# Learning Curves for Onboard Autonomy: Does Each Successive Autonomous-Operations Flight Demonstration Lower the Cost-to-Field of the Next?

**A Doctoral Dissertation**

**Candidate:** JPL_AUTONOMY_EDL_01
**Research category:** Autonomous Systems and Robotics
**Method:** Quantitative. Wright/Henderson log-log experience-curve regression with two-way fixed effects.
**Theoretical anchors:** W. Brian Arthur (increasing returns, learning effects, path dependence) and Joel Mokyr (propositional versus prescriptive useful knowledge, codification as a slope moderator).
**Stage:** Design-stage artifact. No coefficient is fitted on the full dataset; all expected results are explicitly illustrative.
**Date:** 2026-06-15


## Abstract

Stewardship of public investment in space autonomy rests on a belief long held but never measured: that the cost of qualifying onboard spacecraft autonomy for flight falls as the technology matures. That belief deserves testing rather than trust, and this dissertation tests it. It specifies and operationalizes a test of whether the recurring engineering cost to qualify a class of onboard autonomy capability declines as a power-law function of cumulative flight-demonstrated heritage. The unit of analysis is the capability-class development episode: a mission or technology demonstration that fields a defined autonomy capability such as onboard planning, autonomous science target selection, autonomous navigation, autonomous fault detection and recovery, or autonomous entry-descent-and-landing hazard handling. The contribution is a single falsifiable claim. The null hypothesis holds that per-episode autonomy development cost is flat with respect to the cumulative count of prior flight demonstrations within a capability class. The alternative holds that per-episode cost declines along a log-log experience curve with a statistically detectable negative slope after controlling for capability class and decade. The estimator is an ordinary-least-squares regression of the logarithm of normalized development cost on the logarithm of cumulative flight-demonstrated heritage, with capability-class and decade fixed effects, fit to a constructed panel assembled from NASA TechPort project records and technology-readiness histories, NTRS autonomy demonstration reports for Deep Space 1 Remote Agent, the EO-1 Autonomous Sciencecraft Experiment, AEGIS, and Ingenuity, and NASA Instrument Cost Model-class parametric development-cost estimates. The work is presented at the design stage: the data construction, model specification, identification strategy, and threats to validity are fully developed, and an explicitly non-empirical specification of expected results shows how a finding would be read against the hypotheses. No fitted coefficients are reported, because panel assembly and cost normalization are not yet complete. The two anchor frameworks sharpen the test rather than decorate it. Arthur identifies the learning-effect mechanism the slope measures and warns that the realized path is contingent; Mokyr predicts that the slope steepens where the underlying knowledge is mature, codified, and reusable. The intended outcome is a defensible measurement of the autonomy experience-curve slope, or a credible failure to reject the flat-cost null. Either result informs NASA and JPL technology-investment sequencing and the build-or-wait decision that the agency currently answers by assertion.

**Keywords:** onboard autonomy, experience curve, learning rate, heritage reuse, technology readiness level, NASA, Jet Propulsion Laboratory, portfolio sequencing, two-way fixed effects, design-stage measurement.


## Table of Contents

- Abstract
- Table of Contents
- List of Tables
- Chapter 1. Introduction
  - 1.1 The problem: autonomy as an assumed but unmeasured heritage asset
  - 1.2 The gap: two unjoined literatures
  - 1.3 The single falsifiable contribution
  - 1.4 Why it matters for NASA, JPL, and the named stakeholders
  - 1.5 Scope, design-stage posture, and delimitations
  - 1.6 Definitions of key terms
  - 1.7 Roadmap of the dissertation
- Chapter 2. Theoretical Framework
  - 2.1 The Wright experience curve and organizational learning-by-doing
  - 2.2 The experience curve as a validated forecasting object
  - 2.3 Wright versus Moore and the necessity of a time control
  - 2.4 Anchor 1, W. Brian Arthur: increasing returns, learning effects, and path dependence
  - 2.5 Anchor 2, Joel Mokyr: propositional and prescriptive knowledge as a slope moderator
  - 2.6 Heterogeneity of learning rates and the formal case for capability-class fixed effects
  - 2.7 The conceptual model and chapter synthesis
- Chapter 3. Literature Review
  - 3.1 Onboard planning and scheduling: the origin class
  - 3.2 Autonomous science target selection: the cleanest heritage pair
  - 3.3 Autonomous navigation: the deepest heritage chain
  - 3.4 Autonomous fault detection, isolation, and recovery
  - 3.5 Autonomous entry, descent, and landing hazard handling
  - 3.6 The codified-knowledge substrate: the Mokyr moderator
  - 3.7 Synthesis: the gap stated and the propositions that follow
- Chapter 4. Data and Measurement
  - 4.1 The four named datasets
  - 4.2 Unit of analysis and the capability-class taxonomy
  - 4.3 The dependent variable: three-layer cost normalization and reliability flags
  - 4.4 The independent variable: the forward-only cumulative-heritage counting rule
  - 4.5 Fixed effects, the maturation covariate, and the parametric cost-model basis
  - 4.6 Coverage window, validation, ethics and access, and the four material limitations
- Chapter 5. Research Design and Identification
  - 5.1 The estimator: ordinary least squares on the log-log two-way-fixed-effects form
  - 5.2 Identification: what beta is identified off, and the role of each fixed-effect set
  - 5.3 The three pre-registered robustness specifications
  - 5.4 Threats to validity and the design responses
  - 5.5 External validity and the path-dependence caution
  - 5.6 Summary and implications of the design
- Chapter 6. Analysis Plan and Expected Results
  - 6.1 Problem frame for this chapter
  - 6.2 The estimation procedure
  - 6.3 Pre-analysis checks and the fixed decision rule
  - 6.4 Illustrative, explicitly non-empirical expected results
  - 6.5 Falsification, the specified result tables, and reproducibility
- Chapter 7. Discussion
  - 7.1 If H1 holds: the slope as a planning parameter
  - 7.2 If H0 holds: the rationale shifts from cost reduction to capability value
  - 7.3 Rival explanations and the responses to them
  - 7.4 Effect heterogeneity across classes: the Mokyr codification moderator
  - 7.5 Path dependence and the external-validity bound; the launch-cadence analogue
  - 7.6 Synthesis: the argument carried to the decision
- Chapter 8. Conclusion
  - 8.1 The contribution restated: one slope, or one credible failure to find one
  - 8.2 Both outcomes are decision-relevant: the symmetric value of the design
  - 8.3 The anchors sharpened the test
  - 8.4 From design to execution: the remaining build steps and reproducibility commitments
  - 8.5 Closing
- References
- Appendix A. Variable and Data Dictionary, with the Capability-Class Taxonomy
- Appendix B. Derivation of the Three-Layer Cost Normalization
- Appendix C. The Forward-Only Cumulative-Heritage Counting Log
- Appendix D. Pre-Registration of Specifications and the Fixed Decision Rule
- Appendix E. Supplementary Tables and the Triaged Literature
- Appendix F. Instrument and Data-Source Query Details


## List of Tables and Figures

- Table 3.1. Onboard planning and scheduling heritage chain (Chapter 3).
- Table 3.2. Autonomous navigation heritage chain (Chapter 3).
- Table 3.3. Heritage record versus cost measurement, by capability class (Chapter 3).
- Table 4.1. Operationalization of every variable in the estimating equation (Chapter 4).
- Table E.1. Episode inventory schema, design-stage template (Appendix E).
- Table E.2. Capability-class summary schema, design-stage template (Appendix E).
- Table E.3. Specification-results template, all cells empty at design stage (Appendix E).

No empirical figures are populated. Diagnostic plots of log cost against log heritage by capability class are specified in Chapter 6 and are produced only in the build phase. This dissertation contains no fitted-coefficient figures, consistent with its design-stage posture.



