\]](#ref-5). Terrain-relative navigation rested on a long maturation of onboard computer vision and orbital-imagery mapping, including a dedicated test campaign in a relevant environment before it flew [\[8\]](#ref-8), [\[9\]](#ref-9), [\[121\]](#ref-121). The discreteness of these insertions, mission by mission, is what makes them codable as binary covariates and what gives the design its step structure.

### 1.3.2 NASA and JPL as the institutional setting

The institutional setting is NASA's robotic Mars program, executed principally at the Jet Propulsion Laboratory, and that setting shapes both the data and the stakes. JPL builds and operates the landers, holds the EDL engineering heritage tacitly in its teams and codified in its flight-software and test infrastructure, and produces the reconstruction studies that are this dissertation's primary data source [\[72\]](#ref-72), [\[74\]](#ref-74). The agency's forward architectures, Mars Sample Return and human Mars, are studied across multiple NASA centers and sponsored across mission directorates [\[112\]](#ref-112), [\[113\]](#ref-113), [\[122\]](#ref-122), and they are the consumers of any defensible learning rate. The work is therefore not a detached academic exercise; it is addressed to a specific institutional decision community that already owns the data, already faces the requirement-setting problem, and already lacks the quantitative basis the dissertation proposes to supply. That alignment between the analytic question and an existing institutional need is part of what makes the contribution material rather than merely interesting.

## 1.4 The research questions, broken out

The thesis decomposes into a primary research question and a set of subsidiary questions that the chapters answer in turn. Stating them explicitly disciplines the rest of the dissertation, because each subsequent chapter can be checked against the question it is meant to advance.

**Primary research question.** Does the three-sigma Mars landing-ellipse series contract along a measurable learning curve, and is that contraction driven primarily by onboard EDL-guidance technology insertions (guided lifting entry, the range-to-go parachute trigger, terrain-relative navigation) rather than by launch-vehicle and interplanetary injection accuracy?

The primary question has two clauses, and the design must answer both. The first clause is descriptive and measurement-bound: is there a measurable learning curve at all, and what is its form once the series is constructed with the comparability discipline the data demand? The second clause is causal and attributive: conditional on a contraction existing, does it load on the onboard-guidance covariates rather than on the approach-accuracy control? The dissertation's contribution lives almost entirely in the second clause, because the first clause is, qualitatively, not in dispute. What is in dispute, and what the data can settle, is the attribution.

Four subsidiary questions follow.

**RQ1 (measurement construction).** Can a comparable cross-mission three-sigma ellipse series be constructed from the public record, given that the design ellipse is a simulation product whose conventions changed across missions? This is the Kuznetsian comparability question, and it is logically prior to any inference; Chapter 4 answers it.

**RQ2 (functional form and rate).** Conditional on a comparable series, does the contraction take an exponential, log-linear form, and what is the proportional contraction per mission generation, read as \(\exp(\beta_1) - 1\) from the baseline specification? Chapters 5 and 6 specify and plan to answer this.

**RQ3 (technology attribution).** Do the technology covariates \(\text{GuidedEntry}_i\), \(\text{RangeTrigger}_i\), and \(\text{TRN}_i\) carry jointly significant, negative explanatory power for ellipse area once mission sequence and the approach-accuracy control are present, and do the largest fit increments fall at the guided-entry and TRN transitions as the macro-versus-incremental reading predicts? This is the core attributive question and the locus of H1 against H0.

**RQ4 (rival arbitration).** Does the approach-accuracy control delta absorb the technology terms (favoring H0), and does the InSight counterfactual sit on or above the fitted trend (discriminating the technology hypothesis from the pure-time hypothesis)? Chapter 5 sets up the identification that answers RQ4, and Chapter 6 commits in advance to the decision rule that reads it.
These four questions are not independent. RQ1 conditions everything downstream, since a series that fails the comparability test cannot support RQ2 through RQ4. RQ4 is the sharp end: the contribution survives or fails on whether the technology attribution withstands the two controls. The questions are sequenced deliberately, so that the dissertation cannot smuggle a causal claim past an unmet measurement precondition.

## 1.5 The falsifiable contribution stated as H0 and H1

The contribution is one testable proposition, stated verbatim as the competing hypotheses fixed at the design stage. These are reproduced exactly and are not to be paraphrased anywhere in the dissertation.

- **H1 (the contribution):** The three-sigma landing-ellipse semi-major axis (equivalently, ellipse area) for Mars surface missions declines along an exponential learning curve in which the dominant explanatory variables are onboard EDL-guidance technology generations (guided entry, range trigger, and terrain-relative navigation), and in which launch-vehicle and interplanetary injection accuracy, once the standard approach-navigation corrections are accounted for, is not the binding constraint on ellipse size.

- **H0 (the null):** Ellipse contraction is unrelated to onboard EDL-guidance technology generation. Under H0 the technology covariates carry no explanatory power once mission sequence or a time trend is included, and any apparent learning curve is either an artifact of a generic time trend or is driven by approach-navigation improvements rather than onboard guidance.

The proposition is falsifiable in the strict sense the program demands, and the reason deserves to be explicit, because falsifiability is the dividing line between this dissertation and the qualitative narrative it replaces. H1 predicts a specific sign and a specific ordering of effects. It predicts that the gamma coefficients on the three technology indicators are jointly significant and negative, that the delta coefficient on the approach-accuracy control is small and insignificant, and, in the sharper Mokyrian reading, that the largest fit increments arrive at the guided-entry and TRN transitions with a smaller increment at the range trigger. A regression on the assembled data can return coefficients that contradict each of these predictions. If the technology-indicator coefficients are insignificant or wrongly signed, or if the approach-accuracy proxy absorbs their explanatory power, or if the InSight observation sits on rather than above the trend, H1 is rejected and H0 is not. The design is built so that the contradiction, if it comes, is the headline finding rather than a buried negative result. This is what it means to convert a narrative into a hypothesis: not to assert the narrative more forcefully, but to expose it to a test it can fail.

What the contribution is not must also be said plainly. It is not a claim that the strong form of the learning curve, a constant percentage improvement per doubling of a large cumulative production count, holds for Mars landings. Mars landings violate the conditions under which that strong form is reliable: there are fewer than a dozen of them, and each is a bespoke vehicle rather than a unit off a production line [\[18\]](#ref-18), [\[19\]](#ref-19). The defensible claim is weaker and is stated as such throughout. Ordering the missions by guidance generation produces a monotone, roughly log-linear contraction whose steps align with identifiable technology insertions. Whether to call that a learning curve or a technology-generation effect is partly semantic; the falsifiable content lies in the alignment of the steps with the technologies, which the data can confirm or deny regardless of the label.

## 1.6 Significance for NASA, JPL, and named stakeholders

The significance of the work, conditional on its execution, runs along three lines, and each connects the analytic result to a decision a named stakeholder actually makes. The confidence attached to these significance claims is moderate at the design stage. The decision relevance is high if H1 survives, but the magnitude of any single technology's effect is bounded by the small sample, so the value to stakeholders lies in the signed attribution and the disciplined boundary, not in a precise rate.

**Requirements-setting.** Future architectures must specify a landing-accuracy requirement years in advance of any hardware. The Human Mars EDL Architecture Study and the human-class pathfinder assessments make this need explicit: they exist precisely to prioritize technology investments and to characterize where the binding risks lie [\[112\]](#ref-112), [\[113\]](#ref-113), [\[122\]](#ref-122). A quantified learning rate, attributed to specific technologies, lets a project reason about what accuracy is achievable with a given guidance suite rather than negotiating the requirement qualitatively against engineering intuition. The mechanism is direct. A coefficient on a technology indicator is, under H1, an estimate of the proportional ellipse reduction that flying that technology buys, which is exactly the input a requirement-setting trade needs and currently lacks.

**Investment valuation.** If terrain-relative navigation, rather than launch and approach accuracy, is the dominant lever on landing precision, then continued investment in onboard sensing and divert capability has a higher marginal return on precision than equivalent investment in approach navigation, and the augmented model makes that comparison explicit through the contrast between the gamma coefficients and delta. This matters because the two kinds of investment compete for the same constrained technology-development budget across NASA centers, and the supersonic-retropropulsion and high-mass EDL literature shows that the forward edge of the field, where human-class landing lives, will demand hard prioritization among candidate technologies [\[33\]](#ref-33), [\[111\]](#ref-111), [\[122\]](#ref-122). A model that values guidance against approach accuracy on a common quantitative basis is a tool for that prioritization.

**Honest extrapolation.** Following Kuznets, the work is built to separate a genuine secular improvement trend from a one-time level shift produced by a single mission. NASA forecasting culture is prone to extrapolating a step change as if it were a trend, and the consequence of that error is a future requirement set on an unearned expectation. The dissertation is designed to test which the contraction is, and the InSight counterfactual is the instrument that does so. The stakeholder value here is defensive: it protects a future project from reading a precision off a curve that the underlying knowledge chain may not, in fact, support.

The named stakeholder community is concrete. The robotic Mars program at JPL owns the reconstruction data and faces the requirement-setting problem on every mission. The Mars Sample Return campaign, whose retrieval lander must place itself with enough precision to recover a cached sample, is a direct consumer of any landing-precision rate [\[127\]](#ref-127). The human Mars architecture teams, sponsored across the Space Technology, Science, and Human Exploration directorates, are the explicit audience for technology-investment prioritization [\[112\]](#ref-112), [\[122\]](#ref-122). For each, the contribution converts a number that is currently argued into a number that can be estimated and bounded.

## 1.7 Scope and delimitations

The scope is drawn deliberately and narrowly, and stating the delimitations precisely is part of the Kuznetsian discipline the work imposes on itself, because an aggregate is meaningless without a stated boundary of coverage [\[24\]](#ref-24).

**Coverage boundary.** The population is the set of United States-led Mars surface missions that successfully entered, descended, and landed: Viking 1, Viking 2, Mars Pathfinder, MER Spirit, MER Opportunity, Phoenix, MSL/Curiosity, InSight, and Mars 2020/Perseverance. This is the entire relevant population rather than a sample of it, so there is no sampling-frame bias, but there is irreducible small-n, nine to eleven events depending on the inclusion rule for the near-identical Viking and MER pairs.

**Held-out case.** Tianwen-1 (2021) is excluded from estimation and held out as an external-validity reference [\[16\]](#ref-16). It is a successful, contemporaneous, non-US Mars landing, and treating it as out-of-sample lets the dissertation discuss generalization to another agency's program without contaminating the in-sample fit with a vehicle whose reconstruction conventions and reporting are not directly comparable.

**Body and class delimitation.** The fit is on robotic Mars landings only. Lunar precision-landing navigation and human-class EDL appear in the dissertation as forward and out-of-sample reference points, not as in-sample evidence [\[17\]](#ref-17), [\[112\]](#ref-112), [\[113\]](#ref-113), [\[122\]](#ref-122). The work does not claim that a rate fitted on robotic Mars landings transfers mechanically to the Moon or to human-scale vehicles; it claims only what it estimates, on the population it estimates it from.

**Outcome delimitation.** The dependent variable is landing precision as captured by the three-sigma design ellipse, with achieved miss distance from post-landing localization as a parallel construct-validity check. The work does not address landing safety, hazard density at sites, or scientific yield except insofar as the ellipse governs which sites are reachable; those are downstream of the precision metric and outside the boundary.

**Methodological delimitation.** The single most important delimitation, carried forward explicitly from the design, is that the work does not produce a systems or capability architecture. The contribution is an econometric and cliometric measurement-and-attribution claim about a constructed performance series, not the design of a real capability, system, or data or service exchange. The dissertation therefore does not, and must not, force capability-architecture vocabulary onto the argument; the decision relevance described in Section 1.6 is carried in plain prose, not as a traceability artifact. This is a deliberate scope decision, not an omission.

## 1.8 Definitions of key terms

Because the inference depends on the precise meaning of each construct, the key terms are defined here and used identically throughout the dissertation. The variable notation is fixed and is not to be altered in any chapter.

**Landing ellipse.** The probabilistic region on the planetary surface, projected from the EDL performance simulation, inside which the vehicle is expected to touch down at a stated confidence level. Reported as a semi-major axis a and a semi-minor axis b.

**Three-sigma.** The confidence-level convention used throughout, denoting the region that captures the touchdown location with the dispersion taken to three standard deviations. The sigma level is a definitional boundary in the Kuznetsian sense: an ellipse value is meaningless without it, and every reported value in the dissertation carries its sigma level explicitly.

**Ellipse area and the dependent variable.** The primary dependent variable is \(\ln(\text{EllipseArea}_i)\), the natural log of the three-sigma landing-ellipse area, computed as \(\ln(\pi \cdot a \cdot b)\), using the design ellipse reported in the mission's EDL performance study. The secondary dependent variable is ln(achieved miss distance), the log of the distance between the targeted aim point and the actual landing location from post-landing localization.
**Guided (lifting) entry.** Closed-loop control of a lifting entry vehicle, through bank-angle modulation, to null the downrange and crossrange dispersion accumulated during the hypersonic phase. First flown at Mars by MSL [\[3\]](#ref-3), [\[4\]](#ref-4). Coded by the binary indicator \(\text{GuidedEntry}_i\), set to one for MSL and Mars 2020.

**Range trigger (range-to-go parachute trigger).** Parachute-deploy logic that commands deploy based on navigated range-to-go rather than navigated velocity, reducing the parachute-deploy contribution to the footprint. Documented for MSL [\[5\]](#ref-5). Coded by the binary indicator \(\text{RangeTrigger}_i\), set to one for MSL and Mars 2020.

**Terrain-relative navigation (TRN).** Onboard estimation of the vehicle's position by matching descent imagery to a georeferenced onboard reference map, collapsing the position-knowledge error and enabling an autonomous divert away from hazards. First flown at Mars by Mars 2020 via the Lander Vision System [\[8\]](#ref-8), [\[9\]](#ref-9). Coded by the binary indicator \(\text{TRN}_i\), set to one for Mars 2020 only.

**Experience axis.** \(\text{Sequence}_i\), the mission sequence index running from one (Viking 1) through the index of Mars 2020, with a cumulative-landings count as an alternative specification. This is the ordinal generation measure along which the learning curve is read.

**Approach accuracy (the control).** \(\text{ApproachAccuracy}_i\), the reported approach-navigation delivery accuracy, expressed as entry-flight-path-angle or entry-point delivery dispersion, capturing how accurately the vehicle was delivered to the atmospheric interface independent of onboard EDL guidance. This control is what separates the onboard-guidance explanation (H1) from the injection-accuracy explanation (H0).

**Baseline and augmented specifications.** The baseline is \(\ln(\text{EllipseArea}_i) = \beta_0 + \beta_1 \cdot \text{Sequence}_i + \epsilon_i\). The augmented specification is \(\ln(\text{EllipseArea}_i)\) = \(\beta_0 + \beta_1 \cdot \text{Sequence}_i + \gamma_1 \cdot \text{GuidedEntry}_i + \gamma_2 \cdot \text{RangeTrigger}_i + \gamma_3 \cdot \text{TRN}_i + \delta \cdot \text{ApproachAccuracy}_i + \epsilon_i\). The learning-rate reading is \(\exp(\beta_1) - 1\), the proportional change in ellipse area per unit mission sequence.

**Learning curve / experience curve.** The empirical regularity in which a performance or cost metric improves at a roughly constant proportional rate as a function of cumulative experience or successive technology generations [\[18\]](#ref-18), [\[19\]](#ref-19). Adopted here in the weaker, generation-indexed form appropriate to a small number of bespoke, technology-differentiated events.

**Propositional and prescriptive knowledge.** From Mokyr [\[23\]](#ref-23): propositional knowledge is the understanding of why a technique works; prescriptive knowledge is the technique itself. The three EDL levers are treated as discrete additions to the prescriptive landing-technique base, each resting on a maturing propositional base.

## 1.9 The argument spine: why the design follows from the problem

Before the roadmap, the chain of reasoning that takes the problem to the design deserves to be stated plainly, since the design is not an arbitrary methodological choice but a direct consequence of the problem's structure. That reasoning runs through five steps, each defended in turn by the rest of the dissertation.

The first step is that the contraction is real. The two-order-of-magnitude reduction in the ellipse is documented across a coherent mission set from Viking through Mars 2020 [\[1\]](#ref-1), [\[5\]](#ref-5), [\[8\]](#ref-8), and a change of that magnitude across a defined population is a genuine phenomenon, not a measurement artifact. What remains uncertain is not the existence of the contraction, which can be asserted with high confidence, but the precise magnitudes of the earliest values, whose comparability rests on older reconstruction conventions that Chapter 4 must pin down.

The second step is that the contraction matters. The ellipse governs which sites are reachable and feeds the site-selection and forward-architecture decisions that drive multi-billion-dollar programs [\[15\]](#ref-15), [\[112\]](#ref-112), [\[113\]](#ref-113), [\[122\]](#ref-122); a number that gates site admissibility and informs architecture choices on this scale is material by any reasonable standard. Materiality to a future decision does not, by itself, establish that the decision-relevant lever is the one this dissertation names, which is precisely what the empirical work must show.

The third step is that the design isolates the causal mechanism. Modeling discrete technology generations as covariates, in the Mokyrian sense, on a measurement-disciplined series, in the Kuznetsian sense, can recover an attribution that a featureless time trend cannot, because the insertions are discrete and the error sources each lever attacks are physically separable [\[3\]](#ref-3), [\[5\]](#ref-5), [\[8\]](#ref-8), [\[9\]](#ref-9), [\[23\]](#ref-23), [\[24\]](#ref-24). The difficulty is the collinearity of technology with calendar time, which the small sample makes severe, so confidence here is only moderate and is bounded explicitly by Chapter 5's identification argument.

The fourth step is that the design discriminates the technology account from its rivals more sharply than any single-mission study could. InSight is a late mission flown on the calendar of MSL and Mars 2020 but with the guidance heritage of Viking and Phoenix [\[15\]](#ref-15), [\[14\]](#ref-14), so it functions as a within-period counterfactual that breaks the otherwise perfect collinearity of technology with time, and the delivered-entry-state control separates the onboard levers from the injection-accuracy rival [\[13\]](#ref-13). The cost of relying on this counterfactual is that identification then leans heavily on a single observation, which keeps confidence at the moderate level.

The fifth step is that the residual risk is acceptable. The small sample and the collinearity are real, but permutation inference, the nested specifications, and the dual dependent variable bound them, so the design is informative whichever way the coefficient falls. Because the analysis pre-commits to reporting a null result as the headline, and cross-checks the design ellipse against achieved miss distance using the reconstruction methodology and the robustness battery [\[7\]](#ref-7), [\[24\]](#ref-24), it cannot mislead even on a small sample. No small-sample design can deliver a precise point estimate, and none is claimed: confidence is high that the design is honest and only moderate that it will prove decisive. Each step names a mechanism rather than asserting a bare correlation, and where only correlation is available, above all the time-technology collinearity at the heart of the third step, the dissertation says so and downgrades its confidence accordingly.

## 1.10 Roadmap of the dissertation

The dissertation proceeds in eight chapters and a backmatter.

**Chapter 2, Theoretical Framework**, develops the two anchor methodologists and shows that they are not decorative. Mokyr's distinction between propositional and prescriptive knowledge, and between macro-invention and incremental improvement, generates the decision to model discrete technology generations as covariates [\[23\]](#ref-23), [\[22\]](#ref-22). Kuznets's insistence on a stated boundary, valuation convention, and netting rule, and on decomposing a change before theorizing it, generates the decision to construct the series before fitting any slope [\[24\]](#ref-24), [\[20\]](#ref-20), [\[21\]](#ref-21). The chapter builds the conceptual model, technology-generation steps on a measurement-disciplined log series, that the later chapters estimate, and it bridges from firm-level learning curves to a small-n, generation-indexed aerospace series through the organizational-learning literature [\[27\]](#ref-27), [\[28\]](#ref-28).

**Chapter 3, Literature Review**, the longest chapter, surveys the EDL-engineering canon thematically: guided lifting entry and entry-guidance algorithms [\[3\]](#ref-3), [\[4\]](#ref-4), the range trigger and footprint reduction [\[5\]](#ref-5), terrain-relative navigation and the Lander Vision System [\[8\]](#ref-8), [\[9\]](#ref-9), hazard detection and avoidance and precision-landing programs [\[12\]](#ref-12), supersonic retropropulsion and high-mass EDL as the forward edge [\[33\]](#ref-33), [\[111\]](#ref-111), and the mission-context studies that establish the public record. It maps each technology lever to the physical error source it attacks and states the gap precisely: the literature documents each mission and each technology in depth but never as one series and never with formal attribution against the injection-accuracy rival.

**Chapter 4, Data and Measurement**, takes the three named datasets, the NTRS landing-accuracy reconstructions, the TechPort technology-insertion records, and the PDS landing-site localization, and confronts the Kuznetsian comparability problem head-on, because the design ellipse is a simulation product whose conventions changed across missions. It operationalizes every variable in a measurement table, validates the design ellipse against achieved miss distance, and reports coverage and limitations, in particular the irreducible small-n, the simulation-convention drift, and the indicator-time collinearity.

**Chapter 5, Research Design**, presents the estimator and the identification strategy. It defends the log-linear form, writes out the baseline and augmented specifications exactly, builds the nested-specification hierarchy that adds one technology generation at a time, and grounds identification not on asymptotic significance but on the InSight counterfactual [\[15\]](#ref-15) and the approach-accuracy control [\[13\]](#ref-13). It treats every threat to validity, internal, external, construct, and statistical-conclusion, and pairs each with its mitigation.

**Chapter 6, Analysis Plan**, is the pre-registered, design-stage plan. It commits in advance to a single decision rule on H0 against H1, specifies the step-by-step estimation procedure, gives the expected signs with the full mechanism reasoning for each lever, and presents the result tables as specified-but-unpopulated templates. Every number in it is labeled expected or illustrative; no estimate is executed.

**Chapter 7, Discussion**, interprets implications symmetrically under both outcomes. If H1 holds, onboard guidance is the dominant lever and the era's improvement is a genuine Mokyrian learning curve. If H0 holds, the contraction is generic maturation or approach-accuracy-driven, which is itself decision-relevant. The chapter confronts the rival explanations rather than dismissing them and bounds external validity to US-Mars with Tianwen-1 and lunar precision landing as out-of-sample reference points [\[16\]](#ref-16), [\[17\]](#ref-17).

**Chapter 8, Conclusion**, restates the contribution and what stands even if the hypothesis is not confirmed: a defined measurement, a stated boundary, and a falsifiable hypothesis where the literature offered only narrative. It states the honest limitations and lays out a concrete future-research program, beginning with executing the frozen design on the assembled data.

The **backmatter** compiles the full reference list with resolvable identifiers and the appendices: the variable and data dictionary, the learning-rate and log-area derivations, the dataset access paths, and the specified-but-unpopulated specification and coding templates.
The through-line across all eight chapters is the single thesis stated in Section 1.1: that the Mars landing ellipse has contracted along an exponential, roughly log-linear learning curve whose steps align with identifiable onboard EDL-guidance insertions rather than with approach accuracy, and that this dissertation specifies the complete, falsifiable, cliometrically disciplined design needed to test that proposition and convert it into a technology-attributed basis for setting landing-accuracy requirements. The chapters that follow do not restate that thesis; they earn it.


# Chapter 2: Theoretical Framework

## 2.1 The chapter's answer, and the problem it solves

This chapter's thesis is that the two methodological anchors of the dissertation, Joel Mokyr's economic history of technology and Simon Kuznets's discipline of national-accounts measurement, are not ornamental citations imported to give an engineering study a scholarly veneer. They are load-bearing. Each generates one of the two design decisions that separate this dissertation from a naive curve fit, and together they specify, in advance, the form of the conceptual model that Chapters 5 and 6 will estimate. Mokyr generates the decision to model the contraction of the Mars landing ellipse as a sequence of discrete technology generations entered as covariates, rather than as a single smooth trend through time, because if the contraction is genuinely the visible trace of additions to a prescriptive knowledge base, those additions must show up as identifiable steps tied to the missions on which the technologies first flew. Kuznets generates the decision to refuse to report any learning rate before the measurement boundary of the ellipse series, its sigma convention, its valuation as design capability versus achieved accuracy, and its netting rule across heterogeneous simulation products, has been stated and the change decomposed. Stated together as a single claim: the learning-curve apparatus supplies the functional form, Mokyr supplies the causal mechanism that justifies indexing experience by technology generation, and Kuznets supplies the measurement protocol without which the fitted slope is an artifact rather than a finding.

The local problem this chapter addresses moves from the current state of theory, through the state that would be needed, to the gap between them and the consequence of leaving it open. Two mature but disjoint literatures sit on either side of the question. On one side is the experience-curve literature, which has a precise functional form for how a performance or cost metric improves with cumulative experience but which was built on high-volume production data and is silent on what to do when there are fewer than a dozen bespoke events [\[18\]](#ref-18), [\[19\]](#ref-19), [\[29\]](#ref-29), [\[30\]](#ref-30), [\[31\]](#ref-31). On the other side is the economic-history-of-technology literature, which has a rich account of why some techniques improve cumulatively and others stagnate but which offers no estimator [\[23\]](#ref-23), [\[22\]](#ref-22), [\[25\]](#ref-25), [\[26\]](#ref-26). Neither literature, on its own, tells an analyst how to attribute the contraction of a short, technology-differentiated aerospace series to specific onboard technologies while controlling for a rival cause. The desired state is a single conceptual model in which a measurement-disciplined log series of ellipse areas is regressed on a technology-generation experience axis, with each generation interpretable as a discrete addition to a prescriptive base and each reported value carrying its measurement provenance. The gap is that the bridge between firm-level learning curves and a small-n, generation-indexed aerospace series has never been built; the organizational-learning literature that would supply the planks of that bridge [\[27\]](#ref-27), [\[28\]](#ref-28) has not been connected to the EDL problem. Leave the gap open and any learning rate fitted to the Mars ellipses becomes uninterpretable. It could not say whether the contraction reflects technology, generic maturation, or a measurement convention that drifted, and so could not inform the requirement-setting and investment-valuation decisions that motivate the work.

The chapter builds the framework one anchor at a time. Section 2.2 develops the learning and experience curve as an empirical regularity, states its functional form precisely, and is candid about the conditions under which it holds and the conditions, both violated by Mars landings, under which it does not. Section 2.3 develops Mokyr in depth and maps the three EDL technology levers onto his propositional-prescriptive distinction and his macro-versus-incremental distinction. Section 2.4 develops Kuznets in depth and translates his three measurement requirements, boundary, valuation, netting, into operational rules for constructing the ellipse series. Section 2.5 builds the bridge from firm-level learning curves to the small-n aerospace series using the organizational-learning literature. Section 2.6 assembles the conceptual model that the empirical chapters will test and states the central, defensible claim the model embodies, together with the reasoning that supports it and the objections it must answer. A short closing section states the calibrated confidence of the framework and what evidence would raise or lower it.

A scope note governs the whole chapter. A formal architecture traceability, one that would link a strategic objective through capability, operational activity, system function, and data exchange to a measure and a decision, is deliberately out of scope. This dissertation's contribution is an econometric and cliometric account of a constructed performance series, not the design of a system or capability, and forcing capability-architecture vocabulary onto a regression would misrepresent what is being argued. The decision relevance of the work, its bearing on requirement-setting and investment valuation, is carried in plain prose in Chapter 1 and Chapter 7, not as a traceability table here.

## 2.2 The learning curve as an empirical regularity, and its limits

The claim of this section is that the learning curve supplies a defensible functional form for the dissertation's model, but only in a weakened, generation-indexed version that the section derives explicitly rather than borrowing whole; the strong, production-volume form of the regularity is inapplicable to Mars landings and the dissertation does not invoke it.

The learning curve, sometimes called the experience curve in its cost-of-output form, is among the oldest empirical regularities in the study of technological change. In its canonical statement a performance or unit-cost metric improves by a roughly constant percentage with each doubling of cumulative output or experience. When the metric and the cumulative-experience measure are both plotted on logarithmic axes the relationship appears as a straight line, and the slope of that line is conventionally summarized as a learning rate or progress ratio. The encyclopedic and textbook restatements of the theory are explicit that this is a descriptive generalization with a long pedigree across manufacturing, energy, and information technology rather than a derived law [\[29\]](#ref-29), [\[31\]](#ref-31). The functional form follows directly from the constant-percentage statement: if a metric Y improves by a fixed fraction with each doubling of cumulative experience X, then ln(Y) is linear in ln(X), and a regression of the log metric on the log experience measure recovers the rate as the slope. This is the form the dissertation adopts, with the natural logarithm of three-sigma ellipse area as the metric, and it is why a log-linear specification is the natural, not merely convenient, choice for the model.

The reliability of the regularity, and its failure modes, are an active subject of study, and the dissertation's honesty about its own application depends on taking that study seriously. Alberth's review of experience-curve forecasting is the load-bearing source here: it asks directly whether the curve is "myth or magic," and concludes that the relationship is empirically powerful within a technology paradigm but is not a law of nature, can be confounded by scale effects and input-price movements, and must be validated rather than assumed [\[18\]](#ref-18). What matters for this dissertation is not that the curve is unreliable but that its reliability is conditional, and that the conditions are statable. Where the experience axis is a genuine, large count of cumulative production, where the metric is measured the same way throughout, and where no single confounder moves monotonically with experience, the curve is trustworthy; where those conditions fail, a fitted slope can be an artifact of the confounder rather than a measure of learning. Ziegler and Trancik's re-examination of lithium-ion battery improvement is the methodological template for disciplined estimation under these cautions: they separate the contributions of distinct mechanisms to the observed rate, and they are explicit about what the experience axis actually counts, refusing to treat a single composite rate as if it were a primitive [\[19\]](#ref-19). Read together, Alberth and Ziegler and Trancik converge on a single methodological instruction: do not report a learning rate as if the number alone were the finding; report it alongside an account of which mechanism produced it and a statement of what the experience axis measures. That convergence is what justifies the dissertation's nested-specification design, in which technology generations are added one at a time so that the contribution of each is visible rather than buried inside a composite slope.

Two facets of the experience-curve literature bear directly on the Mars application and must be confronted rather than glossed. The first is the kinked or multi-segment curve. Chang's work on predicting the performance of large learning systems via a kinked experience curve establishes that the assumption of a single constant slope across the whole history is often wrong, and that the curve can bend at points where the underlying technology or organization changes [\[30\]](#ref-30). The reading for this dissertation is favorable: a kink is exactly what the Mokyrian story predicts at a technology-generation boundary. If guided entry, the range trigger, and terrain-relative navigation are genuine additions to the prescriptive base, the ellipse curve should not be a single straight line but a piecewise one, with its steepest segments at the missions that first flew those technologies. The kinked-curve literature thus supplies a precedent for the very feature the dissertation expects to find, and it cautions against the naive single-slope fit that would average the kinks away. The second facet is the broader cliometric finding, from Crafts's review of how cliometricians have analyzed technological change, that even genuinely important new technologies often have small initial aggregate effects and that rapid measured productivity gains frequently reflect the removal of inefficiency rather than pure technological advance [\[25\]](#ref-25). Translated to the Mars problem, this is a warning against over-attributing the contraction to the headline technologies and against treating the steepest contraction as automatically the most important macro-invention; some of the early contraction may be the removal of inefficiency, better atmospheric modeling, more faithful simulation, that is not itself a guidance technology, and the design must leave room to detect that.

The honest limit of the learning curve in this application is stated plainly because the dissertation's credibility rests on stating it. The strong form of the regularity is strongest where the experience axis is a large count of cumulative production and where every unit is measured identically. Mars landings violate both conditions decisively. There are fewer than a dozen of them across the entire robotic era, and each is a bespoke vehicle designed to its own requirements rather than a unit coming off a production line. The dissertation therefore does not claim the strong form. It claims something weaker and defensible: that ordering the missions by guidance generation produces a monotone, roughly log-linear contraction whose steps align with identifiable technology insertions. Whether to call that a "learning curve" or a "technology-generation effect" is partly a matter of naming, and the dissertation does not stake its contribution on the label. The falsifiable content lives in the alignment of the steps with the technologies, which the data can confirm or deny regardless of what the relationship is called. Confidence in the functional form is therefore high, since the log-linear form follows from the constant-proportional-change assumption and matches both the physical mechanism and the convention of the source literature, while confidence in any specific rate is held low by design until the series is built and the confounders addressed.

## 2.3 Mokyr in depth: propositional knowledge, prescriptive technique, and the three levers

The claim of this section is that Mokyr's economic history of technology supplies the causal mechanism behind a learning curve, and that his framework, applied to EDL, justifies the central modeling choice of treating the three guidance technologies as discrete, separately-coded covariates rather than as a single index of modernity.

Mokyr's organizing distinction is between propositional knowledge and prescriptive knowledge [\[23\]](#ref-23), [\[22\]](#ref-22). Propositional knowledge is the understanding of why a technique works: the body of natural regularities, theories, and empirical facts about the world that a technique exploits. Prescriptive knowledge is the technique itself, the set of instructions for doing a thing. The dossier that anchors this dissertation states the distinction and its consequence directly: techniques that rest on a deep propositional base are extensible and self-correcting, because the supporting understanding lets each generation of the technique be diagnosed when it fails, improved when it underperforms, and built upon when it succeeds; techniques found by trial and error without an underlying theory tend to stagnate, because there is no map of the space of possible improvements and no way to know why a change helped or hurt [\[23\]](#ref-23). This is the mechanism behind a learning curve in Mokyr's reading. A learning curve is not a free-floating regularity that descends on a technology by magic; it is what extensibility looks like, plotted over time, when the propositional base supporting a technique is wide and deep enough that each generation can be understood and the next one designed rather than stumbled upon. The recent restatement of Mokyr's program, occasioned by the 2025 Nobel award, makes the same point in the language of "useful knowledge": sustained growth depends on the accumulation and the dissemination of both propositional and prescriptive knowledge, and the mechanism by which a technology keeps improving is the feedback between the two [\[26\]](#ref-26). The reading for this dissertation is that a Mars landing technique whose contraction follows a learning curve must, on Mokyr's account, be a technique resting on a maturing propositional base, and the way to test that reading is to ask whether the observed improvements line up with identifiable deepenings of that base.

Carlota Perez's account of techno-economic paradigms is cited in this dissertation as the published restatement of the Mokyrian point, not as a third and independent anchor [\[22\]](#ref-22). Perez describes how a key enabling capability can reorganize a whole domain around itself, becoming the common factor that subsequent innovations exploit. This is the same structure as Mokyr's extensible technique resting on a propositional base, expressed at the level of a domain rather than a single artifact, and the dissertation invokes it to license the claim that onboard guidance, once it crossed a capability threshold, became the organizing axis of Mars EDL improvement in the way Perez's paradigms organize their domains. The bible is explicit that Perez is a restatement and that no new anchor is to be introduced; this section honors that constraint and treats Perez strictly as corroboration of the Mokyrian mechanism.

The application to the three EDL levers is the substantive payoff of the section, and it is where the framework earns the modeling choice it justifies. Each of the three technologies that the dissertation codes as a covariate is, on the Mokyrian reading, a discrete addition to the prescriptive landing-technique base, and each rests on its own identifiable propositional base. Guided lifting entry, the closed-loop modulation of bank angle to null the downrange and crossrange dispersion accumulated during the hypersonic phase, rests on the propositional base of entry aerodynamics and atmospheric reconstruction; the entry-guidance literature that documents this lever is built on exactly that base of aerodynamic understanding [\[3\]](#ref-3), [\[4\]](#ref-4). The range-to-go parachute trigger, which commands parachute deploy on navigated range rather than velocity, rests on the propositional base of onboard navigation, the ability to know how far downrange the vehicle actually is at the deploy decision [\[5\]](#ref-5). Terrain-relative navigation, which collapses the position-knowledge error by matching descent imagery to an onboard reference map and thereby enables an autonomous divert, rests on a third and distinct propositional base, that of computer vision and orbital mapping; the construction of the Mars 2020 onboard reference map from orbital imagery is the visible deepening of that base [\[8\]](#ref-8), [\[9\]](#ref-9). Because each lever exploits a different body of propositional knowledge and attacks a physically distinct error source, the levers are, in principle, separable, and the engineering separability is the direct counterpart of Mokyr's claim that each technique rests on its own base. This separability is what justifies coding three distinct covariates rather than one composite modernity index: a single index would assert that the three levers are interchangeable manifestations of a generic improvement, which the Mokyrian framework denies and which the physics contradicts.

Mokyr's second distinction, between a one-time macro-invention and sustained incremental improvement, gives the framework a further and testable refinement. A macro-invention is a discontinuous leap that opens a new region of the technique space; incremental improvements are the subsequent refinements that work that region. Applied to the three levers, the framework predicts a pattern in the data, not merely a direction. Guided entry was closer to a macro-invention: it changed the entry regime from ballistic to lifting and removed the single dominant early error source in one step. The range trigger was incremental: a refinement of an existing parachute-deploy logic that bought a further, smaller reduction in the footprint [\[5\]](#ref-5). Terrain-relative navigation was again macro: it introduced an entirely new error source to attack, position knowledge, that the prior two levers left wholly untouched, and it enabled a capability, autonomous hazard avoidance, that had no predecessor [\[8\]](#ref-8). The empirical prediction that follows is sharp: the largest single increments in model fit should appear at the guided-entry transition and at the TRN transition, with a smaller increment at the range trigger. If the data show that pattern, the pattern is itself evidence for the Mokyrian macro-versus-incremental distinction, and it strengthens the causal interpretation beyond what a bare correlation could support, because the framework predicted the relative magnitudes in advance rather than rationalizing them after the fact.

The same framework that supplies the mechanism also supplies a boundary condition on the dissertation's external validity, and intellectual honesty requires stating it inside the framework rather than smuggling it into the discussion as an afterthought. Mokyr is emphatic that technological progress is reversible: a technique resting on a knowledge base degrades when the base is not maintained, when the propositional understanding decays, when the engineering teams who hold it tacitly disperse, when the test infrastructure that validated it is dismantled. The landing-precision capability is therefore not a permanent possession that a future program automatically inherits; it is embodied in a chain of propositional knowledge, tacit team expertise, flight-software heritage, and validation infrastructure, and a future program inherits whatever of that chain it can reconstitute. The practical consequence, drawn out fully in Chapter 7 but established here as a property of the framework, is that a learning curve fitted to past missions describes capability that was demonstrated, not capability that is guaranteed to be available, and any extrapolation along the curve implicitly assumes the knowledge chain is intact. The framework thus contains its own warning against mechanical extrapolation, which is the right place for the warning to live.

A statement of the causal mechanism, traced from its driver through to its strategic implication, closes the section. The driver is the insertion of an onboard EDL-guidance technology that rests on a maturing propositional base. The mechanism is that the technology removes a physically distinct error source: guided entry nulls hypersonic dispersion, the range trigger corrects parachute-deploy dispersion, TRN collapses position-knowledge error and enables divert. The observable effect is a discrete drop in the three-sigma ellipse at the first mission to fly the technology. The operational consequence is that a smaller ellipse lets a project target hazardous, scientifically rich terrain that a larger ellipse would have excluded. The strategic implication is that landing precision becomes a quantifiable, technology-attributed design variable for human-Mars and sample-return architectures rather than a qualitatively negotiated number. Where the small sample and the collinearity of technology with time prevent the data from cleanly separating the levers, the framework requires that this be said and the confidence downgraded accordingly; the mechanism is named so that the analysis can test it, not asserted so that the analysis can assume it.

## 2.4 Kuznets in depth: boundary, valuation, netting, and the transient-secular distinction

The claim of this section is that Kuznets's measurement discipline is not a stylistic preference for careful definitions but a set of three operational requirements, each of which converts directly into a rule for constructing the ellipse series, and that without those rules the fitted slope of Section 2.2's functional form would measure a convention rather than a phenomenon.

Kuznets's lifelong insistence, recorded in the anchoring dossier and restated in the modern national-accounts literature, was that an aggregate is meaningless without three things stated explicitly: a boundary of coverage that says what is included and what is excluded, a valuation convention that says how heterogeneous items are made commensurable, and a netting rule that says what is subtracted to avoid double counting [\[24\]](#ref-24), [\[20\]](#ref-20). The discipline that follows is to build long, comparable, decomposed series before theorizing about them, because a series assembled without these conventions stated will silently mix incomparable things and any trend fitted to it will be an artifact of the mixing. The modern restatement of the national-accounts measurement problem makes the same point at length: the apparent simplicity of a single headline number, gross domestic product, conceals a dense layer of boundary, valuation, and netting decisions, and the number's meaning is hostage to those decisions [\[20\]](#ref-20). The lineage matters to this dissertation because the landing ellipse is exactly the kind of constructed measurement Kuznets analyzed. Its reported value is not a directly observed quantity like a length read off a ruler; it is a simulation product whose number depends on the confidence level chosen, on whether it describes the targeting capability or the achieved miss distance, and on the Monte Carlo and atmospheric-modeling conventions used to generate it. A series of such numbers, assembled across missions whose conventions differed, is precisely the kind of aggregate Kuznets warned must not be theorized about until its boundaries are stated.

The three requirements translate into three concrete construction rules, and stating the translation is the operational core of the section. The boundary rule fixes what counts as a member of the series and at what confidence level it is read. The dissertation's boundary is the set of U.S.-led Mars surface missions that successfully entered, descended, and landed, and the convention is the three-sigma design ellipse; every value entering the series must be a three-sigma figure or be converted to one with the conversion stated, because mixing one-sigma and three-sigma figures would manufacture spurious contraction or expansion. The valuation rule fixes how the ellipse is made commensurable across missions whose ellipses were generated by different simulations. The dissertation's valuation convention is the natural logarithm of ellipse area, computed as the log of pi times the semi-major axis times the semi-minor axis, which renders the contraction multiplicative and comparable in proportional terms across the enormous dynamic range from Viking to Mars 2020; the choice of a proportional rather than additive valuation is itself a Kuznetsian decision and is defended as such rather than assumed. The netting rule fixes what must be held separate so that the contraction attributed to one cause is not also being counted under another. The dissertation's netting rule is the approach-accuracy control: the portion of delivery accuracy attributable to launch and interplanetary navigation is held constant by including the reported approach-navigation delivery dispersion as a covariate, so that any contraction loading on the technology indicators is net of approach accuracy rather than confounded with it. These three rules are the direct, operational descendants of Kuznets's three requirements, and they are stated here so that the data chapter can implement them and the analysis chapter can be held to them.
Kuznets's second great distinction, between a transient movement and a secular trend, binds the dissertation's inference as tightly as the measurement requirements bind its construction. A transient movement is a one-time level shift or a cyclical fluctuation; a secular trend is a sustained directional change. Kuznets warned, repeatedly and specifically, against extrapolating a secular claim from a short window in which a single level shift might masquerade as a trend [\[24\]](#ref-24). The warning bites here because the entire series is a short window of nine to eleven events, and because a single dramatic mission, MSL's introduction of guided entry and the sky-crane, or Mars 2020's introduction of TRN, could produce a level shift that a naive fit would read as a continuing secular decline. The framework's response is to build the distinction into the model rather than to hope it away. By coding technology generations as covariates and fitting nested specifications, the design asks directly whether the contraction is a sequence of level shifts at identifiable technology boundaries, which is what the Mokyrian macro-invention story predicts, or a smooth secular decline that would extrapolate forward, or some combination. The transient-secular distinction is therefore not a caveat appended to the conclusion; it is one of the questions the model is built to answer. The honest answer may be that the era's improvement is better described as a small number of large level shifts than as a smooth trend, which would itself matter to any program tempted to extrapolate the curve.

A subtler facet of the Kuznetsian inheritance is the satellite-account measurement lineage, which the dissertation invokes to mark the pedigree of the measurement ethic it applies rather than as direct evidence. Kuznets pioneered the construction of economic measurements from indirect observation when direct measurement was impossible, and the modern descendant of that practice is the use of orbital observation to construct economic series where conventional accounts are weak. Henderson, Storeygard, and Weil's use of satellite night-lights to proxy economic growth is the canonical modern instance, and it stands in this dissertation as a marker of the lineage: the discipline of building a defensible measurement from imperfect, indirect, convention-laden observation is the same discipline whether the observation is a night-light raster or a reconstructed landing ellipse [\[21\]](#ref-21). The dissertation does not lean on this lineage for inference. It cites it to locate the measurement ethic in a tradition and to make explicit that the ellipse series is treated as a constructed economic-style measurement, with all the boundary, valuation, and netting care that tradition demands, rather than as a directly observed physical quantity.