# Chapter 1. Introduction

## 1.0 Overview and central claim

An agency that commits public funds to autonomous spaceflight owes its sponsors an honest account of what each mission's hard-won experience saves the next. For more than two decades, the engineers who build and fly autonomous spacecraft have served a quiet conviction: that what is learned on one mission lightens the burden of the next. That conviction has guided real investment, and it merits the respect of being measured rather than assumed. This dissertation delivers the first fitted measurement of the experience-curve slope of onboard-autonomy qualification cost on cumulative flight-demonstrated heritage: a single number, or a credible failure to find one, that converts the heritage-lowers-cost assumption held inside NASA and the Jet Propulsion Laboratory from an assertion into a planning parameter. That sentence is the thesis of the chapter, and the rest of the chapter develops it rather than building toward it. The claim is not that autonomy heritage lowers cost. The claim is that the size of that effect is a measurable quantity no one has yet measured, that a standard and well-validated estimator can measure it, and that measuring it is decision-relevant for an agency that currently sequences its autonomy investments on the strength of an unquantified belief.

The chapter is organized to make that thesis legible and defensible. Section 1.1 states the problem in full: how the portfolio reasons about heritage today, what a measured learning rate would let it reason instead, the gap between those two states, and what is lost by leaving that gap open. Section 1.2 develops the gap as the failure of two mature literatures to meet. Section 1.3 states the single falsifiable contribution as the null and alternative hypotheses, carried verbatim from the approved prospectus so that the contribution this chapter introduces is identical to the contribution the later chapters test. Section 1.4 establishes significance for NASA, for JPL, and for the named stakeholders who make autonomy-portfolio decisions, anchoring the abstract experience-curve construct to a concrete build-or-wait choice. Section 1.5 fixes the scope and delimitations and is explicit about the design-stage posture of the work, the single most important honesty commitment in the dissertation. Section 1.6 defines the key terms so that the rest of the document can use them without re-deriving them. Section 1.7 maps the dissertation chapter by chapter so a reader knows where each part of the argument lives. The convention throughout is to interpret each source for what its convergence with the others means for the argument, not to list it, and to state the confidence attached to each major claim together with the evidence that would raise or lower it.

## 1.1 The problem: autonomy as an assumed but unmeasured heritage asset

### 1.1.1 The current state

NASA and the Jet Propulsion Laboratory invest in onboard autonomy because autonomy raises the science and operational return of missions that operate under long communication delays and constrained ground contact. The historical record of that investment is unusually clean, because the signal events are individually documented in the primary engineering literature. The Deep Space 1 Remote Agent Experiment in 1999 was the first occasion on which an artificial-intelligence agent controlled a NASA spacecraft, integrating onboard planning, plan execution, and model-based fault recovery into a single flight experiment [\[12\]](#ref-12). The Earth Observing One Autonomous Sciencecraft Experiment moved onboard event detection and replanning out of the laboratory and into routine Earth-observation operations after 2003, improving the timeliness and quality of returned science data [\[24\]](#ref-24). The Autonomous Exploration for Gathering Increased Science capability, AEGIS, placed autonomous science target selection first on the Opportunity rover and then on the Curiosity ChemCam instrument, removing the ground-in-the-loop latency that had limited narrow-field instrument targeting [\[34\]](#ref-34), [\[40\]](#ref-40). The Mars 2020 mission flew autonomous surface navigation at scale on the Perseverance rover and demonstrated the first powered autonomous flight on another planet with the Ingenuity helicopter [\[99\]](#ref-99), [\[7\]](#ref-7).

Inside the agency, each of these episodes is treated not merely as a past success but as a heritage asset. The operative belief is that once a capability has flown, the next mission needing that capability can field it more cheaply, because the design patterns, the verification and validation approaches, and the flight-software components developed for the first demonstration are reusable on the second. This belief is not idle. It is the working assumption that justifies the sequence of investments in the Autonomous Systems and Robotics portfolio, and it is the implicit logic behind technology-readiness-level accounting, which treats the maturation of a capability across demonstrations as a movement up a ladder of decreasing risk and, by tacit extension, decreasing cost. The synthesis review of space autonomy by Gao and Chien [\[42\]](#ref-42) states the belief plainly in its survey of the field: flown heritage lowers the barrier to the next demonstration. The current state, then, is a portfolio governed by a heritage-reuse logic that is everywhere assumed and nowhere measured.

### 1.1.2 The desired state

The desired state is a cost-anchored, falsifiable learning rate for each autonomy capability class. Concretely, it is a number, or a small set of class-specific numbers, that answers the question: by what fraction does the recurring engineering cost to qualify a class of onboard autonomy for flight fall each time the cumulative count of prior in-class flight demonstrations doubles? A program office that possessed such a number could reason about the dollars attached to advancing a capability, not only about its readiness level. It could decide whether an intermediate, lower-stakes demonstration would buy down enough cost to bring a target capability within the cap of a future cost-constrained mission, and it could estimate roughly how many such demonstrations would be required. The desired state is therefore the replacement of an article of faith with a measured slope, accompanied by a confidence interval that honestly bounds what the measurement can and cannot support.

### 1.1.3 The gap

The gap between the current and desired states is the absence of any fitted experience-curve treatment of onboard-autonomy qualification cost. No published study fits a Wright-type log-log model of capability-class autonomy development cost on cumulative flight-demonstrated heritage. The gap is not that the question is uninteresting or that the data are wholly unavailable; it is that the two bodies of knowledge that would have to be joined to answer it have developed in isolation. Section 1.2 develops this gap in detail, because the gap is the load-bearing justification for the dissertation and deserves a full treatment rather than an assertion.
### 1.1.4 The consequence

Leaving the gap open means that portfolio sequencing and build-or-wait decisions in the Autonomous Systems and Robotics portfolio continue to be made by assertion. When a future mission needs a capability that is presently too expensive to qualify within its cost cap, the decision whether to fund an intermediate demonstration that would lower the eventual cost is made today without a quantitative basis. It rests on the qualitative confidence of experienced engineers that heritage helps. That is a real form of knowledge, but not a measured one, and it cannot distinguish a steep cost-decline curve from a shallow one, or a genuine learning effect from a coincident decline in the price of computing. The cost of this is not a hypothetical mis-investment in a single case. It is the systematic absence of a quantitative complement to the technology-readiness-level scale, which tracks the maturity of a capability but not the cost of advancing it, and which is known to be ordinal and non-monetary [\[70\]](#ref-70), [\[81\]](#ref-81). The agency can say a capability has moved from readiness level four to readiness level six; it cannot say what that movement cost, nor what the next movement will cost, in a way that is anchored to a measured rate.

The problem can be summarized in a single line, with its supporting structure made explicit. NASA and JPL govern a multi-decade autonomy-investment program with a heritage-lowers-cost logic that has never been measured, and the cost of that omission is decision-making by assertion. The demonstration record is documented and treated as heritage [\[12\]](#ref-12), [\[24\]](#ref-24), [\[34\]](#ref-34), [\[40\]](#ref-40), [\[99\]](#ref-99), [\[7\]](#ref-7), and the recurring heritage claim is stated explicitly in the field's own synthesis [\[42\]](#ref-42). A belief that drives resource-allocation decisions and is treated as a quantity (cheaper next time) is a quantity that ought to be measured before it is relied upon. The experience-curve tradition in the economics of technology has established for nearly a century that this kind of cost-quantity relationship is measurable, and that beliefs about it are frequently wrong in magnitude even when right in direction [\[94\]](#ref-94). The claim here is that the effect is unmeasured, not that it is absent; the direction may well be as assumed. If engineering judgment about heritage were already precise and calibrated, a formal measurement would add little. The dissertation's premise is that it is not, which Section 1.4 substantiates by showing that the decisions at stake turn on magnitudes that judgment alone cannot supply. Confidence in this problem statement is **high**, because every premise rests on primary documentation rather than on inference, and the only contestable element is whether the unmeasured belief is precise enough in practice to make formal measurement redundant, which is itself an empirical question the dissertation is designed to inform.

## 1.2 The gap: two unjoined literatures

The gap identified in Section 1.1.3 is best understood as the non-intersection of two mature and internally rigorous literatures, each of which supplies exactly half of what an answer requires and neither of which supplies the other half.