The combined force of the two anchors is the point the section closes on, because it is the combination, not either anchor alone, that distinguishes the dissertation from a naive analysis. A naive analyst would regress log ellipse on calendar year, read off a slope, and announce a learning rate. Both anchors reject that move, and they reject it for different and complementary reasons. Mokyr rejects it because it hides the mechanism: a slope through time says nothing about why the contraction happened and so cannot attribute it to a technology or distinguish it from generic maturation. Kuznets rejects it because it hides the boundary: a slope through time treats the ellipse values as comparable when they were generated by drifting conventions, and so risks measuring the drift rather than the phenomenon. The framework's discipline is to satisfy both objections at once, modeling discrete technology generations on a measurement-boundary-stated, decomposed series, and that simultaneous satisfaction is what makes the result, if it survives, defensible rather than merely suggestive. Confidence in the measurement discipline is high; it is a method, not an empirical claim, and the only thing that could lower it is a demonstration that the ellipse series cannot in fact be made comparable across missions, the central evidentiary risk that Chapter 4 confronts directly.

## 2.5 From firm learning curves to a generation-indexed aerospace series

The claim of this section is that the gap between the firm-level learning-curve literature, built on many organizations repeating a process many times, and the dissertation's setting, a handful of unique missions flown by a single agency, can be bridged by the organizational-learning literature, which establishes that improvement rates are heterogeneous, depend on the kind of knowledge being accumulated, and are driven by deliberate learning processes rather than by mere repetition.

The bridge is necessary because a naive transfer of the firm-level learning curve to Mars landings would be illegitimate, and saying so protects the framework from a fair objection. The classical learning curve was estimated on settings, airframe assembly, semiconductor fabrication, surgical procedures, where the same organization performs a recognizable process many times and the cumulative count of repetitions is large. Mars landings have none of these features: there are few events, each is performed once by a bespoke vehicle, and the "organization" is a single agency whose teams and contractors change across decades. If the only mechanism behind a learning curve were the mechanical accumulation of repetitions, the framework could not apply at all, because the repetition count is too small to drive anything. The organizational-learning literature supplies the escape by showing that mechanical repetition is not the operative mechanism.

Pisano, Bohmer, and Edmondson's study of the adoption of minimally invasive cardiac surgery is the first plank of the bridge, and its finding is directly on point: organizations achieve different rates of performance improvement from equivalent levels of cumulative experience, which means that experience alone does not determine the improvement rate and that something about how an organization learns intervenes between experience and improvement [\[27\]](#ref-27). For this dissertation the relevant experience axis is not a raw count of landings but a measure of how far the agency's learning process has advanced, and the technology-generation index proxies that better than the mission count, because each technology generation embodies a deliberate, codified advance in the agency's landing knowledge rather than a passive accumulation of flights. This is the empirical license for the dissertation's primary experience axis being mission-sequence-by-generation rather than cumulative-landings-count, with the cumulative count retained only as an alternative specification to test whether the contraction tracks ordinal generation or raw experience.

Edmondson, Winslow, Bohmer, and Pisano's companion study of tacit versus codified knowledge in performance improvement following technology adoption is the second and more refined plank [\[28\]](#ref-28). Its central finding is that the same technology can present opportunities for improvement along more than one dimension, and that improvement along a dimension driven by codified knowledge proceeds differently from improvement along a dimension that requires tacit knowledge. The relevance to Mars EDL is direct. The three guidance levers differ precisely in the balance of codified and tacit knowledge they embody. Guided entry and the range trigger are heavily codified: they are algorithms, expressible as flight software and validated in simulation, and their insertion is close to the codified-knowledge case in which improvement follows the adoption of a documented technique. Terrain-relative navigation is more tacit-laden: it depends on the agency's hard-won, partly tacit competence in computer vision, onboard mapping, and the integration of a vision system into a real-time descent loop, competence matured through a long sequence of relevant-environment tests and flight demonstrations before it could be flown [\[9\]](#ref-9), [\[10\]](#ref-10). The Edmondson framework predicts that the codified levers should produce more uniform, more transferable improvements, while the tacit-laden lever should produce improvements more dependent on the specific agency's accumulated competence and therefore less automatically transferable to another program. This prediction connects back to the Mokyrian reversibility boundary condition of Section 2.3: the tacit component of TRN is exactly the part of the capability most at risk of degrading if the knowledge chain is not maintained, because tacit knowledge resides in teams rather than in documents and is the first to disperse.

The bridge, assembled, yields a defensible reinterpretation of what the dissertation's experience axis measures. The axis is not cumulative production, which the setting does not have, and it is not calendar time, which Kuznets warns hides the boundary and Mokyr warns hides the mechanism. It is the sequence of technology generations, each of which the organizational-learning literature licenses as a marker of a deliberate, knowledge-embodying advance in the agency's landing competence rather than a passive count of flights. This reinterpretation is what makes a learning-curve apparatus, built for high-volume firm-level data, legitimately applicable to a small-n, generation-indexed aerospace series. Confidence in the bridge is moderate rather than high, because it rests on transferring a finding from cardiac surgery and other firm settings to a unique aerospace setting, a transfer the organizational-learning literature supports in principle but that no one has previously made for EDL. The bridge would be strengthened by any aerospace-specific evidence of heterogeneous learning rates across programs, and weakened by evidence that the agency's landing improvements were in fact driven by passive accumulation rather than deliberate technology insertion, which the InSight counterfactual is partly designed to test.

## 2.6 The conceptual model the empirical work will test

This section shows that the three preceding anchors assemble into a single conceptual model, that the model is fully specified, and that it embodies one defensible, falsifiable proposition which the empirical chapters will estimate and which can be stated completely as a connected argument.

The conceptual model is a measurement-disciplined, technology-generation-indexed learning curve. Its dependent variable is the natural logarithm of the three-sigma landing-ellipse area, \(\ln(\text{EllipseArea}_i)\), constructed under the Kuznetsian boundary, valuation, and netting rules of Section 2.4, with the natural logarithm of achieved miss distance from post-landing localization serving as a secondary dependent variable to guard the construct against divergence between design capability and achieved accuracy. Its experience axis is \(\text{Sequence}_i\), the mission-sequence index reinterpreted through Section 2.5 as a marker of deliberate, knowledge-embodying technology generations rather than a passive count, with a cumulative-landings count retained as an alternative specification. Its technology covariates are the three binary indicators that Section 2.3 justified coding separately: \(\text{GuidedEntry}_i\), set to one for MSL and Mars 2020; \(\text{RangeTrigger}_i\), set to one for MSL and Mars 2020; and \(\text{TRN}_i\), set to one for Mars 2020 only. Its control is \(\text{ApproachAccuracy}_i\), the reported approach-navigation delivery dispersion that implements the Kuznetsian netting rule by holding constant the launch-and-injection contribution to delivery accuracy. The model is fitted by ordinary least squares on the log-linear form, in a hierarchy of nested specifications that add one technology generation at a time so that the contribution of each is visible, with permutation-based and exact inference appropriate to the sample size of nine to eleven.

The baseline specification, written exactly as the bible fixes it, is

\[
\ln(\text{EllipseArea}_i) = \beta_0 + \beta_1 \cdot \text{Sequence}_i + \epsilon_i\qquad\qquad (1)
\]

which recovers a constant proportional contraction per mission generation, with \(\exp(\beta_1) - 1\) read as the proportional change in ellipse area per unit of mission sequence. The augmented specification, also exactly as fixed, is

\[
\ln(\text{EllipseArea}_i) = \beta_0 + \beta_1 \cdot \text{Sequence}_i + \gamma_1 \cdot \text{GuidedEntry}_i + \gamma_2 \cdot \text{RangeTrigger}_i + \gamma_3 \cdot \text{TRN}_i + \delta \cdot \text{ApproachAccuracy}_i + \epsilon_i\qquad\qquad (2)
\]

in which the contribution hypothesis H1 predicts that the gamma coefficients are jointly significant and negative, each technology reducing ellipse area, and that delta is small and insignificant, approach accuracy not being the binding constraint, while the null H0 predicts the gamma coefficients are jointly insignificant once the sequence trend and the approach-accuracy control are present. The mapping of the framework onto these coefficients is exact, and that is what makes the anchors testable rather than decorative. The gamma coefficients are the Mokyrian additions to the prescriptive base; their joint significance and negative sign is the empirical signature of the extensibility mechanism. The relative magnitudes of the gamma coefficients are the Mokyrian macro-versus-incremental prediction; the framework expects the guided-entry term \(\gamma_1\) and the TRN term \(\gamma_3\) to dominate the range-trigger term \(\gamma_2\). The delta coefficient is the Kuznetsian netting rule made operational; its insignificance is what licenses attributing the residual contraction to onboard guidance rather than to approach accuracy. The nested specification structure is the joint demand of Alberth's "report the mechanism, not just the rate" and Kuznets's "decompose before theorizing." Nothing in the model is free of the framework; every term is the operational image of an anchored idea.

The model embodies a single central proposition, which can be set out in full as the chapter's principal argument. The proposition is that the three-sigma Mars landing ellipse contracts along an exponential, roughly log-linear learning curve whose discrete steps align with the insertion of onboard EDL-guidance technologies, and that onboard guidance, not approach accuracy, is the dominant lever. The evidence for it is the documented, large, monotone contraction of the ellipse across the missions [\[1\]](#ref-1), [\[5\]](#ref-5), [\[8\]](#ref-8), together with the alignment of the steepest documented reductions with the missions that first flew guided entry and TRN. What licenses reading that alignment as more than coincidence is the Mokyrian mechanism: a technique resting on a maturing propositional base improves by extensibility, and additions to that base appear as discrete steps tied to the missions on which they first flew, so a contraction whose steps align with technology insertions is the visible trace of that mechanism [\[23\]](#ref-23), [\[22\]](#ref-22), [\[26\]](#ref-26). This principle itself rests on two further bodies of evidence: the economic-history finding that propositional-base extensibility is the general mechanism behind durable technological improvement [\[25\]](#ref-25), [\[26\]](#ref-26), and the organizational-learning finding that improvement rates track deliberate, knowledge-embodying advances rather than passive repetition [\[27\]](#ref-27), [\[28\]](#ref-28), which together justify indexing experience by generation in a small-n setting. The proposition is advanced under an essential and explicitly protected limit: it holds with moderate confidence at the design stage and with at most order-of-magnitude precision on any rate, because the small sample and the collinearity of technology with calendar time prevent a clean separation of the three levers, so the strongest defensible form of the proposition is a signed, ordered, counterfactual-surviving attribution rather than a precise point estimate. The conditions under which the proposition fails are stated in advance and are exactly the conditions that would refute it: if the gamma coefficients are insignificant or wrongly signed, if the approach-accuracy control absorbs them, if the InSight observation sits on rather than above the trend, or if the achieved-miss-distance series fails to show the contraction the design ellipse shows, then the proposition is rejected. Stating those failure conditions concretely and in advance is what makes the proposition falsifiable in the strict sense, and the framework's task is to have generated a proposition that the data can actually overturn.

The line of argument that the dissertation carries across all chapters is established here at the level of theory, and stating its theoretical contribution closes the model. That the problem is real is established at the framework level by the documented contraction's being currently only qualitatively explained, with no joint model and no formal attribution. That the problem is material follows from the ellipse's governing which sites are reachable, a point the framework connects to the Mokyrian observation that capability determines the accessible technique space. That the design addresses the causal mechanism is the entire content of Sections 2.3 and 2.4: modeling discrete technology generations on a measurement-disciplined series isolates the onboard-guidance mechanism in the way Mokyr and Kuznets jointly require. That the design discriminates the technology account from its rivals follows at the framework level from the transient-secular distinction and the netting rule, which together separate the technology hypothesis from the pure-time and approach-accuracy rivals more sharply than any single-mission study could. And that the residual risk is acceptable follows from the framework's own candor about small-n and collinearity and from the design features, the nested specifications, the dual dependent variable, and permutation inference, that bound those risks; the framework is informative whichever way the coefficients fall, because a rejection of H1 is itself a decision-relevant finding about the nature of EDL maturation. The theoretical contribution, then, is not a new theory of technological change but a demonstrated transfer: the marriage of Mokyr's mechanism and Kuznets's measurement discipline to an aerospace performance series, executed with enough fidelity that the resulting model is falsifiable, and the two anchors are shown to generate the design rather than ornament it.

## 2.7 Confidence, and what would move it

The framework rests on three claims held at three different confidence levels, and stating them explicitly is the calibration that intellectual honesty requires of a design-stage argument. The functional-form claim, that a log-linear specification is the right form for the model, is held at high confidence, because it follows deductively from the constant-proportional-change premise, matches the multiplicative physics of fractional error-source reduction, and matches the reporting convention of the source literature [\[18\]](#ref-18), [\[19\]](#ref-19), [\[29\]](#ref-29), [\[31\]](#ref-31); it would be lowered only by evidence that the contraction is additive rather than proportional, which the linear-in-levels robustness check in Chapter 5 is built to detect. The mechanism claim, that the contraction is driven by Mokyrian additions to a prescriptive base rather than by generic maturation, is held at moderate confidence at the design stage, because the framework is internally coherent and the levers map cleanly onto distinct propositional bases [\[23\]](#ref-23), [\[9\]](#ref-9), but the data have not yet been assembled and the central confounder, the collinearity of technology with time, is real; it would be raised by an InSight counterfactual that sits above the trend and by gamma coefficients that survive the approach-accuracy control, and lowered by either failing. The bridge claim, that firm-level learning curves transfer legitimately to a small-n agency series via the organizational-learning literature, is held at moderate-to-low confidence, because the transfer from cardiac surgery and other firm settings to a unique aerospace setting is novel and untested [\[27\]](#ref-27), [\[28\]](#ref-28); it would be raised by aerospace-specific evidence of heterogeneous program learning rates and lowered by evidence that landing improvement was driven by passive accumulation. None of these confidences is a number, because at the design stage a number would be false precision; each is a graded judgment paired with the specific, pre-specified observation that would move it, the honest form for a framework that has specified an experiment it has not yet run. The chapters that follow hold the model fixed and turn to the question this framework cannot answer on its own: whether the data, once disciplined into a comparable series, actually show the pattern the framework predicts.


# Chapter 3: Literature Review

## 3.1 The chapter thesis and the shape of the gap

The entry, descent, and landing (EDL) engineering literature is, by any reasonable standard, mature, careful, and quantitatively rich. It documents each Mars surface mission in depth, it documents each major guidance technology in depth, and it documents the physical error sources that govern where a vehicle touches down. What it has never done, and what this chapter establishes by a thematic reading of the corpus, is assemble those mission-by-mission and technology-by-technology accounts into a single cross-mission performance series and submit that series to a formal attribution test against the rival explanation that the spacecraft are simply being delivered to the top of the Martian atmosphere more accurately. That is the gap. The literature contains every ingredient required to build the dependent variable, to code the technology indicators, and to construct the approach-accuracy control, but it contains no study that does so jointly, and so it leaves the central causal question of this dissertation, whether onboard guidance or approach accuracy is the binding constraint on the landing ellipse, formally unanswered.
Stated in the problem frame the program requires: the current state of the literature is a set of internally excellent but disconnected accounts, each designed to certify or reconstruct a single mission or to advance a single guidance technology; the desired state is a constructed, comparable, decomposed ellipse series fitted with a learning-rate model that attributes the contraction to identifiable technology insertions while holding approach accuracy constant; the gap between them is the absence of any work that joins the EDL-engineering record to the technology-economics learning-curve apparatus and arbitrates among the rival causes; and the consequence of leaving the gap open is that future architectures, Mars Sample Return retrieval and human landing among them, continue to negotiate landing-accuracy requirements qualitatively and cannot value a guidance investment against an approach-navigation investment on a common quantitative basis [\[112\]](#ref-112), [\[113\]](#ref-113), [\[122\]](#ref-122).

This chapter develops that thesis by reading the corpus thematically rather than chronologically. Section 3.2 treats guided lifting entry and the entry-guidance algorithms that implement it. Section 3.3 treats the range-to-go parachute trigger. Section 3.4 treats terrain-relative navigation (TRN) and the Lander Vision System (LVS), including the onboard mapping that makes TRN possible. Section 3.5 treats hazard detection and avoidance and the dedicated precision-landing capability programs that matured these technologies to flight. Section 3.6 treats supersonic retropropulsion and high-mass, human-class EDL as the forward edge of the field. Section 3.7 treats the mission-context and instrument literature that establishes the public record from which the series is reconstructed. Section 3.8 supplies the synthesis: a table mapping each technology lever to the physical error source it attacks, and a second table classifying each major source by what it contributes and where it stops short. Section 3.9 states the gap explicitly and derives the propositions that the remainder of the dissertation tests. Every source is interpreted rather than merely listed. For each major work the review states what it found, the method it used, the limitation that bounds its reach, and the relation it bears to the gap. Citations are numbered to the corpus keys in `research/corpus.jsonl`; seed references retain their dissertation numbering as S-keys and harvested references carry R-keys.

A word on the confidence register that governs the whole chapter. The EDL-engineering claims reviewed here are, individually, high-confidence: they rest on flight data, on Monte Carlo dispersion analyses validated against that flight data, and on peer review in the aerospace literature. The dissertation's own contribution, by contrast, is a design-stage attribution claim of, at best, moderate confidence, bounded by a sample of nine to eleven missions and by the near-collinearity of the technology indicators with calendar time. The literature review is where those two confidence grades meet. The engineering record is strong enough to build the series; it is not, on its own, strong enough to settle the attribution, because no single mission study was designed to. Keeping that distinction visible is the discipline this chapter imposes on itself, and it is the discipline that the anchor methodologists, Mokyr on mechanism and Kuznets on measurement, demand of any claim that a performance series follows a learning curve.

## 3.2 Guided lifting entry and the entry-guidance algorithms

The first and largest of the three technology levers is guided lifting entry: the closed-loop control of a lifting entry vehicle, by bank-angle modulation, to null the downrange and crossrange dispersion that a ballistic capsule accumulates during the long hypersonic phase. The claim this dissertation makes about guided entry is that it removed the dominant component of the early footprint in a single step, which in Mokyr's vocabulary marks it as closer to a macro-invention than to an incremental refinement. The basis for that claim is in the entry-guidance literature, and it is quantitative.

The foundational comparison is Mendeck and colleagues' evaluation of reference-path (Apollo-heritage) versus predictive path-planning guidance for Mars entry [\[3\]](#ref-3). The method is Monte Carlo simulation of an MSL-class vehicle across a dispersed entry state and a dispersed atmosphere, with the figure of merit being the delivered landing dispersion at parachute deploy. The finding is that closed-loop guided entry reduces that dispersion by a large factor relative to an unguided ballistic trajectory, and that the reference-path approach, which tracks a precomputed drag profile and modulates bank to stay on it, is adequate for the MSL accuracy requirement. The limitation, for the dissertation's purposes, is that this is a single-vehicle simulation study: it quantifies the dispersion reduction that guided entry can achieve in principle for one mission class, but it does not place that reduction inside a cross-mission series and so cannot, by itself, tell us how much of the historical ellipse contraction guided entry actually delivered. That is the move from per-mission capability to cross-mission attribution that the gap concerns. The companion paper by Mendeck and Craig documents the specific entry-guidance algorithm flown by MSL, the first Mars mission to fly closed-loop guided lifting entry [\[4\]](#ref-4), and it is the source from which the guided-entry technology indicator is coded at the MSL mission. Its method is design documentation and verification simulation rather than flight reconstruction, and its limitation is the same: it certifies that MSL flew guided entry, which is what the indicator needs, but it does not by construction speak to the counterfactual ellipse MSL would have flown without it.

The system-level context for the guided-entry lever comes from the two MSL EDL papers by Steltzner and by Way [\[1\]](#ref-1), [\[2\]](#ref-2). Steltzner's performance paper describes the MSL architecture and its predicted landing performance, including the guided lifting entry and the parachute deploy that together define the targeting footprint, and it establishes the design landing ellipse for the first guided-entry Mars mission [\[1\]](#ref-1). Way's overview paper describes the full EDL chain, guided entry, supersonic parachute, powered descent, and sky-crane, and is the source for the MSL guidance suite considered as a whole relative to its ballistic-entry predecessors [\[2\]](#ref-2). The method in both is system engineering analysis backed by end-to-end simulation; the value to the dissertation is that they fix the MSL point in the ellipse series and identify which technologies were bundled at that mission. The limitation, again, is the bundling itself: MSL introduced guided entry and the range trigger simultaneously, so neither of these papers can separate the two levers, and the separation is left to the nested-specification design of Chapter 5. Way's later preliminary-assessment paper, reconstructing the MSL EDL simulation against the actual Curiosity landing, confirms that the as-flown performance matched the predicted footprint and so validates the design ellipse as a measurement [\[52\]](#ref-52); its method is post-flight reconstruction, and its limitation for attribution is that one validated mission point does not constitute a series.

Beyond MSL, the entry-guidance literature has continued to develop along two axes that matter for the dissertation's framing even though they postdate the missions in the series. The first is the extension of predictor-corrector guidance to human-scale vehicles. Lugo and colleagues describe a generalized numerical predictor-corrector targeting guidance, originally developed for the Mars Surveyor Program and extended to permit fully generalized aerocapture and EDL guidance design for human-scale Mars missions [\[34\]](#ref-34). The method is algorithm development with trajectory simulation; the finding is that a single guidance formulation can be made general enough to serve precision targeting across a wide vehicle and mission range; the limitation is that this is forward-looking design work, validated in simulation rather than in flight, so it bears on the dissertation as evidence that the propositional base under guided entry is deep and extensible (a Mokyrian point) rather than as a data point in the series. The dual-quaternion six-degree-of-freedom guidance work by Lugo and colleagues extends the same human-scale agenda into a fully coupled translational-rotational formulation [\[43\]](#ref-43), and the mid-lift-to-drag fractional-polynomial powered-descent guidance of Johnson, Lu, and Sostaric addresses the same human-scale precision problem from the terminal-descent side [\[98\]](#ref-98). The second axis is uncertainty quantification under guidance: Duan and Mease analyze Mars entry guidance for high-elevation landing with explicit uncertainty quantification and reduction [\[87\]](#ref-87), Williams and colleagues validate linear-covariance techniques for Mars EDL guidance and navigation performance analysis against Monte Carlo [\[44\]](#ref-44), and Dai, Gao, and Xia develop a nonlinear compound controller for reference-trajectory tracking in Mars atmospheric entry [\[51\]](#ref-51). Amato treats Mars EDL guidance explicitly under dynamic uncertainty [\[64\]](#ref-64), and Mischna's atmospheric-perspective review situates the guidance problem inside the atmospheric variability that drives the dispersion the guidance must null [\[65\]](#ref-65).

There is a further reason the entry-guidance literature matters to the dissertation beyond supplying the guided-entry indicator, and it concerns the relationship between guidance and the atmosphere the guidance must fly through. The dispersion that guided entry nulls is not a fixed quantity; it is set by the variability of the Martian atmosphere and by the fidelity with which that atmosphere can be forecast and reconstructed. Mischna's atmospheric-perspective review of Mars EDL makes the point that the guidance problem and the atmosphere problem are not separable: the same wind, density, and temperature dispersions that the entry vehicle must tolerate are the dispersions that the guidance is designed to absorb [\[65\]](#ref-65). This bears directly on the dissertation's identification problem, because atmospheric-modeling fidelity improved monotonically across the era at the same time as the guidance technologies were inserted, and an improvement in the modeled atmosphere could in principle tighten the design ellipse without any change in onboard guidance at all. The entry-guidance literature does not resolve this confound, because each study holds the atmosphere model fixed and varies the guidance; but in flagging it here the review hands Chapter 5 a specific named threat (atmospheric-fidelity confounding) rather than a vague worry about omitted trends. The mitigation, the approach-accuracy control and the InSight counterfactual, is designed against this class of monotone confound.

A second methodological observation concerns how the entry-guidance literature measures its own success. The figure of merit in [\[3\]](#ref-3), [\[4\]](#ref-4), and [\[52\]](#ref-52) is delivered dispersion at parachute deploy or at landing, computed from a Monte Carlo ensemble that disperses the entry state and the atmosphere and propagates the closed-loop trajectory. This is the same construct the dissertation's dependent variable encodes, the design landing ellipse, which means the entry-guidance literature is not merely adjacent to the measurement but is in fact one of its sources: the design ellipse for a guided mission is the output of precisely the kind of dispersion analysis these papers perform. The implication, which Chapter 4 develops under the Kuznetsian comparability heading, is that the dependent variable inherits whatever simulation conventions the source dispersion analyses used, and those conventions are not constant across missions. The reference-path guidance evaluated in [\[3\]](#ref-3) disperses a different ensemble, under different atmosphere assumptions, than the predictor-corrector formulations of [\[34\]](#ref-34) or the linear-covariance analyses validated in [\[44\]](#ref-44). The review draws the conclusion that the design ellipse is a convention-laden product of the very guidance studies that establish the lever, which is why the dissertation refuses to treat the ellipse numbers as homogeneous and insists on attaching a provenance to each.

The interpretive synthesis of this body is as follows. Guided lifting entry rests on a propositional base, entry aerodynamics, atmospheric reconstruction, and onboard navigation, that the literature shows to be deep, actively developed, and extensible from robotic to human scale and from reference-path to predictor-corrector formulations [\[3\]](#ref-3), [\[4\]](#ref-4), [\[34\]](#ref-34), [\[43\]](#ref-43), [\[44\]](#ref-44), [\[51\]](#ref-51), [\[64\]](#ref-64), [\[87\]](#ref-87). That depth is the engineering counterpart of Mokyr's claim that a technique resting on a wide propositional base is self-correcting, and it is observable in the literature itself: the same guidance lineage that began with reference-path tracking for MSL [\[4\]](#ref-4) now extends to six-degree-of-freedom dual-quaternion formulations for human-scale vehicles [\[43\]](#ref-43), which is what extensibility looks like in the documentary record. The convergence of these sources is that guided entry can null a large component of hypersonic dispersion; what none of them does, because none was designed to, is quantify how much of the realized historical ellipse contraction is attributable to that lever as against the others or against approach accuracy, and none separates the guidance contribution from the coincident atmospheric-fidelity improvement. This evidence supports the per-mission capability claim at high confidence and leaves the cross-mission attribution open, which is the gap.

## 3.3 The range-to-go parachute trigger and footprint reduction

The second lever is the range-to-go parachute trigger, and the dissertation treats it as an incremental rather than a macro innovation: a refinement of existing parachute-deploy logic that bought a further, smaller footprint reduction on top of guided entry. The single most important source for this lever is Way's analysis of the range trigger for MSL [\[5\]](#ref-5). The method is a direct comparison, by Monte Carlo simulation, of a navigated velocity trigger, which commands parachute deploy when the vehicle reaches a target Mach number or velocity, against a range-to-go (Smart Chute) trigger, which commands deploy based on navigated range to the target so as to compensate for how far downrange the vehicle actually is. The finding is that the range trigger reduces the landing footprint relative to the velocity trigger, because it converts a quantity the vehicle knows poorly at deploy (its exact velocity relative to the dispersed atmosphere) into a decision keyed on a quantity the guided vehicle knows comparatively well (its navigated range). This is the primary quantitative source for the range-trigger indicator and for the proposition that the lever attacks a distinct, smaller error source than guided entry does.

The limitation of [\[5\]](#ref-5), for attribution, is twofold. First, it is again a single-mission, single-vehicle study: it quantifies the footprint reduction the range trigger achieves for MSL, not its contribution to the cross-mission series. Second, because MSL flew guided entry and the range trigger together, the realized MSL ellipse reflects both, and Way's paper, which isolates the trigger effect in simulation by holding the rest of the EDL chain fixed, gives the dissertation a basis for expecting the range-trigger increment to be smaller than the guided-entry increment but does not by itself establish that ordering in the realized data. The ordering is a proposition the data must confirm.

The parachute-systems literature supplies the engineering context that makes the range-trigger lever intelligible as an incremental refinement of a mature subsystem. Witkowski and Kandis document the reefing of the MSL parachute, the largest supersonic parachute flown at Mars at the time [\[47\]](#ref-47), and Witkowski's earlier overview of the Phoenix parachute decelerator system describes the heritage decelerator on which the later designs built [\[48\]](#ref-48). The method in both is decelerator-system design and qualification testing; their value to the dissertation is that they establish the parachute as a long-lived, incrementally improved subsystem whose deploy logic, not whose canopy, is the locus of the range-trigger innovation. This is the engineering evidence for the Mokyrian classification of the range trigger as incremental: the macro change (guided entry) altered the flight regime from ballistic to lifting, while the range trigger refined the deploy decision within an existing parachute architecture. Crain and Bishop's analysis of Mars entry navigation from atmospheric interface through parachute deploy supplies the navigation-side context for why a range-keyed trigger is feasible only once the vehicle navigates its range well during entry [\[46\]](#ref-46), which ties the range-trigger lever back to the guided-entry lever it depends on. The reverse dependence is part of why the indicators are collinear, and the review flags it here so that Chapter 5's identification discussion inherits it rather than discovering it.

The interpretive synthesis is that the range trigger is well documented as a footprint-reducing refinement [\[5\]](#ref-5), well situated in a mature parachute and entry-navigation literature [\[46\]](#ref-46), [\[47\]](#ref-47), [\[48\]](#ref-48), and correctly classified as incremental relative to guided entry. The convergence of these sources supports, at high confidence, the claim that the range trigger reduces the footprint and, at moderate confidence, the claim that its effect is smaller than guided entry's; the cross-mission magnitude of its contribution remains, like guided entry's, unattributed in the existing literature.

## 3.4 Terrain-relative navigation and the Lander Vision System

The third lever, terrain-relative navigation, is the one the dissertation treats as a second macro-invention, because it introduced an entirely new error source to attack, position-knowledge error, that the first two levers left untouched, and it enabled a capability, autonomous hazard avoidance and divert, that had no predecessor on a Mars lander. The TRN literature is the richest single theme in the corpus, and the review treats it in three layers: the flight demonstration on Mars 2020, the onboard mapping that makes TRN possible, and the longer development lineage of vision-aided navigation that brought TRN to flight readiness.

The flight-demonstration layer is anchored by Johnson and colleagues' report of the Mars 2020 Lander Vision System flight performance [\[8\]](#ref-8). The method is post-flight analysis of the LVS as it actually operated during the Perseverance descent into Jezero, using the Safe Target Selection algorithm to localize the descent vehicle against an onboard map and to select a reachable safe target. The finding is that TRN, via the LVS, delivered Perseverance to a hazardous, scientifically rich site that earlier systems could not have attempted, by collapsing position-knowledge error and enabling an autonomous divert. This is the primary source for the TRN indicator and for the achieved-precision claim at a hazardous site. Its limitation, for the dissertation, is that it is a single flight, the only Mars mission to date with TRN, so the TRN indicator is identified off one mission, which is why the design leans so heavily on the InSight counterfactual to break the time-technology collinearity rather than on within-TRN variation. The NTRS reconstruction by Dutta and colleagues of the Mars 2020 EDL GNC and Safe Target Selection performance supplements [\[8\]](#ref-8) with the flight-data reconstruction of how LVS localized the vehicle and how STS selected the target within the reachable region [\[78\]](#ref-78); its method is post-flight GNC reconstruction, and it contributes the as-flown detail that the TRN output was handed to the MSL-heritage powered-descent system to execute, which matters because it confirms TRN added a localization-and-divert capability on top of, rather than in place of, the guided-entry and range-trigger chain.

The onboard-mapping layer is essential because TRN is only as good as the reference map it matches against, and the construction of that map is itself a maturing propositional base in Mokyr's sense. Cheng and colleagues document making the onboard reference map from Mars Reconnaissance Orbiter Context Camera (MRO/CTX) imagery for the Mars 2020 LVS [\[9\]](#ref-9). The method is photogrammetric: orthorectify and georeference orbital imagery into a base map against which descent images can be matched in real time. The finding is that such a map can be built at the resolution and accuracy LVS requires; the evidence it supplies the dissertation is that TRN rests on a deep, separately maturing propositional base (orbital mapping and computer vision), which is the Mokyrian extensibility argument made concrete. Tao and colleagues' super-resolution restoration of TGO CaSSIS imagery, demonstrated specifically with the Perseverance landing site, shows the same mapping base advancing on a separate orbital sensor [\[62\]](#ref-62), and Parkes and colleagues' PANGU planet-surface simulation tool shows the test-and-validation infrastructure that lets vision-based navigation be exercised against representative synthetic terrain before flight [\[60\]](#ref-60). The methodological point the review draws from this layer is that TRN is not a self-contained black box; it is the visible flight insertion of a knowledge chain (orbital imaging, map construction, image matching) whose health is a precondition for the capability, which is the basis for the Mokyrian reversibility caution the dissertation carries into its discussion.

The development-lineage layer establishes that TRN reached flight readiness along a long, traceable path, which both strengthens the causal interpretation (the capability did not appear by accident at Mars 2020) and supplies the technology-readiness narrative the indicator encodes. Johnson and colleagues' general approach to terrain-relative navigation for planetary landing lays out the early architecture [\[45\]](#ref-45); Mourikis and colleagues' vision-aided inertial navigation (VISINAV) work, in both its journal and conference forms, develops and experimentally validates the tight integration of inertial measurements with mapped-landmark and opportunistic-feature observations that underlies precise planetary landing [\[53\]](#ref-53), [\[54\]](#ref-54); Johnson and colleagues' real-time TRN test results from a relevant environment for Mars landing demonstrate the system working in representative conditions ahead of Mars 2020 [\[10\]](#ref-10); Trawny and colleagues' flight testing of TRN and large-divert guidance on the Masten Xombie VTVL rocket demonstrates the combined localize-and-divert capability on a real powered-descent vehicle [\[91\]](#ref-91); and Johnson and colleagues' map-relative localization system paper generalizes the LVS approach into a reusable architecture for planetary landing [\[11\]](#ref-11). Woicke and Mooij's stereo-vision TRN using hazard-mapping measurements [\[55\]](#ref-55) and the more recent reviews and demonstrations by Sutterlin and Eapen [\[66\]](#ref-66), Abdhul Rahuman and Lee [\[67\]](#ref-67), Kim and Singh [\[82\]](#ref-82), and Silvestrini [\[79\]](#ref-79) show the field continuing to broaden after Mars 2020. The survey by Wang and colleagues of planetary landings with terrain sensing and hazard avoidance situates the three levers, and TRN especially, within the broader precision-landing field [\[12\]](#ref-12).

It is worth dwelling on the development lineage because it carries an inferential weight the dissertation relies on. The point it establishes is that TRN's appearance at Mars 2020 was the flight insertion of a deliberately matured capability rather than a coincidence of calendar timing. The evidence is the traceable sequence from the early general approach [\[45\]](#ref-45), through the experimentally validated vision-aided inertial navigation of Mourikis and colleagues [\[53\]](#ref-53), [\[54\]](#ref-54), through relevant-environment testing for Mars landing [\[10\]](#ref-10), through the Masten VTVL flight test of combined TRN and large-divert guidance [\[91\]](#ref-91), to the flight system itself [\[8\]](#ref-8), [\[78\]](#ref-78). A capability demonstrated in this order, laboratory to relevant environment to free-flight rocket to mission, is a matured insertion rather than an accident, and the reason that inference holds is the technology-readiness-level convention itself, under which each demonstration step raises the documented readiness of the system, a convention the capability-program literature of Section 3.5 makes explicit [\[114\]](#ref-114), [\[116\]](#ref-116), [\[123\]](#ref-123). The review is careful about the limit of this reasoning: the lineage strengthens the causal reading of the TRN indicator, but it does not quantify the TRN coefficient, because the readiness narrative is about whether TRN could fly, not about how much it shifted the realized ellipse. The objection it must answer is that the readiness lineage and the ellipse contraction could both be expressions of a third monotone factor; the answer is again the InSight counterfactual, which holds the calendar roughly fixed while removing the guidance suite. Mourikis and colleagues' work deserves particular weight in this chain because it is one of the few sources in the corpus that experimentally validates, rather than only simulates, the tight integration of inertial and visual measurements that underlies terrain-relative position estimation [\[53\]](#ref-53), [\[54\]](#ref-54); its method is algorithm development with hardware-in-the-loop and field experiments, its finding is that mapped-landmark and opportunistic-feature observations can be fused with inertial data to yield accurate terrain-relative position, attitude, and velocity in real time, and its limitation for the dissertation is that it predates and is upstream of any Mars flight, so it backs the mechanism without contributing a series observation.

A final interpretive point on the TRN theme concerns the breadth of the post-Mars-2020 literature [\[55\]](#ref-55), [\[66\]](#ref-66), [\[67\]](#ref-67), [\[79\]](#ref-79), [\[82\]](#ref-82) and what that breadth does and does not establish. The recent reviews and demonstrations show TRN diversifying across sensors (stereo vision, lidar, learned features), across bodies (Mars, Moon, asteroids), and across estimation paradigms (filtering, deep regression). This breadth is genuine evidence for Mokyrian extensibility: a technique that proliferates across sensors and bodies is resting on a propositional base wide enough to support variation, which is the property Mokyr attributes to self-correcting techniques. The review is careful not to overclaim from it. None of these post-2020 works is a Mars ellipse observation, and several are terrestrial or simulation studies; they enter the dissertation as backing for the extensibility argument and as external-validity reference points, not as data on the learning curve. The confidence they support is high for the proposition that TRN is a general and extensible capability, and they support no incremental confidence at all on the magnitude of TRN's contribution to the Mars series, which remains identified off the single Mars 2020 flight.

The interpretive synthesis of the TRN theme is the strongest single block of evidence in the chapter for the dissertation's mechanism, and also the sharpest illustration of its identification problem. The convergence of the flight, mapping, and lineage layers establishes at very high confidence that TRN is a real, distinct capability resting on a deep and separately maturing propositional base, that it attacks position-knowledge error that the other two levers cannot touch, and that it enabled autonomous divert with no predecessor [\[8\]](#ref-8), [\[9\]](#ref-9), [\[10\]](#ref-10), [\[11\]](#ref-11), [\[12\]](#ref-12), [\[45\]](#ref-45), [\[53\]](#ref-53), [\[54\]](#ref-54), [\[78\]](#ref-78), [\[91\]](#ref-91). What the literature does not and cannot supply is the cross-mission magnitude of TRN's contribution to the ellipse series, because TRN has flown on exactly one Mars mission. The TRN indicator is therefore identified off a single observation, and the honest confidence on any TRN coefficient is correspondingly bounded. The review states this plainly because it is the central reason the dissertation's strongest defensible claim is a signed, counterfactual-surviving attribution rather than a precise coefficient.

## 3.5 Hazard detection and avoidance and precision-landing capability programs
The fourth theme is the set of dedicated precision-landing and hazard-detection-and-avoidance (HDA) capability programs that matured TRN and divert from laboratory concepts to flight-ready systems. This theme matters to the dissertation for two reasons. The first concerns the technology-readiness trajectory that the technology indicators implicitly assume. The indicators are binary (flew or did not), but the reason a technology could fly on a given mission is that a capability program had carried it to the required readiness level, and that program history is the corroborating record for coding the indicators in the absence of a single citable TechPort document. The second is that the HDA literature treats precision landing and hazard avoidance as a general planetary capability rather than a Mars-specific artifact, which bears on the external-validity discussion the dissertation bounds in Chapter 7.

The principal capability program is Autonomous Landing and Hazard Avoidance Technology (ALHAT), documented across a sequence of papers. Epp and Smith describe the ALHAT charter to place humans and cargo safely, precisely, and repeatedly on the lunar surface with hazard-avoidance capability, with target landing accuracies ranging from hundreds of meters for sortie missions to tens of meters for outpost-class missions [\[120\]](#ref-120). Epp, Robertson, and Brady describe the integrated autonomous guidance, navigation, and control hardware-and-software system and the interdependencies, lander hazard robustness, terrain, lighting, trajectories, sensors, that drive its design [\[114\]](#ref-114). Rutishauser, Epp, and Robertson report the free-flight terrestrial rocket-lander demonstration that carried the ALHAT system toward Technology Readiness Level six [\[123\]](#ref-123), and Epp, Robertson, and Carson report the real-time HDA demonstration for a planetary lander, again maturing the integrated system toward TRL six [\[116\]](#ref-116). Ivanov, Huertas, and Carson develop the probabilistic hazard-detection algorithm at the perception core of ALHAT [\[115\]](#ref-115), and Huertas and colleagues evaluate HDA-algorithm performance for safe lunar landings, addressing the measurement problem of how to characterize such a system's performance for engineers planning landed missions [\[70\]](#ref-70). Robertson's synopsis of precision-landing and hazard-avoidance capabilities for space exploration ties the program record together and articulates the shift from controlled blind landings, where dispersion is set by inertial navigation and surface-relative sensing, to TRN-enabled tight control over the actual landing location [\[85\]](#ref-85).

The method common to this theme is integrated system development with progressively higher-fidelity demonstration, terrestrial rocket flights, relevant-environment tests, and TRL accounting, rather than Mars flight reconstruction. The finding the dissertation draws from it is that the capability the indicators encode was matured deliberately and traceably, which strengthens the causal reading of the TRN indicator: TRN appeared at Mars 2020 because a capability program put it there, not because of a coincident calendar trend. The limitation, for the series, is that these programs are predominantly lunar-targeted and terrestrial-demonstration efforts. They corroborate the readiness narrative but they are not themselves Mars ellipse observations, so they enter the dissertation as backing for the indicators and as external-validity reference points, not as dependent-variable data. Jiang, Li, and Tao's HDA strategy and guidance work [\[41\]](#ref-41), [\[109\]](#ref-109) extends the HDA literature on the algorithmic side, and the more recent terrain-sensing-lidar work of Amzajerdian and colleagues [\[128\]](#ref-128), [\[129\]](#ref-129) and the SURF-based lunar TRN algorithm of Newcomb [\[130\]](#ref-130) show the sensing base for HDA and TRN continuing to advance toward Artemis-era lunar landings.

The interpretive synthesis is that the HDA and capability-program literature supplies the technology-readiness backing for the indicators at high confidence, and supplies the generality-of-capability claim (precision landing and hazard avoidance are not Mars-only) at high confidence, while contributing no direct ellipse-series data. Its relation to the gap is that it explains why a technology could be inserted on a given mission without explaining how much that insertion shifted the realized ellipse, which is the attribution the dissertation supplies.

## 3.6 Supersonic retropropulsion and high-mass, human-class EDL

The fifth theme is the forward edge of the field: supersonic retropropulsion (SRP) and the high-mass, human-class EDL architectures that motivate it. This theme is not in the series, none of the missions from Viking through Mars 2020 used SRP, but it matters to the dissertation in three ways. It establishes why landing precision is a decision-critical variable for the architectures the dissertation's result is meant to inform; it shows the propositional base under EDL continuing to deepen, which is the Mokyrian extensibility argument projected forward; and it demonstrates that the problem is material, because the next generation of missions cannot proceed without a quantified accuracy basis.

Korzun and colleagues' concept for the EDL of high-mass payloads at Mars frames the core difficulty: the Viking-era deceleration technologies do not scale to the tens of metric tons that human exploration requires [\[33\]](#ref-33). Edquist and colleagues develop SRP specifically for future Mars EDL systems, proposing the maturation milestones the technology needs [\[111\]](#ref-111), and Braun, Sforzo, and Campbell advance SRP using Mars-relevant flight data, drawing on the supersonic retropropulsion maneuvers SpaceX performed with the Falcon 9 first stage as the first firings of a rocket into an opposing supersonic freestream [\[40\]](#ref-40). Sreelekshmi and colleagues develop guidance with SRP for Mars pinpoint landing, connecting the deceleration technology to the precision objective [\[126\]](#ref-126). The human-scale architecture studies, Polsgrove and Dwyer-Cianciolo's EDL architecture study overview [\[112\]](#ref-112), the rigid-decelerators follow-on [\[113\]](#ref-113), and Lillard and Olejniczak's human-Mars EDL pathfinder study of technology-development gaps and mitigations [\[122\]](#ref-122), establish the requirements context: human-class missions assume delivery of multiple large payloads to designated locations with landing precision on the order of tens of meters, and the architecture studies exist to prioritize the technology investments that close the gap to that precision [\[86\]](#ref-86).