The first is the experience-curve literature in the economics of technology. Since the airframe cost-quantity studies of the 1930s and 1950s, unit production cost has been observed to fall as a power-law function of cumulative output across many classes of manufactured goods. Thompson [\[94\]](#ref-94) reviews the long history of this regularity and the accumulated evidence that organizational learning-by-doing, not merely static economies of scale, drives it. The modern treatment of the experience curve is not anecdotal but statistical: it is a forecasting object with a quantifiable error distribution. Nagy and colleagues [\[79\]](#ref-79) compare functional forms across many technologies and find that the Wright power law, in which cost is a function of cumulative output, performs well as a predictor. Farmer and Lafond [\[35\]](#ref-35) quantify the predictive accuracy of experience curves and characterize the distribution of their forecast errors, and Lafond and colleagues [\[61\]](#ref-61) extend this to full distributional forecasts. The lesson this dissertation takes from that literature is not a result but a method: the log-log specification is the standard and best-validated functional form, the learning rate is the parameter of interest, and the central empirical hazard is that an observed cost decline reflects a confounder, such as scale, a time trend, or input-price change, rather than cumulative experience. What this literature does not contain is any application to the non-recurring engineering cost of qualifying a software-intensive capability for flight. It is, almost in its entirety, a literature about the unit cost of manufactured hardware.

The second is the spacecraft-autonomy and technology-maturation literature. It documents individual autonomy demonstrations in fine engineering detail, it synthesizes the state of space autonomy across decades [\[42\]](#ref-42), and it analyzes the technology-readiness-level construct used to track maturation, including that construct's known limitations as an ordinal, non-cost measure [\[70\]](#ref-70), [\[81\]](#ref-81). This literature is rich precisely where the first is silent: it knows what autonomy heritage is, how it accrues, and where it is reused. But it describes heritage qualitatively. It does not estimate whether heritage lowers cost along a measurable curve, and it offers no fitted slope, no learning rate, and no confidence interval on the cost effect of accumulated demonstrations.

Read together, the two literatures form the core of the gap argument. Each is necessary and neither is sufficient. The experience-curve literature supplies a validated measurement instrument with no application to autonomy; the autonomy literature supplies a documented heritage chronology with no measurement instrument applied to it. The contribution of this dissertation is to bring the instrument to the chronology: to treat the autonomy demonstration record as the cumulative-output axis of a Wright curve and the qualification cost of each demonstration as the unit cost, and to fit the model that the first literature has validated to the data that the second literature has assembled. Confidence that this gap is genuine, rather than an artifact of an incomplete search, is **high** for the absence of a fitted autonomy experience curve, which a focused corpus assembly did not find, and **moderate** as a permanent claim, because absence of evidence in a search is weaker than a proof of non-existence; the confidence would fall if a prior fitted study were located and would rise as the corpus search broadens without finding one.

## 1.3 The single falsifiable contribution

The dissertation tests exactly one claim, and the discipline of the entire document is that it never quietly expands that claim into several. The claim is stated as a pair of hypotheses, carried here verbatim from the approved prospectus so that the contribution introduced in this chapter is the identical contribution tested in the later chapters.

- **H0 (null):** Per-episode autonomy development cost is flat with respect to the cumulative count of prior flight demonstrations within a capability class. The experience-curve slope coefficient is statistically indistinguishable from zero.
- **H1 (alternative):** Per-episode autonomy development cost declines along a log-log experience curve as cumulative flight-demonstrated heritage accumulates. The slope coefficient is negative and statistically significant after controlling for capability class and decade.

The estimating equation that operationalizes this pair is the two-way fixed-effects log-log regression

\[ \ln(\text{Cost}_{icd}) = \alpha + \beta \, \ln(\text{CumHeritage}_{icd}) + \gamma_c + \delta_d + \epsilon_{icd} \qquad\qquad (1) \]

for episode \(i\) in capability class \(c\) and decade \(d\), where \(\text{Cost}_{icd}\) is normalized development cost, \(\text{CumHeritage}_{icd}\) is cumulative within-class flight-demonstrated heritage, \(\gamma_c\) are capability-class fixed effects, \(\delta_d\) are decade fixed effects, and \(\epsilon_{icd}\) is the error term. The coefficient \(\beta\) is the experience-curve slope and is the single parameter on which the hypotheses turn. The implied learning rate, reported for interpretability, is one minus two raised to the power \(\beta\). A negative \(\beta\) whose confidence interval excludes zero supports H1; a \(\beta\) indistinguishable from zero supports H0.

The contribution is the measurement itself: a fitted slope, a confidence interval, and an explicit accept-or-reject decision on H0 under a decision rule that is fixed before estimation. The integrity of the work depends on the contribution being falsifiable in the strict sense, not merely supportable. A fitted slope that is zero or positive, or whose confidence interval contains zero across the baseline and the pre-registered robustness specifications, rejects H1 and affirms the flat-cost null. The dissertation is therefore constructed so that a null result is a real and reportable finding, not a failed experiment. This symmetry is the chapter's most consequential design choice, and Section 1.4 explains why both branches of the disjunction are decision-relevant rather than only the affirmative one.

The causal logic underneath H1 is named explicitly rather than left as a bare correlation, because an experience curve that cannot name its mechanism measures only an association. Cumulative flight heritage within a capability class is the cause. As demonstrations accumulate, they leave behind codified, reusable design patterns, validated verification approaches, and flight-software components. That reusable stock lowers the recurring non-recurring engineering effort required to qualify the next same-class capability. That cost reduction, if it is regular, appears as a measurable negative log-log slope. And the slope, once measured, becomes a planning parameter for autonomy-portfolio sequencing and build-or-wait decisions. Each link is a claim that the later chapters defend; the sequence is stated here so the reader knows that the regression is not a fishing expedition but the empirical shadow of a specified mechanism. Confidence in the direction of the mechanism is **moderate to high**, grounded in Arthur's identification of learning effects as a genuine source of increasing returns [\[3\]](#ref-3) and in the broad cross-technology validation of the Wright form [\[79\]](#ref-79), [\[94\]](#ref-94); confidence in the magnitude is deliberately **low** at the design stage, because no coefficient has been fitted and the magnitude is precisely what the study exists to measure.

## 1.4 Why it matters for NASA, JPL, and the named stakeholders

The significance of the measurement is concrete, and it is best shown through the decision it informs rather than asserted in the abstract.

### 1.4.1 The build-or-wait decision

The sharpest use of a measured autonomy learning rate is the build-or-wait decision faced by a portfolio office. Suppose a future mission requires an onboard autonomy capability that, on current evidence, would cost more to qualify than the mission's cost cap allows. The portfolio has two options. It can wait, hoping that some other mission will fund a demonstration of that capability and so reduce the cost by the time the future mission needs it. Or it can build, funding an intermediate, lower-stakes demonstration now specifically to buy down the cost of the eventual qualification. The choice between these depends entirely on a quantity that no one currently possesses: the slope of the cost-decline curve for that capability class. The build-or-wait decision cannot be made rationally without a measured slope, because the decision is a comparison between the cost of an intermediate demonstration and the cost reduction it would induce, and the cost reduction is the slope times the change in log cumulative heritage. A decision whose optimum is a function of an unknown parameter is, in the absence of that parameter, a guess. This is the standard logic of experience-curve-based capacity and investment planning in the economics of technology, where the learning rate is the input that determines whether early investment pays for itself [\[94\]](#ref-94), [\[35\]](#ref-35). A measured slope improves the decision; it does not automate it, because the slope is an average within a class and a specific capability may sit above or below the class curve. One might object that experienced engineers already make this call well without a number. The answer is that engineering judgment can supply direction but not the magnitude on which the optimum turns, and the launch-cadence literature provides a concrete space example in which each flight lowers the cost of the next by an amount that only measurement, not intuition, could supply [\[91\]](#ref-91), [\[37\]](#ref-37). Confidence that the decision is real and that a slope would improve it is **high**; confidence about how often the decision arises in practice is **moderate** and would rise with a survey of portfolio decisions, which is outside the present scope.

### 1.4.2 The technology-readiness-level complement

The second significance is the provision of a cost-anchored complement to the technology-readiness-level scale. The readiness-level construct is the agency's standard maturity tracker [\[70\]](#ref-70), but it is ordinal and non-monetary, and its shortcomings as a management instrument are documented [\[81\]](#ref-81). It can record that a capability has advanced from one readiness level to the next, but it cannot attach a cost to that advance, and so it cannot tell a program what the next advance will cost. A measured experience-curve slope supplies the missing dimension: it lets a program reason about the dollars attached to advancing a capability across the demonstration sequence, not only about the readiness level reached. The two instruments are complementary rather than competing; the readiness level tracks where a capability is, and the slope tracks what it costs to move it. The most recent and most capable demonstrations, autonomous navigation at scale on Perseverance and powered autonomous flight on Ingenuity [\[99\]](#ref-99), [\[7\]](#ref-7), are exactly the kind of high-investment advances for which a cost-anchored complement would be most valuable, because the dollars at stake in advancing such capabilities are large enough that an unquantified maturity ladder is an inadequate basis for planning.

### 1.4.3 The symmetric value of the result and the named stakeholders

The significance does not depend on H1 holding. If the slope is measurably negative, it becomes a planning parameter, and the heritage-reuse logic that justifies much autonomy investment is vindicated and quantified. If the slope is indistinguishable from zero, the finding is equally consequential in a different direction: the portfolio rationale must then shift from cost reduction to capability value, and the agency should not rely on heritage-driven cost decline to justify autonomy investment that is better justified by the operational return of the capability itself. Both outcomes inform the same decision-makers. The stakeholders for whom this measurement matters are the Autonomous Systems and Robotics portfolio leadership who set investment sequence, the mission formulation teams who must decide whether a needed capability fits within a cost cap, and the technology-maturation managers who currently account for progress in readiness levels alone. For each of these, a measured slope, or a credible demonstration that no reliable slope exists, changes how a real decision is reasoned. The logic of the whole dissertation can be stated compactly here. The question is genuine and unanswered (Section 1.1), and it carries weight because it governs resource-allocation decisions of consequence (Section 1.4.1). The proposed measurement reaches the causal mechanism rather than a surface correlation (Section 1.3), and it does so with the best-validated estimator the relevant literature offers, developed in the later design chapters. Whatever it concludes is bounded honestly by the design-stage posture and the pre-registered robustness checks (Section 1.5). Each later chapter takes up one link in that argument and discharges it in turn.

## 1.5 Scope, design-stage posture, and delimitations
### 1.5.1 Scope

The study examines onboard autonomy flight demonstrations conducted by NASA and the Jet Propulsion Laboratory, from the Deep Space 1 Remote Agent Experiment in 1999 through the Mars 2020 demonstrations and their immediate successors. The window is chosen deliberately. It is the period over which onboard autonomy moved from isolated single experiments to routine mission elements, and it is therefore the period over which an experience curve, if one exists, should be detectable. The unit of analysis is the capability-class development episode, defined in Section 1.6. The capability classes under study are onboard planning and scheduling, autonomous science target selection, autonomous surface or in-flight navigation, autonomous fault detection and recovery, and autonomous entry-descent-and-landing hazard handling.

### 1.5.2 The design-stage posture

The central honesty commitment of this dissertation is stated here without qualification: the work is presented at the design stage, and no fitted coefficient from the full dataset is reported as an executed result anywhere in the document. The panel assembly and the cost normalization are not yet complete. The data construction, the model specification, the identification strategy, and the threats to validity are fully developed, while every statement about a result is framed as expected or illustrative and is labeled as such. Where the document shows what a result would look like, it does so to demonstrate how the measurement would be read against the hypotheses, not to claim that the measurement has been made. This posture is not a weakness to be apologized for. It is the correct register for a dissertation whose contribution is a fully specified, falsifiable measurement design, and whose execution is the next phase of work. A reader should hold the document to the standard of design rigor, not the standard of an executed empirical result, because the latter is out of scope and is never claimed.

### 1.5.3 Delimitations

Four delimitations bound the study. First, it does not estimate a universal autonomy learning rate. The fixed-effects design estimates an average within-class slope, and the anchor frameworks developed in Chapter 2 give principled reasons to expect that the realized slope is contingent on historical investment sequence rather than a constant of nature [\[5\]](#ref-5). Second, it does not extend to commercial autonomy, terrestrial autonomy, or mission classes outside the demonstration window; the external validity of any finding is confined to NASA and JPL deep-space and planetary autonomy. Third, it does not treat the technology-readiness-level history as a cost proxy. The readiness level is included only as a maturation covariate in a robustness specification, because its ordinal, non-cost character makes it unsuitable as a direct measure of the dependent variable [\[70\]](#ref-70), [\[81\]](#ref-81). Fourth, the study omits any architecture-traceability framing. This is an econometric measurement study, not a systems-architecture study; no real capability, system, or data-service exchange is the subject of the work, and the document therefore does not force architecture vocabulary onto a quantitative cost contribution. The single permitted conceptual exception, developed in the discussion chapter, is that the fitted slope may be described in plain prose as an input to a portfolio decision, which is a statement about how a number would be used, not an architecture chain.

## 1.6 Definitions of key terms

The following terms carry fixed meanings throughout the dissertation. They are defined here so that later chapters may use them without re-derivation, and the definitions hold consistently throughout.

**Episode.** The unit of analysis. An episode is a single mission or technology demonstration that fields one defined onboard autonomy capability. Each episode is assigned to exactly one primary capability class; an episode that fields more than one capability is either split into separate episodes or assigned to its dominant capability with an explicit flag. The episode, not the mission, is the row of the panel, because the same mission may contribute episodes to more than one capability class.

**Capability class.** A coherent functional category of onboard autonomy. The five classes under study are onboard planning and scheduling, autonomous science target selection, autonomous surface or in-flight navigation, autonomous fault detection and recovery, and autonomous entry-descent-and-landing hazard handling. The class is the level at which heritage is counted and at which the fixed effects absorb baseline differences in cost and intrinsic difficulty.

**Cost (dependent variable).** The recurring engineering cost to qualify the capability for flight, expressed in constant-year dollars and normalized by capability scope on a NASA Instrument Cost Model-class basis. The normalization proceeds in three layers: extract the autonomy-specific non-recurring engineering portion from project documentation where possible; impute it via a NICM-class parametric estimate where it is not separately reported [\[89\]](#ref-89); and deflate every figure to constant-year dollars, recording for each observation which layer produced the figure and an associated reliability flag. The variable enters the model as the natural logarithm of normalized cost.

**CumHeritage (independent variable).** Cumulative flight-demonstrated heritage. For each episode, the count of prior flight demonstrations within the same capability class that reached flight operation before the episode's development-start date. The counting rule is forward-only: a prior demonstration counts toward an episode's heritage stock only if it reached flight before that episode's development began, because heritage that arrives later cannot have lowered the earlier development's cost. The first observation in each class has its cumulative count set to one before taking the logarithm, so that the first demonstration is the origin of the curve. The variable enters the model as the natural logarithm of cumulative heritage.

**Capability-class fixed effects (\(\gamma_c\)).** Indicator variables for each capability class, absorbing time-invariant differences in baseline cost and intrinsic difficulty across classes.

**Decade fixed effects (\(\delta_d\)).** Indicator variables for the decade of development start, absorbing economy-wide and agency-wide time trends in cost, tooling, and computing that affect all classes in common.

**Maturation covariate.** The technology-readiness-level at development start, drawn from NASA TechPort, included only in a robustness specification to separate the heritage effect from the simpler effect of beginning a project at a higher maturity.

**Experience-curve slope (\(\beta\)).** The coefficient on the logarithm of cumulative heritage in the estimating equation. It is the single parameter on which the hypotheses turn. The implied learning rate, reported for interpretability, is one minus two raised to the power \(\beta\).

## 1.7 Roadmap of the dissertation

The dissertation is organized so that each chapter discharges one part of the argument introduced in this chapter, and the reader can locate any element of the contribution by chapter.

**Chapter 2, Theoretical Framework,** establishes the experience curve as a measurement object and develops the two anchor lenses that sharpen the test. It sets out the Wright form and the evidence for organizational learning-by-doing [\[94\]](#ref-94), the validation of the log-log form as a forecasting object [\[79\]](#ref-79), [\[35\]](#ref-35), [\[61\]](#ref-61), and the need for a time control to separate genuine learning from a coincident time trend. It then develops the two anchors around their named causal mechanisms: W. Brian Arthur on increasing returns, learning effects, and path dependence [\[3\]](#ref-3), [\[4\]](#ref-4), [\[5\]](#ref-5), who identifies the learning-effect mechanism the curve measures and warns that the realized path is contingent; and Joel Mokyr on propositional versus prescriptive useful knowledge, who supplies the testable moderator that codification and reuse infrastructure should steepen the cost-decline curve.