The method across this theme is concept design, wind-tunnel and flight-data-anchored technology assessment, and architecture trade study. The finding the dissertation needs is that landing precision is, for these architectures, an explicit and binding requirement that drives multi-year technology-investment decisions, which is the evidence that the problem is material. The limitation is that none of this is series data, and SRP in particular postdates the entire population, so the theme enters as forward context and as the strategic motivation for the contribution, not as evidence on the learning curve itself. Lugo and colleagues' integrated precision-landing performance and technology assessment of a human-scale Mars lander using a generalized simulation framework is the clearest bridge: it assesses, in simulation, how candidate technologies trade against a stated landing-precision requirement, which is the requirements-setting and investment-valuation use the dissertation's executed result would serve [\[86\]](#ref-86).

The interpretive synthesis is that the SRP and human-class theme supplies, at high confidence, the materiality of the landing-precision variable for the architectures now under study, and supplies, at moderate confidence, the projection of Mokyrian extensibility forward to human scale; it supplies no series data and is correctly held out of the population.

## 3.7 Mission-context studies, instruments, and the public record

The sixth theme is the mission-context and instrument literature that establishes the public record from which the ellipse series and its controls are reconstructed. This theme is methodologically load-bearing in a way that is easy to underrate: the dependent variable of the dissertation is a constructed measurement, and its comparability across missions depends on the quality and consistency of the reconstruction record. Kuznets's discipline, that an aggregate is meaningless without a stated boundary of coverage and a consistent valuation convention, applies directly, and this theme is where the boundary is set.

The mission-overview record spans the population. Crisp and colleagues describe the Mars Exploration Rover mission, whose cruise and EDL design built on the Mars Pathfinder configuration [\[59\]](#ref-59), and Desai and Knocke provide the MER EDL trajectory analysis from which the MER ellipse is reconstructed [\[57\]](#ref-57). Desai and colleagues' Phoenix EDL papers, the operations analysis [\[36\]](#ref-36) and, with Edquist and Schoenenberger, the Phoenix entry-capsule aerodynamics [\[35\]](#ref-35), together with Desai's Phoenix performance paper [\[6\]](#ref-6), establish Phoenix as a ballistic-entry modern lander whose design ellipse is a within-period control on the guided-versus-unguided contrast. Grotzinger and colleagues describe the MSL mission and science investigation and its Gale crater site [\[39\]](#ref-39), and the Mars 2020 EDL system overview by Nelessen and colleagues describes the architecture that carried Perseverance, including the TRN insertion, building explicitly on MSL heritage [\[121\]](#ref-121). The instrument literature, the Mars 2020 engineering cameras and microphone designed in part to document EDL events [\[38\]](#ref-38), Mastcam-Z [\[42\]](#ref-42), SHERLOC [\[37\]](#ref-37), and ChemCam on MSL [\[56\]](#ref-56), together with the geomorphic mapping of the Mars 2020 landing ellipse at the candidate Northeast Syrtis site [\[68\]](#ref-68) and the reporting of the first Jezero samples [\[61\]](#ref-61), establishes the dense public record that makes each mission's landing and site reconstructable. The MEDLI and MEDLI2 instrumentation papers, Hwang and colleagues' MEDLI2 description [\[73\]](#ref-73), Bose and colleagues' companion MEDLI2 paper [\[76\]](#ref-76), and the reconstructions they enable, are the instrumented backbone of the late-mission reconstructions.

The method across this theme is mission and instrument description and, in the reconstruction subset, statistical trajectory reconstruction from flight data. The finding the dissertation draws is that a consistent, mission-by-mission ellipse series can be reconstructed from the public record, because every mission in the population has a documented design ellipse and, for the instrumented missions, a flight-data reconstruction against which that design ellipse can be checked. The limitation, and it is the Kuznetsian one the dissertation foregrounds, is that the simulation conventions behind the design ellipse changed across missions: the number of sigmas reported, the treatment of the atmosphere, and the Monte Carlo fidelity are not constant from Viking to Mars 2020, so the series is only as comparable as the reconstruction work allows, and Chapter 4 must state the convention attached to each value rather than treating the numbers as homogeneous. This theme is where that comparability problem is sourced, and it is the reason the dissertation refuses to report a single learning rate without first stating the measurement boundary.

The MER pair deserves a separate interpretive note because of the role it plays in the experience-axis specification. Spirit and Opportunity were near-identical vehicles landed weeks apart in 2004 [\[59\]](#ref-59), both built on the Pathfinder cruise-and-entry configuration [\[57\]](#ref-57), and both delivered by ballistic entry with airbag landing into ellipses on the order of one hundred kilometers. For the dissertation this near-identity is not a redundancy to be discarded; it is information. It supplies a within-program replication that lets the design test whether the experience axis is better measured by mission sequence (in which Spirit and Opportunity are two steps) or by program (in which they are one), and it provides a check on measurement noise, because two near-identical vehicles should produce near-identical design ellipses if the reconstruction is consistent. The MER trajectory analysis of Desai and Knocke is the source for both ellipses [\[57\]](#ref-57), and its method, pre-flight trajectory and dispersion analysis, is the same family of Monte Carlo dispersion study that produces the design ellipse throughout the series, which reinforces the comparability the dissertation is building toward while inheriting the convention-drift caveat. The Viking and Pathfinder anchors at the high end of the series rest on older records, and the review flags, consistent with the evidence-gaps section of the shared bible, that the earliest three-sigma ellipse values may sit only in NTRS reconstructions rather than in DOI-bearing journals, so Chapter 4 must source them explicitly and flag any value whose simulation convention cannot be pinned down. The Pathfinder entry reconstruction is available [\[71\]](#ref-71), and the Pathfinder dispersion analysis of Spencer and Braun [\[84\]](#ref-84) supplies the methodological precedent, but the Viking-era values are the most provenance-fragile points in the series and are treated as such.

The instrument literature has a more specific function than establishing a general public record, and the review draws it out because it bears on the secondary dependent variable. The Mars 2020 engineering cameras were designed in part to document the EDL sequence itself [\[38\]](#ref-38), and the dense surface-imaging payload, Mastcam-Z [\[42\]](#ref-42), SHERLOC [\[37\]](#ref-37), and the orbital-imaging base that super-resolution methods sharpen [\[62\]](#ref-62), is what makes post-landing localization against a georeferenced map possible. Post-landing localization is the source of the achieved miss distance that the dissertation uses as a construct-validity check on the design ellipse, so the instrument literature is not merely context; it is the enabling record for the robustness dependent variable. The relation to the gap is that the same imaging that lets a project confirm where a vehicle actually landed, and so compute achieved miss distance, has never been assembled into a cross-mission achieved-accuracy series to be set alongside the design-ellipse series. Building both and comparing them is part of the dissertation's construct-validity design, and the literature supplies the ingredients for both without having combined them.

The interpretive synthesis is that the mission-context and instrument literature supplies, at high confidence, the existence and reconstructability of a mission-by-mission ellipse series and of a parallel achieved-accuracy series, and supplies, at the same time, the explicit warning that the design-ellipse series is a constructed measurement with drifting conventions whose earliest values are the most provenance-fragile. Its relation to the gap is that it makes both series buildable while making clear that no prior work has built either as one comparable, decomposed object, let alone set them against each other.

## 3.8 Synthesis: levers to error sources, and what each literature contributes

The thematic reading above is synthesized in two tables. The first maps each technology lever to the physical error source it attacks, which is the justification for treating the three levers as separate covariates rather than as a single modernity index; the mapping is the engineering counterpart of Mokyr's claim that each technique rests on its own propositional base. The second classifies the major literature themes by what each contributes to the dissertation and where each stops short, which is the explicit statement of the gap in tabular form.

**Table 3.1. Technology levers, the error source each attacks, and the primary literature.**

| Lever | Physical error source attacked | Mokyrian class | Primary sources | What the source establishes | What it does not establish |
|---|---|---|---|---|---|
| Guided lifting entry | Downrange and crossrange dispersion accumulated during the hypersonic phase | Macro-invention (ballistic to lifting) | S03, S04, S01, S02, R034, R052 | Closed-loop bank modulation nulls hypersonic dispersion for an MSL-class vehicle | Cross-mission magnitude of its contribution to the realized series |
| Range-to-go parachute trigger | Dispersion introduced at parachute deploy | Incremental refinement of deploy logic | S05, R046, R047, R048 | Range-keyed deploy reduces footprint relative to a velocity trigger | That its effect is smaller than guided entry's in the realized data |
| Terrain-relative navigation (LVS) | Position-knowledge error; enables autonomous divert | Macro-invention (new error source, new capability) | S08, S09, S10, S11, S12, R045, R053, R054, R078, R091 | TRN collapses position-knowledge error and enables divert at a hazardous site | Magnitude of TRN's contribution (one Mars flight only) |
| Approach navigation (control, not a lever) | Delivered entry-state dispersion at atmospheric interface | Not an onboard EDL lever | R046, R077, R078 (and Chapter 4 reconstructions) | How accurately the vehicle was delivered to the interface | Whether onboard guidance or approach accuracy binds the ellipse |

**Table 3.2. Literature themes, contribution, and gap.**

| Theme (section) | Method common to the theme | Contribution to the dissertation | Where it stops short |
|---|---|---|---|
| Guided entry (3.2) | Monte Carlo dispersion simulation; design documentation | Codes the guided-entry indicator; quantifies per-mission dispersion reduction | No cross-mission series; bundles guided entry with the range trigger at MSL |
| Range trigger (3.3) | Trigger comparison by simulation; decelerator qualification | Codes the range-trigger indicator; classifies the lever as incremental | No realized-series magnitude; collinear with guided entry |
| TRN and LVS (3.4) | Flight reconstruction; mapping photogrammetry; vision-nav development | Codes the TRN indicator; documents the deep propositional base | One Mars flight; TRN coefficient identified off a single observation |
| HDA and capability programs (3.5) | Integrated system demonstration; TRL accounting | Backs the readiness narrative behind the indicators; generality of capability | Lunar/terrestrial, not Mars ellipse data |
| SRP and human-class EDL (3.6) | Concept design; architecture trade study | Materiality of landing precision for future architectures | Postdates the population; not series data |
| Mission context and instruments (3.7) | Mission/instrument description; statistical reconstruction | Makes the series reconstructable; sets the measurement boundary | Conventions drift across missions; no prior work builds it as one series |
| Learning curve and anchors (3.2 framing, Ch. 2) | Empirical regularity estimation; economic-history analysis | Supplies the model form and the mechanism/measurement discipline | No aerospace-specific prior rate; that is the contribution |

Two readings follow from the tables. The first is that the levers are physically separable. Each attacks a distinct error source, so treating them as separate covariates is justified by the engineering, not merely convenient for the statistics; this is what justifies the augmented specification of the model. The reservation is that physical separability does not guarantee statistical separability in a sample of nine to eleven, and the near-collinearity of the indicators with each other and with calendar time means the design must lean on the nested specifications and the InSight counterfactual rather than on a single over-parameterized regression. The second is that every theme contributes a necessary ingredient, an indicator, a control, a dependent-variable record, a readiness narrative, a materiality argument, and not one of them, nor all of them together, performs the cross-mission attribution. The gap is not a gap in any single literature; it is the gap between literatures.

## 3.9 The gap and the propositions that follow

The current state of the literature, read thematically, is a field that has solved, to high standards of rigor, every sub-problem the dissertation depends on while leaving the joint problem untouched. The desired state is a single constructed ellipse series, fitted with a learning-rate model and decomposed by technology generation, that attributes the contraction to identifiable onboard-guidance insertions while holding approach accuracy constant. The gap is the absence of any work that joins the EDL-engineering record to the technology-economics learning-curve apparatus and arbitrates among the rival causes. The consequence of the gap is that the most consequential number in Mars mission design, the landing ellipse, is set qualitatively, and guidance investments cannot be valued against approach-navigation investments on a common quantitative basis [\[86\]](#ref-86), [\[112\]](#ref-112), [\[113\]](#ref-113), [\[122\]](#ref-122).
The causal mechanism the gap obscures, and that the dissertation's design is built to expose, runs as follows. The driver is the insertion of an onboard EDL-guidance technology resting on a maturing propositional base: entry aerodynamics and atmospheric reconstruction under guided entry [\[3\]](#ref-3), [\[4\]](#ref-4), [\[52\]](#ref-52), entry navigation under the range trigger [\[5\]](#ref-5), [\[46\]](#ref-46), and computer vision and orbital mapping under TRN [\[8\]](#ref-8), [\[9\]](#ref-9), [\[45\]](#ref-45), [\[53\]](#ref-53). Each technology removes a physically distinct error source: guided entry nulls hypersonic dispersion, the range trigger corrects parachute-deploy dispersion, and TRN collapses position-knowledge error and enables divert [\[8\]](#ref-8), [\[12\]](#ref-12). The observable effect is a discrete drop in the three-sigma ellipse at the mission that first flies the technology. A smaller ellipse lets a project target hazardous, scientifically rich terrain that a larger ellipse would force it to avoid, as Mars 2020 did at Jezero [\[8\]](#ref-8), [\[61\]](#ref-61), [\[68\]](#ref-68). Landing precision thus becomes a quantifiable, technology-attributed design variable for the human-Mars and sample-return architectures now being studied [\[112\]](#ref-112), [\[122\]](#ref-122), [\[127\]](#ref-127). Where the data permit only correlation, because the indicators are collinear and the sample is small, the dissertation says so and downgrades confidence, per the program's epistemic discipline; the mechanism is named so that the attribution rests on a stated physical pathway rather than on a bare cross-mission correlation.

From the gap and the mechanism, three propositions follow that the remainder of the dissertation operationalizes and tests. They are stated here as the literature-derived basis for the hypotheses fixed in the approved design document, not as restatements of them.

**Proposition 1 (decomposability).** Because each lever attacks a physically distinct error source (Table 3.1), the cross-mission ellipse contraction is decomposable into technology-attributable components rather than being a single undifferentiated trend. The evidence is the entry-guidance, range-trigger, and TRN literatures, which document three separable mechanisms [\[3\]](#ref-3), [\[5\]](#ref-5), [\[8\]](#ref-8), [\[9\]](#ref-9), [\[12\]](#ref-12), and what licenses reading them as decomposable is Mokyr's claim that separable techniques rest on separable propositional bases. Decomposability in physics does not guarantee identifiability in a sample of nine to eleven, and the design must answer the objection that an omitted monotone factor (modeling fidelity, computing) could masquerade as a technology effect. The approach-accuracy control and the InSight counterfactual are built in for that reason.

**Proposition 2 (step structure).** Because guided entry and TRN are macro-inventions and the range trigger is incremental (Table 3.1), the contraction should show its largest discrete drops at the missions that first fly guided entry (MSL) and TRN (Mars 2020), with a smaller drop attributable to the range trigger, rather than a featureless decline. The evidence is the simulation-quantified per-lever effects [\[3\]](#ref-3), [\[5\]](#ref-5), [\[8\]](#ref-8), read through the Mokyrian macro-versus-incremental distinction. MSL bundles guided entry and the range trigger, so the macro and incremental contributions at that mission cannot be fully separated from the realized data alone, and the standing objection is the pure-time-trend explanation, which Proposition 3 confronts directly.

**Proposition 3 (technology over approach accuracy).** Because the levers attack onboard error sources that are physically distinct from delivered entry-state dispersion (Table 3.1, final row), the contraction should load on the technology indicators rather than on the approach-accuracy control, and a late, ballistic, unguided lander should inherit a large ellipse despite its late date. The evidence is the entry-navigation and approach-delivery literature [\[46\]](#ref-46), [\[77\]](#ref-77), [\[78\]](#ref-78) and the InSight landing-site record, which documents a modern lander that deliberately accepted a large flat ellipse without the guidance suite [\[15\]](#ref-15), [\[74\]](#ref-74), [\[14\]](#ref-14); what connects them is the separability of an onboard mechanism from an approach-delivery mechanism as error sources. The InSight discrimination rests on a single observation, so it can distinguish technology from pure time but cannot apportion credit among the three levers. The contrary explanation, that better approach navigation alone shrank the ellipse, is precisely H0, which the augmented specification is built to test by including the approach-accuracy control.

These three propositions are the literature-grounded content of the falsifiable contribution. Proposition 1 motivates modeling discrete technology generations as covariates rather than fitting a single smooth trend. Proposition 2 predicts the specific step pattern the nested specifications will look for. Proposition 3 sets up the contrast between the onboard-guidance explanation (H1) and the approach-accuracy explanation (H0), and identifies the InSight counterfactual as the linchpin that breaks the otherwise perfect time-technology collinearity. None of the three can be settled by any source reviewed in this chapter, because no source assembled the series or ran the attribution; all three are testable on the data that the named datasets, sourced through this literature, make available. That is the precise sense in which the literature supplies every ingredient and leaves the dish uncooked, and it is the gap the dissertation closes.

Before the closing note on confidence, the argument that runs beneath the whole chapter can be stated compactly, because the thematic reading has now supplied the evidence for each step. The problem is real: the ellipse contraction is large, documented across the population, and currently explained only qualitatively and mission by mission [\[1\]](#ref-1), [\[2\]](#ref-2), [\[5\]](#ref-5), [\[6\]](#ref-6), [\[8\]](#ref-8), [\[15\]](#ref-15). The problem is material: the ellipse governs which sites are reachable and is decision-critical for the sample-return and human-Mars architectures whose technology investments are being prioritized now [\[8\]](#ref-8), [\[15\]](#ref-15), [\[86\]](#ref-86), [\[112\]](#ref-112), [\[113\]](#ref-113), [\[122\]](#ref-122), [\[127\]](#ref-127). The design addresses the causal mechanism: modeling discrete technology generations as covariates on a convention-stated series isolates the onboard-guidance mechanism that the entry-guidance, range-trigger, and TRN literatures show to be physically separable [\[3\]](#ref-3), [\[5\]](#ref-5), [\[8\]](#ref-8), [\[9\]](#ref-9), [\[12\]](#ref-12), [\[45\]](#ref-45), [\[52\]](#ref-52), [\[78\]](#ref-78). The design discriminates the technology account from its rivals: the InSight counterfactual plus the approach-accuracy control separate the technology hypothesis from the pure-time and injection-accuracy rivals in a way no single-mission study can, because no single-mission study contains a within-period unguided control [\[14\]](#ref-14), [\[15\]](#ref-15), [\[74\]](#ref-74), [\[77\]](#ref-77). The residual risk is acceptable: the small sample and the indicator collinearity are real and are bounded, not eliminated, by the nested specifications, the permutation inference, and the dual dependent variable, and the design is informative whichever way the coefficients fall [\[7\]](#ref-7), [\[81\]](#ref-81), [\[83\]](#ref-83), [\[84\]](#ref-84). The chapter's contribution to this argument is to have shown, theme by theme, that the evidence for each step exists in the literature and that the gap is the failure to assemble it, not a shortage of it.

A closing note on confidence and on what would move it. The per-mission engineering claims that ground the three propositions are high to very high confidence, resting on flight data and validated dispersion analyses [\[8\]](#ref-8), [\[52\]](#ref-52), [\[78\]](#ref-78). The cross-mission attribution that the propositions point toward is, at the design stage and in advance of execution, moderate confidence at best, bounded by the sample size and the indicator collinearity that this review has flagged repeatedly. What would raise that confidence is exactly what the design provides and the existing literature lacks: a comparable, convention-stated series (raising construct and statistical-conclusion confidence), the InSight counterfactual sitting above the trend (raising internal-validity confidence on Proposition 3), and an achieved-miss-distance series from PDS confirming the design-ellipse contraction (raising construct confidence). What would lower it is a finding that the approach-accuracy control absorbs the technology terms, or that InSight sits on the trend, either of which would shift the weight toward H0. The literature reviewed here cannot deliver any of those tests; it can only, and it does, establish that they are the right tests and that the data to run them exist. The dissertation's contribution begins exactly where this literature ends.


# Chapter 4: Data and Measurement

## 4.1 The chapter thesis

The three named datasets of this dissertation, the NASA Technical Reports Server (NTRS) landing-accuracy reconstructions, the TechPort technology-portfolio records, and the Planetary Data System (PDS) landing-site localization products, jointly yield a comparable cross-mission landing-ellipse series sufficient to test the learning-curve hypothesis, but only if the Kuznetsian comparability problem is confronted explicitly rather than assumed away. That is the claim this chapter establishes and defends. The design landing ellipse is not a direct measurement; it is a simulation product whose generating conventions, the Monte Carlo dispersion model, the assumed atmosphere, the confidence level at which the contour is reported, and the definition of the target itself, changed materially across forty-five years of Mars exploration. A naive analyst who downloads a semi-major axis from each mission's overview paper and regresses the log of the resulting area on mission sequence will obtain a slope, and that slope will look like a learning rate, but it will conflate genuine capability improvement with drift in the measurement convention. Following Kuznets [\[24\]](#ref-24), [\[20\]](#ref-20), this chapter refuses that shortcut. It states, dataset by dataset and variable by variable, what each number means, where it comes from, how it was generated, and what would make two numbers from two different missions non-comparable. The payoff is that when Chapters 5 and 6 fit the model, every value carries its boundary with it, and the inference rests on a series that has been constructed rather than merely collected.

This is the central methodological burden of the dissertation, and it is discharged here rather than in the analysis chapters, because the discipline must precede the estimation. The contribution of the dissertation is a technology-attributed learning rate for Mars landing precision; that contribution is only as credible as the series the rate is fitted to. A reader who does not trust the series will not trust the slope, regardless of how carefully the regression is specified. Chapter 3 established that the engineering literature documents each mission and each technology in depth but never as one comparable series; this chapter shows how to forge the one series from the many records, and where the welds are weakest.

### 4.1.1 Problem frame for the data chapter

The problem this chapter addresses can be stated in the current-state, desired-state, gap, consequence form that organizes the dissertation. The current state is that the landing-ellipse numbers exist, scattered across mission-overview papers, post-flight reconstruction reports, and landing-site-selection studies, each generated under its own conventions and reported for its own purpose, with no common boundary statement. The desired state is a single tabulated series of nine to eleven mission landing events, each carrying a log ellipse area, a log achieved miss distance, three technology indicators, and an approach-accuracy control, all on stated and comparable definitions. The gap is the comparability problem: the records were never built to be compared to one another, and the conventions behind them drifted in ways that, if uncorrected, would bias the fitted slope. The consequence of failing to close that gap is that any learning rate reported would be uninterpretable, because it could not be decomposed into the part driven by capability and the part driven by convention, which is exactly the decomposition Kuznets insists must precede theorizing [\[24\]](#ref-24). This chapter closes the gap by documenting provenance, operationalizing every variable on an explicit scale, validating the design ellipse against the achieved miss distance where both exist, and stating coverage limits honestly rather than burying them.

A note on citation convention for this chapter and its siblings: references are given by their corpus key in brackets, [\[6\]](#ref-6) for the seed entries and [\[71\]](#ref-71) for the harvested entries, each resolving to a real record in `research/corpus.jsonl` with a DOI or, for the NTRS reconstruction reports that have no journal DOI, a resolvable NTRS identifier. No number in this chapter is an executed empirical estimate; every illustrative quantity is labeled as such, consistent with the design-stage guardrail that governs the whole dissertation.

## 4.2 The three named datasets in depth

### 4.2.1 NTRS landing-accuracy reconstructions: the dependent-variable source

The NASA Technical Reports Server is the primary source for the dependent variable, the design landing ellipse, and for the post-flight reconstructed performance that validates it. NTRS (ntrs.nasa.gov) is the agency's open repository of technical reports, conference papers, and journal preprints, queryable through a public citations API at ntrs.nasa.gov/api/citations/search. Access is unrestricted and requires no credential; records are retrieved by free-text and field search and, where a paper was subsequently published in a peer-reviewed venue, cross-referenced to its DOI so that the reconstruction value can be pinned to a citable source. The convergence of the NTRS holdings with the peer-reviewed EDL literature is what makes the dependent-variable series sourceable: most modern missions are documented both as an NTRS reconstruction report and as a journal or conference paper, and the two can be checked against each other.

The provenance of the dependent variable differs across the era in a way that bears directly on comparability, and the difference is worth setting out because it is the concrete form of the Kuznetsian boundary problem. For the modern missions, the design ellipse rests on a documented statistical reconstruction methodology. Karlgaard and colleagues [\[7\]](#ref-7) established the canonical approach: a statistical reconstruction of the Mars EDL trajectory and atmospheric profile from onboard and tracking data, which produces both a best-estimate trajectory and a dispersion characterization. That methodology was applied mission by mission. For MSL, the reconstruction is documented in the NTRS record by Karlgaard, Kutty, Schoenenberger, and Shidner [\[72\]](#ref-72), which implements a Kalman-filter approach to reconstruct the entry, descent, and landing trajectory from flight data. For Mars 2020, the flight-data reconstruction is carried by the MEDLI2 instrumentation suite and reported by Karlgaard and colleagues [\[13\]](#ref-13), with the instrument suite itself described in the NTRS records by Hwang and colleagues [\[73\]](#ref-73) and Bose and colleagues [\[76\]](#ref-76). For InSight, the trajectory and atmosphere reconstruction is documented by Karlgaard, Korzun, Schoenenberger, and colleagues [\[74\]](#ref-74), with a parallel communications-based reconstruction [\[14\]](#ref-14) that supports the InSight control case central to the identification strategy. For Phoenix, the EDL performance is reported by Desai and colleagues [\[6\]](#ref-6), which provides the design ellipse for a late-era ballistic-entry lander.

The earliest missions are where the provenance thins, and the dissertation states this plainly rather than papering over it. Mars Pathfinder's atmospheric entry reconstruction exists in the NTRS record by Kallemeyn, Peng, Braun, and Thurman [\[71\]](#ref-71), which characterizes the spacecraft performance during entry but predates the modern MEDLI-instrumented reconstruction era; its dispersion characterization rests on a sparser instrumentation set. The Viking-era three-sigma ellipse values rest on records older still, generated under 1970s simulation conventions and reported in documents that may exist only in the NTRS archive rather than in any DOI-bearing journal. The position the dissertation takes on this evidence can be stated precisely. The NTRS holdings are sufficient to source a comparable dependent-variable series for the modern missions, Pathfinder through Mars 2020, and a flagged-but-usable series for the Viking pair. They are sufficient because the modern reconstructions [\[6\]](#ref-6), [\[7\]](#ref-7), [\[13\]](#ref-13), [\[14\]](#ref-14), [\[71\]](#ref-71), [\[72\]](#ref-72), [\[73\]](#ref-73), [\[74\]](#ref-74), [\[76\]](#ref-76) apply a documented and largely consistent statistical methodology descended from [\[7\]](#ref-7), and when a set of measurements is generated by the same documented methodology, the resulting values are comparable up to the residual differences in instrumentation and atmosphere model, which can themselves be characterized and bounded. That comparability is not a matter of faith but an estimable quantity, because the reconstruction methodology literature [\[7\]](#ref-7), [\[13\]](#ref-13) explicitly reports the uncertainty contributions of its inputs. The comparability holds with high confidence for MSL, MEDLI2-era Mars 2020, and InSight, with moderate confidence for Pathfinder and Phoenix, and with low confidence for the Viking pair, whose conventions cannot be fully reconstructed from the surviving record. Should the Viking-era ellipse convention turn out to differ from the modern convention in a way that cannot be characterized, those two observations must be down-weighted or reported with a wide provenance band rather than treated as exact, and the dissertation commits to that treatment in advance rather than discovering it after the fact. This bounded position is the honest statement of what the dependent-variable source can and cannot deliver, and it is the discipline Kuznets demanded: state the boundary of coverage before drawing the series [\[24\]](#ref-24).

### 4.2.2 TechPort: coding the technology indicators

The independent variables, the three binary technology indicators for guided lifting entry, the range-to-go parachute trigger, and terrain-relative navigation, are coded from TechPort (techport.nasa.gov), NASA's public technology-portfolio database. TechPort catalogues agency technology-development projects with their technology-readiness-level histories, their sponsoring programs, and, for this dissertation, the missions on which a technology was first demonstrated or infused. Access is open through the TechPort web interface and its data-export facility. The role of TechPort is narrow and specific: it establishes the mission on which each guidance technology first flew, which is the single fact each indicator variable encodes.

TechPort is a portal and database rather than a citable paper, so it does not appear as a DOI-bearing entry in the corpus, and the dissertation is explicit about that to avoid the appearance of a phantom citation. The access path is documented here so the coding is reproducible: each technology's infusion mission is read from its TechPort project record, and that reading is corroborated against the flight-test and development literature already in the corpus, which carries the weight for credibility. The guided-entry indicator, set to one for MSL and Mars 2020, is corroborated by the entry-guidance development and flight literature established in Chapter 3 and by the MSL EDL system-development record by Steltzner, San Martin, and Rivellini [\[75\]](#ref-75), which documents the extension of the EDL performance envelope that guided entry produced. The range-trigger indicator, set to one for MSL and Mars 2020, is corroborated by the range-trigger analysis the dissertation treats as primary in Chapter 3. The TRN indicator, set to one for Mars 2020 only, is corroborated by the most direct possible source: the Perseverance EDL GNC and Safe Target Selection reconstruction by Dutta, Way, Casoliva, and Brugarolas [\[78\]](#ref-78), which documents that Mars 2020 landed using a new TRN capability comprising the Lander Vision System sensor and the Safe Target Selection algorithm, and by the broader hazard-detection-and-avoidance development literature [\[70\]](#ref-70). The coding is therefore doubly sourced: TechPort gives the portfolio-level infusion record, and the reconstruction literature gives the flight confirmation that the technology actually operated on the mission to which the indicator assigns it.

The mechanism behind treating these as separate covariates, carried forward from Chapter 2 and Chapter 3, is what justifies coding three indicators rather than one modernity index. Each technology attacks a physically distinct component of the landing footprint: guided lifting entry nulls the hypersonic downrange and crossrange dispersion, the range trigger corrects the parachute-deploy dispersion, and TRN collapses the position-knowledge error and enables the divert [\[78\]](#ref-78). Because the error sources are physically separable, the indicators are conceptually separable even though, as Section 4.4 will detail, the small sample makes them statistically collinear. The TechPort coding records which technology was present; the regression design records that their effects cannot be fully disentangled in nine to eleven observations; both facts are stated, and neither is allowed to hide the other.

### 4.2.3 PDS: achieved miss distance and the construct-validity safeguard
The third dataset, the Planetary Data System (pds.nasa.gov), supplies the secondary dependent variable: the achieved miss distance, defined as the distance between the targeted aim point and the actual landed location. PDS is NASA's permanent archive for planetary mission data, including the orbital imaging products from which a landed vehicle is identified and localized against a georeferenced base map. After each successful landing, the lander or rover is imaged from orbit (historically by the Mars Reconnaissance Orbiter and its predecessors) and its position is fixed against a controlled cartographic frame; differencing that fixed position against the pre-flight aim point yields the achieved miss distance. Access is open through the PDS archive interface.

The role of PDS is to provide a capability measure independent of the design-ellipse simulation conventions, the construct-validity safeguard on which the dissertation depends. The design ellipse [from NTRS] is a prediction; the achieved miss distance [from PDS] is an outcome. If the two series tell the same story of contraction, the construct "landing precision" is well measured and the inference does not hinge on the simulation conventions. If they diverge, that divergence is itself a finding about the relationship between predicted and achieved performance, and it is reported rather than suppressed. What makes the achieved miss distance valuable is its independence from the simulation pipeline. It is computed from orbital imagery and cartography, sharing none of the Monte Carlo dispersion modeling, atmosphere assumptions, or confidence-level conventions that generate the design ellipse, so the two measures fail in independent ways and their agreement is informative.

The PDS series has one structural limitation that the dissertation states up front: achieved miss distances are precise but few, because only a successful landing produces a localizable lander, and there are at most nine to eleven of those. A definitional subtlety also requires naming. The relevant aim point against which the miss is measured changed character across the era. For an unguided ballistic lander, the aim point is the center of the targeted ellipse, and the miss distance is essentially the realized draw from the dispersion. For a guided lander with an autonomous divert such as Mars 2020, the realized landing location is the result of an onboard decision to avoid hazards, so the "miss" relative to a single pre-flight aim point conflates targeting error with deliberate divert [\[78\]](#ref-78). The achieved-miss-distance series must therefore record, for each mission, whether the landing location reflects passive dispersion or active retargeting, and the two cannot be compared as if they were the same quantity. This is documented in the data dictionary and flagged in the robustness analysis rather than averaged over.

## 4.3 Unit of analysis and the operationalization of every variable

### 4.3.1 Unit of analysis

The unit of analysis is the mission landing event, carried forward exactly from the prospectus. The population is the set of U.S.-led Mars surface missions that successfully completed entry, descent, and landing: Viking 1 (1976), Viking 2 (1976), Mars Pathfinder (1997), Mars Exploration Rover Spirit (2004), MER Opportunity (2004), Phoenix (2008), Mars Science Laboratory / Curiosity (2012), InSight (2018), and Mars 2020 / Perseverance (2021). Tianwen-1 (2021) is held out as an external-validity case [\[16\]](#ref-16). This is the entire relevant population of successful U.S. Mars landings, not a sample drawn from it, which has two consequences the dissertation treats as load-bearing. First, there is no sampling-frame bias and no question of representativeness, because the data are a census. Second, the count is irreducibly small, nine to eleven events depending on whether the near-identical within-program pairs (the two Vikings, the two MERs) are counted as one observation or two. That smallness is the central statistical-conclusion threat, addressed in the research-design chapter and acknowledged here as a property of the data rather than a fixable deficiency.

The within-program pairs deserve a specific note because they shape the unit-of-analysis decision. Viking 1 and Viking 2 flew the same EDL architecture to the same class of ellipse, and Spirit and Opportunity likewise shared the MER airbag-lander architecture. Treating each landing as an independent observation overstates the effective sample size, because the second member of each pair carries little new information about the technology-performance relationship. The design therefore specifies, as an alternative experience unit, the program rather than the individual mission, collapsing each pair to a single observation; this is reported as a robustness specification in Chapter 5. Recording the pair structure in the data is part of the measurement task, not the analysis task, and it belongs here.

### 4.3.2 Constructing the dependent variable: from a reported ellipse to a comparable log area

Before the measurement table, the construction of the primary dependent variable warrants a step-by-step account, because it is the single most consequential transformation in the dissertation and because the Kuznetsian discipline lives or dies in its details. The reported quantity in a mission EDL performance study is rarely "the ellipse area." It is, depending on the document, a semi-major and semi-minor axis pair in kilometers, a downrange-by-crossrange dispersion box, a circular targeting radius, or a verbal characterization ("approximately twenty kilometers"). The construction maps each of these reported forms onto a single comparable scalar, the natural log of the three-sigma ellipse area in square kilometers, through a documented sequence of decisions.

The first decision is the confidence-level normalization. An ellipse reported at one or two sigma is converted to a three-sigma equivalent under the stated assumption of a bivariate normal dispersion, for which the linear axes scale with the sigma multiple; a two-sigma semi-axis is multiplied by three-halves to reach the three-sigma semi-axis, and the area by the square of that factor. This conversion is exact only if the underlying dispersion is genuinely Gaussian, which the reconstruction methodology [\[7\]](#ref-7) treats as a working assumption rather than a proven fact, so the conversion is recorded as an applied transformation and the missions requiring it are flagged. The second decision is the axis-to-area computation: where both semi-axes are reported, the area is pi times their product; where only a circular radius is reported, the semi-minor axis is set equal to the semi-major and the resulting area is pi times the radius squared, with the equal-axis assumption flagged. The third decision is the log transform, taken last, which converts the multiplicative comparability of areas into the additive comparability the regression uses and stabilizes the variance across a series spanning several orders of magnitude. The reasoning for this ordering is that the convention normalizations must be applied in the natural (un-logged) units where they are physically defined, and the log taken only once the value is on a common three-sigma area basis.

This construction is where the dissertation's measurement ethic is most visibly Kuznetsian. A careless analyst reads "twenty kilometers" off the MSL paper and "few kilometers" off the Mars 2020 paper and treats their ratio as a precision gain, without noticing that the MSL number may be a three-sigma semi-major axis while the Mars 2020 number may be an effective targeting region that already accounts for the autonomous divert [\[78\]](#ref-78). The construction sequence above forces those two numbers onto the same definitional footing before they are compared, and it records every transformation applied to each, so that the ratio of two cells in the final series is a ratio of like quantities. Confidence in this construction is high for the missions whose papers report axes and sigma levels explicitly (MSL, Mars 2020, InSight, Phoenix) and moderate for those whose ellipse must be inferred from a verbal characterization or an older convention (the Viking pair, and to a lesser degree Pathfinder [\[71\]](#ref-71)).

### 4.3.3 The measurement table

The complete operationalization of every variable is given in the measurement table below. Each row states the construct, its operational definition, the named source dataset, and the scale on which it is measured. This table is the heart of the chapter and the object the appendix data dictionary will reproduce in full. It is the explicit boundary statement Kuznets demands [\[24\]](#ref-24): no variable enters the regression without a stated definition and a stated source.

| Construct | Operational definition | Source | Scale |
|---|---|---|---|
| Landing precision (primary DV) | \(\ln(\text{EllipseArea}_i)\) = natural log of the three-sigma design landing-ellipse area, computed as \(\ln(\pi \cdot a \cdot b)\) where a is the design semi-major axis and b the semi-minor axis reported in the mission EDL performance study | NTRS reconstruction / EDL performance reports [\[6\]](#ref-6), [\[7\]](#ref-7), [\[13\]](#ref-13), [\[14\]](#ref-14), [\[71\]](#ref-71), [\[72\]](#ref-72), [\[74\]](#ref-74) | Continuous, log of square kilometers |
| Landing precision (secondary DV) | \(\ln(\text{MissDistance}_i)\) = natural log of the achieved miss distance between targeted aim point and localized landed position | PDS landing-site localization | Continuous, log of kilometers |
| Experience axis (primary) | \(\text{Sequence}_i\) = ordinal mission sequence index, 1 for Viking 1 through the index of Mars 2020 | Mission record (public chronology) | Ordinal integer, 1..n |
| Experience axis (alternative) | Cumulative-landings count at the time of mission i | Mission record (public chronology) | Count integer |
| Guided lifting entry | \(\text{GuidedEntry}_i\) = 1 if the mission flew closed-loop guided lifting entry, else 0 (1 for MSL and Mars 2020) | TechPort infusion record, corroborated by [\[75\]](#ref-75), [\[78\]](#ref-78) | Binary {0,1} |
| Range-to-go trigger | \(\text{RangeTrigger}_i\) = 1 if the mission used a range-to-go parachute-deploy trigger, else 0 (1 for MSL and Mars 2020) | TechPort infusion record, corroborated by Ch. 3 range-trigger source | Binary {0,1} |
| Terrain-relative navigation | \(\text{TRN}_i\) = 1 if the mission flew terrain-relative navigation with autonomous divert, else 0 (1 for Mars 2020 only) | TechPort infusion record, corroborated by [\[78\]](#ref-78) | Binary {0,1} |
| Approach-navigation accuracy (control) | \(\text{ApproachAccuracy}_i\) = reported approach-navigation delivery accuracy at atmospheric interface (entry-flight-path-angle dispersion or entry-point delivery dispersion) | NTRS navigation/reconstruction reports [\[77\]](#ref-77), [\[78\]](#ref-78), [\[13\]](#ref-13) | Continuous, degrees or kilometers (standardized) |
| Program identifier (pairing) | Program to which mission i belongs, used to collapse within-program pairs in the alternative experience unit | Mission record | Categorical |
| Divert flag (DV provenance) | Indicator that the achieved landing location reflects active TRN divert rather than passive dispersion | [\[78\]](#ref-78) | Binary {0,1} |

Three operationalization choices in this table require defense beyond the one-line definition, and each is given a short, reasoned treatment because the credibility of the series depends on them.

The first is the choice of ellipse **area** rather than semi-major axis as the dependent variable. The prospectus fixes both the area form, \(\ln(\pi \cdot a \cdot b)\), and the semi-major-axis form as equivalent, and the dissertation reports area. The reasoning is that area captures the full two-dimensional targeting region, the operationally meaningful quantity, the set of surface points the vehicle might reach, whereas the semi-major axis alone discards the cross-axis information. Where a mission reports only a circular or single-dimension targeting radius, the semi-minor axis is set equal to the semi-major axis and the convention is flagged in the data dictionary so the assumption is visible. The area form also has the variance-stabilizing property exploited in Chapter 5: because the early ellipses are enormous and the late ones tiny, the log of area compresses a range spanning several orders of magnitude into a tractable linear scale.

The second is the operationalization of the approach-accuracy control, the variable that does the heavy identification work and is also the hardest to assemble. The construct is "how accurately the vehicle was delivered to the atmospheric interface, independent of onboard EDL guidance." The operational proxy is the reported entry-flight-path-angle dispersion or entry-point delivery dispersion from the navigation record. For MSL this is sourced from the navigation-results record by Martin-Mur, Kruizinga, Burkhart, and colleagues [\[77\]](#ref-77), which documents the delivery to Gale Crater; for Mars 2020 the relevant delivery accuracy is reported alongside the GNC reconstruction [\[78\]](#ref-78), [\[13\]](#ref-13). The honest limitation, flagged in Section 4.4 and again in the evidence-gap register, is that this control is not reported as a single tabulated cross-mission series anywhere in the literature; it is reconstructed from per-mission navigation reports that use somewhat different conventions for stating delivery accuracy (some report entry-flight-path-angle three-sigma in degrees, others report an entry-point miss in kilometers). Standardizing these onto a common scale is part of the measurement task, and any mission whose delivered-state dispersion is not in the public record is flagged as a coverage limitation rather than imputed, because imputing the control variable would defeat the identification it is meant to enable.

The third is the binary coding of the technology indicators, which deliberately discards within-technology variation. A guided-entry implementation on Mars 2020 was not identical to the one on MSL; the TRN of Mars 2020 was a first flight, not a mature capability. Coding each lever as a simple present/absent indicator treats these as homogeneous, which is a real simplification. The reasoning for accepting it is twofold. The mechanism the dissertation tests is the **insertion** of a discrete capability that removes a distinct error source, not the gradual refinement of that capability, and a binary indicator is the natural encoding of an insertion event in the Mokyrian macro-versus-incremental frame [\[23\]](#ref-23), [\[22\]](#ref-22). And the sample is far too small to support continuous within-technology covariates; with nine to eleven observations, adding graded technology variables would exhaust the degrees of freedom immediately. The binary coding is therefore a deliberate, sample-constrained choice, defended rather than defaulted to, and its cost (the inability to detect within-technology learning) is stated.

## 4.4 Data quality, validation, and coverage limitations

### 4.4.1 Validation against known values

The dissertation's principal internal validation is the cross-check of the design ellipse against the achieved miss distance, possible precisely because the two come from independent pipelines (NTRS simulation reconstructions versus PDS orbital cartography). The validation logic is a convergence argument. The design-ellipse series is a trustworthy measure of landing precision if, for the missions where both a design ellipse from NTRS and an achieved miss distance from PDS exist, the two are consistent in a specific sense: each achieved landing falls comfortably inside its design ellipse, and the rank ordering of missions by design-ellipse size matches their rank ordering by achieved miss distance. When a predicted dispersion and an independently measured outcome agree in rank and are mutually consistent in magnitude across the population, the prediction is validated against the outcome, because two independent measurement chains that converge are unlikely to share an error. The reconstruction methodology [\[7\]](#ref-7), [\[13\]](#ref-13) is designed to produce dispersions that bound the realized performance, so consistency between the design ellipse and the realized landing is the methodology working as specified; the InSight reconstruction [\[74\]](#ref-74), [\[14\]](#ref-14) notes that several InSight performance metrics fell near the boundaries of their predictions, itself an instance of the predicted dispersion being checked against realized performance. This validation is strong where PDS localizations exist and the divert flag is zero (passive dispersion), and weaker for Mars 2020, where the active TRN divert [\[78\]](#ref-78) means the achieved location is not a clean draw from the design dispersion. Should the achieved-miss-distance series fail to show the contraction the design-ellipse series shows, the construct validity of the design ellipse is impugned, and the dissertation commits in advance (Chapter 5) to reporting that contradiction as a primary finding rather than discarding the secondary DV.

A second validation, weaker but available, is the agreement of the NTRS reconstruction with its peer-reviewed published counterpart for the missions documented in both. Where the MSL design ellipse can be read both from the NTRS Kalman-filter reconstruction [\[72\]](#ref-72) and from the published EDL performance and system literature, agreement between the two checks that the value transcribed into the series is not an artifact of a single document. This is a transcription-and-consistency check rather than an independent-measurement check, and it is recorded as such; it guards against reading error and version drift, not against shared methodological bias.

### 4.4.2 The comparability problem stated concretely

The Kuznetsian comparability problem is not abstract; it has concrete, enumerable sources, and the dissertation lists them so the reader can judge their severity. The design ellipse from mission to mission can differ because of: the confidence level at which the contour is drawn (one, two, or three sigma); the atmosphere model assumed in the Monte Carlo dispersion (which improved markedly across the era, in part because of the MEDLI and MEDLI2 instrumentation [\[73\]](#ref-73), [\[76\]](#ref-76) that measured the atmosphere in flight and fed back into later models); the number of Monte Carlo draws and the dispersion inputs included; whether the reported ellipse describes the targeting capability before the final approach maneuvers or after them; and, for Mars 2020, whether the reported targeting region accounts for the autonomous divert that can move the landing point within the descent [\[78\]](#ref-78). The mitigation is uniform and stated in advance: every ellipse value entering the series is recorded with its sigma level and its simulation-convention provenance, all values are converted to a common three-sigma area where the conversion is defensible, and any value whose convention cannot be pinned down is carried with an explicit provenance band rather than as a point. This is the operational meaning of the Kuznets rule that no aggregate is reported without its boundary [\[24\]](#ref-24); in this dissertation the rule is enforced at the level of the individual cell of the data table.

The severity of this problem is not uniform across the series, and the confidence statement reflects that. For the MEDLI2-era missions (Mars 2020) and the directly comparable modern reconstructions (MSL, InSight), the conventions are documented well enough that comparability holds with high confidence. For Phoenix [\[6\]](#ref-6) and Pathfinder [\[71\]](#ref-71), comparability holds with moderate confidence, because the reconstruction is documented but the atmosphere model and dispersion inputs are of an earlier generation. For the Viking pair, comparability holds with low confidence, because the surviving 1970s record may not fully specify the convention; these two observations anchor the high end of the series and are the most influential on the fitted slope precisely because they are the largest ellipses, so their provenance uncertainty propagates directly into the learning-rate estimate. The dissertation's response is to report the slope both with and without the Viking pair as a sensitivity check, so the dependence of the headline rate on the least-comparable observations is visible rather than hidden.
### 4.4.3 Known biases by dataset

Each dataset carries characteristic biases, and naming them is part of the data-quality accounting. The NTRS dependent-variable source has a survivorship character: it documents successful landings in depth and failed attempts only partially. That is harmless for this dissertation, because the population is defined as successful landings, but it means the series cannot speak to the precision a failed vehicle would have achieved. The NTRS source also carries a recency-of-instrumentation bias: later missions are reconstructed from richer flight data (MEDLI, MEDLI2) than earlier ones [\[73\]](#ref-73), [\[76\]](#ref-76), so the dispersion characterizations of later missions are tighter and better-founded than those of earlier missions. Measurement quality is therefore correlated with mission sequence, the very axis the regression uses. This correlation does not by itself bias the slope, but it does mean the residual variance is heteroskedastic across the sequence, smaller for recent missions and larger for old ones, which the inference plan in Chapter 5 must accommodate.

The TechPort independent-variable source has a portfolio-curation bias: it records technologies that NASA chose to track as portfolio projects, which is essentially complete for the major EDL guidance levers but could in principle omit a minor capability developed informally. For the three levers this dissertation codes, the corroborating flight literature [\[75\]](#ref-75), [\[78\]](#ref-78) closes this gap, so the bias is negligible for the variables actually used. The PDS secondary-DV source carries the divert-conflation bias already discussed [\[78\]](#ref-78) and the few-observations limit; it is also subject to cartographic-frame drift, because the controlled base map against which landers are localized has itself been revised over the era, so a miss distance computed against an older frame is not strictly comparable to one computed against a newer frame. The mitigation is to localize all landers against the most recent common cartographic frame where the archived imagery allows it, and to flag any mission for which only an older-frame localization survives.

### 4.4.4 Coverage limitations

The coverage limitations are stated as properties of the data, not as apologies. First, the series is the full population of successful U.S. Mars landings, so it is a census with no sampling bias, but it is irreducibly small, nine to eleven observations, and no amount of careful measurement can change that count. Second, the design ellipse is a simulation product whose conventions drifted, which is the comparability problem detailed above; provenance recording mitigates this but does not eliminate it. Third, the technology indicators are nearly collinear with the sequence index, because the technologies were inserted monotonically over time; the data on their own cannot separate a perfectly time-aligned technology effect from a pure time trend, which is why the InSight counterfactual [\[14\]](#ref-14), [\[15\]](#ref-15), [\[74\]](#ref-74) and the approach-accuracy control [\[77\]](#ref-77), [\[78\]](#ref-78), [\[13\]](#ref-13) are built into the design as the instruments that break the collinearity. Fourth, the achieved-miss-distance series from PDS is precise but sparse and partly divert-conflated, so it functions as a robustness check on the primary DV rather than as a co-equal series. Fifth, the approach-accuracy control is not available as a single tabulated series and must be reconstructed from heterogeneous per-mission navigation reports [\[77\]](#ref-77), [\[78\]](#ref-78), with any missing mission flagged rather than imputed. Each limitation is reported with the result it affects, consistent with the dissertation's discipline of attaching every boundary to its number.

A specific evidence gap, flagged honestly per the expansion plan, concerns the InSight reconstruction depth. The InSight control case is the linchpin of identification, and the corpus carries a small number of InSight-specific reconstruction records: the trajectory-and-atmosphere reconstruction by Karlgaard and colleagues [\[74\]](#ref-74) and the communications-based reconstruction [\[14\]](#ref-14), together with the landing-site-selection record [\[15\]](#ref-15) that documents the large flat Elysium Planitia ellipse InSight accepted. The achieved miss distance and the precise ellipse provenance for InSight will be confirmed against the primary NTRS record [\[74\]](#ref-74) during the execution phase; if a specific value is needed and is not in the corpus, the dissertation states the gap rather than fabricating the number. The same discipline governs the Viking-era ellipse provenance, which must be sourced explicitly from the NTRS archive and flagged where the simulation convention cannot be determined, and the approach-accuracy values, which are assembled from the navigation record [\[77\]](#ref-77), [\[78\]](#ref-78) mission by mission and flagged where the public record is silent.

### 4.4.5 The data-assembly workflow and its quality-control protocol

The constructed series is not assembled in a single pass; it is built through a documented workflow with quality-control gates at each stage, and stating that workflow is part of making the measurement auditable. The workflow has four stages. In the first, retrieval, the dependent-variable records are pulled from NTRS by mission, capturing for each the reported ellipse form, its sigma level, its simulation-convention statement, and its bibliographic identifier; the technology-indicator infusion records are pulled from TechPort; and the achieved-miss-distance products are pulled from PDS where they exist. The gate at this stage is completeness: every mission in the population must yield at least a dependent-variable record, and any mission lacking one is escalated as a coverage failure rather than dropped silently.

In the second stage, normalization, each retrieved ellipse is passed through the construction sequence of Section 4.3.2, recording every transformation applied. The gate here is convention traceability: no normalized value enters the series without a logged provenance string stating its original form, its original sigma level, and the transformations applied to reach the common three-sigma log area. In the third stage, corroboration, each technology indicator is checked against the flight-reconstruction literature [\[75\]](#ref-75), [\[78\]](#ref-78) and each approach-accuracy value against its navigation report [\[77\]](#ref-77), [\[78\]](#ref-78), [\[13\]](#ref-13); the gate is double sourcing, meaning a technology indicator coded from TechPort alone, without flight-literature confirmation that the technology operated on the mission, is flagged as single-sourced and treated as provisional. In the fourth stage, validation, the design-ellipse series is cross-checked against the achieved-miss-distance series per Section 4.4.1, and the rank-consistency and within-ellipse-containment checks are run; the gate is the validation pass, and any mission failing it is annotated in the data dictionary with the nature of the discrepancy.

The reasoning behind staging the workflow this way, with explicit gates, is that the comparability problem is best controlled at the point where each value is created rather than diagnosed after the whole series is assembled. A provenance error in the Viking ellipse, for instance, is far cheaper to catch at the normalization gate, where the analyst examines the Viking record in isolation, than after the regression has been fitted and the slope looks suspicious. This is the operational discipline that the Kuznets anchor demands: build the comparable, decomposed series first, with the boundary of each component recorded as it is constructed, and only then fit a slope to it [\[24\]](#ref-24). The workflow is documented in the appendix instrument-and-query section so that an independent analyst can rerun it against the same public archives and obtain the same series, gate failures and all.

## 4.5 Ethics, access, and reproducibility

The ethics and access posture of this dissertation is simple and favorable, which is worth stating because it removes a class of objections that burden other empirical work. All three datasets are public and open. NTRS, TechPort, and PDS are NASA repositories available without credential, fee, or institutional gatekeeping; the citations API for NTRS, the data-export facility for TechPort, and the archive interface for PDS are all open access. There are no human subjects, no personally identifiable information, no proprietary or export-controlled data, and no privacy or consent considerations, because the unit of analysis is a spacecraft landing event documented in the public technical record. The data are engineering reconstructions of physical events, not observations of people.

Reproducibility is correspondingly strong and is engineered in rather than asserted. Because every variable in the measurement table resolves to a named public source and a documented access path, an independent researcher can reconstruct the entire dataset from the dissertation and the public archives. The appendix data dictionary reproduces the measurement table with, for each mission and each variable, the specific source record (NTRS identifier or DOI, TechPort project identifier, or PDS product identifier) from which the value was drawn, the convention under which it was reported, and any standardization applied. The dissertation also commits to publishing the assembled coding sheet, the mission-by-mission table of ellipse values with their sigma levels and provenance bands, the technology-indicator coding with its TechPort and flight-literature corroboration, and the approach-accuracy control with its per-mission source and standardization, so that the constructed series is auditable cell by cell. This is the reproducibility counterpart of the Kuznetsian boundary discipline: not only is every aggregate reported with its boundary, but the boundary itself is published in a form that lets another analyst check it.

One access caveat is recorded for completeness. While all three datasets are open, constructing the comparable series from them requires judgment, the standardization of differing approach-accuracy conventions, the conversion of differing sigma levels to a common three-sigma area, the determination of which Viking-era records best represent the design convention, and that judgment is part of the contribution and is documented rather than automated. Two careful analysts could in principle make slightly different standardization choices; the dissertation mitigates this by stating its choices explicitly and by reporting the sensitivity of the headline learning rate to the choices that matter most (the Viking provenance and the divert-conflation treatment), so that the irreducible judgment is bounded and visible rather than concealed inside an opaque pipeline.

## 4.6 What this chapter establishes for the argument

This chapter has carried the prospectus data section forward and elaborated it into a full measurement design, and in doing so it advances two parts of the dissertation's argument at the level of the data. That the problem is real, the ellipse contraction is large and currently only qualitatively explained, is supported here by the provenance demonstration: the design-ellipse values exist across the full population from Viking to Mars 2020 [\[6\]](#ref-6), [\[7\]](#ref-7), [\[13\]](#ref-13), [\[14\]](#ref-14), [\[71\]](#ref-71), [\[72\]](#ref-72), [\[74\]](#ref-74), they span several orders of magnitude, and they have never been assembled into one comparable series, which is the concrete form of "real but unmodeled." That the design addresses the causal mechanism is supported here by the operationalization: the three technology indicators are coded from TechPort and corroborated by flight reconstructions [\[75\]](#ref-75), [\[78\]](#ref-78) to map directly onto the three physically distinct error sources the mechanism names, and the approach-accuracy control [\[77\]](#ref-77), [\[78\]](#ref-78), [\[13\]](#ref-13) is operationalized specifically to separate the onboard-guidance mechanism from the injection-accuracy rival.

The measurement design also makes explicit the residual risk that the closing stage of the argument must show is acceptable. The risks are small-n, simulation-convention drift, indicator-time collinearity, divert conflation in the secondary DV, and the heterogeneity of the approach-accuracy control. None is fatal, and each is paired here with a stated mitigation: census coverage removes sampling bias and leaves only the irreducible small count; provenance recording and the with-and-without-Viking sensitivity bound the convention drift; the InSight counterfactual and the approach-accuracy control are built into the data to break the collinearity; the divert flag isolates the conflated observations; and the heterogeneous control values are standardized and flagged. The confidence statement for the chapter as a whole is therefore moderate-to-high: the dependent-variable series can be constructed with high confidence for the modern missions and moderate confidence for the earliest, the technology indicators can be coded with high confidence given the double sourcing, and the approach-accuracy control can be assembled with moderate confidence subject to the coverage flags. The evidence that would raise this confidence is direct retrieval of the Viking-era simulation-convention documentation and a complete public tabulation of per-mission delivered-entry-state dispersions; the evidence that would lower it is a discovery that the achieved-miss-distance series and the design-ellipse series diverge in rank, which the dissertation is built to detect and to report.

The series is constructable, its boundaries are stated, its biases are named, and its access is open. Chapter 5 takes this constructed series and specifies the identification strategy that turns it into a test of the learning-curve hypothesis, leaning, as the data chapter has prepared it to, on the InSight counterfactual and the approach-accuracy control rather than on asymptotic significance that nine to eleven observations cannot support.


# Chapter 5: Research Design and Identification

## 5.1 The chapter thesis

Identification in this dissertation does not come from sample size, and it does not come from asymptotic significance. It comes from two design features that are physically and historically prior to any regression: the InSight counterfactual and the approach-accuracy control. The thesis of this chapter is that these two features, and only these two, break the time-technology collinearity that would otherwise make the central hypothesis untestable on nine to eleven observations. A naive analyst, presented with a sequence of Mars landing ellipses that fall monotonically across forty-five years, would regress the log of ellipse area on the calendar year, read off a slope, and call it a learning rate. That analysis is worthless for the question this dissertation asks, because every candidate cause of the contraction also rises monotonically across the same forty-five years: onboard guidance, modeling fidelity, atmospheric knowledge, flight-computer throughput, and approach-navigation accuracy all improved together, and a slope on time cannot tell them apart. The design described in this chapter is constructed to introduce variation that the calendar does not contain. InSight (2018) is a late mission that deliberately did not carry the guidance suite, so it sits at a high sequence index with low technology indicators and is the one observation in the population whose technology generation is decoupled from its date [\[15\]](#ref-15). The delivered-entry-state dispersion, reconstructed mission by mission, holds constant the part of accuracy that launch and interplanetary navigation supply, so that whatever contraction loads on the technology indicators after this control is in cannot be the injection-accuracy story [\[7\]](#ref-7), [\[13\]](#ref-13). The estimator is log-linear ordinary least squares, chosen for reasons of measurement convention and variance behavior that this chapter defends rather than assumes; the inference is permutation-based and exact rather than asymptotic, because with this many observations no other inference is honest. Everything else in this chapter, the nested-specification hierarchy, the robustness battery, the power analysis, the pre-registration commitment, and the computational plan, exists to make that single identification argument auditable and to state in advance what evidence would raise or lower confidence in it.

This is a design-stage chapter. The model is fully specified and the data sources are named and accessible, but no regression has been run on the assembled dataset. Where a number appears, it is illustrative, chosen to show the shape of an expected result or the form of an output table, and it is labeled as such. The strongest claim this design can support, once executed, is a signed, order-of-magnitude, counterfactual-surviving attribution, not a precise point estimate of a learning rate. The chapter is built to make that claim defensible, and to make a contradicting result equally informative.

## 5.2 The problem this chapter addresses

The local problem of Chapter 5 is an identification problem, and it is worth framing in the same current-state, desired-state, gap, consequence structure that governs the dissertation as a whole, because the design decisions follow from the gap rather than from methodological taste.
The current state is that the cross-mission ellipse contraction is, in the engineering literature, attributed to specific technologies one mission at a time, in papers that were never designed to arbitrate among rival causes [\[1\]](#ref-1), [\[5\]](#ref-5), [\[8\]](#ref-8). Each of those papers is correct about its own mission. None of them, alone or in aggregate, identifies the marginal contribution of onboard guidance against the rival contributions of approach-navigation accuracy and generic temporal maturation, because each studies a single point on a one-dimensional time axis along which everything moved together. The desired state is a single estimator, fit to the full constructed series, whose coefficients can be read as the contribution of identifiable technology insertions net of the approach-navigation control and net of a generic sequence trend. What separates the two is the absence of identifying variation. In the raw data, technology generation is an almost deterministic function of mission sequence, because the technologies were inserted monotonically and irreversibly: MSL added guided entry and the range trigger, Mars 2020 added terrain-relative navigation, and nothing was ever removed. Leave this gap unfilled, and a project manager deciding whether to fund terrain-relative navigation for a future lander, or to spend the same money on approach navigation, still cannot read the answer off the data, because the data as ordinarily presented cannot separate the two effects. This chapter specifies the design that supplies the missing variation and thereby converts an unidentified description into an identified, falsifiable test.

The chapter is candid that this is a hard identification problem and that the design does not make it disappear. It makes it tractable, bounded, and honest. The InSight counterfactual is a single observation, and a single observation cannot carry the full weight of an econometric identification claim in the way a randomized assignment or a large natural experiment could. The approach-accuracy control is a measured covariate assembled from reconstruction records of varying completeness, not a clean instrument. The chapter states these limits in the same breath as the design that exploits them, and it builds a robustness battery whose entire job is to show how much of the conclusion depends on each fragile element. The standard here is not that the identification is airtight, which would be a false claim on this sample, but that every threat to identification is named, paired with a mitigation, and reported alongside the result rather than buried.

## 5.3 The estimator and why it is chosen

The estimator is ordinary least squares on a log-linear model of three-sigma ellipse area against mission sequence and the technology covariates, with a linear-in-levels specification carried as a robustness check. The measurement convention of the learning-curve literature and the variance structure of the dependent variable dictate this choice, not convenience.

Three properties of the problem drive the choice. The first is the multiplicative reading. Taking the natural logarithm of ellipse area makes the model multiplicative in levels, so a slope coefficient reads as a constant proportional contraction per unit of the experience axis. This is the natural form for a learning curve and the form in which the experience-curve literature reports its rates: a learning rate is a percentage reduction per doubling of cumulative experience, a statement about ratios, not differences [\[18\]](#ref-18), [\[19\]](#ref-19). To estimate a learning rate is to estimate a slope in log space; to estimate it in levels would be to estimate something else and then mislabel it. The second property is variance stabilization. The dependent variable spans more than two orders of magnitude, from Viking-class targeting regions in the high hundreds of kilometers along the major axis to a few-kilometer effective targeting region for Mars 2020 [\[5\]](#ref-5), [\[8\]](#ref-8). In a levels regression, the squared residuals of the two enormous early ellipses would dominate the loss function, the fit would be driven almost entirely by Viking and the Mars Exploration Rovers, and the recent, decision-relevant missions would fit poorly. The log transform compresses that range and gives each generation comparable leverage on the fit, which is what is wanted when the scientific interest is concentrated in the recent, small-ellipse end of the series. The third property is the match between the functional form and the physical mechanism. Each technology lever removes a fraction of a physically distinct error source rather than a fixed number of kilometers: guided entry nulls a proportion of the hypersonic dispersion, the range trigger removes a proportion of the parachute-deploy dispersion, and terrain-relative navigation collapses a proportion of the position-knowledge error [\[3\]](#ref-3), [\[4\]](#ref-4), [\[5\]](#ref-5), [\[8\]](#ref-8). A mechanism that acts multiplicatively on distinct error budgets is more naturally represented in logs than in levels, so the functional form is not merely statistically convenient; it is the form the engineering physics implies.

These three properties together justify the estimator. When the quantity of scientific interest is a proportional rate of change and the dependent variable has multiplicative error structure spanning orders of magnitude, the log-linear model recovers the parameter of interest in its native units and distributes fit fairly across the range, a standard choice in the cliometric and experience-curve tradition that the two anchors supply [\[18\]](#ref-18), [\[19\]](#ref-19), [\[24\]](#ref-24). The experience-curve literature is explicit that the relationship is reported and estimated in logarithmic space and that the rate is a proportional one [\[18\]](#ref-18); the lithium-ion reassessment demonstrates the discipline of estimating such a rate carefully, in logs, while separating the contributions of distinct mechanisms and being precise about what the experience axis measures [\[19\]](#ref-19). The dispersion-analysis precedents that construct the underlying ellipse and miss-distance quantities, for the Mars Exploration Rovers, Phoenix, and Mars Pathfinder, report dispersions whose magnitudes confirm the order-of-magnitude span that motivates the variance-stabilization argument [\[81\]](#ref-81), [\[83\]](#ref-83), [\[84\]](#ref-84).

The log form is the primary specification, not the only admissible one. Its cost is that it cannot represent a true zero and presumes proportional rather than additive contraction. Both costs are acceptable here, because no ellipse is zero and the mechanism is closer to multiplicative than additive, but the presumption is testable and will be tested. Were the linear-in-levels specification to disagree materially with the log specification, in the sign or ordering of the technology terms rather than merely in scale, that disagreement would be reported as a finding and not suppressed, and it would lower confidence in the proportional-rate interpretation. The robustness battery in Section 5.8 includes the levels re-fit precisely so that this objection can be adjudicated rather than asserted away.

Ordinary least squares, rather than a more elaborate estimator, is appropriate because the design is a single cross-section of fully observed events with no panel structure, no repeated measurement, and no selection into the sample: the population is the complete robotic-era record of successful United States Mars landings, not a draw from it. There is no sampling frame from which to correct a selection bias, because every member of the population is in the dataset. There is no clustering or serial dependence to model in the conventional sense, because each mission is a distinct vehicle and a distinct landing event rather than a repeated draw from a stationary process; the only dependence is the deliberate near-duplication within the Viking pair and the Mars Exploration Rover pair, which the design handles by a program-level experience-unit specification described in Section 5.5 rather than by a variance correction. Generalized least squares, fixed effects, and instrumental-variables estimators are all considered and rejected in this chapter at the points where they would naturally arise, and the rejection is reasoned rather than reflexive.

## 5.4 The specifications written out

The two specifications are reproduced here exactly as the dissertation's governing design states them, because the entire identification argument is a statement about the difference between them.

The baseline specification is

\[
\ln(\text{EllipseArea}_i) = \beta_0 + \beta_1 \cdot \text{Sequence}_i + \epsilon_i\qquad\qquad (1')
\]

where \(\ln(\text{EllipseArea}_i)\) is the natural logarithm of the three-sigma landing-ellipse area, \(\ln(\pi \cdot a \cdot b)\), with a the semi-major axis and b the semi-minor axis of the design ellipse from mission i's EDL performance study, and \(\text{Sequence}_i\) is the mission sequence index running from one for Viking 1 through the index of Mars 2020. The coefficient \(\beta_1\) recovers a constant proportional contraction rate per unit of mission sequence, and the learning-rate reading is \(\exp(\beta_1) - 1\), the proportional change in ellipse area per unit of the experience axis. The baseline is the descriptive backbone: it asserts nothing about mechanism and is expected, on the qualitative record, to return a large negative \(\beta_1\) and a high in-sample fit. The baseline is not the test of the contribution. It is the foil against which the augmented specification is read, because a high-fitting baseline trend is exactly what H0 predicts and exactly what a generic-maturation explanation would also produce.

The augmented specification is

\[
\ln(\text{EllipseArea}_i) = \beta_0 + \beta_1 \cdot \text{Sequence}_i + \gamma_1 \cdot \text{GuidedEntry}_i + \gamma_2 \cdot \text{RangeTrigger}_i + \gamma_3 \cdot \text{TRN}_i + \delta \cdot \text{ApproachAccuracy}_i + \epsilon_i\qquad\qquad (2')
\]

where \(\text{GuidedEntry}_i\) is one for MSL and Mars 2020 and zero otherwise, \(\text{RangeTrigger}_i\) is one for MSL and Mars 2020 and zero otherwise, \(\text{TRN}_i\) is one for Mars 2020 only, and \(\text{ApproachAccuracy}_i\) is the reported approach-navigation delivery accuracy for mission i, the entry-flight-path-angle or entry-point delivery dispersion that captures how accurately the vehicle was delivered to the atmospheric interface independent of onboard EDL guidance [\[7\]](#ref-7), [\[13\]](#ref-13). H1 makes two joint predictions about this specification: the gamma coefficients are jointly significant and negative, meaning each technology generation reduces ellipse area, and delta is small and statistically insignificant, meaning approach accuracy is not the binding constraint. H0 makes the complementary prediction: once the sequence trend and the approach-accuracy control are present, the gamma coefficients carry no joint explanatory power, and any apparent technology effect is absorbed either by the sequence trend or by the approach control.

The notation and the variable definitions here are fixed by the dissertation's design and are used identically in every chapter; no chapter is permitted to redefine the experience axis, the technology indicators, or the control. The augmented model as written is over-parameterized for the available degrees of freedom, with five slope parameters and an intercept against nine to eleven observations, which is why this chapter does not estimate it as a single horse race. The next section describes the nested hierarchy that is the actual estimation strategy, and Section 5.6 explains why the full augmented equation is written out at all when it will not be fit in one piece.

## 5.5 The nested-specification hierarchy

The technology effect is estimated not by fitting the full augmented equation in one over-parameterized regression but by a hierarchy of nested specifications, adding one technology generation at a time and reporting the incremental fit and the exact test on each added term, because that is the only way to extract attributable components from a series in which the covariates are nearly collinear.

With three technology indicators that are themselves nearly collinear, and all nearly collinear with the sequence index, a single regression that enters all of them simultaneously would distribute the explained variance arbitrarily across highly correlated regressors, producing coefficient estimates with enormous variance inflation and signs that flip on the addition or removal of a single observation. That is not identification; it is noise dressed as estimation. The nested approach instead asks a sequence of well-posed questions. Starting from the baseline sequence-only model, the first nested step adds GuidedEntry and asks whether the fit improves beyond what the sequence trend already explains and whether the added term is negative. The second step adds RangeTrigger and asks the same question conditional on guided entry. The third adds TRN. The fourth adds the approach-accuracy control and asks whether its entry changes the technology terms, which is the H0-versus-H1 adjudication. At each step the reported quantities are the change in adjusted R-squared, which measures incremental fit penalized for the added parameter, and an exact or permutation-based test on the added coefficient, which measures whether the increment is distinguishable from what reordering the labels would produce. This is the cliometric move of decomposing an aggregate change into attributable components in a stated order rather than asserting all components at once [\[24\]](#ref-24).

When regressors are nearly collinear and the sample is small, sequential addition with incremental-fit reporting is the disciplined alternative to a single multicollinear regression, because it isolates the marginal explanatory content of each generation conditional on the ones already entered, and it makes the order of entry, and therefore the analyst's priors, explicit and auditable rather than hidden inside a variance-covariance matrix. This decompositional approach has direct precedent: the dispersion-analysis literature that constructs the underlying quantities proceeds in exactly this spirit, attributing landing dispersion to its physical error sources one component at a time rather than reporting an undifferentiated total [\[81\]](#ref-81), [\[83\]](#ref-83), [\[84\]](#ref-84), [\[87\]](#ref-87), and Kuznets's measurement discipline supplies the methodological precedent: build the series, state the boundary, and decompose the change before announcing a slope [\[24\]](#ref-24).

The order of entry is not neutral, and a different order could in principle redistribute the incremental fit. The pre-registered order, sequence, then guided entry, then range trigger, then TRN, then approach accuracy, is chosen on physical and chronological grounds (it follows the order in which the levers entered service and the order of their expected effect size) and is frozen in advance so that it cannot be reverse-engineered to a desired result. If the incremental fit were highly sensitive to the order of entry, that sensitivity would itself be reported, because it would mean the data cannot apportion credit among the levers and that only the joint technology effect is identified. The design anticipates that the range trigger, entering between two larger-effect levers and itself an incremental rather than macro invention, may not be separately identifiable, and it states in advance that a non-significant separate range-trigger increment is consistent with H1 provided the guided-entry and TRN increments survive.

An illustrative template of the nested-specification table follows. The entries are placeholders that show the form the output will take, not estimates; no row has been computed.

| Specification | Sequence coeff. | Added technology term | Approach control | Incremental adj. R-squared | Exact test on added term |
|---|---|---|---|---|---|
| Baseline (sequence only) | negative, large | none | absent | reference | not applicable |
| + Guided entry | smaller | guided-entry negative | absent | positive increment | reported |
| + Range trigger | smaller | range-trigger negative | absent | smaller increment | reported |
| + TRN | smaller | TRN negative, largest | absent | larger increment | reported |
| + Approach-accuracy control | little changed under H1 | technology terms retain sign under H1 | present | small change under H1 | reported |

The decisive row is the last one. Under H1 the technology terms retain their sign and joint significance when the approach control enters, and the control itself is small and insignificant; under H0 the control absorbs the technology terms, which lose joint significance. The plan commits in advance to reporting whichever pattern appears, and Section 5.10 records that commitment formally.
The program-versus-mission experience-unit specification belongs here as a deliberate variant of the experience axis. The Viking pair and the Mars Exploration Rover pair are near-identical vehicles flown within the same program, and counting each as a separate unit of experience risks overstating the effective sample and understating the standard errors. One nested specification therefore collapses each program pair to a single program-level observation, reducing the experience axis from a mission count to a program count, and re-estimates. If the technology effect survives the loss of the within-pair degrees of freedom, that is evidence that the result is not an artifact of double-counting near-duplicate vehicles. If it does not survive, that fragility is reported.

## 5.6 Identification strategy

This is the core of the chapter, and it is stated as a formal argument because the dissertation's credibility rests on it more than on any single estimate.

The marginal effect of onboard EDL-guidance technology on ellipse area is identified, to the limited but real degree that nine to eleven observations allow, by three features acting together: the across-mission step structure of the indicators, the InSight within-period counterfactual, and the approach-accuracy control. Of these, the InSight counterfactual is the binding identifying assumption, and the design's identification is no stronger than the validity of the InSight case.

The first feature is the across-mission step structure. Guided entry, the range trigger, and terrain-relative navigation entered service on identifiable missions, creating discrete jumps in the indicator variables at MSL and at Mars 2020 [\[4\]](#ref-4), [\[5\]](#ref-5), [\[8\]](#ref-8). If the ellipse series shows matching discrete drops at exactly those missions, and not at the missions where no lever was added, the alignment of the steps with the technologies is evidence for attribution beyond what a smooth trend would explain. This is the engineering analog of an event-study logic: the technology is the event, the ellipse drop is the response, and the test is whether the response is concentrated at the event rather than spread across the calendar. Standing alone, step structure is weak, because the steps coincide with time and the small sample offers few non-event missions to serve as a counterfactual baseline. It becomes strong only when joined to feature two.

The second feature is the InSight counterfactual. InSight landed in 2018, six years after MSL had demonstrated guided entry and the range trigger, yet it deliberately flew a Phoenix-heritage ballistic, unguided entry into a large, flat ellipse in Elysium Planitia, because its seismology and heat-flow science did not require precision and its budget did not justify the guidance suite [\[15\]](#ref-15). InSight therefore occupies a position in the data that the calendar alone cannot produce: a late mission with the date of the modern era and the guidance of the Viking-Phoenix era. It is the single observation in the population whose technology generation is decoupled from its sequence index. The identifying logic is a within-period comparison. If ellipse contraction were a pure time trend or a product of generic EDL maturation, InSight should have inherited a small ellipse for free, because by 2018 the calendar and the general state of the art had advanced as far as they had for MSL. It did not; it accepted a large ellipse by design [\[15\]](#ref-15). The InSight residual, the vertical distance of its observed ellipse from the fitted sequence trend, is therefore the design's central identifying quantity. A large positive InSight residual, the ellipse sitting well above the trend line, is the signature of a technology effect and against a pure-time effect; an InSight residual near zero, the ellipse sitting on the trend, is the signature of generic maturation and against the technology hypothesis. This single observation breaks the otherwise near-perfect time-technology collinearity, and the design's identification claim is exactly as strong, and exactly as fragile, as this one case.

The third feature is the approach-accuracy control. Even if the technology effect is distinguished from generic time, it must still be distinguished from the rival explanation that the vehicles were simply delivered to the atmospheric interface more accurately by improving launch and interplanetary navigation, which is the substance of H0. The approach-accuracy control, the reconstructed delivered entry-state dispersion for each mission, holds that part of accuracy constant [\[7\]](#ref-7), [\[13\]](#ref-13). Once it is in the model, any residual contraction loading on the technology indicators is, by construction, attributable to what happened after atmospheric interface, that is, to onboard EDL guidance, rather than to getting to the interface more precisely. The delta coefficient on this control is the direct test of H0: a large, significant delta that absorbs the technology terms supports H0; a small, insignificant delta with the technology terms intact supports H1.

These three features work together. A monotone confound is broken by any observation that violates the monotone relationship, and a rival mechanism is excluded by conditioning on a measured proxy for that mechanism. InSight violates the time-technology monotonicity; the approach-accuracy control conditions on the injection-accuracy mechanism. Together they convert an unidentified trend into a conditionally identified technology effect, under the assumption that the InSight case is a valid counterfactual and the approach proxy is an adequate measure of the rival mechanism. The supporting record bears this out. The InSight landing-site selection record documents that the large Elysium ellipse was a deliberate choice tied to mission requirements and not a capability ceiling, which is what makes it a valid counterfactual rather than a failure [\[15\]](#ref-15); the communications-based InSight EDL reconstruction provides the independent flight-data basis for placing InSight's achieved performance in the series [\[14\]](#ref-14); and the MEDLI2 reconstruction and the statistical-reconstruction methodology provide the delivered-entry-state quantities that operationalize the approach control [\[7\]](#ref-7), [\[13\]](#ref-13).

Identification here is conditional and partial, not unconditional and complete. It is conditional on the InSight counterfactual being a genuine technology-off case rather than confounded by some InSight-specific factor (a different target latitude, atmosphere, or arrival geometry), and it is partial in that even when the joint technology effect is identified, the apportionment of credit among the three individual levers is only weakly identified, because the three indicators co-move and InSight lacks all three at once. The principal objection is that InSight's large ellipse reflects not the absence of the guidance suite but some idiosyncratic feature of the InSight mission, in which case the counterfactual is contaminated and the identification fails. The design confronts this directly through the drop-InSight robustness re-fit (Section 5.8), which shows exactly how much of the conclusion depends on InSight, and through the achieved-miss-distance dependent variable, which provides a capability measure for InSight that is independent of the pre-flight ellipse convention. If the conclusion collapses when InSight is removed, the chapter will report that the result is InSight-dependent and will downgrade confidence accordingly, rather than presenting a fragile identification as robust.

The honest summary of the identification strategy is therefore this. The design does not claim to identify a clean causal effect in the sense a randomized experiment would. It claims to construct, from a complete population of expensive engineering events, the one within-period counterfactual the historical record happens to contain, and to condition on the one rival mechanism the reconstruction record happens to measure, and thereby to make a falsifiable, attribution-bearing test possible where the existing literature offered only mission-by-mission narrative. Confidence in the identification is **moderate** by design, and the chapter states what would raise it (a second technology-off late mission, which only future flights can supply) and what would lower it (evidence that InSight's ellipse is confounded by an InSight-specific factor).

A note on why instrumental variables are not used, since a reader trained in econometrics will ask. An instrument for technology insertion would have to be a variable that shifts the probability a mission carried guided entry, the range trigger, or TRN, but that affects ellipse area only through that channel. No such instrument exists in this setting, and inventing one would be worse than useless. Technology insertion was decided by program managers on the basis of mission science requirements, budget, and the maturity of the technology, all of which are plausibly correlated with the achievable ellipse directly. There is no lottery, no policy discontinuity, and no exogenous cost shifter that assigned guidance suites to missions. The design therefore does not pretend to instrument what cannot be instrumented; it relies instead on the within-period counterfactual and the conditioning control, and it states their limits openly. This is the correct posture for a small-n historical-engineering identification problem, and claiming an instrument here would fabricate identifying variation rather than discover it.

## 5.7 Threats to validity

Each of the four classical validity threats is stated as a claim about a specific danger to the inference, paired with the specific design feature that mitigates it. No threat is asserted without its mitigation, and no mitigation is claimed to be complete where it is not.

### 5.7.1 Internal validity

**Threat.** The dominant internal-validity threat is the time-technology collinearity already discussed: any unobserved factor that improved monotonically across the era, modeling fidelity, atmospheric knowledge, flight-computer throughput, or institutional learning, could load on the technology indicators and masquerade as a technology effect, because all of these moved with the calendar and so with the indicators. A coefficient that the analyst reads as the effect of guided entry could in truth be the effect of better atmospheric models that happened to arrive at the same mission.

**Mitigation.** The InSight counterfactual and the approach-accuracy control are the primary defenses, and the nested-specification reporting is the secondary one. InSight holds the calendar roughly fixed at the modern era while turning the guidance off, so a confound that tracks only the calendar cannot explain InSight's large ellipse [\[15\]](#ref-15). The approach control removes the specific monotone confound, delivery accuracy, that is most plausibly correlated with the technology and most directly a rival cause [\[7\]](#ref-7), [\[13\]](#ref-13). The nested reporting makes any residual confound visible: if the technology increments shrink toward zero as plausible monotone controls enter, that is reported as evidence that the indicators were proxying for general maturation. Confidence that the internal-validity threat is controlled is **moderate**; it would rise with a second technology-off mission and fall if the achieved-miss-distance series failed to show the contraction the design ellipse shows.

**A further internal-validity threat requires a dedicated treatment before the analysis plan is finalized: the approach-accuracy control may be a bad control rather than a clean pre-treatment confounder.** The concern has two distinct forms, and both bear on whether conditioning on \(\text{ApproachAccuracy}_i\) identifies the onboard-technology effect or distorts it.

The first form is post-treatment contamination. The same EDL engineering program that drove each technology generation also improved approach navigation, because the teams, the propositional knowledge base, and the institutional infrastructure were shared. If approach accuracy is itself a product of the same maturation process that produced guided entry and terrain-relative navigation, conditioning on it may partial out part of the technology effect rather than isolating an independent rival mechanism. A variable that partially mediates the treatment, rather than purely preceding it, is a post-treatment control in the sense that directed acyclic graph causal reasoning formalizes: including it in the regression attenuates the coefficient on the upstream cause, and the attenuation is not bias correction but bias induction. The delivered-entry-state precision reported in the MEDLI2 reconstruction and the statistical-reconstruction literature is the operationalization of \(\text{ApproachAccuracy}_i\) [\[7\]](#ref-7), [\[13\]](#ref-13), but whether that precision improved because of exogenous navigation advances or because the same program culture that drove TRN adoption also drove tighter delivery targeting is not resolved in those sources; it must be examined before the control is treated as clean.

The second form is collider bias. If unobserved mission ambition, the willingness to target scientifically rich but technically demanding sites, is a common cause of both higher technology adoption (ambitious sites require more precise delivery) and tighter approach accuracy (more investment flows to missions with ambitious targets), then conditioning on \(\text{ApproachAccuracy}_i\) conditions on a collider, opening a back-door path from mission ambition to the technology block that the conditioning introduces rather than closes. The result would be a spurious negative association between approach accuracy and the technology indicators even in the absence of any true causal link, and the coefficient pattern that the decision rule reads as failure to reject H0, a significant delta that absorbs the technology terms, could in principle be produced by collider bias rather than by the injection-accuracy rival.

**The design's response is a pre-registered diagnostic rather than a design-stage arbitration.** The causal status of \(\text{ApproachAccuracy}_i\) cannot be fully adjudicated without the assembled data and, ultimately, without more detailed program-history reconstruction than this dissertation undertakes. The binding commitment entered here is that the analysis plan treats the bad-control question as one of the key diagnostics of the study, not an afterthought. The robustness battery (Section 5.8) commits in advance to a control-free re-fit that drops \(\text{ApproachAccuracy}_i\) entirely and reports the technology learning rate without it; the comparison between the with-control and without-control specifications is then a direct measure of the degree to which conditioning on approach accuracy changes the technology attribution. A large downward revision of the technology coefficients when the control is added is consistent with the post-treatment or collider reading and would be flagged as a concern; a negligible change would be consistent with the control being clean. The pre-registration prevents selective reporting: both specifications are committed to the output regardless of how the comparison falls. This does not resolve the causal question, but it turns it into an observable and reported diagnostic rather than an unexamined assumption. Confidence that the approach-accuracy control is a clean pre-treatment confounder is **lower than moderate** until the post-treatment and collider threats are assessed on the assembled series; the design-stage rating of **moderate** for internal validity above reflects the package of mitigations including both the InSight counterfactual and the forthcoming diagnostic re-fit, not a presumption that the control is unambiguously pre-treatment.

### 5.7.2 External validity

**Threat.** The model is fit on United States Mars landings only. Generalization to other agencies, other planetary bodies, or human-scale vehicles is not automatic and could fail, because the propositional knowledge base, the institutional channel, and the physics of a much larger vehicle differ.

**Mitigation.** External validity is deliberately bounded rather than over-claimed. Tianwen-1 is held out of the estimation sample entirely and used only as an out-of-sample, non-United-States reference point for discussion [\[16\]](#ref-16); lunar autonomous precision-landing navigation is cited to establish that terrain-relative navigation is a general planetary capability rather than a Mars-specific artifact, again as a reference point and not as in-sample evidence [\[17\]](#ref-17). The human-scale precision-landing assessments that target fifty-meter delivery for twenty-five-tonne payloads are treated as the forward edge of the capability and as a destination the curve might or might not reach, not as observations the curve is fit to [\[86\]](#ref-86). The chapter states explicitly that the fitted curve describes demonstrated United States robotic Mars capability and is not warranted to extrapolate to human-class vehicles or to other programs without the supporting knowledge chain in place. Confidence in external validity beyond the estimation population is **low**, and the chapter says so.

**A Phase 1 precondition on Tianwen-1 in-sample inclusion.** The Conclave panel noted a structural gap that the design should acknowledge directly: the entire technology-versus-time discrimination currently rests on InSight as the single technology-off comparison point, and a second technology-discordant observation would substantially strengthen the identification backbone. Tianwen-1 (2021) is the natural candidate because it flew guided entry in a different value network, outside the NASA/JPL apparatus, in the modern era, with a published reconstructed entry trajectory and a landing site positioned in Utopia Planitia [\[16\]](#ref-16). Bringing it in-sample as a second comparison point would convert a single-case discrimination into a two-point comparison that survives the loss of either point and would partially separate the technology attribution from any NASA-institutional confound. The binding precondition that Phase 1 execution must settle before any in-sample inclusion is **cross-agency reconstruction-convention commensurability**: whether Tianwen-1's dispersion and atmosphere reconstruction conventions are harmonizable with the NTRS Karlgaard-lineage series that supplies the primary ellipse values [\[7\]](#ref-7), [\[13\]](#ref-13), or whether bringing it in-sample re-imports the Kuznetsian comparability problem across agencies with different simulation traditions. If the reconstruction conventions are not harmonizable at the stated sigma level and provenance discipline the series requires, Tianwen-1 remains a discussion reference point only; if they are harmonizable, the design-stage scope decision to hold it out should be revisited in the execution phase. This is stated here as an open Phase 1 obligation, not as a design change, consistent with the pre-registration commitment that no in-sample addition is made without meeting the comparability standard.

### 5.7.3 Construct validity
**Threat.** The construct of interest is landing precision, and the primary dependent variable, the three-sigma design ellipse, is an imperfect proxy for it. The design ellipse is a pre-flight simulation product whose conventions changed across missions: the sigma level, the error sources modeled, and the simulation assumptions [\[7\]](#ref-7). A contraction in the reported design ellipse could in part reflect a change in how the ellipse was computed rather than a change in achievable precision. This is the Kuznetsian comparability problem.

**Mitigation.** The design carries a second, independent dependent variable: the achieved miss distance from Planetary Data System post-landing localization, the distance between the targeted aim point and the actual landing location. This is a realized capability measure computed the same way for every mission and independent of the pre-flight simulation conventions. The full analysis is re-run on this construct, and agreement between the design-ellipse result and the achieved-miss-distance result is the construct-validity test. If both series show the same contraction and the same technology attribution, the construct holds across the choice of proxy; if they diverge, that divergence is reported and the simulation-convention drift is implicated. Every reported ellipse value carries its sigma level and its simulation-convention provenance, so that no value is compared across missions without its boundary attached, per the measurement discipline the dissertation adopts from Kuznets [\[24\]](#ref-24). Confidence that the construct is adequately measured is **moderate to high** when both dependent variables agree and **low** when they diverge. The design forces that distinction to be made rather than hidden.