**Chapter 3, Literature Review,** carries the empirical spine. Organized by capability class, it establishes the demonstration chronology that feeds the cumulative-heritage variable, tracing onboard planning and scheduling from Remote Agent to the EO-1 Autonomous Sciencecraft Experiment [\[12\]](#ref-12), [\[24\]](#ref-24), [\[23\]](#ref-23), autonomous science target selection through the AEGIS sequence [\[34\]](#ref-34), [\[40\]](#ref-40), autonomous navigation through Perseverance and Ingenuity [\[99\]](#ref-99), [\[7\]](#ref-7), and the codified-knowledge substrate that the Mokyr moderator predicts should matter. It converts the qualitative heritage-lowers-cost claim that recurs in the field's synthesis [\[42\]](#ref-42) into the testable quantity the dissertation measures.

**Chapter 4, Data and Measurement,** specifies the four named sources, the unit of analysis, and the variable construction, with the three-layer cost-normalization procedure developed as a construct-validity argument. It is explicit that the dependent variable is the most delicate construct in the study and that the normalization makes the measurement error auditable rather than eliminating it [\[89\]](#ref-89), [\[70\]](#ref-70), [\[81\]](#ref-81).

**Chapter 5, Research Design,** defends the estimator choice and makes the identification strategy the heart of the argument. It explains why ordinary least squares on the log-log two-way fixed-effects form is preferred to a non-linear or Bayesian estimator in a small panel, what the slope is identified off, the three pre-registered robustness specifications, and the full four-way threats-to-validity matrix covering internal, external, construct, and statistical-conclusion validity [\[14\]](#ref-14), [\[61\]](#ref-61), [\[5\]](#ref-5).

**Chapter 6, Analysis Plan,** lays out the five-step estimation procedure, the two mandatory pre-analysis checks on fixed-effects feasibility and small-panel influence, the fixed decision rule on the null, and the explicitly non-empirical illustrative expected-results block, together with a statement of what a falsifying result looks like and the reproducibility commitments that make the eventual measurement auditable [\[79\]](#ref-79), [\[61\]](#ref-61).

**Chapter 7, Discussion,** works through the implications under each branch of the disjunction, the three rival explanations for a negative slope and the responses to them, the effect heterogeneity that the Mokyr framework predicts as the most policy-relevant secondary finding, the path-dependence bound on external validity that Arthur's framework imposes, and the launch-cadence analogue as a concrete example of the learning-effect mechanism [\[91\]](#ref-91), [\[37\]](#ref-37), [\[5\]](#ref-5).

**Chapter 8, Conclusion,** restates the central finding, summarizes the contribution as a single measurement under either branch of the disjunction, restates the design-stage honesty, names the steps required to convert the design into execution, and closes the argument.
The chapter has now defended its opening thesis. The problem is a portfolio governed by an unmeasured heritage-lowers-cost logic; the gap is the failure of the experience-curve and autonomy literatures to meet; the contribution is a single falsifiable slope measured by a validated estimator on a constructed panel; the significance is a real build-or-wait decision and a cost-anchored complement to the readiness-level scale; the scope is bounded and the design-stage posture is explicit; the terms are fixed; and the remaining chapters are mapped. Chapter 2 supplies the theoretical foundation that makes the slope a measurement of a named mechanism rather than a bare correlation.



# Chapter 2. Theoretical Framework

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

This chapter delivers the conceptual model that the empirical work of this dissertation will test, and it states that model in one sentence at the outset. The recurring engineering cost to qualify a class of onboard spacecraft autonomy capability for flight is predicted to fall as a power-law function of the cumulative count of prior flight demonstrations within that capability class, because each demonstration codifies reusable design patterns, verification approaches, and flight-software components that lower the cost of the next demonstration of the same class. That predicted decline is identified cleanly only when calendar-time trends and intrinsic class difficulty are absorbed by fixed effects. Everything in this chapter serves to justify, qualify, and bound that one sentence. The estimating equation that operationalizes it is fixed and carried unchanged from the prospectus:

\[ \ln(\text{Cost}_{icd}) = \alpha + \beta \, \ln(\text{CumHeritage}_{icd}) + \gamma_c + \delta_d + \epsilon_{icd} \qquad\qquad (1) \]

for episode \(i\) in capability class \(c\) and decade \(d\), where \(\text{Cost}_{icd}\) is normalized development cost, \(\text{CumHeritage}_{icd}\) is cumulative within-class flight-demonstrated heritage, \(\gamma_c\) are capability-class fixed effects, \(\delta_d\) are decade fixed effects, and \(\epsilon_{icd}\) is the error term. The single parameter \(\beta\) is the experience-curve slope, and the implied learning rate is \(1 - 2^{\beta}\). The two hypotheses turn on \(\beta\): H0 holds that \(\beta\) is statistically indistinguishable from zero (cost is flat with respect to heritage); H1 holds that \(\beta\) is negative and statistically significant after controlling for capability class and decade.

The theoretical problem this chapter addresses begins from a bifurcated state of knowledge. A mature, century-old, well-validated body of experience-curve theory explains and forecasts how unit production cost falls with cumulative output, but it was built almost entirely on manufactured hardware and almost never on the non-recurring engineering cost of qualifying a software-intensive capability. A separate body of innovation theory, principally the increasing-returns work of W. Brian Arthur and the useful-knowledge work of Joel Mokyr, explains why cumulative experience should lower cost and predicts when the decline will be steep and when it will be flat, but that theory has not been operationalized as a fitted measurement on autonomy qualification cost. What this chapter must supply is a single theoretical apparatus that gives the functional form to estimate, names the causal mechanism that the slope measures, predicts the conditions under which the slope steepens or flattens, and names, in advance, the principal threat to a clean reading of the slope. No prior work has assembled these pieces into a testable model of autonomy qualification cost. Until they are assembled, NASA and the Jet Propulsion Laboratory continue to make portfolio-sequencing and build-or-wait decisions on the heritage-lowers-cost assumption without any theory that says how much, how fast, or under what conditions cost should fall, and therefore without any way to know whether the assumption is even directionally correct.

This chapter closes that gap at the level of theory. It does four things, in order. First, it establishes the Wright experience curve as a measurement object, develops the organizational learning-by-doing that gives it its causal content, and shows why the log-log form is the right functional commitment (Section 2.1). Second, it presents the modern econometric evidence that the experience curve is a validated forecasting object across many technologies, which is what licenses borrowing the form for a new domain (Section 2.2). Third, it draws the Wright-versus-Moore distinction and derives from it the formal necessity of a time control, the single most consequential specification decision in the design (Section 2.3). Fourth, it develops the two anchor frameworks around their explicit causal mechanisms: Arthur on increasing returns, learning effects, and path dependence (Section 2.4), and Mokyr on propositional versus prescriptive knowledge and codification as a slope moderator (Section 2.5). It then synthesizes the heterogeneity of empirically observed learning rates into the formal case for capability-class fixed effects (Section 2.6) and closes by drawing the threads together into the conceptual model the empirical chapters will carry (Section 2.7). The chapter does not force architecture vocabulary onto what is an econometric measurement study; no DoDAF, BEA, or capability-architecture chain is constructed here.

A note on register and epistemic discipline is owed before the argument begins. This is a design-stage dissertation. No coefficient has been fitted on the full dataset. Every statement about what the slope will be is conditional and explicitly labeled as expected or illustrative. The theory developed here predicts a direction and a set of moderators; it does not assert a result. Where the chapter states a causal claim, it names the mechanism that carries the cause through to the predicted effect, and where only correlation is available it says so and lowers the confidence accordingly. Confidence levels are stated as low, moderate, high, or very high, and each is tied to what evidence would raise or lower it.