### 5.7.4 Statistical-conclusion validity

**Threat.** With nine to eleven observations and five candidate slope parameters, conventional asymptotic inference is unreliable. Normal-theory standard errors, t-tests, and F-tests assume large-sample behavior that this dataset cannot supply. Reported p-values from such tests would be untrustworthy, and a result that crossed a conventional significance threshold by that route would be an artifact of the wrong reference distribution rather than evidence.

**Mitigation.** The design does not use asymptotic inference for its primary claims. It uses exact and permutation-based tests, described in detail in Section 5.9, that construct the reference distribution by reassigning the technology labels across missions and recomputing the statistic, so that the achieved value is compared against the full distribution of values obtainable by relabeling rather than against a normal approximation. Effect sizes are reported with permutation-based intervals that are honestly wide, and the chapter commits in advance not to claim that the small-sample regression proves a learning rate. The strongest claim the inference will support is a sign, an order of magnitude, and an attribution that survives the InSight counterfactual and the permutation reference distribution. Confidence in the statistical conclusion is bounded by the sample and is **moderate** at best. The design's contribution to statistical-conclusion validity is not to manufacture power the data lack but to report uncertainty honestly and to use the exact inference appropriate to the n.

## 5.8 The robustness battery

The robustness battery is the operational expression of the chapter's central discipline: every fragile element of the identification is stress-tested by a pre-specified re-fit, and the result of each re-fit is reported whether or not it is convenient. Four re-fits are committed in advance.

**Re-fit one: the achieved-miss-distance dependent variable.** The entire nested hierarchy is re-estimated with ln(achieved miss distance) from Planetary Data System localization replacing ln(design ellipse area) as the dependent variable. This is at once the construct-validity safeguard of Section 5.7.3 and a robustness check on the headline result. The interpretive rule is fixed: if the technology attribution holds on both dependent variables, the result holds across the construct; if it holds on the design ellipse but vanishes on the achieved miss distance, the chapter reports that the contraction may be partly a simulation-convention artifact and downgrades the confidence in H1 accordingly. The known limitation is that achieved miss distances are precise but few, since only successful landings produce them and a few early localizations are coarse, so this re-fit has even fewer effective degrees of freedom than the primary one and its inference is correspondingly weaker [\[85\]](#ref-85).

**Re-fit two: drop InSight.** The model is re-estimated with the InSight observation removed, and the technology coefficients with and without InSight are compared directly. Because InSight is the linchpin of identification, this re-fit quantifies exactly how much of the conclusion the single counterfactual carries. If the technology attribution is similar with and without InSight, the result does not hinge on one point and confidence rises; if the attribution collapses or reverses when InSight is removed, the chapter reports that the identification is InSight-dependent. That is an honest statement of the design's central vulnerability rather than a defeat, because a design that names its own load-bearing assumption is more credible than one that conceals it.

**Re-fit three: program rather than mission as the experience unit.** As described in Section 5.5, the Viking pair and the Mars Exploration Rover pair are collapsed to single program observations and the model is re-estimated on the program-level experience axis. This tests whether the technology effect survives the removal of the within-pair near-duplicate degrees of freedom and guards against the standard errors being understated by double-counting near-identical vehicles.

**Re-fit four: linear-in-levels functional form.** The augmented specification is re-estimated in levels rather than logs, as committed in Section 5.3. The comparison is on the sign and ordering of the technology terms, not on their scale, which necessarily differs between forms. Material divergence in sign or ordering is reported and lowers confidence in the proportional-rate interpretation; agreement in sign and ordering, with the expected scale difference, confirms that the headline result is not an artifact of the log transform.

**Re-fit five: control-free specification (approach-accuracy control dropped).** The nested hierarchy is re-estimated through all three technology stages but the approach-accuracy control is omitted entirely, so that the learning rate is reported without conditioning on \(\text{ApproachAccuracy}_i\). This re-fit is a committed member of the battery for a specific reason: as Section 5.7.1 now documents, the approach-accuracy control may be a post-treatment variable or a collider rather than a clean pre-treatment confounder. Adding a control that is post-treatment attenuates the upstream coefficient; conditioning on a collider opens a spurious back-door path. Either mechanism could cause the with-control specification to under-state the technology effect or, in the collider case, to produce a pattern that superficially resembles failure to reject H0. The control-free re-fit reports the technology learning rate as it stands without that potential distortion, and the comparison between re-fit five and the primary with-control specification is the direct diagnostic: a large downward revision of the technology coefficients when the control enters is the signature of post-treatment attenuation or collider bias, and it is reported as such rather than as confirmation of H0. A negligible change between re-fit five and the primary specification is consistent with the control being a clean pre-treatment covariate and strengthens confidence in the H0-versus-H1 adjudication. Both the with-control and control-free learning-rate estimates are presented in the executed study regardless of how the comparison falls. The control-free estimate is the rate attributed to onboard technology generation without conditioning on approach accuracy at all, and it is a valid primary object of interest independent of its role as a diagnostic.

Two further re-fits are noted as available but secondary. The cumulative-landings count may replace the ordinal sequence index on the experience axis, to test whether contraction tracks cumulative experience or ordinal technology generation; H1 is agnostic between these, so this is a descriptive variant rather than a test of the contribution. A leave-one-out re-estimation, removing each mission in turn, may also be run to identify any single mission whose removal materially changes the technology coefficients, a generic small-sample influence diagnostic appropriate to a population this size. The dispersion-analysis precedents confirm that influence of this kind is a real concern in landing-dispersion modeling, where a single anomalous reconstruction can shift an estimated dispersion, and so the leave-one-out diagnostic is a recognized safeguard in this literature [\[81\]](#ref-81), [\[83\]](#ref-83), [\[84\]](#ref-84), [\[89\]](#ref-89).

The robustness battery is not a search for a specification that confirms H1. It is a fixed set of re-fits, committed before estimation, whose collective purpose is to map the boundary of the conclusion: to state, for each fragile element, how far the result depends on it. The reporting rule is that all six re-fits appear in the executed study regardless of outcome, and that any re-fit which overturns the primary result becomes a reported finding in its own right.

## 5.9 Power and minimum-detectable-effect analysis

A conventional power calculation is the wrong tool for this design, and presenting one as if it justified the sample would be misleading. The honest analysis is a minimum-detectable-effect calculation that states, given the fixed population of nine to eleven events, how large a technology effect would have to be for the permutation test to distinguish it from chance. That quantity is large, which the chapter reports as a limitation rather than a strength.

Statistical power is the probability of detecting an effect of a given size at a given significance level for a given sample size, and it is ordinarily used to choose a sample size before data collection. Here the sample size cannot be chosen: the population of successful United States robotic Mars landings is fixed at nine to eleven events, every one of which is in the dataset, and no design decision can add an observation. A power analysis whose output is a recommended sample size is therefore moot. The analysis that remains meaningful is its inverse: holding the sample at the population size, and holding the permutation test's reference distribution fixed, what is the smallest technology effect, the smallest gamma coefficient, that the test could reliably distinguish from the distribution of effects obtainable by relabeling the technology indicators across missions? This is the minimum detectable effect, and it is the honest characterization of what the design can and cannot see.

When the sample size is fixed by the population and inference is by permutation, the minimum-detectable-effect under the permutation reference distribution is the correct expression of the design's sensitivity, because it answers the question the analyst can act on (how large must the effect be to be visible) rather than the question the analyst cannot (how large must the sample be). The permutation reference distribution for a technology indicator is generated by the number of distinct ways the indicator's ones can be assigned across the missions, which for an indicator that is one on two missions out of nine is a small and enumerable set of assignments. With so few distinct relabelings, the achievable resolution of the test is coarse, and the smallest detectable effect is correspondingly large. This is not a defect of the method. It is a faithful report of the information content of nine to eleven observations, and it is consistent with the dispersion-analysis literature's own caution that small numbers of reconstructed events support order-of-magnitude rather than precise inference [\[81\]](#ref-81), [\[83\]](#ref-83).

The following figures are illustrative and are presented only to convey the order of magnitude of the design's sensitivity; none has been computed from the assembled data, and all are placeholders. To illustrate the form the calculation takes, with two missions carrying guided entry out of nine, the guided-entry indicator can be relabeled in a number of distinct ways on the order of the binomial coefficient nine-choose-two, which is thirty-six, so the permutation test for that single indicator has on the order of a few dozen distinct reference values and can resolve only effects large enough to place the observed statistic in the extreme tail of that small set. For the TRN indicator, which is one on a single mission out of nine, there are only nine distinct single-mission relabelings, so the TRN effect must be very large, the Mars 2020 ellipse must be a clear outlier below the trend, for the test to register it. These illustrative counts show why the design's resolution is coarse for individual levers and why the chapter relies on the joint technology effect and the InSight residual, which pool information across the indicators, rather than on the separate identification of each lever.

A reader might object that a design with such low resolution cannot be worth executing. The answer is that the design is not aimed at a precise rate. It is aimed at a sign, an order of magnitude, and a counterfactual-surviving attribution, and the minimum-detectable-effect for that coarser claim is well within the range the qualitative record suggests the true effect occupies: the ellipse contracted by more than two orders of magnitude across the era, and an effect of that size sits far in the tail of any permutation distribution the small sample can generate. The design can see a large effect clearly and a small effect not at all, and since the scientific question is whether onboard guidance is a large lever, the design is matched to the question even though its resolution for fine apportionment is poor. The chapter reports this matching explicitly and does not overstate what the minimum-detectable-effect permits.

The practical consequence for the executed study is a reporting rule. Every coefficient will be accompanied by its permutation-based interval, and where that interval is wide enough that the minimum detectable effect exceeds the estimated effect, the chapter will state plainly that the design cannot distinguish that particular coefficient from zero, rather than reporting a point estimate as if it were resolved. Confidence in any individual lever's separate effect is **low** by construction; confidence in the joint technology effect and in the InSight-based discrimination of technology from time is **moderate**, and the power analysis is the formal statement of why those two confidence levels differ.

## 5.10 The pre-registration commitment

The design is pre-registered in the strong sense: the specifications, the order of nested entry, the decision rule on H0 and H1, the robustness battery, and the reporting discipline are all frozen in this document before the regression is run on the assembled data, so that the analysis cannot be reverse-engineered to a desired result and a null finding is as reportable as a positive one.

The credibility of a small-sample study is most threatened by analytic flexibility. With nine to eleven observations and several admissible specifications, an analyst free to choose the specification after seeing the data could almost always find one that supports a prior, and the resulting estimate would carry no evidential weight. Pre-registration removes that flexibility by fixing the analytic choices in advance. This chapter freezes, in order, the dependent variable (ln design ellipse area, with ln achieved miss distance as the pre-committed robustness dependent variable), the experience axis (mission sequence index, with the cumulative-landings count and the program-level count as pre-committed variants), the technology indicators and the approach control exactly as defined, the nested order of entry, the exact and permutation-based inference, the five-plus-two robustness re-fits (including the pre-committed control-free specification that drops the approach-accuracy control, described in Section 5.8 and reflecting the bad-control diagnostic obligation documented in Section 5.7.1), and the decision rule below.

**The fixed decision rule.** The rule is committed verbatim and is symmetric between the hypotheses. Support for H1 is declared if and only if, in the augmented specification, the gamma coefficients are jointly significant and negative under the permutation test and the delta coefficient on approach accuracy is small and insignificant, and this pattern survives the drop-InSight and achieved-miss-distance re-fits at least in sign and ordering. A failure to reject H0 is declared if the technology terms lose joint significance once the sequence trend and the approach-accuracy control are present, or if the approach-accuracy control absorbs the technology terms, or if the InSight residual sits on rather than above the trend. Any other pattern, for example joint technology significance that does not survive dropping InSight, is reported as a partial or InSight-dependent result with explicitly downgraded confidence, not forced into one of the two clean verdicts.
A decision rule fixed before the data are seen converts an exploratory exercise into a confirmatory test. The mapping from data pattern to conclusion is specified independently of the data and so cannot be tuned to them. This is the standard justification for pre-registration, and it applies with special force at small n, where analytic flexibility is most dangerous. The dissertation's measurement anchor reinforces it: Kuznets's insistence on stating the boundary, the valuation convention, and the decomposition before theorizing is, in modern terms, a pre-registration ethic, and the chapter's commitment to attach a sigma level and a simulation-convention provenance to every reported value is the operational form of that ethic [\[24\]](#ref-24).

Pre-registration freezes the confirmatory analysis; it does not forbid exploration. Any analysis run beyond the frozen plan, for instance an unanticipated diagnostic suggested by a surprising residual, will be reported as explicitly exploratory and post hoc, clearly separated from the confirmatory results, and will not be allowed to alter the verdict on H0 or H1.

One might object that freezing the plan before assembling the final data risks committing to a specification the data later reveal to be inappropriate, for example if a variable proves unmeasurable for some mission. The design accommodates this by pre-committing not only the specifications but also the handling of missing values: any mission whose approach-accuracy control or whose ellipse provenance cannot be sourced from the public record is flagged as a coverage limitation, and the model is reported both with that mission excluded and, where defensible, with the value treated as missing rather than imputed. This is committed in advance, so that the missing-data handling is itself part of the frozen plan and not a degree of freedom.

## 5.11 The computational and software plan

The computational plan is deliberately modest, because the analysis is small and the premium is on transparency and reproducibility rather than on computational scale.

The data assembly is the heaviest part of the workflow, not the estimation. The ellipse series is assembled from the NASA Technical Reports Server reconstruction records, retrieved through the public citations interface at ntrs.nasa.gov/api/citations/search and cross-referenced by DOI to the published versions, recording for each mission the reported three-sigma semi-major and semi-minor axes and the simulation convention behind each value [\[7\]](#ref-7), [\[13\]](#ref-13), [\[14\]](#ref-14). The technology indicators are coded from TechPort insertion history, recording the mission on which each lever first flew. The approach-accuracy control is assembled from the reconstruction records that report delivered entry-state dispersion [\[7\]](#ref-7), [\[13\]](#ref-13). The achieved-miss-distance series is computed from Planetary Data System localization products. Each assembled value is stored with its provenance, its source record, its sigma level, and its simulation convention, in a single coding sheet that is itself a deliverable, so that the constructed series can be audited value by value against its source.

The estimation is ordinary least squares on at most eleven observations, which is computationally trivial and can be performed in any standard statistical environment. The plan specifies a scripted, version-controlled workflow in a reproducible-research framework: a single analysis script that ingests the coding sheet, constructs the log transforms, fits the nested hierarchy and the robustness re-fits in their frozen order, runs the permutation tests, and emits the specification table and the diagnostic figure, with a fixed random seed for the permutation resampling so that the reported reference distributions are exactly reproducible. The permutation inference is the only step with any computational content, and even it is small: the reference distributions for the individual indicators are enumerated exactly rather than approximated by sampling, because, as Section 5.9 noted, the number of distinct relabelings of an indicator across nine to eleven missions is small enough to enumerate in full. Exact enumeration is preferred to Monte Carlo permutation wherever the relabeling set is small, because it removes the Monte Carlo error entirely and makes the inference deterministic given the data.

The modern guidance and optimization literature is cited here as forward context rather than as a tool of this analysis, because it bears on where the capability is heading even though it plays no role in the small regression. GPU-accelerated sequential convex programming and fast Monte Carlo dispersion analysis are the computational frontier of powered-descent guidance and of the dispersion studies that construct landing footprints for future high-mass vehicles [\[88\]](#ref-88); probabilistic and learning-based terrain-relative navigation is the frontier of the position-knowledge lever this dissertation treats as a discrete indicator [\[82\]](#ref-82). These methods will define the error budgets of the human-scale landers whose fifty-meter precision targets sit beyond the fitted curve [\[86\]](#ref-86). None of them is used to estimate the curve, and the chapter is explicit that importing them would be a category error: the contribution is a measurement-and-attribution claim about a constructed historical series, estimated by a transparent small-sample regression, not a guidance-design exercise.

The deliverables of the computational plan are four: the audited coding sheet with full provenance for every value, the analysis script with its fixed seed and enumerated permutation sets, the populated specification table and diagnostic figure, and a short robustness appendix reporting all six re-fits. All four are committed to a version-controlled repository, so that the entire path from public source records to reported result is reproducible by a third party. This reproducibility commitment is the computational counterpart of the pre-registration commitment: the former freezes the analytic choices, the latter makes the executed analysis auditable end to end.

## 5.12 The dispersion-analysis precedent for constructing the series

A final element of the design deserves its own treatment, because the credibility of the dependent variable rests on it. The constructed ellipse series inherits its conventions from an established dispersion-analysis tradition, and that lineage is what makes the cross-mission series defensible rather than an ad hoc collection of numbers pulled from heterogeneous papers.

The three-sigma ellipse for each mission is not invented by this dissertation. It is the standard output of mission-specific landing-dispersion analyses that share a common Monte Carlo methodology, and locating each value in that tradition is what licenses comparing them across missions, subject to the convention caveats already stated.

Landing-dispersion analysis is a mature subfield with a consistent core method: a Monte Carlo over the dominant entry uncertainties, atmospheric variability, delivered entry state, aerodynamic and mass properties, that propagates each sampled case to the surface and reports the resulting footprint as a covariance ellipse at a stated sigma level. This method is documented for the Mars Exploration Rovers [\[81\]](#ref-81), for Phoenix [\[83\]](#ref-83), for Mars Pathfinder [\[84\]](#ref-84), and, for a sample-return capsule that establishes the method's reach beyond Mars surface landers, for the Stardust capsule [\[89\]](#ref-89). The statistical-reconstruction methodology that converts flight data into a reconstructed trajectory and atmosphere, the basis for the post-flight ellipse and for the delivered-entry-state control, is itself a documented, reusable method rather than a per-mission improvisation [\[7\]](#ref-7). Because the same family of Monte Carlo dispersion and reconstruction methods produced the ellipses across missions, the values are members of one methodological family, and the cross-mission comparison is a comparison of like with like at the level of method even where the specific modeled error sources evolved.

A series is comparable across units when its members are produced by a common, documented measurement procedure, even if the procedure was refined over time, provided the refinements are tracked. The dispersion-analysis tradition supplies that common procedure, and the design's requirement to record each value's simulation convention tracks the refinements. The dispersion-analysis precedents converge on a shared Monte Carlo footprint methodology across vehicles as different as the Mars Exploration Rovers, Phoenix, Pathfinder, and Stardust, which confirms that the method is a stable, transferable measurement procedure rather than a mission-specific artifact [\[81\]](#ref-81), [\[83\]](#ref-83), [\[84\]](#ref-84), [\[89\]](#ref-89), and the high-elevation entry-guidance work with explicit uncertainty quantification and reduction shows that the uncertainty budgeting underlying these ellipses is itself a studied, formal exercise and not a black box [\[87\]](#ref-87).

The methodological family is common, but the modeled error sources and the simulation assumptions evolved, which is exactly the Kuznetsian comparability caveat. The design does not claim the ellipses are identically constructed; it claims they are constructed within one tradition whose evolution is documented and recorded with each value, so that comparability is a tracked and stated property rather than an assumed one. If, on assembling the series, a mission's ellipse proves to rest on a convention so different that it cannot be reconciled with the others, the value is flagged and the analysis is reported both with and without it, in keeping with the missing-data and coverage-limitation rules of the pre-registration. The earliest missions, Viking and Mars Pathfinder, are the most exposed to this risk, because their ellipse provenance rests on older records whose conventions are hardest to pin down, and the design treats any such value as a candidate for the flag rather than forcing it into the series unexamined.

The interpretive payoff of the dispersion-analysis lineage is that it connects the dependent variable to the same physical decomposition the technology covariates encode. The Monte Carlo footprint is built from distinct error sources, and the three technology levers each attack a distinct error source: guided entry the hypersonic dispersion, the range trigger the parachute-deploy dispersion, TRN the position-knowledge error [\[3\]](#ref-3), [\[4\]](#ref-4), [\[5\]](#ref-5), [\[8\]](#ref-8). The dependent variable and the independent variables are therefore two views of the same error budget, the ellipse aggregating the error sources and the indicators marking the technologies that removed them, which is the deepest reason the regression is a coherent test of the mechanism rather than a coincidental correlation of two trends. This coherence is the engineering counterpart of Mokyr's claim that each technology rests on its own propositional base, and it is what raises confidence that a surviving technology coefficient reflects a real mechanism rather than a spurious alignment of two monotone series [\[23\]](#ref-23).

## 5.13 Chapter summary and how it advances the argument

This chapter has specified the complete empirical design and argued its identification formally. The estimator is log-linear ordinary least squares, chosen because the quantity of interest is a proportional contraction rate and the dependent variable has multiplicative error spanning orders of magnitude, with a linear-in-levels re-fit carried as a committed robustness check. The baseline and augmented specifications are written out exactly as the dissertation's design fixes them, and the augmented model is estimated not as a single multicollinear horse race but through a frozen nested hierarchy that adds one technology generation at a time and reports incremental fit and exact inference at each step. Identification rests on three features acting together, the across-mission step structure, the InSight within-period counterfactual, and the approach-accuracy control, of which the InSight counterfactual is the binding and most fragile assumption, and the chapter has stated plainly that the design's identification claim is exactly as strong as the validity of that one case [\[15\]](#ref-15), [\[7\]](#ref-7), [\[13\]](#ref-13). All four classical validity threats have been paired with their mitigations, with confidence levels stated and the evidence that would raise or lower each confidence named. The robustness battery stress-tests every fragile element, and the drop-InSight re-fit in particular quantifies how much of the conclusion the single counterfactual carries. The power analysis has been reframed honestly as a minimum-detectable-effect analysis, because the sample size is fixed by the population and cannot be chosen, and it has reported that the design can see a large joint effect clearly and individual levers only weakly. The analysis is pre-registered in the strong sense, with a fixed and symmetric decision rule, and the computational plan is a transparent, reproducible, exact-enumeration workflow whose deliverables make the path from public source records to reported result auditable end to end.

The chapter contributes the design half of the dissertation's overall argument. The introduction and the data chapters establish that the problem is real and material: the ellipse contraction is large, well documented, and decision-critical for Mars Sample Return retrieval and human Mars architectures [\[1\]](#ref-1), [\[5\]](#ref-5), [\[8\]](#ref-8), [\[86\]](#ref-86). This chapter establishes the next two steps. It establishes that the design addresses the causal mechanism, by conditioning on the approach-navigation rival and modeling discrete technology generations as covariates on a measurement-disciplined series, so that a surviving technology coefficient is attributable to onboard guidance rather than to delivery accuracy or to generic time [\[7\]](#ref-7), [\[13\]](#ref-13), [\[24\]](#ref-24). And it establishes that the design discriminates the technology account from a qualitative single-mission study, because the InSight counterfactual plus the approach control separate the technology hypothesis from the pure-time and injection-accuracy rivals in a way that no single-mission paper can [\[15\]](#ref-15). The final step, that the residual risk is acceptable, is the candid posture of the whole chapter: the small-n and collinearity risks are real and are not eliminated, but they are bounded by permutation inference, the nested hierarchy, the dual dependent variable, and a robustness battery that names its own load-bearing assumption, and the design is informative whichever way the coefficient falls, because a null result, reported under the frozen decision rule, is as much a contribution to the requirements-setting question as a positive one. Confidence in the overall design is **moderate to high** as a design, and deliberately **moderate** as to what its execution can resolve, and the chapter has been written to make that distinction, between the quality of the design and the resolution of the data, visible rather than blurred. The analysis plan that operationalizes this design, step by step and with its expected signs and illustrative output tables, is the subject of Chapter 6.


# Chapter 6: Analysis Plan and Expected Results

## 6.1 The chapter thesis

This chapter commits the dissertation, in advance of seeing the estimates, to a single decision rule on the hypothesis pair and to reporting whichever pattern the assembled data reveal. That is the answer this chapter delivers, and every subsection develops it. The plan fixes the order of operations, the estimator, the inferential procedure, the robustness battery, and the rule that maps a configuration of coefficients onto a verdict for H1 or H0, before any number is computed on the constructed series. The expected signs are stated with their mechanism reasoning so that the prediction is genuinely falsifiable rather than retrofitted, and the form of every output table is shown with illustrative placeholders so that a reader can see exactly what the executed study would produce without mistaking the placeholders for findings. This is the pre-registration that converts the design of Chapter 5 into an executable protocol.

The reason this matters is specific to the small-sample, design-stage character of the work. A study with nine to eleven observations and three nearly collinear technology indicators is exactly the kind of study in which an analyst, seeing the data first, can find a specification that supports a preferred conclusion. The discipline that protects the contribution is not statistical power, which is modest by construction, but the prior commitment to a fixed analysis and a fixed decision rule. The chapter therefore reads as a contract: here is the procedure, here is the rule, here is what each outcome will be taken to mean, and here is the reporting standard that attaches a measurement boundary to every reported value per Kuznets. Nothing in the chapter is an executed estimate. The regression has not been run on the assembled dataset, and the chapter says so at every point where a number appears.

### 6.1.1 Problem frame for this chapter

The current state is a research design, set out in Chapter 5, that names an estimator, an identification strategy resting on the InSight counterfactual and the approach-accuracy control, and a battery of validity threats with their mitigations. The desired state is a protocol so completely specified that a second analyst, handed the named datasets and this chapter, would produce the same tables, run the same tests, and reach the same verdict from the same coefficients, with no degrees of freedom left for the result to be steered after the fact. The gap is that a design is not yet a protocol: a design says what model to fit, while a protocol says in what order, with what inferential machinery, under what decision rule, and with what reporting of provenance. The consequence of leaving that gap unclosed is that the small-sample fragility the design honestly acknowledges would become an opening for exactly the post-hoc specification search the dissertation is built to foreclose. This chapter closes the gap by writing the protocol down and freezing it.
## 6.2 The pre-registered estimation procedure

The procedure has six ordered steps. The ordering is itself a commitment, because several of the steps could in principle be reordered to flatter a result, and fixing the sequence removes that latitude. Each step is stated here as it will be executed, and the corpus evidence that grounds the method choice is interpreted for what it contributes rather than merely cited.

**Step 1. Assemble the ellipse series.** From the NTRS reconstruction records and the published EDL performance studies, record for each of the eight to eleven missions in the population the reported three-sigma landing-ellipse semi-major axis a and semi-minor axis b, together with the simulation convention that produced them. The MSL range-trigger study reports the footprint reduction the range trigger buys and so anchors the MSL design ellipse [\[5\]](#ref-5); the Mars 2020 Lander Vision System flight-performance study anchors the post-TRN targeting region [\[8\]](#ref-8); the InSight reconstruction and the InSight site-selection record anchor the late, ballistic, large-ellipse control point [\[15\]](#ref-15). Each value is entered with its provenance attached. Where a Viking-era or Pathfinder-era three-sigma value rests only on an older NTRS record whose simulation convention cannot be pinned down, the value is flagged as a comparability caveat rather than silently harmonized, which is the Kuznetsian discipline made operational at the data-entry stage. This step produces no estimate. It produces the raw series and its provenance ledger.

**Step 2. Compute the dependent variables.** Form the primary dependent variable \(\ln(\text{EllipseArea}_i) = \ln(\pi \cdot a_i \cdot b_i)\) for each mission, and, separately, the secondary dependent variable \(\ln(\text{achieved miss distance})\) from the PDS localization products where they exist. The log transform is not an analytic convenience adopted at this step. Chapter 5 defended it as the form that makes the model multiplicative in levels, stabilizes the variance across ellipses that span several orders of magnitude, and matches the convention in which the experience-curve literature reports rates. The two dependent variables are carried in parallel from this point forward, so the construct-validity check (design ellipse versus achieved accuracy) is built into the procedure rather than appended to it. Where PDS yields a miss distance for only a subset of missions, the secondary-DV analysis is run on that subset and the reduced coverage is reported, not imputed.

**Step 3. Code the covariates.** Set the experience axis \(\text{Sequence}_i\) to the mission sequence index, 1 for Viking 1 through the index of Mars 2020, and form the alternative cumulative-landings count for the robustness specification. Code the three binary technology indicators from TechPort insertion history and the corroborating flight literature: \(\text{GuidedEntry}_i\) equal to one for MSL and Mars 2020, \(\text{RangeTrigger}_i\) equal to one for MSL and Mars 2020, and \(\text{TRN}_i\) equal to one for Mars 2020 only. Code the control \(\text{ApproachAccuracy}_i\) from the reported approach-navigation delivery accuracy, the entry-flight-path-angle or entry-point delivery dispersion, drawn from the navigation and reconstruction records. The coding sheet is frozen at this step and reported in full in the backmatter, so the mapping from mission to indicator vector is auditable and fixed before any regression is run.

**Step 4. Fit the baseline.** Estimate \(\ln(\text{EllipseArea}_i) = \beta_0 + \beta_1 \cdot \text{Sequence}_i + \epsilon_i\) by ordinary least squares. This recovers the constant proportional contraction per mission generation: \(\exp(\beta_1) - 1\) is the proportional change in ellipse area per unit of sequence, the learning-rate reading fixed in the notation. The baseline is fit first and reported first, because the headline contraction is not in dispute and the baseline establishes the magnitude against which the technology terms must earn their incremental fit. The baseline is not the test of the contribution; it is the backdrop for it.

**Step 5. Fit the nested augmented specifications.** Add the technology generations one at a time, in their historical order, and at each step record the incremental adjusted R-squared and the test on the added term: baseline, then plus guided entry, then plus the range trigger, then plus TRN, then plus the approach-accuracy control. This nested hierarchy is the cliometric decomposition of the aggregate contraction into attributable components, and it is the deliberate alternative to throwing all four covariates into one over-parameterized regression that the small sample cannot support. The full augmented specification, \(\ln(\text{EllipseArea}_i) = \beta_0 + \beta_1 \cdot \text{Sequence}_i + \gamma_1 \cdot \text{GuidedEntry}_i + \gamma_2 \cdot \text{RangeTrigger}_i + \gamma_3 \cdot \text{TRN}_i + \delta \cdot \text{ApproachAccuracy}_i + \epsilon_i\), is reported as the final row, but it is read through the lens of the incremental fits rather than as a standalone over-fit.

**Step 6. Run the robustness battery and the inference.** Re-fit the whole hierarchy on the achieved-miss-distance dependent variable; re-fit excluding the InSight observation to quantify how much of the identification leans on that single counterfactual; re-fit with program rather than mission as the experience unit, collapsing the Viking pair and the MER pair; and re-fit the linear-in-levels form as the functional-form check. For every coefficient of interest, compute permutation-based and exact inference appropriate to n equal to nine to eleven rather than relying on asymptotic standard errors. The precedent for constructing per-mission distributions is established in the Mars Exploration Rover and Phoenix dispersion studies and in the Pathfinder entry reconstruction, which show that a per-mission landing distribution is a routinely reconstructed object rather than a novel one [\[81\]](#ref-81), [\[83\]](#ref-83), [\[84\]](#ref-84). The GPU-accelerated Monte Carlo framework for powered-descent guidance demonstrates that fast, large-sample resampling of EDL trajectory outcomes is now computationally cheap, the forward-method backdrop against which the permutation inference on the constructed series is unremarkable [\[88\]](#ref-88).

The inferential machinery invoked in step 6 warrants a note here, because it is the part of the procedure most easily underspecified and most consequential for a sample of this size. With nine to eleven observations, the asymptotic theory that licenses the usual t and F statistics does not apply: the sampling distributions of the coefficients are not well approximated by their normal limits, and a p-value read off a t-table would misstate the evidence. The plan therefore replaces asymptotic inference with a permutation procedure. For the joint test on the technology block, the procedure holds the sequence index and the approach-accuracy control fixed, permutes the assignment of the technology indicator vectors across missions, re-fits the augmented specification under each permutation, and accumulates the distribution of the joint test statistic under the null that the technology labels are exchangeable. The observed statistic is then located in that permutation distribution to yield an exact, finite-sample significance level that makes no asymptotic assumption. Because the indicator structure is nested (every TRN mission is a guided-entry mission), the permutation set is restricted to label assignments that respect the historical monotonicity where that structure is part of the maintained hypothesis, and is left unrestricted in a second pass to show how much of the significance depends on the monotone structure itself. Reporting both passes is part of the honesty the small sample demands: the restricted pass tests the technology effect given the generation ordering, the unrestricted pass tests whether the ordering alone is doing the work.

The plan also fixes, in advance, the power and minimum-detectable-effect reasoning that bounds what the test can find. A formal power calculation in the usual sense is not meaningful for a fixed, full-population series of nine to eleven points, because there is no resampling of units from a larger population; the population is the entire robotic-era record of successful U.S. Mars landings. What is meaningful, and what the plan commits to reporting, is a minimum-detectable-effect statement framed against the permutation distribution: given the fixed design matrix and the permutation null, how large must a technology coefficient be, in log-area units, for the observed statistic to fall outside, say, the upper fifth percentile of the permutation distribution. That threshold is a property of the design matrix alone and can be computed before the dependent variable is touched, which makes it a genuinely pre-registerable quantity. Reporting it serves two purposes: it tells the reader in advance how large an effect the design can resolve, and it guards against the small-sample temptation to read a non-significant correctly-signed coefficient as a confirmation. If the minimum detectable effect is large relative to the order-of-magnitude drops the public record suggests, the plan says so, and the discrimination then rests more heavily on the InSight counterfactual than on the joint significance test, which is exactly the contingency the tie-breaking provision of Section 6.3 anticipates.

That this six-step procedure is sufficient to test the contribution can be argued directly. The ordered procedure, executed on the named datasets, returns the evidence needed to decide between H1 and H0, because it produces a constructed series with provenance, two dependent variables, a frozen covariate coding, a nested decomposition, and small-sample-appropriate inference, which jointly map onto every term in the augmented specification. A hypothesis that predicts a specific sign and ordering of coefficients is tested by estimating those coefficients with inference calibrated to the sample size, which is what steps 4 through 6 do. This follows the cliometric tradition, in which an aggregate change is decided not by a single regression but by a disciplined decomposition into attributable components with the measurement boundary stated [\[24\]](#ref-24), and the experience-curve tradition, in which a rate is estimated only after the experience axis and the metric are pinned down [\[18\]](#ref-18), [\[19\]](#ref-19). The procedure is informative, not dispositive: with this sample size it returns a sign, an order of magnitude, and a counterfactual-survival verdict, not a precise point estimate with a narrow interval. The objection it must survive is that the procedure could appear to support H1 through collinearity alone, with the technology indicators standing in for an omitted monotone trend; step 5's nested reporting and step 6's InSight-excluded re-fit are the explicit defenses against that objection, and they are run regardless of how the baseline falls. Confidence in the sufficiency of the procedure is high; confidence in the precision of any single coefficient it will produce is, by design, low, and the chapter never conflates the two.

## 6.3 The fixed decision rule

The decision rule is stated now, before any estimate, and is binding. It is written so a reader can apply it mechanically to the executed coefficient table and reach the same verdict the candidate would.

**Support for H1 requires both of the following.** First, the technology coefficients \(\gamma_1\), \(\gamma_2\), \(\gamma_3\) are jointly significant under the permutation test and each carries a negative sign, meaning each technology generation is associated with a reduction in log ellipse area conditional on the sequence trend. Second, the approach-accuracy coefficient delta is small in magnitude and statistically insignificant, meaning the delivered-entry-state accuracy does not absorb the contraction once the technology indicators are present. The conjunction matters: a negative, jointly significant technology block with an insignificant approach control is the precise coefficient signature that the onboard-guidance mechanism predicts and that the injection-accuracy rival does not.

**Failure to reject H0 follows from either of the following.** First, the technology terms are jointly insignificant once the sequence trend or a time trend is included, so that ordering the missions by date explains the contraction as well as ordering them by guidance generation. Second, the approach-accuracy control delta is significant and absorbs the explanatory power of the technology block, so that the contraction loads on getting to the atmospheric interface more accurately rather than on onboard guidance. Either configuration is reported as a failure to reject the null, and either is itself decision-relevant: the first says the era's improvement is generic maturation, the second says approach navigation is the binding lever, and both redirect where a future project should spend to shrink an ellipse.

The rule includes a tie-breaking provision for the configuration the small sample makes most likely: a technology block that is correctly signed but not jointly significant at conventional permutation thresholds. In that case the verdict is "consistent with but not confirming H1," and the InSight counterfactual is given decisive weight. If the technology block is correctly signed, the approach control is insignificant, and the InSight residual sits visibly above the fitted trend, the study reports qualified support for H1 with the qualification stated plainly. If the technology block is correctly signed but the InSight residual sits on the trend, the study reports that the design cannot discriminate the technology hypothesis from generic maturation on this sample, which is an honest non-result rather than a hidden one. This provision is what prevents the small sample from forcing a binary verdict the data cannot bear.

The decision rule deserves a defense of its own. The rule correctly partitions coefficient configurations into verdicts on H1 and H0 because the augmented specification was constructed so that each hypothesis predicts a distinct, observable coefficient pattern, stated in the notation as "gamma jointly significant and negative, delta small and insignificant" for H1 and "gamma absorbed by sequence or by delta" for H0. A falsifiable hypothesis is one that names in advance which observable patterns confirm and which disconfirm it, and the rule is that naming made explicit, in keeping with the falsifiability requirement restated in the dissertation's own contribution section, which demands a specific sign and ordering that the data can contradict. Because the thresholds are permutation-based and the sample is small, the rule's boundary cases are handled by the explicit tie-breaking provision rather than by a brittle p-value cutoff. The objection the rule must withstand is that a determined analyst could relabel a boundary case to taste; the freezing of the rule before estimation, together with the published tie-breaking provision, is the defense. Confidence that the rule is well posed is high, because it is derived directly from the pre-existing notation rather than invented here.

## 6.4 Expected signs and their mechanisms

The plan predicts signs, not magnitudes, and it predicts them with a mechanism for each so the prediction is anchored in physics and economic history rather than in hope. The mechanism chain for each lever follows the dissertation's named-mechanism structure: driver, mechanism, observable effect, operational consequence, strategic implication. Stating the chain in full is what distinguishes a falsifiable mechanistic prediction from a bare expectation of correlation. Where only correlation can be recovered from the small sample, the text says so and downgrades the confidence accordingly.

### 6.4.1 Guided lifting entry

The expected sign of \(\gamma_1\) is negative and, among the three levers, expected to carry one of the two largest increments in fit. The driver is the insertion of closed-loop bank-angle guidance on a lifting entry vehicle, first flown at Mars by MSL [\[5\]](#ref-5), [\[72\]](#ref-72). The mechanism is that closed-loop modulation nulls the downrange and crossrange dispersion accumulated during the long hypersonic phase, the dispersion that a ballistic capsule cannot correct because it flies wherever its delivered entry state and the atmosphere carry it. The observable effect is a discrete drop in the three-sigma ellipse at the MSL mission, the first to fly the technology, and not before. The operational consequence is that a project can target a constrained site rather than accepting only uniformly bland terrain across hundreds of kilometers. The strategic implication is that the dominant early error source, hypersonic dispersion, becomes a controlled rather than an accepted quantity. In Mokyr's vocabulary, guided entry is closer to a macro-invention than to an incremental refinement: it changed the entry regime from ballistic to lifting and removed the dominant error source in one step, resting on a maturing propositional base of entry aerodynamics and atmospheric reconstruction. Expected sign: negative, large. Confidence in the sign: high. Confidence in the magnitude: low, because guided entry and the range trigger entered on the same mission and the sample cannot cleanly separate their individual contributions, a collinearity the chapter states rather than hides.

### 6.4.2 The range-to-go parachute trigger

The expected sign of \(\gamma_2\) is negative and expected to carry the smallest of the three technology increments. The driver is the replacement of a velocity-based parachute-deploy command with a range-to-go command, the Smart Chute logic flown on MSL [\[5\]](#ref-5). The mechanism is narrower than guided entry: it attacks the dispersion introduced specifically at parachute deploy by deploying earlier or later to compensate for how far downrange the vehicle has actually flown, leaving the hypersonic-phase dispersion to the guided-entry lever. The observable effect is an additional, smaller drop in the ellipse at MSL, on top of the guided-entry drop, which is why the two cannot be cleanly separated on a single shared mission. The operational consequence is a further tightening of the achievable footprint that the range-trigger study quantifies directly [\[5\]](#ref-5). The strategic implication is modest by comparison: it is a refinement of an existing deploy logic rather than a new regime. In Mokyr's terms the range trigger is incremental, a self-correcting improvement to a technique already in the prescriptive base. Expected sign: negative, small. Confidence in the sign: moderate, lower than for guided entry because the range trigger's effect is the hardest to isolate, sharing its mission with guided entry and contributing the smaller share of a joint drop. The chapter treats the MSL increment as primarily a guided-entry-plus-range-trigger bundle and is explicit that the within-bundle apportionment is the weakest inference in the study.

**A qualification on the range trigger's propositional base.** The Mokyrian characterization of the range trigger as incremental rests on the presumption that its propositional base, the body of engineering knowledge documenting where the velocity-based deploy logic leaves residual dispersion and how a range-to-go command eliminates it, is dated and independently codified in the literature. For guided entry and TRN, that codification is strong: the entry-guidance propositional base runs through the MSL guidance, navigation, and control record and the MEDLI2 reconstruction lineage, and the TRN propositional base runs through the Lander Vision System development and the Johnson et al. flight-performance record [\[8\]](#ref-8), [\[10\]](#ref-10), [\[11\]](#ref-11). For the range trigger, the primary external documentation is the Way 2011 range-trigger study itself [\[5\]](#ref-5), and a dated, independent account of the propositional base for the range trigger's development, analogous to the aerodynamics and reconstruction literature that backs guided entry, has not been identified in the Phase 0 corpus search. This leaves the range trigger as the weakest of the three levers in the Mokyrian sense: its existence as a technology insertion is well-documented, its mechanism is physically clear, but the external propositional-base record that would date its maturation trajectory and confirm it as a codified technique ready for broad adoption is thinner than for the other two levers. The Phase 1 execution should specifically search for independent range-trigger propositional-base documentation, including MSL Aerospace Conference proceedings and the entry-guidance development literature, to close this gap before the Mokyrian ordering of the three levers is treated as fully validated.

### 6.4.3 Terrain-relative navigation
The expected sign of \(\gamma_3\) is negative, and this lever should carry the largest single technology increment, observed at the Mars 2020 transition. The driver is the insertion of terrain-relative navigation and the Lander Vision System, with autonomous Safe Target Selection, first flown at Mars by Mars 2020 [\[8\]](#ref-8). The mechanism is distinct from both prior levers. TRN attacks the position-knowledge error itself, the gap between where the vehicle believes it is and where it actually is, by matching descent imagery to an onboard reference map, and it then enables an autonomous divert away from hazards that no prior lever could perform. The observable effect is a drop in the ellipse at Mars 2020 that no other mission shares, which makes TRN the most cleanly identified of the three levers, since Mars 2020 is the only mission carrying it. The operational consequence is the delivery of Perseverance into a hazardous site that earlier systems could not have attempted at all, not merely a smaller ellipse within a benign site. The strategic implication is that landing precision becomes decoupled from site benignity: the vehicle can be placed away from local hazards rather than requiring their absence across the whole ellipse. In Mokyr's terms TRN is again macro. It introduced an entirely new error source to attack and a capability with no predecessor, resting on the maturing propositional base of computer vision and onboard orbital mapping. Expected sign: negative, largest. Confidence in the sign is high, and confidence in the relative magnitude is higher here than for the MSL bundle precisely because TRN is borne by a single mission and is not confounded with a co-flown lever. The forward literature on autonomous powered-descent and divert guidance, including the convex-optimization powered-descent line and the flight-tested terrain-relative-navigation-plus-large-divert testbed, establishes that the divert capability TRN enables is a mature and actively extended technology rather than a one-off, which strengthens the Mokyrian reading of TRN as an extensible addition to the prescriptive base rather than an isolated trick [\[91\]](#ref-91), [\[92\]](#ref-92), [\[95\]](#ref-95), [\[108\]](#ref-108).

### 6.4.4 The approach-accuracy control

The expected sign and significance of delta is the hinge of the whole test. Under H1, delta is expected to be small and insignificant: once the technology indicators are present, the delivered-entry-state accuracy should not absorb the contraction, because the dissertation's thesis is that onboard guidance, not injection accuracy, is the binding constraint on ellipse size. The mechanism reasoning is that approach navigation determines how accurately the vehicle arrives at the atmospheric interface, but the onboard guidance determines how much of any arrival error is corrected during entry, descent, and landing. A vehicle delivered precisely to the interface but flying ballistically still inherits a large ellipse, while a vehicle delivered with ordinary accuracy but flying guided entry with TRN can null much of that error. The MSL navigation results and the MEDLI2 reconstruction supply the delivered-accuracy values that operationalize this control [\[13\]](#ref-13), [\[77\]](#ref-77). The observable effect predicted under H1 is that the technology coefficients retain their sign and rough magnitude when delta enters the specification, which is the last row of the nested hierarchy. The contrary prediction, which the rule treats as failure to reject H0, is that delta is significant and the technology block collapses when it enters. Expected under H1: delta small and insignificant. Confidence is moderate, because the approach-accuracy series is the control variable most likely to have coverage gaps in the public record, and any mission whose delivered-accuracy value cannot be sourced is flagged as a coverage limitation rather than imputed, which weakens the control to the extent the gaps bind.

### 6.4.5 Forward-method context: why the divert capability is extensible

The expected signs above are predictions about a historical series, but the plan also fixes how that series is to be read against the forward edge of the guidance literature, because the Mokyrian reading of the result depends on whether the levers rest on an extensible propositional base or were one-off achievements. The plan treats the powered-descent and convex-optimization guidance literature as forward context, not as in-sample evidence, and uses it to adjudicate one question: is the TRN-plus-divert capability that drove the Mars 2020 ellipse drop a mature, extensible technology that future architectures will inherit and improve, or a bespoke artifact unlikely to transfer? The convex-programming approach to powered-descent guidance, which reformulates the nonconvex minimum-fuel pinpoint-landing problem as a solvable second-order cone program, is the foundational result establishing that the divert trajectory can be computed onboard with deterministic convergence [\[92\]](#ref-92), [\[95\]](#ref-95). The subsequent line of pseudospectral and sequential-convex-programming work extends that result to higher-fidelity vehicle models, free final time, and continuous-time constraint satisfaction, and the survey of autonomous powered-descent guidance methods documents the field's trajectory toward tens-of-meters accuracy as a maturing standard rather than a one-time demonstration [\[93\]](#ref-93), [\[99\]](#ref-99), [\[107\]](#ref-107), [\[108\]](#ref-108). The flight-tested terrain-relative-navigation-and-large-divert testbed on a vertical-takeoff vertical-landing rocket shows the integrated capability demonstrated end to end before its Mars insertion, the technology-readiness evidence that the lever rested on a base mature enough to fly [\[91\]](#ref-91). Interpreted for the dissertation's argument, this convergence means that if \(\gamma_3\) is large and negative, the result is not the signature of a lucky one-off; it is the signature of an extensible addition to the prescriptive base, exactly the Mokyrian pattern in which a technique resting on a deep propositional foundation (convex optimization, computer vision, onboard mapping) is self-correcting and carries forward. This reading is forward context that strengthens the causal interpretation of a confirmed H1; it is not used to estimate any coefficient, and the plan is explicit that the historical series, not the forward literature, decides the hypothesis.

The composite expected pattern, stated as the prediction the executed study will confirm or reject, is therefore: a large negative baseline sequence coefficient; the largest incremental fits at the MSL transition (guided entry plus range trigger, jointly) and at the Mars 2020 transition (TRN); a negative, jointly significant technology block; a small and insignificant approach-accuracy control; and an InSight observation sitting above the fitted trend. Any departure from this pattern is informative, and several departures would falsify the contribution outright, which is the property the program requires.

## 6.5 The design of the illustrative simulation

To show the form of the output without producing an executed estimate, the plan specifies an illustrative simulation whose sole purpose is to display the shape of the tables and the figure the executed study will populate. The simulation is described here so that a reader understands exactly what it is and, more to the point, what it is not. It is not a fit to the assembled data, it is not a power simulation calibrated to expected effect sizes, and it produces no number that should be read as a finding. It is a layout device.

The illustrative construction proceeds as follows. Take the qualitative ordering the public record supports, ellipses falling from a Viking-class major dimension in the high hundreds of kilometers, to roughly one hundred kilometers at MER, to roughly twenty kilometers at MSL after guided entry and the range trigger, to a few kilometers of effective targeting at Mars 2020 after TRN [\[5\]](#ref-5), [\[8\]](#ref-8). Place these order-of-magnitude anchors on the log axis purely to show the monotone descent and the steps at MSL and Mars 2020. Draw a straight line through them to show what a baseline log-linear fit would look like as a line, and mark where an InSight point with a deliberately large, late ellipse would sit above that line. The resulting figure is captioned, in the executed study, as the single primary figure: log ellipse area against mission generation, with the fitted line and the InSight residual marked. In the design-stage document the same figure is shown with every value labeled illustrative and with no axis number presented as an estimate.

The illustrative coefficient table takes the form below. It is the form the executed nested hierarchy will populate. The entries are placeholders that describe the expected qualitative behavior, not estimates; no cell contains a computed number, and the table is specified-but-unpopulated by design.

| Specification | Sequence coefficient | Technology terms | Approach control | Adjusted R-squared |
|---|---|---|---|---|
| Baseline (sequence only) | negative, large [illustrative] | none | not entered | high [illustrative] |
| + Guided entry | smaller [illustrative] | guided-entry term negative [illustrative] | not entered | higher [illustrative] |
| + Range trigger | smaller [illustrative] | range-trigger term negative, small [illustrative] | not entered | higher [illustrative] |
| + TRN | smaller [illustrative] | TRN term negative, largest [illustrative] | not entered | highest [illustrative] |
| + Approach-accuracy control | little changed [illustrative] | technology terms retain sign [illustrative] | small, insignificant [illustrative] | highest [illustrative] |

A companion robustness table is specified with the same discipline: rows for the achieved-miss-distance dependent variable, the InSight-excluded re-fit, the program-as-experience-unit re-fit, and the linear-in-levels form, with columns for the technology-block sign, the technology-block joint-significance verdict, and the InSight residual position. Every cell is a placeholder describing the expected qualitative behavior and labeled illustrative; no cell is an estimate. Showing both tables in the design-stage document makes the reporting commitment concrete: this is the exact shape of the output, and the executed study will fill these cells with permutation-based estimates and verdicts and nothing else, so that there is no room to add or drop a specification after the fact.

That an illustrative simulation belongs in a design-stage analysis plan is itself defensible. Displaying the table form with labeled placeholders strengthens rather than weakens the pre-registration, because a frozen table template removes a degree of freedom: the executed study cannot quietly add a flattering specification or drop an unflattering one without the addition or deletion being visible against the template. Pre-registration works by fixing the analysis surface in advance, and a table template is part of that surface, in keeping with the cliometric requirement that the form of a constructed measurement and its decomposition be stated before inference [\[24\]](#ref-24). The illustrative values carry no evidential weight and are labeled at every occurrence to prevent their being mistaken for results. The device must survive the risk that a careless reader confuses placeholders with estimates; the explicit "[illustrative]" label in every cell and the repeated statement that the regression has not been executed are the defense. Confidence that the device is appropriate is high; it is standard pre-registration practice translated to this setting.

## 6.6 The event-study and profile interpretation of the InSight residual

The single most consequential reading in the executed study is the position of the InSight observation relative to the fitted trend, and the plan specifies in advance how that reading will be made and bounded. The interpretation is structured as a one-case event-study profile: InSight is the late mission whose calendar matches MSL and Mars 2020 but whose guidance suite matches Viking and Phoenix, and its residual from the sequence-only fit is the quantity that discriminates the technology hypothesis from the pure-time hypothesis.

The logic is as follows. InSight launched and landed in 2018, six years after MSL demonstrated guided entry and the range trigger, yet it deliberately flew a Phoenix-heritage ballistic, unguided entry into a large, flat ellipse in Elysium Planitia, because its seismology and heat-flow science did not require precision and its budget did not justify the guidance suite [\[15\]](#ref-15), [\[74\]](#ref-74). Under the pure-time hypothesis, in which contraction is generic maturation that any late mission inherits for free, InSight should sit on or near the sequence-only trend, having a small ellipse simply by virtue of its late date. Under the technology hypothesis, InSight should sit visibly above the trend, carrying the large ellipse that a modern program accepts when it chooses not to fly the technology levers. The residual is therefore a direct test: a large positive residual is evidence for the technology hypothesis and against generic maturation; a near-zero residual is evidence that time alone, or whatever improved monotonically over the era, explains the contraction.

What the InSight case proves and does not prove is stated with care, because the design leans heavily on one observation and overstating its weight would be the most damaging error the dissertation could make. What it can establish is the discrimination between the technology hypothesis and the pure-time hypothesis: if the residual is large, the contraction cannot be a featureless function of the calendar, because here is a late mission that did not inherit the small ellipse. What it cannot establish is the apportionment of credit among guided entry, the range trigger, and TRN, because InSight lacks all three at once and is a single point; it speaks to whether the technology block matters, not to which lever within the block matters most. That apportionment is the work of the nested specifications, and it is where the small sample bites hardest. The plan therefore treats the InSight residual as the linchpin of the technology-versus-time discrimination and declines to use it for within-block apportionment.

The event-study interpretation has a known fragility that the plan confronts rather than glosses. Because identification leans on this one counterfactual, the robustness battery includes the InSight-excluded re-fit specifically to quantify the dependence: re-fitting the entire hierarchy without InSight shows how much of the technology-block significance and how much of the discrimination survives when the linchpin is removed. If the technology block collapses without InSight, the study reports that the identification is counterfactual-dependent to a degree that the reader should weigh, an honest statement of a real limitation. If the technology block largely survives without InSight, drawing then on the across-mission step structure alone, the study reports that the InSight case reinforces rather than carries the identification. Either way the dependence is measured and reported, not assumed away. The dispersion-analysis precedents confirm that a per-mission ellipse and its reconstruction are well-defined objects for InSight as for the others, so the InSight point is as solid a datum as the rest of the series, and its leverage comes from its design-counterfactual position rather than from any fragility in its measurement [\[81\]](#ref-81), [\[83\]](#ref-83), [\[84\]](#ref-84).

The logic of the InSight reading can be made explicit. A large positive InSight residual discriminates the technology hypothesis from the pure-time hypothesis, because InSight holds the calendar of the precision era while lacking its guidance suite, so its ellipse is, by construction, the ellipse a modern program gets without the levers [\[15\]](#ref-15). A counterfactual case sharing the confound (time) but not the treatment (technology) isolates the treatment's contribution, the standard logic of a within-period control, and it is the design's own identification argument from Chapter 5, where InSight is named the linchpin that breaks the otherwise perfect time-technology collinearity. One observation discriminates a hypothesis pair but cannot apportion an effect, and the residual's leverage is real but singular. The reading must survive the possibility that InSight is anomalous for reasons unrelated to its guidance choice, for instance an atmospheric or delivery peculiarity; the reconstruction record, which reports that InSight's performance metrics fell near the boundaries of prediction with an uprange and crossrange landing, is examined to confirm that its large ellipse reflects the accepted design choice rather than an off-nominal entry [\[74\]](#ref-74). Confidence in the discrimination, conditional on a large residual, is moderate-to-high; confidence in any apportionment from InSight alone is low and is not claimed.

## 6.7 The robustness battery in detail

The robustness re-fits named in step 6 are not afterthoughts; each one targets a specific threat to validity from Chapter 5, and the plan fixes in advance what a given re-fit would show under H1 and under H0 so that the battery cannot be read selectively after the fact. Four re-fits are specified.

The achieved-miss-distance re-fit addresses the construct-validity threat that the design ellipse, a pre-flight simulation product, may diverge from the capability actually demonstrated. Re-running the entire nested hierarchy with \(\ln(\text{achieved miss distance})\) from PDS localization as the dependent variable tests whether the technology block retains its sign and significance on a measure that is independent of the simulation conventions and is, in Kuznets's terms, a different valuation of the same underlying construct. Under H1 the two dependent variables should tell the same qualitative story; a technology block that is significant on the design ellipse but vanishes on the achieved miss distance would warn that the contraction lives in the simulation conventions rather than in demonstrated capability, and the plan commits to reporting that divergence as a construct-validity failure rather than suppressing the weaker of the two. The known limitation is coverage: PDS yields a miss distance only for successful landings and only where the localization product exists, so the secondary-DV hierarchy runs on a reduced set, and the reduction is reported with the result.

The InSight-excluded re-fit, already discussed in Section 6.6, is restated here as a formal member of the battery because its role is to quantify the identification's dependence on the single counterfactual. The plan reports the technology-block sign, joint-significance verdict, and the change in incremental fit with InSight removed, so that the reader sees exactly how much of the conclusion survives the loss of the linchpin.

The program-as-experience-unit re-fit addresses a subtle threat in the experience axis. The Viking pair and the MER pair are near-identical vehicles flown within the same program, which means the mission-sequence index treats two nearly redundant events as two independent steps of experience. Collapsing each pair to a single program-level observation, and re-indexing the experience axis by program rather than by mission, tests whether the contraction tracks genuine generational change or is partly an artifact of counting redundant vehicles as separate generations. Under H1 the technology block should be robust to this recoding, because the levers entered at the program level (MSL, Mars 2020) regardless of how the redundant pairs are counted. A technology effect that depends on counting the Viking and MER pairs as four steps rather than two would be fragile, and the plan flags that contingency in advance.

The linear-in-levels re-fit is the functional-form check. The log specification was defended in Chapter 5 on three grounds, multiplicative reading, variance stabilization, and learning-curve convention, but the choice is not innocent, and a model in levels would weight the enormous early residuals very differently. Re-fitting the augmented specification in levels, with ellipse area rather than its logarithm as the dependent variable, tests whether the sign and ordering of the technology effects are an artifact of the transform. Divergence between the two forms is reported rather than suppressed; convergence strengthens the claim that the multiplicative reading is not merely a convenient lens.
The control-free re-fit is the bad-control diagnostic committed in Section 5.7.1 and Section 5.8. It runs the nested hierarchy through all three technology stages without entering \(\text{ApproachAccuracy}_i\) at any step, and it reports the resulting technology learning-rate estimate alongside the primary with-control estimate. The purpose is to measure how far conditioning on approach accuracy changes the attribution. If the technology block is substantially larger in the control-free specification than in the primary specification, that difference is the empirical signature of post-treatment attenuation or collider bias; it is reported as such and given weight in the final assessment of confidence. If the two specifications return a learning rate and a technology-block pattern that are substantively similar, the control is consistent with being a clean pre-treatment covariate, and the primary specification carries its standard interpretation. Under H1, the technology block in the control-free specification should be somewhat larger than in the primary specification, because removing a partial mediator of the technology effect un-attenuates the upstream coefficients, but the sign and ordering should be the same. Under H0, the control-free specification should show a technology block that is larger but still non-significant, because the injection-accuracy rival rather than the technology levers drives the contraction. The control-free learning rate, the rate without conditioning on approach accuracy at all, is itself a valid and reported quantity regardless of the comparison, since it characterizes the technology attribution under the weakest covariate assumption.

A sixth, lighter check sits alongside the battery: the cumulative-landings alternative to the ordinal experience axis. Replacing the sequence index with a cumulative-landings count tests whether the contraction tracks ordinal generation, as the technology hypothesis implies, or cumulative experience in the strict learning-curve sense, in which what matters is the running total of landings rather than which generation a mission belongs to. The two axes are highly correlated in this series but not identical, and the comparison is diagnostic. A contraction that loads on cumulative count rather than on the generation indicators would point toward a genuine production-style learning effect; a contraction that loads on the generation indicators with the cumulative count adding little would point toward the discrete-technology reading the dissertation favors. The experience-curve literature is explicit that the choice of experience axis is consequential and must be disciplined rather than assumed [\[18\]](#ref-18), [\[19\]](#ref-19), which is why the alternative axis is run rather than waved away.

The case that the battery, taken together, defends the contribution against its principal threats can be stated plainly. The four re-fits plus the experience-axis check jointly close the construct-validity, identification-dependence, experience-counting, and functional-form threats from Chapter 5, because each re-fit is matched one-to-one to a named threat and its pre-specified expected behavior under H1 and H0. A threat to validity is neutralized when a pre-registered re-fit shows the conclusion is insensitive to the threat, or, failing that, when the sensitivity is measured and reported, which is the cliometric decomposition ethic of interrogating a constructed series from several measurement angles before any single inference is trusted [\[24\]](#ref-24). The battery reduces but cannot eliminate the small-sample fragility; several re-fits run on the same eleven points are not independent replications, and the plan says so. The battery must also survive the charge that a determined critic could attribute any surviving effect to the shared monotonicity of all the covariates; the unrestricted permutation pass and the program-as-experience-unit re-fit are the specific defenses against that attribution. Confidence that the battery is well matched to the threats is high; confidence that it manufactures robustness the data lack is, correctly, withheld.

## 6.8 The Kuznetsian transient-versus-secular reading

One interpretive commitment deserves its own statement because it is the second anchor's most direct bite on the analysis plan. Kuznets drew a sharp line between a transient movement, a one-time level shift, and a secular trend, a sustained directional change, and warned against extrapolating a secular claim from a short window. The Mars ellipse series is a short window by any standard, and the plan therefore commits in advance to distinguishing, in the executed study, whether the contraction is best read as a secular learning curve or as a sequence of discrete level shifts at the technology-insertion missions.

This is not a merely semantic distinction; it changes what the result licenses. If the augmented model shows a meaningful negative sequence coefficient even after the technology indicators are absorbed, the reading is that there is a genuine secular component, a background improvement that proceeds between technology insertions, consistent with the strong form of a learning curve. If instead the sequence coefficient collapses once the technology indicators enter and the contraction is fully accounted for by discrete drops at MSL and Mars 2020, the reading is that the series is a staircase of level shifts rather than a smooth secular decline, which is the discrete-technology-generation reading the dissertation's mechanism predicts. The plan commits to reporting which of these two patterns the coefficients show and to bounding the extrapolation accordingly. A staircase of level shifts does not license projecting the trend forward to a future mission, because the next drop waits on the next technology insertion rather than on the passage of time, whereas a genuine secular component would license a more continuous projection within the bounds of the knowledge chain remaining intact. This is the Kuznetsian discipline operationalized as an explicit choice between two readings, decided by the data, with the extrapolation each reading permits stated in advance.

## 6.9 Reporting discipline and provenance

The plan fixes the reporting standard so that the executed study cannot present a number without its measurement boundary, which is the Kuznetsian discipline carried through to the output. Three commitments are binding.

First, every ellipse value reported carries its sigma level and its simulation-convention provenance. A three-sigma value and a one-sigma value are not interchangeable, and a design ellipse from one mission's simulation convention is not strictly comparable to another's. The series is only as comparable as the reconstruction work allows, and the plan requires that limit to travel with each number rather than be buried in a methods footnote. Where a value's convention cannot be pinned down, the value appears with that caveat attached, and the comparability gap is reported as a limitation on the inference rather than smoothed over. This is the operational form of the Kuznets requirement that an aggregate is meaningless without a stated boundary of coverage and a valuation convention.

Second, the output is fixed in advance to a single nested-specification table, a single primary figure (log ellipse area against mission generation with the fitted line and the InSight residual marked), the robustness table, and a short robustness appendix. The learning coefficient is reported as a point estimate with a permutation-based interval, and the study states in one sentence whether the sign and ordering predicted by H1 held. Fixing the output set in advance is part of the pre-registration: the executed study reports these objects and does not add a result not specified here without flagging it as a post-hoc addition.

Third, the reporting is symmetric across outcomes. If the data contradict H1, that contradiction is the headline finding and is reported as such, not relegated to a robustness aside. The design is built to be informative whichever way the coefficient falls, and the reporting honors that by giving a null result the same prominence a confirmation would receive. The decision rule of Section 6.3 is applied to the executed coefficients exactly as written, and the resulting verdict, support for H1, failure to reject H0, or the qualified "consistent with but not confirming" case, is stated plainly with its basis in the coefficient pattern.

This reporting discipline is necessary, not merely tidy. Attaching provenance to every value and fixing the output set are what make the small-sample result trustworthy, because a small-sample, multi-convention series is exactly the setting in which an undisciplined report can present a fragile number as a robust one by omitting its boundary and selecting its specification, and stating the boundary and freezing the output removes both omissions. This is the direct application of the cliometric measurement tradition that the dissertation adopts as its second anchor, in which the construction and decomposition of a series precede and discipline any inference drawn from it [\[24\]](#ref-24). Provenance discipline cannot manufacture precision the data lack; it makes the imprecision honest and visible rather than hiding it. The discipline must survive the charge that exhaustive provenance reporting could be dismissed as hedging. The response is that the right standard is calibrated epistemic modality, and reporting a sign and an order of magnitude with their boundaries attached is the correctly calibrated claim for this evidence grade, neither overclaiming a precise rate nor underclaiming a real and large contraction. Confidence in the appropriateness of the discipline is very high, because it is the direct application of the dissertation's own measurement anchor.

## 6.10 How the analysis plan advances the argument

This chapter advances the dissertation's argument to the point where the design is shown to address the causal mechanism and to discriminate the technology account from its rivals on the strength of a frozen, falsifiable protocol. The introductory and literature chapters established that the problem is real and that it is material; this chapter carries the point that the design addresses the causal mechanism, by specifying the nested decomposition that isolates the onboard-guidance levers as discrete covariates on a measurement-disciplined series [\[24\]](#ref-24), [\[5\]](#ref-5), [\[8\]](#ref-8), and the point that the design discriminates the technology account from its rivals, by specifying the InSight counterfactual and the approach-accuracy control as the explicit means of separating the technology hypothesis from the pure-time and injection-accuracy rivals [\[13\]](#ref-13), [\[15\]](#ref-15). That the residual risk is acceptable is carried in the chapter's repeated, calibrated acknowledgment that the small sample and the collinearity bound the precision of any single coefficient, that the procedure returns a sign and an order of magnitude rather than a precise rate, and that the identification's dependence on the InSight counterfactual is measured by the InSight-excluded re-fit rather than assumed away. This chapter contributes the two middle stages of the argument and the design-stage residual-risk language.

Consistent with the dissertation's explicit scope decision, this chapter establishes the causal mechanism and bounds the confidence of the resulting inference; it does not produce a systems or capability architecture. The contribution is a cliometric measurement-and-attribution claim about a constructed performance series, not the design of a real capability or data exchange, so no strategic-objective-to-capability-to-system-function-to-data-exchange traceability is forced onto the argument. The decision relevance, requirements-setting and investment valuation, is carried in plain prose in the introductory and discussion chapters rather than as an architecture table. The analysis plan is, accordingly, an econometric protocol, and its assurance rests on the falsifiability of the decision rule and the provenance of the series, not on any systems-architecture vocabulary.

## 6.11 Summary of the chapter's commitments

The chapter has fixed, in advance of any estimate, the following: a six-step estimation procedure with a frozen order of operations; a binding decision rule that maps coefficient configurations onto verdicts for H1 and H0, with an explicit tie-breaking provision for the most likely small-sample case; expected signs for each technology lever and the approach control, each justified by a full mechanism chain from driver to strategic implication, with confidence calibrated separately for sign and magnitude; an illustrative simulation that displays the table and figure forms with every value labeled illustrative and no cell populated by an estimate; an event-study reading of the InSight residual that states precisely what one counterfactual can and cannot prove and measures the identification's dependence on it; and a reporting discipline that attaches sigma level and simulation-convention provenance to every value, fixes the output set, and treats a null result with the same prominence as a confirmation. Every number in the chapter is expected or illustrative and is labeled as such; the regression has not been executed on the assembled dataset; and the result tables are specified-but-unpopulated by design. What the chapter delivers is not a finding but the frozen protocol and decision rule that make the eventual finding, in either direction, a disciplined and falsifiable one.


# Chapter 7: Discussion

## 7.1 The chapter thesis

This dissertation is designed to be informative whichever way its central coefficient falls. The discussion that follows makes that claim concrete. If the augmented specification returns jointly significant, negative technology coefficients and a small, insignificant approach-accuracy coefficient, then the Mars landing ellipse contracted along a genuine Mokyrian learning curve whose dominant lever is onboard entry-descent-and-landing (EDL) guidance, and the contraction is a quantitative, technology-attributed design variable that NASA and JPL can use to set landing-accuracy requirements and to value guidance investments against approach-navigation investments on a common basis. If instead the technology terms are absorbed by the sequence trend or by the approach-accuracy control, then the contraction is generic maturation or delivery-accuracy improvement, and that finding is itself decision-relevant, because it would redirect a future program's marginal dollar away from onboard sensing and toward interplanetary navigation. The chapter develops both readings symmetrically, traces each back to the two anchor frameworks (Mokyr for mechanism, Kuznets for measurement), states the policy and mission consequences for the human-Mars and sample-return architectures now under study, confronts three rival explanations on their merits, and bounds the external-validity claim explicitly to U.S.-led Mars landings, with Tianwen-1 and lunar precision-landing held out as reference points rather than as in-sample evidence. The single most important interpretive commitment is the one stated in the analysis plan and honored here: the design produces a decision either way, and the discussion refuses to treat H0 as a failure.

A note on the epistemic status of everything in this chapter is required at the outset, because the chapter interprets outcomes that have not yet been observed. This is a design-stage dissertation. The regression has not been executed on the assembled dataset, and no coefficient, interval, or fit statistic reported anywhere in the work is an empirical estimate. The discussion therefore proceeds conditionally throughout: it specifies what each outcome would mean if it were observed, what confidence that interpretation would carry given the design's known limits, and what additional evidence would raise or lower that confidence. Where the chapter uses the indicative mood for compactness, the conditional is always intended. The interpretive apparatus is real and complete; the numbers it would operate on are not yet in hand.

## 7.2 The problem this chapter addresses

The current state of the field is a documented but uninterpreted contraction. The EDL engineering literature establishes beyond dispute that the Mars landing ellipse has shrunk by more than two orders of magnitude from Viking to Mars 2020, and it documents each of the candidate causal technologies (guided lifting entry, the range-to-go parachute trigger, terrain-relative navigation) in mission-specific depth [\[16\]](#ref-16), [\[121\]](#ref-121). What the literature does not do is tell a project manager what the contraction means for a decision. It does not say whether buying terrain-relative navigation for a future lander will shrink that lander's ellipse, or whether the same budget spent on approach navigation would do as well, because no study arbitrates the rival causes against one another.

The desired state is an interpretation that is decision-ready under either outcome. A program office that must commit to a landing-accuracy requirement years before launch, or that must rank candidate guidance investments, needs to know not only that ellipses have shrunk but why, and needs that answer in a form that survives the small-sample and collinearity problems the design confronts. The gap between the current and desired states is interpretive, not merely empirical: even once the regression is run, its coefficients do not speak for themselves. They must be read against the two anchor frameworks, against the rival explanations, and against the boundary of the population on which they were estimated. Supplying that reading, for both the H1 and the H0 outcome, is the work of this chapter. The consequence of leaving the gap unfilled is that the human-Mars and Mars Sample Return architectures continue to negotiate landing precision qualitatively, valuing guidance technologies by engineering judgment and program heritage rather than against a measured, attributed rate [\[112\]](#ref-112), [\[113\]](#ref-113), [\[122\]](#ref-122), [\[127\]](#ref-127). The discussion exists to convert whichever coefficient the data return into an interpreted, framework-anchored, decision-usable claim.
## 7.3 Implications if H1 holds

### 7.3.1 The primary claim and the argument that supports it

If the augmented specification returns jointly significant and negative technology coefficients with a small, insignificant approach-accuracy coefficient, then onboard EDL guidance, not launch or approach-navigation accuracy, is the dominant lever on Mars landing precision, and the era's improvement is a learning curve in Mokyr's sense rather than a generic time trend.

Three pieces of design-stage evidence would jointly underwrite that conclusion. The first is the discrete-step structure: guided entry, the range trigger, and TRN entered the flight record on identifiable missions, so under H1 the ellipse series shows matching discrete drops at those missions and a comparatively flat profile elsewhere. The second is the InSight counterfactual: a late mission carrying the calendar of MSL and Mars 2020 but the unguided, ballistic, Phoenix-heritage entry of an earlier era sits visibly above the trend, which time alone cannot explain. The third is the approach-accuracy control: with delivered entry-state dispersion held constant, the residual contraction loads on the technology indicators rather than on the proxy for reaching the atmosphere more accurately.

Each technology lever attacks a physically distinct component of the footprint, so an indicator that turns on with that lever and carries a negative, significant coefficient is a credible signature of that lever's effect rather than of a coincident trend. Guided lifting entry nulls the hypersonic downrange and crossrange dispersion that a ballistic capsule cannot correct; the range trigger removes the parachute-deploy dispersion by commanding deploy on navigated range rather than velocity; TRN collapses the position-knowledge error and enables an autonomous divert that has no predecessor capability [\[121\]](#ref-121). Because the error sources are separable in physics, their statistical separation, to the degree the small sample permits, is interpretable as causal attribution rather than as curve description. This reading rests on the Mokyrian account of why a technique with a deep propositional base is extensible and self-correcting. Guided entry rested on maturing entry aerodynamics and atmosphere reconstruction; the range trigger on navigation; TRN on computer vision and onboard mapping. Each lever is a discrete addition to the prescriptive landing-technique base, supported by its own propositional foundation, which is why each can produce an identifiable, attributable step rather than a featureless decline. The hazard-detection-and-avoidance and precision-landing development lineage that preceded and accompanied these flight insertions corroborates that the levers were the product of a sustained, propositionally grounded technology program rather than of incidental improvement [\[114\]](#ref-114), [\[116\]](#ref-116), [\[120\]](#ref-120), [\[123\]](#ref-123).

The strongest defensible reading under H1 is a signed, order-of-magnitude, counterfactual-surviving attribution, not a precise point estimate of any single lever's learning rate. The confidence band is moderate at best, set by the irreducible small-n (nine to eleven events) and the near-collinearity of the technology indicators with the sequence index. It is honest about the fact that the InSight case discriminates the technology hypothesis from the pure-time hypothesis but cannot by itself apportion credit among the three levers. The conclusion would be overturned if a residual monotone confound (modeling fidelity, atmospheric knowledge, onboard computing) drove the contraction and merely happened to track the technology indicators, or if the achieved-miss-distance dependent variable failed to reproduce the contraction the design ellipse shows. That objection is bounded, not eliminated, by the InSight control and the dual dependent variable; it is not bounded away entirely, and the discussion does not pretend otherwise.

### 7.3.2 Requirements-setting under H1

If H1 holds, the most immediate use is in setting landing-accuracy requirements. The mechanism that makes this use legitimate is specific and worth naming. Driver: a program commits to a guidance suite. Mechanism: each technology in that suite removes a quantified fraction of a distinct error source, and the attributed coefficients describe how much. Observable effect: an expected ellipse size conditional on the suite. Operational consequence: a project can specify a defensible accuracy requirement rather than negotiating one qualitatively. Strategic implication: landing precision becomes a design variable with a measured, attributed value rather than an engineering-judgment assertion. This chain converts the learning curve from a retrospective description into a forward design input. The human-Mars EDL architecture studies, which must specify accuracy for vehicles far heavier than any flown, are the natural consumer of such a requirement basis, because they confront precisely the problem of valuing a guidance capability that has not yet flown at scale [\[112\]](#ref-112), [\[113\]](#ref-113), [\[122\]](#ref-122).

The qualifier on this use is the one Kuznets would insist on and that Section 7.6 develops: a requirement read off the fitted curve is valid only within the measurement boundary on which the curve was built, which is sub-twenty-tonne, robotic, U.S.-led Mars entry. Extrapolating the rate to a forty-tonne human-class vehicle entering with supersonic retropropulsion is an out-of-boundary projection, not a within-sample inference, and the supersonic-retropropulsion development literature shows that the high-mass regime introduces error sources and control-authority questions the robotic series never exercised [\[111\]](#ref-111), [\[126\]](#ref-126). The requirement-setting use is therefore strong for sample-return-class robotic landers and weak, explicitly flagged as extrapolation, for human-class vehicles.

### 7.3.3 Investment valuation under H1

The second use under H1 is investment valuation, and it carries the sharpest decision content. If the TRN coefficient is the largest of the three and the approach-accuracy coefficient is insignificant, then the marginal return on landing precision from continued investment in onboard sensing and divert capability exceeds the marginal return from investment in approach navigation. This is a comparative claim the existing literature cannot make, because no prior study placed onboard guidance and approach accuracy in the same model. The mechanism is again nameable: investing in TRN matures the propositional base (descent imaging, onboard map construction, real-time hazard detection) that the divert capability rests on; the matured base produces a tighter effective targeting region; the tighter region is worth a quantified amount in expanded site access. The ongoing development of higher-resolution terrain-sensing lidar and multi-functional altimetry and TRN payloads is exactly the kind of investment whose precision return this model would value [\[128\]](#ref-128), [\[129\]](#ref-129), and the lunar TRN lineage shows that the underlying sensing-and-matching capability is a transferable, cumulatively improving technology rather than a one-mission artifact [\[117\]](#ref-117), [\[118\]](#ref-118), [\[119\]](#ref-119).

The confidence in this valuation use is moderate and asymmetric. The sign and ordering of the technology coefficients, if H1 holds, would be robust to the small sample in the sense that the InSight counterfactual and the nested-specification reporting make a wrongly signed or absorbed coefficient visible. The magnitude, the actual exchange rate between a guidance dollar and an approach-navigation dollar, is far less certain, because magnitude is exactly what the collinear, small-n design estimates worst. The honest statement to a program office is therefore directional: under H1, prefer onboard guidance at the margin, but do not treat the fitted magnitude as a precise price.

### 7.3.4 The Mokyrian extensibility reading under H1

If H1 holds, the deepest interpretive payoff is that the contraction is evidence of extensibility in Mokyr's exact sense. A technique resting on a wide propositional base can be diagnosed, improved, and built upon across generations; a technique found by trial without underlying theory stagnates. The Mars landing record under H1 is the former: each lever was understood well enough to be improved deliberately, and each rested on its own theoretical foundation, which is why the improvements compounded into a roughly log-linear curve rather than petering out. The macro-versus-incremental texture supports this reading. If the largest fit increments come at the guided-entry transition (a macro-invention that changed entry from ballistic to lifting) and the TRN transition (a macro-invention that introduced an entirely new error source to attack and a divert capability with no predecessor), with a smaller increment at the range trigger (an incremental refinement of existing deploy logic), then the pattern of increments is itself a second, independent signature of the Mokyrian distinction, observed in the fit rather than asserted from the history. That convergence, fit increments matching the macro-or-incremental character of each lever, would raise confidence in the causal reading above what the bare coefficients alone could support, because it is a prediction the framework makes that the data could have refused.

## 7.4 Implications if H0 holds

### 7.4.1 The H0 claim, argued symmetrically

Intellectual honesty forbids treating H0 as a failed H1. The discussion therefore gives the null its own full argument rather than a dismissive paragraph.

If the technology coefficients are jointly insignificant once the sequence trend and the approach-accuracy control are present, or if the approach-accuracy coefficient absorbs the technology terms, then the ellipse contraction is not attributable to onboard EDL-guidance generation and is instead generic maturation or delivery-accuracy improvement.

Two design-stage patterns would underwrite H0. The first: the InSight observation sits on the trend rather than above it, meaning a late mission inherited a small ellipse without the guidance suite, which is what a pure-time or generic-maturation story predicts. The second: the approach-accuracy control carries a large, significant coefficient and drives the technology indicators toward zero, meaning the contraction tracks how accurately vehicles were delivered to the atmospheric interface rather than what they did once there.

Approach-navigation accuracy is a physically sufficient channel for ellipse contraction on its own: a vehicle delivered to a tighter entry-flight-path-angle and entry-point dispersion begins its descent with less error to correct, so a monotone improvement in delivery accuracy could in principle produce the entire observed contraction without any onboard-guidance contribution. The Tianwen-1 record, in which a capable powered-descent GNC system was paired with its own delivery and guidance chain, is a reminder that the delivery-and-descent system is an integrated whole whose pieces are not trivially separable [\[16\]](#ref-16), [\[124\]](#ref-124). The discipline here is again Kuznets's, in the form of the warning against attributing a secular movement to one component before the components have been decomposed and controlled. If the approach-accuracy control absorbs the technology terms, the disciplined conclusion is that the aggregate change was misattributed by the qualitative literature, not that the technologies did nothing, but that the data cannot credit them once the delivery channel is held constant.

An H0 result under this design is a statement about attributability given the available controls and the small sample, not a proof that onboard guidance is causally inert. The most that H0 supports is that, within the measured series and with the approach-accuracy proxy in the model, the technology indicators do not carry independent explanatory power. The confidence is moderate and is limited in the same direction as under H1: the collinearity that makes it hard to credit the technologies cleanly also makes it hard to exonerate them cleanly. H0 would itself be undercut if the approach-accuracy proxy were a poor measure of true delivery accuracy, so that it absorbed technology variance only because it was correlated with the era rather than because it captured the delivery channel. The construct-validity threat runs in both directions, and the discussion flags that an H0 driven by a weak proxy is weaker evidence than an H0 driven by a strong one.

### 7.4.2 Why an H0 result is decision-relevant, not a null finding

The practical content of H0 is the inverse of H1 and is no less useful. If approach accuracy, not onboard guidance, is the binding constraint, then a future program's marginal precision dollar is better spent on interplanetary and approach navigation than on onboard sensing and divert. The mechanism is symmetric to the H1 case: investing in approach navigation tightens the delivered entry state; the tighter entry state leaves less dispersion for any descent system to correct; the smaller residual dispersion is the source of the contraction. For an architecture office choosing between funding a next-generation TRN sensor and funding a tighter approach-navigation campaign, an H0 result is a direct instruction to prefer the latter at the margin. This is precisely why the design commits in advance to reporting whichever pattern the data show: the null is a finding with a clear investment implication, and burying it would forfeit half the value of the study.

There is a subtler H0 implication for forecasting culture. If the contraction is generic maturation rather than a sequence of attributable technology steps, then the future trajectory of landing precision is governed by whatever drives that maturation (cumulative institutional experience, computing, atmospheric models) rather than by a roadmap of nameable guidance insertions. A program that assumed it could buy a known ellipse reduction by buying a known technology would, under H0, be mistaken, and would need to forecast precision from the slower, more diffuse maturation trend instead. That correction is worth having before a multibillion-dollar architecture commits to a precision assumption.

## 7.5 The theoretical contribution back to each anchor framework
### 7.5.1 Back to Mokyr: macro versus incremental, extensibility, and reversibility

The dissertation borrows from Mokyr and also returns something to the framework. The return is worth stating, since a candidate's contribution is judged partly on what it gives back to its anchors. Mokyr establishes his macro-versus-incremental distinction and his extensibility thesis largely on terrestrial industrial cases. The Mars landing series, under either outcome, is an unusually clean test bed for both, because it is a small, fully enumerated population of discrete, expensive, technology-differentiated events whose propositional and prescriptive foundations are individually documented. If H1 holds and the fit increments align with the macro-or-incremental character of each lever, the dissertation supplies a quantitative instance of the distinction operating in a setting Mokyr never examined, which strengthens its claim to generality. If H0 holds, the contribution is different but real: a cautionary case in which an apparent learning curve dissolves once the supporting delivery channel is controlled. That case illustrates Mokyr's own insistence that a smooth improvement curve is not self-explanatory and may not reflect the extensibility of the named technique at all.

The reversibility and lock-in reading is the strongest theoretical return, and it is outcome-robust. Mokyr's warning that technological progress is reversible when the supporting knowledge base or institutional channel atrophies applies to Mars landing precision with unusual directness. The capability is not a permanent possession of the agency. It is embodied in a chain of propositional knowledge, the tacit expertise of the engineering teams, the flight-software heritage, and the test infrastructure that validated TRN before flight. The free-flight terrestrial rocket-lander demonstrations and the hazard-detection field campaigns that matured the precision-landing base are the kind of capability-sustaining infrastructure whose loss would break the chain [\[116\]](#ref-116), [\[123\]](#ref-123). The theoretical point this dissertation contributes is that a learning curve fitted to past missions describes capability that was demonstrated, not capability that will be available, and that any forward use of the curve assumes the knowledge chain is intact. This converts Mokyr's reversibility thesis from a historical observation into an operative boundary condition on a forecasting tool, a sharper use than the framework usually receives.

### 7.5.2 Back to Kuznets: level-shift versus secular trend, and the discipline of the constructed series

The contribution back to Kuznets is methodological, and it is independent of which hypothesis the data support. Kuznets's central distinction, between a transient level-shift movement and a genuine secular trend, is the interpretive hinge on which this dissertation's forecasting value turns. NASA forecasting culture is prone to extrapolating a single dramatic step, the Mars 2020 TRN demonstration, as if it were a trend, and to assuming the next mission inherits the last mission's ellipse for free. The design tests which the contraction is, and the InSight control is the instrument that does it. A late mission with a large ellipse is direct evidence that the late-era small ellipses are technology-conditional level shifts, not a free secular trend that any modern mission inherits. The contribution back to Kuznets is a worked example of his level-shift-versus-trend discipline applied to an engineering performance series rather than to a national-income aggregate, showing that the cliometric measurement ethic transfers to domains far from its origin.

The second Kuznetsian return concerns the constructed-series discipline itself. Kuznets insisted that an aggregate is meaningless without a stated boundary of coverage, a valuation convention, and a netting rule, and that one must build a long, comparable, decomposed series before theorizing. The landing-ellipse series is exactly such a constructed measurement. Its reported value depends on the sigma level, on whether it describes targeting capability or achieved miss distance, and on simulation conventions that drifted across missions. By stating those conventions explicitly, by carrying the sigma level and simulation provenance with every value, and by validating the design ellipse against the achieved miss distance, the dissertation contributes a template for applying Kuznetsian construction discipline to aerospace performance metrics, which are routinely compared across programs without the boundary statements Kuznets would demand. That template stands whether the regression supports H1 or H0, because it concerns how the series is built, not what slope it yields.

## 7.6 Policy and mission implications for NASA, JPL, and stakeholders

### 7.6.1 Mars Sample Return retrieval

The Mars Sample Return campaign is the nearest-term consumer of a technology-attributed landing-precision rate. The implication is conditional on the outcome but useful either way. The campaign architecture requires retrieving cached samples and delivering hardware to specified locations with accuracy that governs which retrieval concepts are feasible [\[127\]](#ref-127). Under H1, the model supplies a defensible basis for specifying the retrieval lander's accuracy requirement as a function of its guidance suite, and for valuing the inclusion of TRN-class capability against its cost and mass. Under H0, the model warns the campaign that precision is bought through delivery accuracy rather than onboard guidance, which would reweight the retrieval-lander trade toward approach-navigation investment. The mechanism by which the result matters is the same in both cases. The retrieval concept's feasibility depends on a landing-accuracy number, and the dissertation supplies the first attributed basis for that number rather than an engineering-judgment assertion. The confidence is moderate and is bounded to robotic, sub-twenty-tonne Mars entry, the regime MSR retrieval occupies, so the within-boundary use is among the strongest the model supports.

### 7.6.2 Human Mars entry, descent, and landing

The human-Mars implication is real, but it must be stated with its boundary attached, because human-class EDL sits outside the series on which any rate would be estimated. The human-Mars architecture studies confront a vehicle an order of magnitude heavier than any robotic lander, entering with rigid or deployable decelerators and, in many concepts, supersonic retropropulsion, which introduces control-authority and error-source questions the robotic record never exercised [\[112\]](#ref-112), [\[113\]](#ref-113), [\[122\]](#ref-122), [\[111\]](#ref-111). The honest implication is twofold. First, the attributed learning rate is an input to human-Mars precision planning only as an out-of-boundary reference, not as a within-sample prediction, and the dissertation flags this rather than letting a program read a human-class ellipse off a robotic curve. Second, the framework, the practice of attributing precision to specific onboard-guidance technologies on a measurement-disciplined series, transfers even where the estimated rate does not, so a future human-class series could be constructed and fitted by the same method once enough human-class or high-mass entries exist. The mechanism distinction matters here. Supersonic retropropulsion and pinpoint-landing guidance for high-mass vehicles attack the footprint through partly different physics than the robotic levers, so the human-class rate is an empirical question the robotic curve cannot answer, only frame [\[126\]](#ref-126).