## 2.1 The Wright experience curve and organizational learning-by-doing

The first commitment of this chapter is foundational. The experience curve is not a loose metaphor for "things get cheaper over time" but a specific, falsifiable functional relationship between unit cost and cumulative output, and that specific form is the object this dissertation imports. The relationship to be imported is precise: unit cost falls by a constant percentage each time cumulative output doubles, a relationship that is linear in logarithms and whose slope encodes a single learning rate.

The empirical regularity was first documented by Wright [\[116\]](#ref-116) in the airframe manufacturing context, where the labor hours required to build an aircraft fell predictably as the cumulative number of that airframe produced rose; it has since been documented across a wide range of manufactured goods. Thompson [\[94\]](#ref-94) provides the authoritative modern review of the cost-quantity relationship, tracing it from Wright's original airframe studies through the decades of accumulated evidence and assembling the case that the relationship holds across industries rather than as an artifact of any single one. Fusfeld [\[41\]](#ref-41) and Ayres [\[6\]](#ref-6) situate the progress function within the broader family of technological-forecasting tools and confirm that the cumulative-output formulation has been treated as a forecasting instrument, not a descriptive curiosity, since at least the 1970s. A relationship that recurs across many independent industries, that has survived decades of scrutiny, and that takes the same mathematical form in each setting is a candidate for a general law of cost behavior rather than a local accident, and on that reasoning a general regularity may legitimately be imported into a new domain as a hypothesis to be tested. Pervasive empirical regularities transfer as hypotheses, not as facts, into adjacent domains.

What licenses reading the curve causally rather than merely descriptively is the demonstration in Thompson [\[94\]](#ref-94) that the cost-quantity relationship is driven, at least in substantial part, by organizational learning-by-doing rather than by static scale economies alone. This gives the curve a causal interpretation. If the only thing happening were that larger production runs spread fixed costs over more units, the relationship would be a statement about scale and would not transfer to a setting, like autonomy qualification, where each "unit" is a distinct first-of-kind engineering episode rather than a mass-produced item. Thompson's contribution is to separate the learning component, in which the organization gets better at the task through repetition and accumulates transferable know-how, from the pure scale component. It is the learning component that the present study claims transfers to autonomy. The mechanism by which learning-by-doing operates at the level of the engineering organization is documented directly by von Hippel and Tyre [\[49\]](#ref-49), who studied how learning by doing is actually done on novel process equipment and found that real improvement comes from the identification and resolution of problems that could not be anticipated before the equipment was used. Their finding is mechanistically important for this dissertation: the cost reduction from experience is not automatic but is realized through a concrete organizational process of encountering, diagnosing, and codifying problems that only surface in use. That is the process by which a flown autonomy demonstration is expected to lower the cost of the next one in its class.

The relationship holds with high confidence for the manufactured-hardware unit cost on which it was established. Its transfer to non-recurring engineering cost for a software-intensive flight capability is a hypothesis held at moderate confidence, because the import crosses two boundaries at once: from hardware to software, and from recurring unit cost to non-recurring qualification cost. That caution is preserved, not dropped. This chapter establishes the form as a defensible thing to estimate, not as a thing already known to hold for autonomy. The relationship could fail to transfer if autonomy qualification cost is dominated not by repeatable, learnable engineering tasks but by irreducible first-of-kind invention that does not benefit from prior demonstrations, or if each autonomy demonstration is so idiosyncratic that the "cumulative output within a class" construct does not correspond to a coherent repeated task. That objection is taken seriously and is exactly what the flat-cost null H0 represents: a world in which the experience-curve form, however well it fits hardware, does not describe autonomy qualification because the underlying repeated-task condition is not met.

The causal mechanism implied by this section can be traced from cause to consequence. Repeated execution of a defined engineering task within a capability class produces problem identification, diagnosis, and the codification of solutions into reusable patterns and components; that codified stock lowers the engineering effort required for the next execution of the task; the lowered effort appears as a negative slope of log cost on log cumulative count; and that slope, once measured, becomes a planning parameter for sequencing autonomy investments. The first link, that repetition produces codified problem-solving, is established at high confidence by von Hippel and Tyre [\[49\]](#ref-49); the last link, that the measured slope becomes a planning parameter, is the contribution this dissertation seeks to deliver and is at this stage a design commitment, not a result.

## 2.2 The experience curve as a validated forecasting object

The second claim establishes that the log-log experience curve is not merely a recurrent description but a forecasting model whose predictive performance has been measured, characterized, and compared against rival functional forms. This converts the decision to estimate the Wright form from a matter of convention into a matter of evidence.

Among the candidate functional forms for the cost-experience relationship, the Wright power law, the log-log form, is the best-validated for prediction across a large and heterogeneous set of technologies, and that record of validation is the reason for adopting it here rather than a competing form. Nagy, Farmer, Bui, and Trancik [\[79\]](#ref-79) compared functional forms across a large library of technologies and found the Wright power law performs at least as well as, and generally better than, the alternatives, including the time-based exponential (Moore) form, for out-of-sample prediction. Farmer and Lafond [\[35\]](#ref-35) then quantified how predictable technological progress actually is, characterizing the error distribution of experience-curve forecasts so that a forecaster can attach a calibrated uncertainty band to a projection rather than a point estimate. Lafond and colleagues [\[61\]](#ref-61) extended this to a method for making full distributional forecasts, so that the experience curve yields not just an expected cost but a probability distribution over future cost. Way and colleagues [\[102\]](#ref-102) pushed the formalization further, treating technologies that follow experience curves as assets in a portfolio and showing how standard portfolio theory must be modified when the assets are technologies whose cost paths follow experience curves. That is direct evidence that the experience curve is now treated as a rigorous quantitative object suitable for optimization, not a back-of-envelope heuristic.

A functional form that has been competitively tested against its rivals on out-of-sample prediction across many technologies, and that has had its forecast error distribution explicitly quantified, is the form a careful analyst should adopt when entering a new domain, because adopting it inherits both the validated structure and the known error behavior. One should prefer the functional form with demonstrated out-of-sample performance and characterized uncertainty over forms chosen for analytic convenience. That preference is well founded here because the body of work just cited was produced by overlapping research groups using consistent, transparent, and reproducible methods on shared technology-cost datasets, and the central finding, that the Wright form predicts well and its errors are characterizable, has been reproduced and extended rather than contradicted across the sequence from Nagy [\[79\]](#ref-79) through Farmer and Lafond [\[35\]](#ref-35) to Lafond [\[61\]](#ref-61) and Way [\[102\]](#ref-102). Convergence of independent extensions on a stable central result is more persuasive than a single study, however careful.

Two cautions bound the import. First, the predictive validation was conducted on technologies with substantially more observations per technology than the autonomy panel will have, so the forecasting performance is established at very high confidence for data-rich settings and at lower confidence for the small-panel setting of this study. This motivates a later design commitment: in a small panel the experience curve is used as an estimand whose slope is the object of inference, not as a forecasting engine pushed many doublings into the future, because the validated forecasting performance does not automatically carry to extrapolation from tens of observations. Second, the validated forecasting performance could be irrelevant to this study if autonomy qualification cost simply does not follow any of the forms tested, in which case the comparative superiority of the Wright form among those tested provides no guarantee. Again this is the content of H0. Adopting the best-validated general form is nonetheless the correct prior precisely because it gives the flat-cost null a fair and pre-committed functional form to be tested against, rather than allowing the analyst to search over forms until one fits.

The interpretive payload of this section is that the choice of the log-log form is evidence-based, not decorative. Every source in the cluster is read for what its convergence means for the argument rather than listed: the cumulative message of Nagy [\[79\]](#ref-79), Farmer and Lafond [\[35\]](#ref-35), Lafond [\[61\]](#ref-61), and Way [\[102\]](#ref-102) is that the experience curve has graduated from a stylized fact to an instrument with measured predictive accuracy and quantified error, which is the standard a measurement-focused dissertation needs from its imported functional form. Confidence that the Wright form is the right thing to estimate is therefore high; confidence that it will fit the autonomy data is deliberately withheld and assigned to the empirical test.

## 2.3 Wright versus Moore and the necessity of a time control

The third point is the most consequential single point in the theoretical framework, because it determines a specification decision that the validity of the entire study rests on. The experience curve and a simple time trend are observationally entangled, and separating genuine learning from a coincident time trend is not optional but structurally required. Because cumulative output and calendar time typically move together, a regression of cost on cumulative experience without a time control cannot distinguish learning-by-doing from any cost decline that would have happened anyway with the passage of time, and so a credible experience-curve estimate must include a time control; in this design that control is the set of decade fixed effects.