### 7.6.3 Requirements culture and investment governance

The broadest policy implication is about how NASA and JPL set and govern precision requirements. The current practice negotiates landing accuracy qualitatively and values guidance investments by heritage and engineering judgment. Whichever hypothesis the data support, the dissertation contributes a method for putting that negotiation on a quantitative, attributed footing: a model that says which lever moves the ellipse, by roughly how much, and within what boundary. The strategic implication is that precision becomes a governable design variable with a measured rate and an explicit uncertainty rather than a contested qualitative claim. This is the plain-prose decision relevance the design carries deliberately outside any architecture-traceability formalism. The contribution is an econometric measurement-and-attribution claim about a constructed series, and its decision value is carried in the requirements-and-investment language of this section, not forced into a capability-architecture vocabulary that the contribution does not concern.

## 7.7 Full engagement with rival explanations

The credibility of the attribution rests on confronting the rivals rather than dismissing them. Three are material, and each is treated as a hypothesis the data could vindicate.

### 7.7.1 The pure-time-trend rival

The first rival holds that the contraction reflects generic maturation, better modeling, more atmospheric knowledge, faster onboard computing, rather than the specific named technologies, and that the technology indicators are merely proxies for the calendar. This rival is serious precisely because everything in the era improved monotonically, so a time trend and the technology indicators are nearly collinear. The design's answer is the InSight counterfactual, and the strength of that answer must be stated honestly. InSight flew late, with the calendar of the precision era, but with a deliberately unguided, ballistic, Phoenix-heritage entry into a large, flat ellipse, because its science did not require precision and its budget did not justify the suite. If the InSight ellipse is large, the pure-time-trend rival is falsified, because a late mission did not inherit a small ellipse for free; if the InSight ellipse is small, the rival is supported. This discrimination rests on a single observation, so it separates the technology hypothesis from the pure-time hypothesis but cannot apportion credit among the levers, and the confidence it confers is moderate rather than high. The objection that would weaken the InSight defense is the possibility that InSight's large ellipse reflects a deliberate site choice rather than a capability limit, which is the third rival, addressed below.

### 7.7.2 The approach-accuracy rival, which is H0 itself

The second rival is not external to the hypothesis structure; it is the null. It holds that better interplanetary and approach navigation, not onboard EDL guidance, shrank the ellipse by delivering vehicles to a tighter entry state. The design confronts this rival directly by including the delivered entry-state dispersion as a control, so that any contraction loading on the technology indicators is net of the delivery channel. The honest treatment, carried forward from the H0 discussion, is that this rival may win, and that its winning would be a finding rather than a defeat. The mechanism by which approach accuracy could be the true cause is physically sufficient: a tighter delivered entry state leaves less dispersion for any descent system to correct. The discrimination between this rival and H1 is only as good as the approach-accuracy proxy, which is the construct-validity limit on the whole exercise. The Tianwen-1 integrated GNC and powered-descent record is a useful external reminder that delivery and descent are coupled, and that crediting one channel over the other requires a control good enough to separate them [\[16\]](#ref-16), [\[124\]](#ref-124). Confidence in the discrimination is therefore conditional on the quality of the delivered-entry-state series assembled at execution, which is flagged as a coverage risk rather than assumed away.

### 7.7.3 The site-selection-endogeneity rival

The third rival holds that ellipses shrank because projects chose more forgiving targets over time, so the apparent capability gain is a selection artifact. The first response is that this rival is partly the reverse of the truth, because the later, precision-capable missions chose harder, more hazardous, scientifically richer sites that earlier systems could not have attempted at all, so site difficulty rose with capability rather than falling. The second and decisive response is the achieved-miss-distance dependent variable from PDS localization, which measures how close the vehicle landed to its aim point and is therefore a capability measure independent of how forgiving the chosen site was. If the contraction appears in the achieved-miss-distance series as well as in the design-ellipse series, the site-selection rival cannot explain it, because miss distance does not depend on site forgiveness. Achieved miss distances are precise but few, since only successful landings produce them, so the miss-distance series is a sparser instrument than the design-ellipse series and the rebuttal it provides to the site-selection rival is correspondingly bounded. The hazard-detection-and-avoidance lineage is relevant context here, because the capability to land safely in hazardous terrain is exactly what decoupled site difficulty from ellipse size, and that decoupling is what makes the site-selection rival largely a description of capability rather than a confound to it [\[114\]](#ref-114), [\[115\]](#ref-115), [\[116\]](#ref-116).

### 7.7.4 What confronting the rivals establishes

Taken together, the three rivals are not eliminated; they are bounded. The InSight control bounds the pure-time rival, the approach-accuracy control bounds the H0 delivery rival, and the achieved-miss-distance dependent variable bounds the site-selection rival. None of the three defenses is individually decisive given the small sample, and the discussion states that plainly. Their combined effect is to make the attribution, if it survives, more credible than any single-mission engineering study could make it, and to make a null result, if that is what emerges, an interpretable statement about attributability rather than an artifact of an unconfronted rival. This is how the design discriminates the technology account from its rivals, stated at the level of honesty the work demands: well enough to be informative, not well enough to be certain.

## 7.8 External-validity statement

### 7.8.1 The boundary, stated plainly
The model would be fit on U.S.-led successful Mars surface landings from Viking 1 through Mars 2020, with Tianwen-1 held out. That population is the entire boundary of any in-sample claim. Generalization beyond it, to other bodies, other agencies, other entry regimes, or human-scale vehicles, is not automatic and is not asserted. This is the Kuznetsian boundary statement applied to external validity: the rate, if estimated, describes the contraction within a stated coverage, and a reader who carries it outside that coverage is making a projection the data do not support. The discussion treats the out-of-boundary cases below as reference points, never as evidence that the in-sample rate generalizes.

### 7.8.2 Tianwen-1 as an out-of-sample reference

Tianwen-1 is the most directly comparable out-of-sample case: a non-U.S. Mars lander that flew a capable, partly autonomous GNC and powered-descent system into its own targeted region [\[16\]](#ref-16), [\[124\]](#ref-124). It is held out because including it would conflate the U.S. technology-generation sequence with a different agency's integrated development path, a separation the design is not built to make. As a reference point, Tianwen-1 is informative in one direction. It demonstrates that the precision-landing capability is not a U.S.-specific artifact and that an independent program reached a comparable capability through its own propositional and prescriptive base, consistent with the Mokyrian reading that the capability rests on transferable, theory-grounded foundations. What Tianwen-1 cannot do is serve as in-sample confirmation of the U.S. learning rate, because its missions are not coded into the U.S. technology sequence and its delivery-and-descent system is a different integrated whole. The discussion uses it to illustrate the generality of the capability, not to extend the estimated curve.

### 7.8.3 Lunar precision landing as an out-of-sample reference

The lunar precision-landing literature is the second reference frontier, and it is informative for a different reason. Terrain-relative navigation, hazard detection, and crater-based navigation have a substantial and cumulatively improving lunar lineage, from early TRN overviews and lidar-based approaches to recent integrated optical TRN and SURF-based lunar algorithms [\[17\]](#ref-17), [\[117\]](#ref-117), [\[118\]](#ref-118), [\[119\]](#ref-119), [\[125\]](#ref-125), [\[130\]](#ref-130). This lineage establishes that the onboard-sensing-and-matching capability at the heart of the TRN lever is a general planetary-landing technology rather than a Mars-specific one, which supports the Mokyrian claim that the lever rests on a broad propositional base extensible across bodies. The honest limit is that the lunar regime differs from Mars in the variables that matter most to the footprint: there is no atmosphere to fly a guided lifting entry through, no parachute to trigger, and a different gravity and descent profile, so the lunar cases exercise the TRN lever in isolation from the guided-entry and range-trigger levers. They are therefore evidence that one of the three levers generalizes as a capability, not evidence that the Mars learning rate generalizes as a number. The discussion draws exactly that bounded inference and no more: the technology is general; the estimated rate is not transferred.

### 7.8.4 Human-class and high-mass entry as a boundary, not a reference

The final external-validity statement is that human-class and high-mass Mars entry lies outside the boundary in a way that even the reference cases do not bridge. Supersonic retropropulsion, rigid and deployable decelerators, and pinpoint guidance for vehicles an order of magnitude heavier than the robotic series attack the footprint through partly different physics and introduce error sources the robotic record never exercised [\[111\]](#ref-111), [\[112\]](#ref-112), [\[113\]](#ref-113), [\[122\]](#ref-122), [\[126\]](#ref-126). The robotic learning curve frames the human-class question, by supplying the method and the variables, but cannot answer it, because the human-class regime is an out-of-boundary extrapolation rather than a reference case with a shared capability. The disciplined conclusion is that the dissertation's external validity ends at robotic, sub-twenty-tonne, U.S.-led Mars entry, that the method transfers to human-class and lunar series once those series exist and can be constructed with the same Kuznetsian discipline, and that any near-term human-Mars precision planning that leans on the robotic rate is using it as an explicitly flagged out-of-boundary reference and not as a within-sample prediction.

## 7.9 Confidence summary and the evidence that would move it

The discussion closes by stating, for the chapter's principal interpretive claims, the confidence each would carry and what would raise or lower it, calibrated to the design-stage evidence grade.

The claim that the contraction is real and large is very high confidence, resting on a uniform and uncontested documentary record across the EDL literature [\[16\]](#ref-16), [\[121\]](#ref-121); nothing in the analysis could lower it, because it is the premise, not the inference.

The claim that, under H1, onboard guidance is the dominant lever is moderate confidence by design, set by small-n and collinearity. It would rise toward high if the achieved-miss-distance series reproduced the design-ellipse contraction, if the InSight residual sat clearly above the trend, and if the fit increments matched the macro-or-incremental character of each lever, and it would fall toward low if the approach-accuracy control absorbed the technology terms or if a residual monotone confound tracked the indicators.

The claim that, under H0, the contraction is delivery-driven or generic is symmetric in confidence and in the evidence that would move it. It would rise with a strong, well-measured approach-accuracy proxy that retained significance across specifications and fall if that proxy proved to be a mere era correlate.

The theoretical contributions back to Mokyr (reversibility as a boundary condition) and Kuznets (level-shift versus trend, and constructed-series discipline) are high confidence and outcome-robust, because they concern how the series is built and interpreted rather than which slope it yields. The external-validity claims are stated as bounds rather than estimates, and their confidence is the confidence that the boundary is correctly drawn, which is high, rather than the confidence that any rate generalizes, which the discussion declines to assert at all. This calibrated, two-sided, framework-anchored interpretation is the chapter's contribution: it makes the design decision-usable before a single coefficient is estimated, and it makes the eventual estimation interpretable whichever way it falls.


# Chapter 8: Conclusion

## 8.1 The chapter thesis

The lasting contribution of this dissertation is not a learning rate but an instrument for measuring one: a defined dependent variable with a stated boundary, a falsifiable hypothesis with a fixed decision rule, and an identification strategy that survives the small sample it must work in, assembled where the prior literature offered only a qualitative narrative of shrinking Mars landing ellipses. That instrument stands whether or not the hypothesis it is built to test is ultimately confirmed. This is the answer this final chapter delivers, and every section develops it. The dissertation proposed one testable proposition, that the three-sigma Mars landing-ellipse area contracts along an exponential learning curve driven primarily by onboard entry-descent-and-landing guidance technology, guided lifting entry, the range-to-go parachute trigger, and terrain-relative navigation, rather than by launch-vehicle injection accuracy, and it laid out a complete, executable, and pre-registered plan to decide that proposition against the historical record. The chapter restates that contribution precisely, separates the part of it that depends on confirmation from the part that does not, states the limitations without softening them, and specifies a concrete program of future work whose first and most important step is to execute the frozen design on the assembled data.

A conclusion to a design-stage dissertation has a particular obligation that a conclusion to an executed study does not. It cannot report a finding, because the regression has not been run on the assembled dataset, and it must resist the temptation to write as if it had. What it can do, and what this chapter does, is account honestly for what has been built, what that construction is worth independent of the eventual estimate, and what remains to be done to convert the design into a result. The chapter introduces no new evidence and fits no new model. It synthesizes the argument the preceding seven chapters developed and hands forward a defined, bounded, and falsifiable measurement problem in a state where a successor analyst can pick it up and finish it.

### 8.1.1 Problem frame for this chapter

The current state, entering this chapter, is a fully specified design: a constructed ellipse series with its provenance discipline, a baseline and an augmented specification written out exactly, an identification strategy resting on the InSight counterfactual and the approach-accuracy control, a robustness battery, and a pre-registered decision rule, none of which has yet been executed against the numbers. The desired state is a clear statement of what this body of work contributes, what survives if the central hypothesis fails, where the design is genuinely fragile, and what the path to a completed empirical result looks like. The gap is that a design, however complete, is easily mistaken either for a finished study, which overclaims, or for a mere proposal, which underclaims; both readings miss what a frozen, falsifiable, measurement-disciplined design actually delivers. The consequence of leaving that gap unclosed is that the dissertation's real contribution, the reframing of a qualitative narrative as a falsifiable measurement problem, would be lost between the result it does not yet have and the proposal it is no longer. This chapter closes the gap by stating the contribution at the correct altitude and bounding it honestly.

## 8.2 Restatement of the contribution

The contribution is a single falsifiable proposition and the apparatus required to test it, and each component bears restating at the level of precision the program demands. The proposition is H1: the three-sigma landing-ellipse semi-major axis, and equivalently the ellipse area, for U.S.-led Mars surface missions declines along an exponential learning curve in which the dominant explanatory variables are the onboard EDL-guidance technology generations, and in which launch-vehicle and interplanetary injection accuracy, once the standard approach-navigation corrections are accounted for, is not the binding constraint on ellipse size. The null, H0, is that ellipse contraction is unrelated to onboard EDL-guidance technology generation, so that the technology covariates carry no explanatory power once mission sequence or a time trend is included, and any apparent learning curve is either an artifact of a generic time trend or is driven by approach-navigation improvements rather than onboard guidance.

The novelty is precise and bears restating, because it is easy to mistake the contribution for the well-documented observation that Mars landings have grown more precise. That observation is not in dispute and is not the contribution. The contribution is the joining of two literatures that had not been joined: the EDL engineering literature, which documents each mission and each guidance technology in depth but never as one series and never with formal attribution against the injection-accuracy rival [\[1\]](#ref-1), [\[5\]](#ref-5), [\[8\]](#ref-8), and the technology-economics learning-curve apparatus, which has a mature method for describing how a performance metric improves as a function of successive technology generations but had never been applied to planetary landing precision [\[18\]](#ref-18), [\[19\]](#ref-19). The dissertation treats the cross-mission ellipse sequence as one constructed measurement, fits a learning-rate model to it with the cliometric discipline Kuznets demanded, and attributes the contraction to identifiable technology insertions with the causal logic Mokyr supplies.

The two anchors are not ornamental, and restating their role clarifies what was actually built. From Mokyr came the decision to model discrete technology generations as covariates rather than to fit a single smooth trend, because if the contraction is genuinely driven by additions to the prescriptive knowledge base, those additions should appear as identifiable steps tied to the missions on which the technologies first flew, not as a featureless decline [\[23\]](#ref-23). From Kuznets came the refusal to report a single learning rate without first stating the measurement boundary, the sigma level, the design-versus-achieved distinction, the simulation-convention provenance, and decomposing the change before fitting any slope [\[24\]](#ref-24). These two decisions, taken together, distinguish the design from a naive curve fit for a clear reason: a naive analyst would regress log ellipse on year and announce a slope, which hides the mechanism (Mokyr's objection) and hides the boundary (Kuznets's objection), whereas the design recovers the mechanism through the covariate structure and protects the boundary through the provenance discipline. This rests on the cliometric tradition's repeated demonstration that an aggregate change is decided by a disciplined decomposition into attributable components, not by a single regression [\[24\]](#ref-24), [\[20\]](#ref-20), [\[21\]](#ref-21). The design delivers this discipline; it does not, at the design stage, deliver the estimate that discipline would produce. Confidence that the apparatus is the right apparatus for the question is very high, because it is the direct application of two methodological traditions to a problem they were built to handle; confidence in any particular numerical result remains, by construction, withheld until execution.

The covariate structure also encodes something physical that bears restating, because the separability of the three indicators is the engineering fact that licenses treating them as distinct causes rather than as one undifferentiated index of modernity. Each lever attacks a physically distinct component of the footprint, and that is why the design models them separately rather than collapsing them. Guided lifting entry attacks the downrange and crossrange dispersion accumulated during the long hypersonic phase, where a ballistic capsule flies wherever its delivered entry state and the atmosphere carry it, by modulating bank angle in closed loop to null that dispersion in real time [\[5\]](#ref-5). The range-to-go parachute trigger attacks a different and smaller component, the dispersion introduced at parachute deploy, by replacing a velocity-based deploy command with a range command that deploys earlier or later to compensate for how far downrange the vehicle has actually flown [\[5\]](#ref-5). Terrain-relative navigation attacks a third component the first two cannot touch, the position-knowledge error itself, the gap between where the vehicle believes it is and where it actually is, by matching descent imagery to an onboard reference map and then enabling an autonomous divert away from hazards [\[8\]](#ref-8). Because each lever addresses a physically distinct error source, their effects are in principle separable, and the regression is designed to recover that separation to the extent the small sample allows. This physical separability is the engineering counterpart of Mokyr's claim that each technology rests on its own propositional base, entry aerodynamics for guided entry, deploy dynamics for the range trigger, computer vision and onboard mapping for TRN, and it is the reason the contribution is a three-covariate attribution rather than a single-trend description.

## 8.3 What holds even if the hypothesis is not confirmed
The strongest argument in a design-stage conclusion is the one that establishes value independent of the result, and the dissertation was built so that this argument is genuinely available rather than rhetorical. Three things hold whether the executed regression confirms H1, fails to reject H0, or returns the qualified middle case.

The first is a defined measurement. Before this work, the landing ellipse was discussed mission by mission as a number whose meaning was assumed rather than stated. The dissertation defines it as the natural logarithm of the three-sigma ellipse area, \(\ln(\pi \cdot a \cdot b)\), sourced from the mission EDL performance study, with a secondary dependent variable, the log achieved miss distance from PDS localization, defined to guard the construct against the risk that design and achieved accuracy diverge. That pair of definitions, with the rule that every value carries its sigma level and simulation-convention provenance, is a measurement contribution that survives any regression outcome, because it makes the series comparable, or makes its incomparabilities explicit, in a way the prior literature did not. That this is a real contribution and not bookkeeping follows from the Kuznetsian principle that an aggregate means nothing without a stated boundary [\[24\]](#ref-24); a series that a second analyst can reconstruct from the same records to the same values is a measurement instrument, and that instrument exists now whether or not the slope through it confirms the hypothesis.

The second is a stated boundary. The dissertation fixes the population as the U.S.-led successful Mars surface landings from Viking 1 through Mars 2020, holds Tianwen-1 out for external-validity discussion [\[16\]](#ref-16), names the experience axis as mission sequence with a cumulative-landings alternative, and codes the technology indicators against TechPort insertion history. Stating the boundary is the disciplined move that prevents the most common error in this kind of analysis, extrapolating a secular claim from a short window, against which Kuznets warned directly [\[24\]](#ref-24). A boundary that is wrong can be corrected; a boundary that is unstated cannot even be debated. The boundary stands independent of the estimate.

The third is a falsifiable hypothesis where the literature offered only narrative. H1 predicts a specific sign and a specific ordering of effects: the technology coefficients jointly significant and negative, the approach-accuracy coefficient small and insignificant, and the InSight observation sitting above rather than on the trend. A regression on the assembled data can return coefficients that contradict every one of these predictions. The decision rule that maps coefficient configurations onto a verdict for H1 or H0 is fixed in advance, and the dissertation commits to reporting whichever pattern the data show, with a null result given the same prominence as a confirmation. This falsifiability is the contribution that most clearly transcends the result, because it is precisely the property that a narrative lacks. The claim that converting a narrative into a falsifiable proposition is itself a contribution rests on the program's own standard: a proposition the data can refute is a scientific object, and a narrative the data cannot refute is not. None of these three contributions depends on the sign of any coefficient; they are properties of the design, not of the estimate, so they hold under non-confirmation. If H1 is rejected, the dissertation will have established, with the same apparatus, that the contraction is generic maturation or approach-accuracy-driven rather than onboard-guidance-driven, and that finding is itself decision-relevant for an agency choosing where to invest, which is the symmetric-informativeness property the design was built to guarantee.

## 8.4 Honest limitations

The design's limitations are real, and the program's standard requires that they be stated at full strength rather than acknowledged and waved past. Four are material, and each was carried openly through the preceding chapters rather than discovered at the end.

The first and most binding is small n. The population is the entire relevant set of successful U.S. Mars landings, nine to eleven events depending on inclusion rules, which removes sampling-frame bias but leaves an irreducibly small sample. With this many observations, conventional asymptotic inference is unreliable, which is why the design uses exact and permutation-based tests and reports a sign and an order of magnitude rather than a precise point estimate with a narrow interval. The honest statement is the one the design chapter already made: no small-sample regression on this series will prove a learning rate, and the strongest defensible claim is a signed, order-of-magnitude, counterfactual-surviving attribution. This is a permanent feature of the problem, not a remediable defect of the method, because analysis cannot enlarge the population; only future missions enlarge it.

The second is collinearity. The three technology indicators are nearly collinear with each other and with the time trend, because the technologies were inserted monotonically over the era. Guided entry and the range trigger entered on the same mission, MSL, so the design is explicit that their individual contributions cannot be cleanly separated and treats the MSL increment as a guided-entry-plus-range-trigger bundle whose internal apportionment is the weakest inference in the study. TRN is more cleanly identified because Mars 2020 is the only mission carrying it, but the broader hazard that any monotone era-long improvement, modeling fidelity, atmospheric knowledge, computing, could masquerade as a technology effect is the dominant internal-validity threat. The nested-specification reporting makes any residual confound visible rather than hidden, but it cannot dissolve the collinearity, and the conclusion does not pretend otherwise.

The third is simulation-convention drift. The design ellipse is a simulation product whose conventions changed across missions, so the constructed series is only as comparable as the reconstruction work allows. This is the Kuznetsian comparability problem stated plainly, and the dissertation's response, flagging any value whose convention cannot be pinned down rather than silently harmonizing it, manages the problem but does not eliminate it. Where a Viking-era or Pathfinder-era three-sigma value rests only on an older record whose convention is irrecoverable, the comparability gap travels with the number as a stated limitation on the inference.

The fourth is single-counterfactual dependence. Identification of the technology effect against the pure-time rival leans heavily on one observation, InSight, a late, ballistic, unguided lander deliberately targeted to a large flat ellipse in Elysium Planitia [\[15\]](#ref-15). InSight is the linchpin precisely because it breaks the otherwise perfect time-technology collinearity: it has the calendar of MSL and Mars 2020 but the guidance of Viking and Phoenix, so if its ellipse is large, time alone cannot explain the contraction. Resting identification on a single case is fragile, and the design's response, re-fitting the whole hierarchy with InSight excluded to quantify exactly how much of the identification depends on it, measures the fragility rather than removing it. The honest claim, restated here, is that the InSight case discriminates the technology hypothesis from the pure-time hypothesis but cannot by itself apportion credit among the three levers, and that the apportionment is where the small sample bites hardest. None of these four limitations is fatal, and each is reported with the result rather than buried, but stating them plainly is part of what makes the eventual finding, in either direction, trustworthy.

## 8.5 A concrete future-research program

The future-research program has a clear ordering, and its first item is not an extension but a completion: execute the frozen design on the assembled data. The remaining items extend the work outward only after that completion is in hand.

### 8.5.1 Execute the frozen design on the full data

The immediate and highest-priority work is to run the pre-registered analysis plan against the constructed series. Concretely, this means assembling the ellipse series from the NTRS reconstruction records and published EDL performance studies, recording for each mission the three-sigma semi-major and semi-minor axes with the simulation convention that produced them; computing the log ellipse area and, separately, the log achieved miss distance from PDS localization; coding the three technology indicators from TechPort insertion history and the approach-accuracy control from the reconstruction records [\[13\]](#ref-13); fitting the baseline and the nested augmented specifications by OLS, one technology generation at a time, with permutation-based inference appropriate to n equal to nine to eleven; and running the robustness battery, the achieved-miss-distance dependent variable, the InSight-excluded re-fit, the program-rather-than-mission experience unit, and the linear-in-levels functional-form check. Because the analysis plan and decision rule are frozen, this execution is a matter of disciplined data assembly and computation rather than new design, and the design-stage character of the dissertation is what makes the completion a well-defined task rather than an open-ended one.

The reading of the executed estimates is itself fixed in advance, which is what makes the completion mechanical rather than interpretive. The learning-rate coefficient on mission sequence is read as \(\exp(\beta_1) - 1\), the proportional change in ellipse area per unit of mission sequence, so that a confirmed H1 produces a large negative sequence coefficient in the baseline whose magnitude shrinks as the technology indicators enter the nested hierarchy and absorb the contraction. The decisive reading is not that headline contraction, which the public record already leads us to expect and which is not in dispute, but the two diagnostics that discriminate the hypotheses: whether the InSight observation sits above the fitted trend, as H1 predicts for a late mission carrying the calendar but not the guidance of its contemporaries, or on it, as the pure-time rival predicts; and whether the approach-accuracy coefficient delta stays small and insignificant when it enters last, as H1 predicts, or absorbs the technology block, which the decision rule treats as failure to reject H0. The collapse or survival of the technology coefficients when delta is added is therefore the single most informative line of the executed nested table, and the plan reads it exactly as written rather than after seeing it. This advance commitment is the safeguard against the specific small-sample hazard that an analyst seeing nine observations first could find a specification that flatters a preferred conclusion; freezing the reading removes that degree of freedom.

The principal practical obstacles to the completion are the known coverage gaps the data chapter flagged: the Viking and Pathfinder ellipse provenance that may sit only in older NTRS records, the InSight reconstruction depth on which identification leans, the TechPort access path for the insertion history, and the per-mission approach-accuracy values that are reported across several records but not as a single tabulated series. Each of these is a sourcing task, to be handled by retrieving from the named datasets at execution time and flagging any irrecoverable value as a coverage limitation, never by imputation. This completion is tractable because every variable is operationally defined and every source is named and public; its value is nonetheless bounded by the same small-n and collinearity limits the design already acknowledges, so the executed result will be a disciplined sign-and-order-of-magnitude attribution, not a precise rate.

### 8.5.2 Widen the experience axis and the covariate set

A second line of work, available without new missions, is to enrich the experience axis and the covariate structure within the existing population. The dissertation's primary experience axis is mission sequence, with a cumulative-landings alternative; a natural extension is to test whether contraction tracks ordinal generation, cumulative experience, or a more granular measure of the accumulated propositional base, the volume of reconstruction and test work preceding each mission, which is the Mokyrian quantity the indicators only proxy [\[23\]](#ref-23). The organizational-learning literature supplies the template for separating mere cumulative experience from the collective learning processes that convert experience into improvement at different rates, and for distinguishing the codified from the tacit components of that learning [\[27\]](#ref-27), [\[28\]](#ref-28). Applying that distinction to the EDL series, by coding whether each technology's improvement rested on codified knowledge that transfers cleanly or tacit knowledge held by particular teams, would sharpen the Mokyrian reading, and the existing data can partly support it.

### 8.5.3 Extend to lunar and human-class precision landing once data exist

The third line is genuine extension beyond the current boundary, and it is deliberately staged to follow the in-sample completion rather than to substitute for it. Terrain-relative navigation is not a Mars-specific artifact; it is a general planetary-landing capability, demonstrated for lunar precision landing and surveyed as such [\[17\]](#ref-17), [\[117\]](#ref-117), [\[118\]](#ref-118), [\[119\]](#ref-119). As the Artemis-era lunar landers and their terrain-sensing-lidar and crater-navigation systems accumulate a flight record [\[128\]](#ref-128), [\[129\]](#ref-129), [\[130\]](#ref-130), a parallel lunar ellipse series becomes constructible, and the same learning-curve apparatus can be fit to it, with the Mars result as an out-of-sample prior rather than as in-sample evidence. The human-class Mars architecture is the further extension: the human Mars EDL architecture studies and the supersonic-retropropulsion development line define the forward edge at which payload mass, not guidance, becomes the binding constraint, and at which the relationship between guidance technology and achievable precision may change qualitatively [\[112\]](#ref-112), [\[113\]](#ref-113), [\[122\]](#ref-122), [\[111\]](#ref-111). These extensions are worth staging behind the completion for a clear reason: an apparatus validated on the in-sample Mars series is a credible instrument to carry to the Moon and to human-scale Mars, whereas an apparatus carried outward before it is validated in-sample would compound the small-sample fragility across bodies. Each extension inherits its own boundary and its own provenance discipline, per Kuznets, and the Mars learning curve does not extrapolate mechanically to other bodies; it is a reference point, not a prediction [\[24\]](#ref-24).

### 8.5.4 Maintain the knowledge chain as a research object

A fourth line follows from the Mokyrian reading of reversibility and deserves to be named as research rather than as caveat. The landing-precision capability is embodied in a chain of propositional knowledge, the engineering teams who hold it tacitly, the flight-software heritage, and the test infrastructure that validated each lever before flight; if any link atrophies between missions, a future program inherits not the demonstrated ellipse but whatever capability it can reconstitute [\[23\]](#ref-23). Treating the integrity of that knowledge chain as a measurable covariate, rather than as an unstated assumption behind any forward projection of the curve, is a research program in its own right, and it is the one that most directly serves the requirements-setting use the dissertation was built to enable, because it converts the warning that a fitted curve describes capability that was demonstrated, not capability that is guaranteed, into something an agency can monitor.

## 8.6 How the dissertation's argument stands at its conclusion

This chapter closes the dissertation's argument by carrying its final part, that the residual risk is acceptable and managed, to a full statement, having inherited the earlier parts from the preceding chapters. That the problem is real, the large and currently only qualitatively explained ellipse contraction, was established in the introduction and literature chapters [\[1\]](#ref-1), [\[5\]](#ref-5), [\[8\]](#ref-8). That the problem is material, that the ellipse governs which sites are reachable and is decision-critical for sample return and human Mars, was carried through the discussion of requirements-setting and investment valuation [\[15\]](#ref-15). That the design addresses the causal mechanism, by modeling discrete technology generations as covariates on a Kuznets-disciplined series, and that it discriminates the technology account from its rivals, through the InSight counterfactual and the approach-accuracy control, were carried by the design and analysis chapters [\[24\]](#ref-24), [\[15\]](#ref-15). This concluding chapter carries the residual-risk point explicitly: the small sample, the collinearity, the simulation-convention drift, and the single-counterfactual dependence are real but bounded, by permutation inference, by nested specifications, by the dual dependent variable, and by the InSight-excluded re-fit, and the design is informative whichever way the coefficient falls. This chapter contributes the residual-risk point in its calibrated, design-stage form and confirms that the argument is complete: the problem is real, the problem is material, the design addresses the mechanism, it discriminates that account from its rivals, and the residual risk is acceptable.

Consistent with the dissertation's explicit scope decision, this chapter develops the mechanism and the confidence at the heart of the argument, and it does not produce a systems or capability architecture. The contribution is a cliometric measurement-and-attribution account of a constructed performance series, not the design of a real capability, system, or data exchange, so no capability-to-system traceability is forced onto the argument. The decision relevance, requirements-setting and investment valuation, is carried in plain prose, here and in the introductory and discussion chapters, rather than as a traceability table, and the conclusion rests on the falsifiability of the decision rule and the provenance of the series, not on any systems-architecture vocabulary.
## 8.7 Closing

This dissertation began with a number that everyone in Mars mission design knows and that no one had measured as a series: the landing ellipse, the probabilistic region into which a vehicle is committed years before launch, whose contraction across the robotic era from Viking-class regions in the high hundreds of kilometers to a few-kilometer effective targeting region for Mars 2020 is the quiet history of the discipline. The qualitative story of that contraction has been told many times. What was missing was a quantitative one, and what the quantitative one lacked was an arbiter among its rival causes. The dissertation supplies the design for that arbiter. It defines the measurement, states the boundary, names the mechanism through Mokyr and disciplines the series through Kuznets, fixes a falsifiable hypothesis and a pre-registered decision rule, and rests its identification on the InSight counterfactual and the approach-accuracy control rather than on asymptotic significance the small sample cannot support.

Executed, the work would give NASA and JPL a quantitative, technology-attributed basis for setting landing-accuracy requirements and for valuing guidance investments against approach-navigation investments on a common scale, in the human-Mars and sample-return architectures now under study [\[15\]](#ref-15). Not executed, it still contributes a defined measurement, a stated boundary, and a falsifiable hypothesis where the literature previously offered only a qualitative narrative [\[18\]](#ref-18), [\[23\]](#ref-23), [\[24\]](#ref-24). The proposition may survive its test or fail it; the design was built to be informative either way, and the apparatus to decide it is now in hand. What remains is to assemble the series, run the frozen plan, and report whichever pattern the data reveal. The contribution is the instrument and the discipline; the verdict awaits the data. If the work serves its purpose, it will do so quietly, as a faithful steward of the capability that many hands have built, in the service of those who will guide the next vehicles down to the Martian surface.
## References

The list is given in corpus-key order: seed references first (S01 to S24), then harvested references (R025 to R130). Every entry is real and drawn from the candidate corpus, and each carries a clickable DOI or, for the NASA Technical Reports Server (NTRS) reconstruction reports that have no journal DOI, a resolvable NTRS citation URL. Every in-text citation in the chapters resolves to an entry below; the two Hall-of-Shoulders dossiers [\[23\]](#ref-23) and [\[24\]](#ref-24) are local methodological research instruments and name their underlying canonical works in the entry.

The list is given in corpus-key order: seed references first (S01 to S24), then harvested references (R025 to R130). Each entry is real and drawn from `research/corpus.jsonl`.

### Seed references (S01 to S24)

1. <span id="ref-1"></span>**[S01]** Steltzner, A., et al. (2007). "Mars Science Laboratory: Entry, Descent, and Landing System Performance." *IEEE Aerospace Conference*. DOI: [10.1109/aero.2007.352821](https://doi.org/10.1109/aero.2007.352821).
2. <span id="ref-2"></span>**[S02]** Way, D. W., et al. (2008). "Mars Science Laboratory: Entry, Descent, and Landing System Overview." *IEEE Aerospace Conference*. DOI: [10.1109/aero.2008.4526283](https://doi.org/10.1109/aero.2008.4526283).
3. <span id="ref-3"></span>**[S03]** Mendeck, G. F., et al. (2008). "Entry Guidance Performance for Mars Precision Landing." *Journal of Guidance, Control, and Dynamics* 31(4). DOI: [10.2514/1.36950](https://doi.org/10.2514/1.36950).
4. <span id="ref-4"></span>**[S04]** Mendeck, G. F., & Craig, L. (2011). "Entry Guidance for the 2011 Mars Science Laboratory Mission." *AIAA*. DOI: [10.2514/6.2011-6639](https://doi.org/10.2514/6.2011-6639).
5. <span id="ref-5"></span>**[S05]** Way, D. W. (2011). "On the use of a range trigger for the Mars Science Laboratory Entry, Descent, and Landing." *IEEE Aerospace Conference*. DOI: [10.1109/aero.2011.5747242](https://doi.org/10.1109/aero.2011.5747242).
6. <span id="ref-6"></span>**[S06]** Desai, P. N., et al. (2011). "Entry, Descent, and Landing Performance of the Mars Phoenix Lander." *Journal of Spacecraft and Rockets* 48(5). DOI: [10.2514/1.48239](https://doi.org/10.2514/1.48239).
7. <span id="ref-7"></span>**[S07]** Karlgaard, C. D., et al. (2007). "Statistical Reconstruction of Mars Entry, Descent, and Landing Trajectories and Atmospheric Profiles." *AIAA*. DOI: [10.2514/6.2007-6192](https://doi.org/10.2514/6.2007-6192).
8. <span id="ref-8"></span>**[S08]** Johnson, A. E., et al. (2022). "Mars 2020 Lander Vision System Flight Performance." *AIAA SciTech*. DOI: [10.2514/6.2022-1214](https://doi.org/10.2514/6.2022-1214).
9. <span id="ref-9"></span>**[S09]** Cheng, Y., et al. (2021). "Making an Onboard Reference Map From MRO/CTX Imagery for Mars 2020 Lander Vision System." *Earth and Space Science* 8(4). DOI: [10.1029/2020ea001560](https://doi.org/10.1029/2020ea001560).
10. <span id="ref-10"></span>**[S10]** Johnson, A. E., et al. (2015). "Real-Time Terrain Relative Navigation Test Results from a Relevant Environment for Mars Landing." *AIAA*. DOI: [10.2514/6.2015-0851](https://doi.org/10.2514/6.2015-0851).
11. <span id="ref-11"></span>**[S11]** Johnson, A. E., et al. (2023). "Implementation of a Map Relative Localization System for Planetary Landing." *Journal of Guidance, Control, and Dynamics* 46(4). DOI: [10.2514/1.g006780](https://doi.org/10.2514/1.g006780).
12. <span id="ref-12"></span>**[S12]** Wang, T., et al. (2022). "Planetary landings with terrain sensing and hazard avoidance: A review." *Advances in Space Research* 71(8). DOI: [10.1016/j.asr.2022.11.024](https://doi.org/10.1016/j.asr.2022.11.024).
13. <span id="ref-13"></span>**[S13]** Karlgaard, C. D., et al. (2022). "Mars Entry, Descent, and Landing Instrumentation 2 Trajectory, Aerodynamics, and Atmosphere Reconstruction." *Journal of Spacecraft and Rockets* 59(4). DOI: [10.2514/1.a35440](https://doi.org/10.2514/1.a35440).
14. <span id="ref-14"></span>**[S14]** (NASA/JPL) (2021). "Reconstruction of Entry, Descent, and Landing Communications for the InSight Mars Lander." *Journal of Spacecraft and Rockets* 58(4). DOI: [10.2514/1.a34892](https://doi.org/10.2514/1.a34892).
15. <span id="ref-15"></span>**[S15]** Golombek, M., et al. (2016). "Selection of the InSight Landing Site." *Space Science Reviews* 211. DOI: [10.1007/s11214-016-0321-9](https://doi.org/10.1007/s11214-016-0321-9).
16. <span id="ref-16"></span>**[S16]** Yu, Z., et al. (2021). "The Tianwen-1 Guidance, Navigation, and Control for Mars Entry, Descent, and Landing." *Space: Science & Technology* 2021. DOI: [10.34133/2021/9846185](https://doi.org/10.34133/2021/9846185).
17. <span id="ref-17"></span>**[S17]** (2020). "Crater Navigation System for Autonomous Precision Landing on the Moon." *Journal of Guidance, Control, and Dynamics* 43(8). DOI: [10.2514/1.g004850](https://doi.org/10.2514/1.g004850).
18. <span id="ref-18"></span>**[S18]** Alberth, S. (2008). "Forecasting technology costs via the experience curve - Myth or magic?" *Technological Forecasting and Social Change* 75(7). DOI: [10.1016/j.techfore.2007.09.003](https://doi.org/10.1016/j.techfore.2007.09.003).
19. <span id="ref-19"></span>**[S19]** Ziegler, M. S., & Trancik, J. E. (2021). "Re-examining rates of lithium-ion battery technology improvement and cost decline." *Energy & Environmental Science* 14. DOI: [10.1039/d0ee02681f](https://doi.org/10.1039/d0ee02681f).
20. <span id="ref-20"></span>**[S20]** Landefeld, J. S., Seskin, E. P., & Fraumeni, B. M. (2008). "Taking the Pulse of the Economy: Measuring GDP." *Journal of Economic Perspectives* 22(2). DOI: [10.1257/jep.22.2.193](https://doi.org/10.1257/jep.22.2.193).
21. <span id="ref-21"></span>**[S21]** Henderson, J. V., Storeygard, A., & Weil, D. N. (2012). "Measuring Economic Growth from Outer Space." *American Economic Review* 102(2). DOI: [10.1257/aer.102.2.994](https://doi.org/10.1257/aer.102.2.994).
22. <span id="ref-22"></span>**[S22]** Perez, C. (2009). "Technological revolutions and techno-economic paradigms." *Cambridge Journal of Economics* 34(1). DOI: [10.1093/cje/bep051](https://doi.org/10.1093/cje/bep051).
23. <span id="ref-23"></span>**[S23]** Mokyr, J. (dossier) (2002). "Hall of Shoulders dossier: Joel Mokyr, economic history of technology." Hall of Shoulders research instrument; methodological anchor restating the propositional-versus-prescriptive-knowledge distinction and the macro-versus-incremental-invention line drawn in *The Gifts of Athena: Historical Origins of the Knowledge Economy* (Princeton University Press, 2002). Local record: `hall_of_shoulders/brains/mokyr/dossier.md`.
24. <span id="ref-24"></span>**[S24]** Kuznets, S. (dossier) (1941). "Hall of Shoulders dossier: Simon Kuznets, cliometric measurement." Hall of Shoulders research instrument; methodological anchor restating the boundary-valuation-netting requirement and the transient-versus-secular distinction developed in *National Income and Its Composition, 1919-1938* (NBER, 1941). Local record: `hall_of_shoulders/brains/kuznets/dossier.md`.