The reasoning rests on a formal result. The Wright formulation makes cost a function of cumulative output. The Moore formulation makes cost a function of calendar time. Nagy and colleagues [\[79\]](#ref-79) showed formally that where cumulative output grows exponentially in time, the Wright and Moore formulations are observationally near-equivalent: the same cost path is consistent with a learning story driven by output and with a pure-time story driven by the calendar, and the data alone cannot adjudicate between them without additional structure. Kayal [\[55\]](#ref-55) and Farrell [\[36\]](#ref-36) reinforce, from the technological-forecasting side, that the pace of technological progress can be modeled either as a function of effort and experience or as a function of time, and that the two are easily confounded when experience accumulates steadily. Wei, Smith, and Sohn [\[103\]](#ref-103) add the load-bearing empirical wrinkle that learning rates estimated from retrospective experience curves are frequently not constant and are correlated with deployment programs, which means that part of what a naive experience-curve regression attributes to cumulative output is actually attributable to policy-driven or program-driven surges in deployment that occur at particular times. That finding makes the confounding between experience and time concrete and dangerous rather than merely theoretical.
If two explanatory stories imply the same observable cost path, then any estimate that omits a feature distinguishing them says nothing about which story is true. An estimate meant to support a causal claim about learning must therefore include the feature, here a time control, that breaks the observational equivalence. A parameter is credibly estimated only when the design supplies the variation needed to separate it from its confounds. This requirement has two foundations. The formal demonstration in Nagy [\[79\]](#ref-79) that the equivalence is exact under exponential output growth is a mathematical property rather than an empirical contingency, which makes the case for a time control a matter of logic. The empirical demonstration in Wei [\[103\]](#ref-103) that real learning rates wobble with deployment timing shows that the confound is not hypothetical in actual technology-cost data.

Decade fixed effects break the confounding between cumulative heritage and economy-wide or agency-wide time trends that operate at the decade scale, such as the secular cheapening of computing across the 1990s, 2000s, 2010s, and 2020s. They do not break a confounding that operates within a single decade, because variation faster than the decade is absorbed into neither the fixed effect nor cleanly into the slope. The control is strong against between-decade time trends and only partial against within-decade time trends, a limitation carried forward to the threats-to-validity analysis. Confidence that the decade control removes the dominant time confound is moderate-to-high; confidence that it removes all time confounding is deliberately not claimed. One might object that, with a small panel, decade fixed effects consume so many degrees of freedom that the cure is worse than the disease, leaving too little residual variation to estimate the slope. This is a real tension, not a strawman. The design's response, developed in the research-design chapter and previewed here, is that the decade is the coarsest time partition that still plausibly captures the dominant time trend, the computing-cost decline, while consuming the fewest degrees of freedom; a finer partition such as a year or five-year bin would be cleaner in theory but is infeasible in a panel of tens of observations. The choice of the decade as the time-control granularity is therefore a deliberate bias-variance compromise, made explicit and defended, not an unexamined default.

This section is the hinge of the framework. The bias it must control follows a clear path: cumulative heritage and calendar time co-move, so a slope estimated without a time control mixes learning with time trend, which produces a downward-biased and uninterpretable coefficient that cannot support a causal reading. The design answers by absorbing decade-scale time variation into fixed effects, which yields a slope identified from within-decade variation in heritage and therefore interpretable as learning rather than as calendar, the credible measurement the study is built to produce. The direction of the residual bias is worth stating because it favors the study's credibility: the forward-only heritage-counting rule and the partial within-decade time confound both tend to attenuate the slope toward zero, so a slope that survives as significantly negative is a conservative finding. Confidence in the direction of this bias is moderate, grounded in the structure of the design rather than in executed results.

## 2.4 Anchor 1, W. Brian Arthur: increasing returns, learning effects, and path dependence

The first anchor supplies the causal mechanism that the experience-curve slope measures and, in the same breath, the principal threat to reading that slope as a universal constant. Arthur's contribution is double-edged in a way that sharpens the test rather than decorating it: it predicts the direction of H1 and names the objection the design must answer.

The argument runs as follows. Learning effects are one of the mechanisms that generate increasing returns to a technology, they are the mechanism an experience curve measures, and so Arthur's framework predicts a negative experience-curve slope as the expected signature when learning effects dominate. But Arthur's framework also holds that increasing-returns systems are path-dependent and non-ergodic, so the realized slope reflects the historical sequence of which capability classes received early investment, which bounds any claim that the measured slope is an intrinsic property of the technology. Arthur [\[3\]](#ref-3) identified four self-reinforcing mechanisms through which a technology becomes more attractive the more it is adopted, and learning effects, in which each use of a technique lowers the cost or raises the performance of the next use, is one of the four. The experience curve is, in Arthur's terms, the empirical trace of the learning-effect mechanism: a constant-percentage cost reduction per doubling of use is exactly what "each use lowers the cost of the next" produces when the lowering is proportional. Arthur's book-length treatment [\[4\]](#ref-4) develops the consequences of increasing returns for the economy, establishing that systems with increasing returns do not converge to a unique efficient equilibrium but lock in to one of several possible outcomes depending on early, sometimes small, historical events. His later synthesis [\[5\]](#ref-5) restates this within complexity economics and is explicit that such systems are non-ergodic: the realized path is not a draw from a stationary distribution but a contingent history, so the same technology family could, under a different early sequence of investments, have realized a different cost trajectory.

If learning effects are a mechanism of increasing returns, and if the experience curve is the measurable trace of learning effects, then estimating the experience-curve slope is estimating the strength of that mechanism, and a theory that says the mechanism operates is a theory that predicts the slope's sign. The identity between learning effect as a causal mechanism and experience-curve slope as its observable is the most important conceptual bridge in the dissertation, because it is what licenses interpreting a fitted \(\beta\) as evidence about a causal mechanism rather than as a description of co-movement. That identity rests on more than a single author's conjecture. A broad and well-developed path-dependence literature has elaborated and tested Arthur's framework across domains, giving it standing as a mature theory. Pierson [\[85\]](#ref-85) carried increasing returns and path dependence into the study of politics, demonstrating the framework's reach beyond technology markets. Cowan and Gunby [\[25\]](#ref-25) supplied a documented case of lock-in to an inferior technology in agricultural pest control, showing that the lock-in mechanism Arthur described is observed in the wild and not only modeled. Kenney [\[56\]](#ref-56) traced path dependence in the industrial clustering of Silicon Valley and Route 128, and Dobusch and Schussler [\[28\]](#ref-28) reviewed the positive-feedback mechanisms that generate path dependence across technology markets, regional clusters, and organizations, consolidating the mechanism into a reviewed theory. A second strand has worked to make path dependence testable rather than merely narrated. Vergne and Durand [\[98\]](#ref-98) identify the missing link between the theory and the empirics of path dependence and argue that the concept must be operationalized in a way that is falsifiable rather than invoked after the fact, and David [\[27\]](#ref-27), a founder of the empirical path-dependence program, defends the program against its critics and argues for a disciplined "historical economics" that takes sequence seriously without abandoning rigor. Hotte [\[50\]](#ref-50) further demonstrates that absorptive capacity coevolves with technology diffusion, a path-dependent feedback directly relevant to why a capability class with early investment accumulates its own self-reinforcing advantage.

Arthur's framework predicts the sign of the slope at moderate-to-high confidence: if learning effects are present and dominant, the slope is negative. It does not predict the magnitude of the slope, and it cautions that the magnitude is contingent on history. The framework therefore supports H1 as the theoretically expected direction while withholding any prediction that the magnitude is a stable, transferable constant. That caution is essential and preserved: this dissertation does not claim, and Arthur's theory does not license, that a fitted learning rate for one capability class is a universal autonomy learning rate. The path-dependence half of Arthur's framework is itself the answer to a naive reading of the result. If the realized cost path is contingent on which classes received early investment, then a measured slope could reflect the lucky early concentration of investment in a learnable class rather than an intrinsic learnability of autonomy, and a different historical sequence would have produced a different slope. The design cannot eliminate this, because the history is what it is, but it responds in three ways named here and developed in the design chapter: capability-class fixed effects so the slope is estimated within rather than across the historically privileged classes; an influence diagnostic so that a slope driven by a single early high-leverage episode is flagged as non-robust; and an explicit external-validity bound in the discussion stating that the estimated slope is an average within-class learning rate for the realized history, not a universal constant. Vergne and Durand's [\[98\]](#ref-98) insistence on falsifiable operationalization is the standard the design holds itself to here: path dependence is treated as a testable threat with stated responses, not as an all-purpose excuse.