### Harvested references (R025 to R130)

25. <span id="ref-25"></span>**[R025]** Crafts, N. (2010). "The contribution of new technology to economic growth: lessons from economic history." *Revista de Historia Económica / Journal of Iberian and Latin American Economic History*. DOI: [10.1017/s0212610910000157](https://doi.org/10.1017/s0212610910000157).
26. <span id="ref-26"></span>**[R026]** Chen, K., & Huang, W. (2026). "Joel Mokyr's Contribution to the Nobel Prize: The Economic Growth Theory of Knowledge Dissemination and Institutional Innovation." *Journal of Statistics and Economics*. DOI: [10.62517/jse.202611107](https://doi.org/10.62517/jse.202611107).
27. <span id="ref-27"></span>**[R027]** Pisano, G. P., Bohmer, R. M. J., & Edmondson, A. C. (2001). "Organizational Differences in Rates of Learning: Evidence from the Adoption of Minimally Invasive Cardiac Surgery." *Management Science* 47(6). DOI: [10.1287/mnsc.47.6.752.9811](https://doi.org/10.1287/mnsc.47.6.752.9811).
28. <span id="ref-28"></span>**[R028]** Edmondson, A. C., Winslow, A. B., Bohmer, R. M. J., & Pisano, G. P. (2003). "Learning How and Learning What: Effects of Tacit and Codified Knowledge on Performance Improvement Following Technology Adoption." *Decision Sciences* 34. DOI: [10.1111/1540-5915.02316](https://doi.org/10.1111/1540-5915.02316).
29. <span id="ref-29"></span>**[R029]** Colley, J. L. (2009). "Learning Curve/Experience Curve Theory." *SSRN Electronic Journal*. DOI: [10.2139/ssrn.1421681](https://doi.org/10.2139/ssrn.1421681).
30. <span id="ref-30"></span>**[R030]** Chang, Y. S. (2013). "Next Frontier of Experience Curve Application: Predicting Performance of Large Learning Systems by Kinked Experience Curve." *SSRN Electronic Journal*. DOI: [10.2139/ssrn.2305140](https://doi.org/10.2139/ssrn.2305140).
31. <span id="ref-31"></span>**[R031]** (2011). "Experience Curve." *Encyclopedia of Machine Learning*. DOI: [10.1007/978-0-387-30164-8_292](https://doi.org/10.1007/978-0-387-30164-8_292).
32. <span id="ref-32"></span>**[R032]** (2020). "Aggregate National Accounts, SNA 1993: Gross domestic product (Edition 2019/1)." *OECD National Accounts Statistics*. DOI: [10.1787/13294c60-en](https://doi.org/10.1787/13294c60-en).
33. <span id="ref-33"></span>**[R033]** Korzun, A. M., Dubos, G. F., Iwata, C., & Stahl, B. (2009). "A concept for the entry, descent, and landing of high-mass payloads at Mars." *Acta Astronautica*. DOI: [10.1016/j.actaastro.2009.10.003](https://doi.org/10.1016/j.actaastro.2009.10.003).
34. <span id="ref-34"></span>**[R034]** Lugo, R., Powell, R. W., & Dwyer-Cianciolo, A. M. (2020). "Overview of a Generalized Numerical Predictor-Corrector Targeting Guidance with Application to Human-Scale Mars Entry, Descent, and Landing." *AIAA Scitech 2020 Forum*. DOI: [10.2514/6.2020-0846](https://doi.org/10.2514/6.2020-0846).
35. <span id="ref-35"></span>**[R035]** Edquist, K. T., Desai, P. N., & Schoenenberger, M. (2011). "Aerodynamics for Mars Phoenix Entry Capsule." *Journal of Spacecraft and Rockets*. DOI: [10.2514/1.46219](https://doi.org/10.2514/1.46219).
36. <span id="ref-36"></span>**[R036]** Prince, J. L., Desai, P. N., Queen, E. M., & Grover, M. R. (2011). "Entry, Descent, and Landing Operations Analysis for the Mars Phoenix Lander." *Journal of Spacecraft and Rockets*. DOI: [10.2514/1.46563](https://doi.org/10.2514/1.46563).
37. <span id="ref-37"></span>**[R037]** Bhartia, R., Beegle, L. W., DeFlores, L., & Abbey, W. (2021). "Perseverance's Scanning Habitable Environments with Raman and Luminescence for Organics and Chemicals (SHERLOC) Investigation." *Space Science Reviews* 217. DOI: [10.1007/s11214-021-00812-z](https://doi.org/10.1007/s11214-021-00812-z).
38. <span id="ref-38"></span>**[R038]** Maki, J. N., Gruel, D., McKinney, C., & Ravine, M. A. (2020). "The Mars 2020 Engineering Cameras and Microphone on the Perseverance Rover: A Next-Generation Imaging System for Mars Exploration." *Space Science Reviews* 216. DOI: [10.1007/s11214-020-00765-9](https://doi.org/10.1007/s11214-020-00765-9).
39. <span id="ref-39"></span>**[R039]** Grotzinger, J. P., Crisp, J. A., Vasavada, A. R., & Anderson, R. C. (2012). "Mars Science Laboratory Mission and Science Investigation." *Space Science Reviews* 170. DOI: [10.1007/s11214-012-9892-2](https://doi.org/10.1007/s11214-012-9892-2).
40. <span id="ref-40"></span>**[R040]** Braun, R. D., Sforzo, B., & Campbell, C. (2017). "Advancing Supersonic Retropropulsion Using Mars-Relevant Flight Data: An Overview." *AIAA SPACE and Astronautics Forum and Exposition*. DOI: [10.2514/6.2017-5292](https://doi.org/10.2514/6.2017-5292).
41. <span id="ref-41"></span>**[R041]** Jiang, X., Li, S., & Tao, T. (2016). "Innovative hazard detection and avoidance strategy for autonomous safe planetary landing." *Acta Astronautica*. DOI: [10.1016/j.actaastro.2016.02.028](https://doi.org/10.1016/j.actaastro.2016.02.028).
42. <span id="ref-42"></span>**[R042]** Bell, J. F., Maki, J. N., Mehall, G., & Ravine, M. A. (2021). "The Mars 2020 Perseverance Rover Mast Camera Zoom (Mastcam-Z) Multispectral, Stereoscopic Imaging Investigation." *Space Science Reviews* 217. DOI: [10.1007/s11214-020-00755-x](https://doi.org/10.1007/s11214-020-00755-x).
43. <span id="ref-43"></span>**[R043]** Lugo, R. A., Dutta, S., & Manwell, M. (2025). "Application of a Dual-Quaternion Six Degree-of-Freedom Guidance to Human-Scale Mars Entry, Descent, and Landing." *AIAA SCITECH 2025 Forum*. DOI: [10.2514/6.2025-1893](https://doi.org/10.2514/6.2025-1893).
44. <span id="ref-44"></span>**[R044]** Williams, J., Brandenburg, W. E., Woffinden, D., & Putnam, Z. R. (2022). "Validation of Linear Covariance Techniques for Mars Entry, Descent, and Landing Guidance and Navigation Performance Analysis." *AIAA SCITECH 2022 Forum*. DOI: [10.2514/6.2022-0745](https://doi.org/10.2514/6.2022-0745).
45. <span id="ref-45"></span>**[R045]** Johnson, A., Ansar, A., Matthies, L., & Trawny, N. (2007). "A General Approach to Terrain Relative Navigation for Planetary Landing." *AIAA Infotech@Aerospace 2007 Conference and Exhibit*. DOI: [10.2514/6.2007-2854](https://doi.org/10.2514/6.2007-2854).
46. <span id="ref-46"></span>**[R046]** Crain, T., & Bishop, R. (2002). "Mars Entry Navigation: Atmospheric Interface Through Parachute Deploy." *AIAA Atmospheric Flight Mechanics Conference and Exhibit*. DOI: [10.2514/6.2002-4501](https://doi.org/10.2514/6.2002-4501).
47. <span id="ref-47"></span>**[R047]** Witkowski, A., & Kandis, M. (2010). "Reefing the Mars Science Laboratory parachute." *2010 IEEE Aerospace Conference*. DOI: [10.1109/aero.2010.5447012](https://doi.org/10.1109/aero.2010.5447012).
48. <span id="ref-48"></span>**[R048]** Witkowski, A. (2007). "Mars 2007 Scout Phoenix Parachute Decelerator System Program Overview." *2007 IEEE Aerospace Conference*. DOI: [10.1109/aero.2007.352817](https://doi.org/10.1109/aero.2007.352817).
49. <span id="ref-49"></span>**[R049]** Maddock, R., Henning, A., & Samareh, J. (2016). "Passive vs. parachute system architecture for robotic sample return vehicles." *IEEE Aerospace Conference*. DOI: [10.1109/aero.2016.7500545](https://doi.org/10.1109/aero.2016.7500545).
50. <span id="ref-50"></span>**[R050]** Chavez-Clemente, D., Younse, P., Ballantoni, C. W., & Backus, S. (2025). "Development of the Robotic Transfer Assembly System for Mars Sample Return." *Journal of Spacecraft and Rockets*. DOI: [10.2514/1.a36200](https://doi.org/10.2514/1.a36200).
51. <span id="ref-51"></span>**[R051]** Dai, J., Gao, A., & Xia, Y. (2016). "Mars atmospheric entry guidance for reference trajectory tracking based on robust nonlinear compound controller." *Acta Astronautica*. DOI: [10.1016/j.actaastro.2016.12.013](https://doi.org/10.1016/j.actaastro.2016.12.013).
52. <span id="ref-52"></span>**[R052]** Way, D. W. (2013). "Preliminary assessment of the Mars Science Laboratory entry, descent, and landing simulation." *IEEE Aerospace Conference*. DOI: [10.1109/aero.2013.6497404](https://doi.org/10.1109/aero.2013.6497404).
53. <span id="ref-53"></span>**[R053]** Mourikis, A. I., Trawny, N., Roumeliotis, S. I., & Johnson, A. (2009). "Vision-Aided Inertial Navigation for Spacecraft Entry, Descent, and Landing." *IEEE Transactions on Robotics* 25. DOI: [10.1109/tro.2009.2012342](https://doi.org/10.1109/tro.2009.2012342).
54. <span id="ref-54"></span>**[R054]** Mourikis, A. I., Trawny, N., Roumeliotis, S. I., & Johnson, A. (2007). "Vision-Aided Inertial Navigation for Precise Planetary Landing: Analysis and Experiments." *Robotics: Science and Systems III*. DOI: [10.15607/rss.2007.iii.019](https://doi.org/10.15607/rss.2007.iii.019).
55. <span id="ref-55"></span>**[R055]** Woicke, S., & Mooij, E. (2017). "Terrain Relative Navigation for Planetary Landing Using Stereo Vision Measurements Obtained from Hazard Mapping." DOI: [10.1007/978-3-319-65283-2_39](https://doi.org/10.1007/978-3-319-65283-2_39).
56. <span id="ref-56"></span>**[R056]** Maurice, S., Clegg, S. M., Wiens, R. C., & Gasnault, O. (2016). "ChemCam activities and discoveries during the nominal mission of the Mars Science Laboratory in Gale crater, Mars." *Journal of Analytical Atomic Spectrometry* 31. DOI: [10.1039/c5ja00417a](https://doi.org/10.1039/c5ja00417a).
57. <span id="ref-57"></span>**[R057]** Desai, P. N., & Knocke, P. (2007). "Mars exploration rovers entry, descent, and landing trajectory analysis." *The Journal of the Astronautical Sciences* 55. DOI: [10.1007/bf03256527](https://doi.org/10.1007/bf03256527).
58. <span id="ref-58"></span>**[R058]** Mitcheltree, R., Kellas, S., Dorsey, J. T., & Desai, P. N. (1998). "A passive Earth-entry capsule for Mars Sample Return." DOI: [10.2514/6.1998-2851](https://doi.org/10.2514/6.1998-2851).
59. <span id="ref-59"></span>**[R059]** Crisp, J. A., Adler, M., Matijevic, J., & Squyres, S. W. (2003). "Mars Exploration Rover mission." *Journal of Geophysical Research: Atmospheres* 108. DOI: [10.1029/2002je002038](https://doi.org/10.1029/2002je002038).
60. <span id="ref-60"></span>**[R060]** Parkes, S., Dunstan, M., Martin, I., & Mendham, P. (2006). "PANGU: Planet and Asteroid Natural Scene Generation Utility (57th International Astronautical Congress)." DOI: [10.2514/miaf06](https://doi.org/10.2514/miaf06).
61. <span id="ref-61"></span>**[R061]** Simon, J. I., Hickman-Lewis, K., Cohen, B. A., & Mayhew, L. E. (2023). "Samples Collected From the Floor of Jezero Crater With the Mars 2020 Perseverance Rover." *Journal of Geophysical Research: Planets* 128. DOI: [10.1029/2022je007474](https://doi.org/10.1029/2022je007474).
62. <span id="ref-62"></span>**[R062]** Tao, Y., Conway, S. J., Müller, J.-P., & Putri, A. R. D. (2021). "Single Image Super-Resolution Restoration of TGO CaSSIS Colour Images: Demonstration with Perseverance Rover Landing Site and Mars Science Targets." *Remote Sensing* 13(9). DOI: [10.3390/rs13091777](https://doi.org/10.3390/rs13091777).
63. <span id="ref-63"></span>**[R063]** Bos, B. J., Donovan, D. L., Capone, J., & Wang, C. (2024). "Vision System for the Mars Sample Return Capture Containment and Return System (CCRS)." *Aerospace* 11(6). DOI: [10.3390/aerospace11060456](https://doi.org/10.3390/aerospace11060456).
64. <span id="ref-64"></span>**[R064]** Amato, D. (2021). "Mars Entry, Descent, and Landing guidance under dynamic uncertainty." DOI: [10.52843/cassyni.r8xwrv](https://doi.org/10.52843/cassyni.r8xwrv).
65. <span id="ref-65"></span>**[R065]** Mischna, M. (2024). "Mars Atmospheric Entry, Descent, and Landing: An Atmospheric Perspective." *Oxford Research Encyclopedia of Planetary Science*. DOI: [10.1093/acrefore/9780190647926.013.125](https://doi.org/10.1093/acrefore/9780190647926.013.125).
66. <span id="ref-66"></span>**[R066]** Sutterlin, G. E., & Eapen, R. T. (2023). "Vision-based Terrain Relative Navigation for Planetary Landing." *2023 Regional Student Conferences*. DOI: [10.2514/6.2023-71389](https://doi.org/10.2514/6.2023-71389).
67. <span id="ref-67"></span>**[R067]** Abdhul Rahuman, M. A., & Lee, K. (2026). "A Review of Vision-Based Terrain Relative Navigation for Planetary Descent and Landing of Space Vehicles." *International Journal of Aeronautical and Space Sciences*. DOI: [10.1007/s42405-026-01139-1](https://doi.org/10.1007/s42405-026-01139-1).
68. <span id="ref-68"></span>**[R068]** Sun, V., & Stack, K. M. (2018). "Geomorphic Mapping of the Basement Unit Within the Northeast Syrtis Mars 2020 Landing Ellipse." *Geological Society of America Abstracts with Programs*. DOI: [10.1130/abs/2018rm-313607](https://doi.org/10.1130/abs/2018rm-313607).
69. <span id="ref-69"></span>**[R069]** Palaszewski, B. (2012). "Entry, Descent, and Landing with Propulsive Deceleration: Supersonic Retropropulsion Wind Tunnel Testing." *50th AIAA Aerospace Sciences Meeting*. DOI: [10.2514/6.2012-401](https://doi.org/10.2514/6.2012-401).
70. <span id="ref-70"></span>**[R070]** Huertas, A., Johnson, A., Werner, R. A., & Maddock, R. A. (2010). "Performance evaluation of hazard detection and avoidance algorithms for safe Lunar landings." *IEEE Aerospace Conference*. DOI: [10.1109/aero.2010.5447022](https://doi.org/10.1109/aero.2010.5447022).
71. <span id="ref-71"></span>**[R071]** Kallemeyn, P. H., Peng, C. Y., Braun, R. D., & Thurman, S. W. (1998). "Mars Pathfinder Atmospheric Entry Reconstruction." NASA Technical Reports Server. URL: [https://ntrs.nasa.gov/citations/20210005354](https://ntrs.nasa.gov/citations/20210005354).
72. <span id="ref-72"></span>**[R072]** Karlgaard, C. D., Kutty, P., Schoenenberger, M., & Shidner, J. D. (2013). "Mars Science Laboratory Entry, Descent, and Landing Trajectory and Atmosphere Reconstruction." NASA Technical Reports Server. URL: [https://ntrs.nasa.gov/citations/20130010087](https://ntrs.nasa.gov/citations/20130010087).
73. <span id="ref-73"></span>**[R073]** Hwang, H. H., Bose, D., White, T. R., & Wright, H. S. (2016). "Mars 2020 Entry, Descent and Landing Instrumentation 2 (MEDLI2)." NASA Technical Reports Server. URL: [https://ntrs.nasa.gov/citations/20160008391](https://ntrs.nasa.gov/citations/20160008391).
74. <span id="ref-74"></span>**[R074]** Karlgaard, C. D., Korzun, A. M., Schoenenberger, M., & Bonfiglio, E. P. (2020). "Mars InSight Entry, Descent, and Landing Trajectory and Atmosphere Reconstruction." NASA Technical Reports Server. URL: [https://ntrs.nasa.gov/citations/20200002910](https://ntrs.nasa.gov/citations/20200002910).
75. <span id="ref-75"></span>**[R075]** Steltzner, A. D., San Martin, A. M., & Rivellini, T. P. (2015). "Mars Science Laboratory Entry, Descent and Landing System Development Challenges and Preliminary Flight Performance." NASA Technical Reports Server. URL: [https://ntrs.nasa.gov/citations/20150012005](https://ntrs.nasa.gov/citations/20150012005).
76. <span id="ref-76"></span>**[R076]** Bose, D., Wright, H., White, T., & Schoenenberger, M. (2016). "Mars 2020 Entry, Descent and Landing Instrumentation (MEDLI2)." NASA Technical Reports Server. URL: [https://ntrs.nasa.gov/citations/20160000694](https://ntrs.nasa.gov/citations/20160000694).
77. <span id="ref-77"></span>**[R077]** Martin-Mur, T. J., Kruizinga, G. L., Burkhart, P. D., & Wong, M. C. (2015). "Mars Science Laboratory Navigation Results." NASA Technical Reports Server. URL: [https://ntrs.nasa.gov/citations/20150005583](https://ntrs.nasa.gov/citations/20150005583).
78. <span id="ref-78"></span>**[R078]** Dutta, S., Way, D. W., Casoliva, J., & Brugarolas, P. (2021). "Mars 2020 Perseverance EDL GNC Safe Target Selection Reconstruction." NASA Technical Reports Server. URL: [https://ntrs.nasa.gov/citations/20230005646](https://ntrs.nasa.gov/citations/20230005646).
79. <span id="ref-79"></span>**[R079]** Silvestrini, S. (2026). "Deep visual odometry and pose reconstruction through single image depth map and triangulation for terrain relative navigation." *Astrodynamics*. DOI: [10.1007/s42064-025-0282-4](https://doi.org/10.1007/s42064-025-0282-4).
80. <span id="ref-80"></span>**[R080]** Neukart, F. (2024). "Towards sustainable horizons: A comprehensive blueprint for Mars colonization." *Heliyon*. DOI: [10.1016/j.heliyon.2024.e26180](https://doi.org/10.1016/j.heliyon.2024.e26180).
81. <span id="ref-81"></span>**[R081]** Knocke, P., Wawrzyniak, G., Kennedy, B., & Desai, P. (2004). "Mars Exploration Rovers Landing Dispersion Analysis." *AIAA/AAS Astrodynamics Specialist Conference and Exhibit*. DOI: [10.2514/6.2004-5093](https://doi.org/10.2514/6.2004-5093).
82. <span id="ref-82"></span>**[R082]** Kim, I., & Singh, S. K. (2024). "Probabilistic regression for autonomous terrain relative navigation via multi-modal feature learning." *Scientific Reports* 14. DOI: [10.1038/s41598-024-81377-z](https://doi.org/10.1038/s41598-024-81377-z).
83. <span id="ref-83"></span>**[R083]** Bonfiglio, E., Adams, D., Craig, L. E., & Spencer, D. (2008). "Landing Site Dispersion Analysis and Statistical Assessment for the Mars Phoenix Lander." DOI: [10.2514/6.2008-7348](https://doi.org/10.2514/6.2008-7348).
84. <span id="ref-84"></span>**[R084]** Spencer, D., & Braun, R. (1996). "Mars Pathfinder atmospheric entry: Trajectory design and dispersion analysis." *Journal of Spacecraft and Rockets* 33. DOI: [10.2514/3.26819](https://doi.org/10.2514/3.26819).
85. <span id="ref-85"></span>**[R085]** Robertson, E. A. (2016). "Synopsis of Precision Landing and Hazard Avoidance (PL&HA) Capabilities for Space Exploration." NASA Technical Reports Server. URL: [https://ntrs.nasa.gov/citations/20160014508](https://ntrs.nasa.gov/citations/20160014508).
86. <span id="ref-86"></span>**[R086]** Lugo, R. A., Dwyer Cianciolo, A., Williams, R. A., & Dutta, S. (2021). "Integrated Precision Landing Performance and Technology Assessments of a Human-Scale Mars Lander Using a Generalized Simulation Framework." NASA Technical Reports Server. URL: [https://ntrs.nasa.gov/citations/20210024549](https://ntrs.nasa.gov/citations/20210024549).
87. <span id="ref-87"></span>**[R087]** Duan, G., & Mease, K. D. (2018). "Mars Entry Guidance for High Elevation Landing with Uncertainty Quantification and Reduction." *22nd AIAA International Space Planes and Hypersonics Systems and Technologies Conference*. DOI: [10.2514/6.2018-5315](https://doi.org/10.2514/6.2018-5315).
88. <span id="ref-88"></span>**[R088]** Chari, G. M., Kamath, A. G., Elango, P., & Açıkmese, B. (2024). "Fast Monte Carlo Analysis For 6-DoF Powered-Descent Guidance via GPU-Accelerated Sequential Convex Programming." *AIAA SCITECH 2024 Forum*. DOI: [10.2514/6.2024-1762](https://doi.org/10.2514/6.2024-1762).
89. <span id="ref-89"></span>**[R089]** Desai, P. N., Mitcheltree, R., & Cheatwood, F. M. (1999). "Entry Dispersion Analysis for the Stardust Comet Sample Return Capsule." *Journal of Spacecraft and Rockets* 36. DOI: [10.2514/2.3467](https://doi.org/10.2514/2.3467).
90. <span id="ref-90"></span>**[R090]** Gigerenzer, G., & Gaissmaier, W. (2010). "Heuristic Decision Making." *Annual Review of Psychology* 62. DOI: [10.1146/annurev-psych-120709-145346](https://doi.org/10.1146/annurev-psych-120709-145346).
91. <span id="ref-91"></span>**[R091]** Trawny, N., Benito, J., Tweddle, B., & Bergh, C. (2015). "Flight testing of terrain-relative navigation and large-divert guidance on a VTVL rocket." *AIAA SPACE 2015 Conference and Exposition*. DOI: [10.2514/6.2015-4418](https://doi.org/10.2514/6.2015-4418).
92. <span id="ref-92"></span>**[R092]** Açıkmese, B., & Ploen, S. (2007). "Convex Programming Approach to Powered Descent Guidance for Mars Landing." *Journal of Guidance, Control, and Dynamics* 30(5). DOI: [10.2514/1.27553](https://doi.org/10.2514/1.27553).
93. <span id="ref-93"></span>**[R093]** Song, Y., Miao, X., & Gong, S. (2020). "Adaptive powered descent guidance based on multi-phase pseudospectral convex optimization." *Acta Astronautica*. DOI: [10.1016/j.actaastro.2020.12.019](https://doi.org/10.1016/j.actaastro.2020.12.019).
94. <span id="ref-94"></span>**[R094]** Açıkmese, B., Scharf, D. P., Blackmore, L., & Wolf, A. A. (2008). "Enhancements on the Convex Programming Based Powered Descent Guidance Algorithm for Mars Landing." *AIAA/AAS Astrodynamics Specialist Conference and Exhibit*. DOI: [10.2514/6.2008-6426](https://doi.org/10.2514/6.2008-6426).
95. <span id="ref-95"></span>**[R095]** Açıkmese, A., & Ploen, S. (2005). "A Powered Descent Guidance Algorithm for Mars Pinpoint Landing." *AIAA Guidance, Navigation, and Control Conference and Exhibit*. DOI: [10.2514/6.2005-6288](https://doi.org/10.2514/6.2005-6288).
96. <span id="ref-96"></span>**[R096]** Simplício, P., Marcos, A., & Bennani, S. (2019). "Guidance of Reusable Launchers: Improving Descent and Landing Performance." *Journal of Guidance, Control, and Dynamics* 42. DOI: [10.2514/1.g004155](https://doi.org/10.2514/1.g004155).
97. <span id="ref-97"></span>**[R097]** Eren, U., Dueri, D., & Açıkmese, B. (2015). "Constrained Reachability and Controllability Sets for Planetary Precision Landing via Convex Optimization." *Journal of Guidance, Control, and Dynamics* 38. DOI: [10.2514/1.g000882](https://doi.org/10.2514/1.g000882).
98. <span id="ref-98"></span>**[R098]** Johnson, B. J., Lu, P., & Sostaric, R. R. (2020). "Mid Lift-to-Drag Rigid Vehicle 6-DoF Performance for Human Mars Entry, Descent, and Landing: A Fractional Polynomial Powered Descent Guidance Approach." *AIAA Scitech 2020 Forum*. DOI: [10.2514/6.2020-1513](https://doi.org/10.2514/6.2020-1513).
99. <span id="ref-99"></span>**[R099]** Sagliano, M. (2018). "Pseudospectral Convex Optimization for Powered Descent and Landing." *Journal of Guidance, Control, and Dynamics* 41. DOI: [10.2514/1.g002818](https://doi.org/10.2514/1.g002818).
100. <span id="ref-100"></span>**[R100]** Sagliano, M. (2018). "Generalized hp Pseudospectral Convex Programming for Powered Descent and Landing." *2018 AIAA Guidance, Navigation, and Control Conference*. DOI: [10.2514/6.2018-1870](https://doi.org/10.2514/6.2018-1870).
101. <span id="ref-101"></span>**[R101]** Gao, D. (2026). "Obstacle avoidance guidance for Mars powered descent using convex optimization and elevation angle." *Acta Astronautica*. DOI: [10.1016/j.actaastro.2026.05.059](https://doi.org/10.1016/j.actaastro.2026.05.059).
102. <span id="ref-102"></span>**[R102]** Su, W. (2026). "Convex programming based method for stochastic optimal powered descent guidance with Wasserstein terminal cost." *Advances in Space Research*. DOI: [10.1016/j.asr.2026.01.060](https://doi.org/10.1016/j.asr.2026.01.060).
103. <span id="ref-103"></span>**[R103]** Sun, Y. (2025). "Olfactory-based powered descent guidance for Mars methane plume exploration in long-time-average wind environment." *Advances in Space Research*. DOI: [10.1016/j.asr.2024.09.043](https://doi.org/10.1016/j.asr.2024.09.043).
104. <span id="ref-104"></span>**[R104]** Briden, J. (2025). "Transformer-Based Tight Constraint Prediction for Efficient Powered Descent Guidance." *Journal of Guidance, Control, and Dynamics*. DOI: [10.2514/1.g008302](https://doi.org/10.2514/1.g008302).
105. <span id="ref-105"></span>**[R105]** Briden, J. (2025). "Tight Constraint Prediction of Six-Degree-of-Freedom Transformer-based Powered Descent Guidance." *AIAA SciTech Forum 2025*. DOI: [10.2514/6.2025-2774](https://doi.org/10.2514/6.2025-2774).
106. <span id="ref-106"></span>**[R106]** Uzun, S., Açıkmese, B., & Carson, J. M. (2025). "Sequential Convex Programming for 6-DoF Powered Descent Guidance with Continuous-Time Compound State-Triggered Constraints." *AIAA SCITECH 2025 Forum*. DOI: [10.2514/6.2025-1895](https://doi.org/10.2514/6.2025-1895).
107. <span id="ref-107"></span>**[R107]** Yang, R., & Liu, X. (2020). "Fuel-optimal powered descent guidance with free final-time and path constraints." *Acta Astronautica*. DOI: [10.1016/j.actaastro.2020.03.025](https://doi.org/10.1016/j.actaastro.2020.03.025).
108. <span id="ref-108"></span>**[R108]** Song, Z., Wang, C., Theil, S., & Seelbinder, D. (2020). "Survey of autonomous guidance methods for powered planetary landing." *Frontiers of Information Technology & Electronic Engineering* 21. DOI: [10.1631/fitee.1900458](https://doi.org/10.1631/fitee.1900458).
109. <span id="ref-109"></span>**[R109]** Jiang, X., Li, S., & Tao, T. (2016). "Innovative hazard detection and avoidance guidance for safe lunar landing." *Proceedings of the Institution of Mechanical Engineers, Part G: Journal of Aerospace Engineering* 230. DOI: [10.1177/0954410015625671](https://doi.org/10.1177/0954410015625671).
110. <span id="ref-110"></span>**[R110]** Azimov, D. M., & Bishop, R. H. (2025). "P63-64 Powered Descent Guidance." *Springer Aerospace Technology*. DOI: [10.1007/978-3-031-91088-3_8](https://doi.org/10.1007/978-3-031-91088-3_8).
111. <span id="ref-111"></span>**[R111]** Edquist, K. T., Korzun, A. M., Dyakonov, A., & Studak, J. W. (2014). "Development of Supersonic Retropropulsion for Future Mars Entry, Descent, and Landing Systems." *Journal of Spacecraft and Rockets* 51. DOI: [10.2514/1.a32715](https://doi.org/10.2514/1.a32715).
112. <span id="ref-112"></span>**[R112]** Polsgrove, T., & Dwyer-Cianciolo, A. M. (2016). "Human Mars Entry, Descent and Landing Architecture Study Overview." *AIAA SPACE 2016*. DOI: [10.2514/6.2016-5494](https://doi.org/10.2514/6.2016-5494).
113. <span id="ref-113"></span>**[R113]** Polsgrove, T., Dwyer-Cianciolo, A. M., Robertson, E. A., & Percy, T. K. (2018). "Human Mars Entry, Descent, and Landing Architecture Study: Rigid Decelerators." *2018 AIAA SPACE and Astronautics Forum and Exposition*. DOI: [10.2514/6.2018-5192](https://doi.org/10.2514/6.2018-5192).
114. <span id="ref-114"></span>**[R114]** Epp, C., Robertson, E. A., & Brady, T. (2008). "Autonomous Landing and Hazard Avoidance Technology (ALHAT)." *Proceedings - IEEE Aerospace Conference*. DOI: [10.1109/aero.2008.4526297](https://doi.org/10.1109/aero.2008.4526297).
115. <span id="ref-115"></span>**[R115]** Ivanov, T., Huertas, A., & Carson, J. M. (2013). "Probabilistic Hazard Detection for Autonomous Safe Landing." *AIAA Guidance, Navigation, and Control (GNC) Conference*. DOI: [10.2514/6.2013-5019](https://doi.org/10.2514/6.2013-5019).
116. <span id="ref-116"></span>**[R116]** Epp, C., Robertson, E., & Carson, J. M. (2014). "Real-Time Hazard Detection and Avoidance Demonstration for a Planetary Lander." *AIAA SPACE 2014 Conference and Exposition*. DOI: [10.2514/6.2014-4312](https://doi.org/10.2514/6.2014-4312).
117. <span id="ref-117"></span>**[R117]** Johnson, A. E., & Montgomery, J. F. (2008). "Overview of Terrain Relative Navigation Approaches for Precise Lunar Landing." *2008 IEEE Aerospace Conference*. DOI: [10.1109/aero.2008.4526302](https://doi.org/10.1109/aero.2008.4526302).
118. <span id="ref-118"></span>**[R118]** Johnson, A., & Ivanov, T. (2011). "Analysis and Testing of a LIDAR-Based Approach to Terrain Relative Navigation for Precise Lunar Landing." *AIAA Guidance, Navigation, and Control Conference*. DOI: [10.2514/6.2011-6578](https://doi.org/10.2514/6.2011-6578).
119. <span id="ref-119"></span>**[R119]** Steffes, S. R., Monterroza, F., Benhacine, L., & Mario, C. (2019). "Optical Terrain Relative Navigation Approaches to Lunar Orbit, Descent and Landing." *AIAA Scitech 2019 Forum*. DOI: [10.2514/6.2019-1178](https://doi.org/10.2514/6.2019-1178).
120. <span id="ref-120"></span>**[R120]** Epp, C., & Smith, T. B. (2007). "Autonomous Precision Landing and Hazard Detection and Avoidance Technology (ALHAT)." *2007 IEEE Aerospace Conference*. DOI: [10.1109/aero.2007.352724](https://doi.org/10.1109/aero.2007.352724).
121. <span id="ref-121"></span>**[R121]** Nelessen, A., Sackier, C., Clark, I. G., & Brugarolas, P. (2019). "Mars 2020 Entry, Descent, and Landing System Overview." *2019 IEEE Aerospace Conference*. DOI: [10.1109/aero.2019.8742167](https://doi.org/10.1109/aero.2019.8742167).
122. <span id="ref-122"></span>**[R122]** Lillard, R. P., & Olejniczak, J. (2017). "Human Mars EDL pathfinder study: Assessment of technology development gaps and mitigations." *2017 IEEE Aerospace Conference*. DOI: [10.1109/aero.2017.7943587](https://doi.org/10.1109/aero.2017.7943587).
123. <span id="ref-123"></span>**[R123]** Rutishauser, D., Epp, C., & Robertson, E. A. (2012). "Free-Flight Terrestrial Rocket Lander Demonstration for NASA's Autonomous Landing and Hazard Avoidance Technology (ALHAT) System." DOI: [10.2514/6.2012-5239](https://doi.org/10.2514/6.2012-5239).
124. <span id="ref-124"></span>**[R124]** Huang, X., Xu, C., Hu, J., & Li, M. (2022). "Powered-descent landing GNC system design and flight results for Tianwen-1 mission." *Astrodynamics* 6. DOI: [10.1007/s42064-021-0118-9](https://doi.org/10.1007/s42064-021-0118-9).
125. <span id="ref-125"></span>**[R125]** Parracino, G. P., Ceresoli, M., Silvestrini, S., & Lavagna, M. (2024). "Integrated Optical Terrain Relative Navigation for Autonomous Lunar Landing." *IAF Astrodynamics Symposium*. DOI: [10.52202/078368-0033](https://doi.org/10.52202/078368-0033).
126. <span id="ref-126"></span>**[R126]** Sreelekshmi, J., Navin, M. S., & Shenil, P. S. (2019). "Guidance with Supersonic Retropropulsion for Mars Pinpoint Landing." *International Conference on Computational Collective Intelligence*. DOI: [10.1109/iccci.2019.8822162](https://doi.org/10.1109/iccci.2019.8822162).
127. <span id="ref-127"></span>**[R127]** Muirhead, B., Nicholas, A., Edwards, C., & Umland, J. (2021). "Mars Sample Return Campaign Concept Architecture." DOI: [10.3847/25c2cfeb.a57d10a4](https://doi.org/10.3847/25c2cfeb.a57d10a4).
128. <span id="ref-128"></span>**[R128]** Amzajerdian, F., Gragossian, A., Brewster, P. F., & Heppler, J. M. (2025). "Long-Distance Altimetry and Terrain Relative Navigation Capabilities of a Multi-Functional Imaging Lidar for Precision Safe Landing." NASA Technical Reports Server. URL: [https://ntrs.nasa.gov/citations/20250011015](https://ntrs.nasa.gov/citations/20250011015).
129. <span id="ref-129"></span>**[R129]** Amzajerdian, F., Gragossian, A., Bulyshev, A., & Brewster, P. F. (2025). "High Resolution Terrain Sensing Lidar for Precision Navigation and Safe Landing of Space and Aerial Vehicles." NASA Technical Reports Server. URL: [https://ntrs.nasa.gov/citations/20250000978](https://ntrs.nasa.gov/citations/20250000978).
130. <span id="ref-130"></span>**[R130]** Newcomb, A. J. (2025). "Development of a SURF-Based Lunar Terrain Relative Navigation Algorithm." NASA Technical Reports Server. URL: [https://ntrs.nasa.gov/citations/20250007724](https://ntrs.nasa.gov/citations/20250007724).



## Appendix A: Variable and Data Dictionary

This appendix tabulates every variable that enters the planned regression, giving for each its symbol and name, its construct, its operational definition, its named source dataset, and the scale on which it is measured. The table is the explicit boundary statement the second anchor demands [\[24\]](#ref-24): no variable enters the estimation without a stated definition and a stated source, and the dictionary is the auditable object that guarantees the body chapters did not smuggle in an undefined quantity. The notation reproduces the prospectus exactly; the operational definitions carry forward the construction decisions documented in Chapter 4.

| Symbol | Construct | Operational definition | Source | Scale |
|---|---|---|---|---|
| \(\ln(\text{EllipseArea}_i)\) | Targeting precision (predicted) | Natural log of the three-sigma design ellipse area, computed as \(\ln(\pi \cdot a \cdot b)\) where a is the semi-major and b the semi-minor axis, after confidence-level normalization to three-sigma and any equal-axis or verbal-characterization adjustment | NTRS reconstructions [\[6\]](#ref-6), [\[7\]](#ref-7), [\[13\]](#ref-13), [\[71\]](#ref-71), [\[72\]](#ref-72), [\[73\]](#ref-73), [\[74\]](#ref-74), [\[75\]](#ref-75), [\[76\]](#ref-76), [\[77\]](#ref-77), [\[78\]](#ref-78) | Continuous; \(\ln(\text{km}^2)\) |
| \(\ln(\text{MissDistance}_i)\) | Targeting precision (achieved); robustness DV | Natural log of achieved miss distance, the great-circle distance between the pre-flight aim point and the orbitally localized landed position; flagged for whether it reflects passive dispersion or active divert | PDS localization | Continuous; ln(km) |
| \(\text{Sequence}_i\) | Experience axis (ordinal) | Mission sequence index, 1 = Viking 1 through the index of Mars 2020; alternative cumulative-landings count for the experience-axis robustness check | Mission record | Integer 1–9 (to 11 with pairs split) |
| \(\text{GuidedEntry}_i\) | Guided-lifting-entry generation | Binary indicator, 1 for MSL and Mars 2020 (closed-loop bank-angle modulation), 0 otherwise | TechPort + [\[3\]](#ref-3), [\[4\]](#ref-4), [\[75\]](#ref-75) | Binary {0,1} |
| \(\text{RangeTrigger}_i\) | Range-to-go parachute-trigger generation | Binary indicator, 1 for MSL and Mars 2020 (range-to-go Smart Chute deploy), 0 otherwise | TechPort + [\[5\]](#ref-5) | Binary {0,1} |
| \(\text{TRN}_i\) | Terrain-relative-navigation generation | Binary indicator, 1 for Mars 2020 only (Lander Vision System plus Safe Target Selection), 0 otherwise | TechPort + [\[8\]](#ref-8), [\[78\]](#ref-78) | Binary {0,1} |
| \(\text{ApproachAccuracy}_i\) | Launch-and-approach delivery accuracy (control) | Reported approach-navigation delivery dispersion at the atmospheric interface (entry-flight-path-angle or entry-point dispersion) | NTRS navigation results [\[13\]](#ref-13), [\[77\]](#ref-77) | Continuous; degrees or km |

Three dictionary entries warrant a note that the table cannot hold. First, the primary dependent variable is a constructed scalar, not a downloaded one; its provenance band (high confidence for MSL, Mars 2020, InSight, and Phoenix; moderate for Pathfinder; low for the Viking pair) travels with each value per the Chapter 4 construction recipe, and the dictionary row records the recipe, not a single number. Second, the achieved-miss-distance row carries a definitional flag that is not a nuisance but a finding-in-waiting: for an autonomous-divert mission the miss relative to a single aim point conflates targeting error with deliberate hazard avoidance [\[78\]](#ref-78), so the robustness DV is not strictly the same construct across the unguided and guided eras, and the dictionary marks that discontinuity rather than averaging over it. Third, the technology indicators are doubly sourced, coded from the TechPort infusion record and corroborated by the flight-reconstruction literature, so a single cell in the coding sheet rests on two independent witnesses, which is the safeguard against a mis-coded indicator silently driving the result.

## Appendix B: Derivations

Two derivations underlie the dissertation's quantitative claims and are set out here so the algebra is inspectable rather than asserted.

**B.1 The learning-rate reading.** The baseline specification is \(\ln(\text{EllipseArea}_i) = \beta_0 + \beta_1 \cdot \text{Sequence}_i + \epsilon_i\). Exponentiating the fitted relation gives \(\text{EllipseArea} = \exp(\beta_0) \cdot \exp(\beta_1 \cdot \text{Sequence})\), so a unit increment in the sequence index multiplies the ellipse area by \(\exp(\beta_1)\). The proportional change in ellipse area per unit of mission sequence is therefore \(\exp(\beta_1) - 1\), a negative quantity under H1 because the ellipse contracts. To express the contraction in the classical learning-curve form, where the metric improves by a constant percentage per doubling of cumulative experience, the experience axis is re-expressed in doublings: if the sequence variable measures log-base-two of cumulative landings, the per-doubling progress ratio is \(2^{\beta_1}\) and the learning rate is \(1 - 2^{\beta_1}\). The dissertation reports the per-sequence form \(\exp(\beta_1) - 1\) as primary because the population is a small count of discrete, technology-differentiated events rather than a high-volume production run, and the per-doubling form only as the bridge to the experience-curve literature [\[18\]](#ref-18), [\[19\]](#ref-19).

**B.2 Log-area construction from reported axes.** A design ellipse reported at sigma level s with semi-axes (a_s, b_s) is first normalized to three-sigma under the bivariate-normal assumption, for which linear axes scale with the sigma multiple: \(a_3 = a_s \cdot (3/s)\) and \(b_3 = b_s \cdot (3/s)\), so the three-sigma area is \(\text{Area}_3 = \pi \cdot a_3 \cdot b_3 = \pi \cdot a_s \cdot b_s \cdot (3/s)^2\). Where only a circular targeting radius r is reported, the equal-axis convention sets a = b = r and \(\text{Area} = \pi \cdot r^2\), with the equal-axis assumption flagged. The dependent variable is then \(\ln(\text{Area}_3)\), the log taken last so that the convention normalizations are applied in the physical (un-logged) units where they are defined and the log performs only the variance-stabilizing, multiplicative-to-additive transform the regression requires. Each transformation applied to each mission is recorded, so the ratio of two cells in the final series is a ratio of like quantities.

## Appendix C: Instrument and Query Details

The three datasets are public and require no credential; this appendix documents the access path for each so the series is reproducible. **NTRS** (the dependent variable and the approach-accuracy control) is queried through the public citations API at `ntrs.nasa.gov/api/citations/search`, by free-text and field search, with each reconstruction report cross-referenced to its journal DOI where one exists. The specific reconstruction records used are enumerated by corpus key in Appendix D and resolve to the NTRS citation URLs given in references [\[71\]](#ref-71)–[\[78\]](#ref-78), [\[85\]](#ref-85), [\[86\]](#ref-86), [\[128\]](#ref-128)–[\[130\]](#ref-130). **TechPort** (the three technology indicators) is read through the `techport.nasa.gov` web interface and its data-export facility; each technology's infusion mission is taken from its TechPort project record and corroborated against the flight-test and reconstruction literature. TechPort is a portal rather than a citable paper, so it carries no DOI in the corpus, and that absence is documented here rather than papered over with a phantom citation. **PDS** (the achieved-miss-distance robustness DV) is accessed through the `pds.nasa.gov` archive interface; the localized landing position is differenced against the pre-flight aim point to yield the miss distance, with the orbital-imaging provenance recorded per mission.

## Appendix D: Supplementary Tables

Two supplementary tables are specified here. Both are design-stage objects: the first is a coding sheet whose technology-indicator cells are fixed by the public record, and the second is a results template whose cells are deliberately unpopulated because the regression has not been executed.

**Table D.1: Mission-by-mission technology-indicator coding sheet.** The indicator cells are determined facts from the flight record; the dependent and control columns are placeholders to be filled from the named datasets at execution time.

| Mission (year) | Sequence | GuidedEntry | RangeTrigger | TRN | \(\ln(\text{EllipseArea})\) | ApproachAccuracy |
|---|---|---|---|---|---|---|
| Viking 1 (1976) | 1 | 0 | 0 | 0 | (to source, low conf.) | (to source) |
| Viking 2 (1976) | 2 | 0 | 0 | 0 | (to source, low conf.) | (to source) |
| Mars Pathfinder (1997) | 3 | 0 | 0 | 0 | (to source, mod. conf.) | (to source) |
| MER Spirit (2004) | 4 | 0 | 0 | 0 | (to source) | (to source) |
| MER Opportunity (2004) | 5 | 0 | 0 | 0 | (to source) | (to source) |
| Phoenix (2008) | 6 | 0 | 0 | 0 | (to source) | (to source) |
| MSL / Curiosity (2012) | 7 | 1 | 1 | 0 | (to source) | (to source) |
| InSight (2018) | 8 | 0 | 0 | 0 | (to source, control) | (to source) |
| Mars 2020 / Perseverance (2021) | 9 | 1 | 1 | 1 | (to source) | (to source) |
| Tianwen-1 (2021) | held out | 1 | 0 | 1 | external validity | external validity |

The coding sheet makes the identification structure visible at a glance. InSight (sequence 8) carries zeros across all three indicators despite its late date, which is the within-period counterfactual that breaks the time-technology collinearity [\[15\]](#ref-15); the technology indicators rise monotonically only because the levers were inserted in order, and that monotonicity is precisely the collinearity the nested specifications and the InSight case are designed to manage.

**Table D.2: Nested-specification results template (unpopulated by design).** The specified-but-empty cells display the form the executed output will take; under the design-stage guardrail no cell is filled with an estimate.

| Specification | \(\beta_1\) (Sequence) | \(\gamma_1\) (GuidedEntry) | \(\gamma_2\) (RangeTrigger) | \(\gamma_3\) (TRN) | \(\delta\) (ApproachAccuracy) | Adj. \(R^2\) | Incremental F |
|---|---|---|---|---|---|---|---|
| Baseline | (   ) | n/a | n/a | n/a | n/a | (   ) | n/a |
| + GuidedEntry | (   ) | (   ) | n/a | n/a | n/a | (   ) | (   ) |
| + RangeTrigger | (   ) | (   ) | (   ) | n/a | n/a | (   ) | (   ) |
| + TRN | (   ) | (   ) | (   ) | (   ) | n/a | (   ) | (   ) |
| + ApproachAccuracy | (   ) | (   ) | (   ) | (   ) | (   ) | (   ) | (   ) |

The template encodes the pre-registered decision rule: H1 is supported if the gamma columns fill with jointly significant negative coefficients while delta stays small and insignificant, and H0 is the standing position if the gamma terms are absorbed once the sequence trend and the approach-accuracy control are present. Reporting the incremental F and adjusted \(R^2\) as each generation enters is the cliometric decomposition move, attributing the aggregate contraction to its components one technology at a time rather than asserting them jointly [\[24\]](#ref-24). The emptiness of the cells is the chapter's honesty made structural: the dissertation delivers a frozen, falsifiable protocol and the verifiable series it will be fitted to, not a result.