The mechanism Arthur supplies runs as a chain. Repeated adoption and use of an autonomy capability within a class triggers the learning-effect mechanism of increasing returns, in which each use lowers the cost of the next; that produces a proportional reduction in qualification cost per doubling of cumulative use; that reduction is measured as a negative log-log slope \(\beta\); and the magnitude of that slope is path-dependent and reflects the historical investment sequence. The first half of this chain, learning effects produce a negative slope, is the engine of H1. The second half, the magnitude is contingent, is the discipline that keeps the contribution honest. Confidence that learning effects, if present, produce a negative slope is high on theoretical grounds; confidence that the magnitude is intrinsic rather than historical is intentionally low, by Arthur's own argument.

## 2.5 Anchor 2, Joel Mokyr: propositional and prescriptive knowledge as a slope moderator

The second anchor does what Arthur's cannot: it predicts not just the sign of the slope but the conditions under which the slope should be steep and the conditions under which it should be flat. Mokyr's distinction between propositional and prescriptive knowledge [\[115\]](#ref-115) supplies the dissertation's testable moderator and converts a single-number contribution into a contribution with a policy-relevant secondary finding about heterogeneity across capability classes.

The proposition is this. A technique becomes cheap to reproduce and extend only when it rests on a wide base of propositional knowledge, knowledge of why it works, that allows each prescriptive advance, knowledge of how to do it, to be codified, corrected, and reused rather than rediscovered. The autonomy capability classes whose underlying propositional base is mature, and whose prescriptive knowledge is codified into reusable infrastructure, are therefore predicted to show a steeper cost-decline slope than classes that re-implement each demonstration from scratch.

Mokyr's central analytical move, developed in his account of the growth of useful knowledge [\[115\]](#ref-115), is the separation of propositional knowledge (the theoretical, explanatory base, "knowledge of what and why") from prescriptive knowledge (technique, recipe, "knowledge of how"). His historical argument is that techniques discovered by trial without an underlying explanatory base tend to stagnate, because without the propositional base there is no principled way to know why a technique works, how to extend it, or how to fix it when it fails, whereas techniques anchored in a wide propositional base improve rapidly and cumulatively because the base makes each advance self-correcting and transferable [\[115\]](#ref-115). What licenses transferring this lens to autonomy is the documented existence of mature formal propositional bases for several autonomy capability classes, the formal theory of automated planning and scheduling, of estimation and navigation, and of verification, which means the precondition Mokyr identifies for cumulative cost decline is differentially present across the classes rather than uniformly present or absent. If cumulative, cheapening improvement requires a propositional base that makes prescriptive knowledge codifiable and self-correcting, and if that base is present to differing degrees across autonomy capability classes, then the rate of cost decline should differ across classes in proportion to the maturity and codification of their propositional base. The principle is Mokyr's: propositional depth governs the codifiability and hence the reusability of technique, applied here to a domain where propositional depth varies by class.

Independent empirical evidence, in domains other than space autonomy, supports the testable content of the lens, that codification and reuse infrastructure are associated with steeper or faster cost decline. The clean-energy and emerging-technology cost literature provides several such cases. Glenk, Meier, and Reichelstein [\[44\]](#ref-44) document the cost dynamics of clean energy technologies and show that cost decline is faster where the underlying engineering knowledge is shared and the technology is modular and reusable. Lane and colleagues [\[62\]](#ref-62) forecast technology shares under cost uncertainty in a way that depends on the codified, transferable character of the production technology. Yao, Benson, and Chueh [\[108\]](#ref-108) assess technology roadmaps for sodium-ion batteries and show how the realized cost trajectory depends on the maturity of the shared knowledge base relative to the incumbent lithium-ion base. Myny and colleagues [\[78\]](#ref-78) provide a micro-level demonstration in which a thin-film microprocessor with print-programmable memory becomes feasible because a codified, reusable fabrication base had matured. Hotte [\[50\]](#ref-50), cited above for path dependence, applies here as well, because absorptive capacity, the codified base that lets an organization absorb and reuse new knowledge, is the Mokyr precondition operationalized. None of these are space-autonomy studies, and that distance is acknowledged: they support the general proposition that codification steepens cost decline, and the dissertation transfers the proposition as a hypothesis about autonomy classes, not as an established fact about them.

This moderator is held at moderate confidence as a prediction about autonomy. The direction, more codification means steeper decline, is well supported in general; the application to specific autonomy classes depends on a mapping from the maturity of each class's propositional base and the existence of reuse infrastructure such as the Core Flight System to the steepness of its class-specific slope, and that mapping is itself a measurement the design proposes rather than assumes. The moderator is a prediction to be tested as effect heterogeneity across capability classes, not a result. It could fail in two ways, both taken seriously. First, the propositional base could be mature but the prescriptive knowledge could still fail to be codified in practice, because each flight project re-implements from scratch for reasons of risk aversion, schedule, or organizational fragmentation rather than for lack of theory; in that case a class with a deep propositional base would still show a flat slope, and the moderator would be confounded by organizational practice rather than knowledge maturity. Second, Mokyr's own Cardwell caution [\[114\]](#ref-114), that technological progress is historically reversible and that no society or organization has maintained continuous progress indefinitely, warns that codification can decay, so a class that was codified in one decade may not remain so, and the slope could flatten over time even within a class with a mature base. The design responds to the first concern by including a reuse-infrastructure consideration in the heterogeneity analysis rather than treating propositional maturity alone as the moderator, and to the second by the decade fixed effects, which absorb decade-scale shifts in the codification environment. Neither response is complete, and both are flagged.

The Mokyr mechanism runs as a chain. The maturity and codification of a capability class's propositional knowledge base, embodied in reuse infrastructure such as shared flight-software frameworks, determines whether each demonstration's prescriptive advances are codified into self-correcting reusable patterns or rediscovered from scratch; that sets how much of the prior demonstration's effort transfers to the next; that transfer appears as a steeper or flatter class-specific experience-curve slope; and that slope tells the portfolio which capability classes will reward sequenced investment and which will not. This is what elevates the dissertation from a single pooled slope to a model with predicted, testable heterogeneity, and it is the most policy-relevant secondary product the design can yield. Confidence that codification steepens the slope in general is moderate-to-high from the supporting literature; confidence about which specific autonomy classes will show the steepest slopes is low and assigned to the empirical heterogeneity test.

## 2.6 Heterogeneity of learning rates and the formal case for capability-class fixed effects

The two anchors converge on a single specification implication, and this section states it as the bridge from theory to estimator. Arthur warns that the realized slope is path-dependent and class-specific; Mokyr predicts that the slope's steepness varies with the codification of each class's knowledge base. Both imply that there is no single autonomy learning rate, only a family of class-specific rates, and that pooling them naively would estimate a meaningless average that no class actually exhibits. The formal response is to estimate a within-class slope using capability-class fixed effects, which is the direct, theory-mandated expression of the experience-curve tradition's own rule that a learning rate transfers only within a coherent technology family.

Because empirically observed learning rates are heterogeneous across technologies and settings, and because both anchor frameworks predict that the autonomy slope is class-specific, the correct estimand is a within-capability-class slope identified with capability-class fixed effects, not a single pooled slope across heterogeneous classes.

The empirical heterogeneity of learning rates is well documented. Bhattacharya and colleagues [\[14\]](#ref-14) examined cost-quantity relations across Indian industries and found that learning rates are diverse, with different industries showing materially different rates of learning by doing, which establishes that a single learning rate is the exception rather than the rule when heterogeneous activities are pooled. Wei, Smith, and Sohn [\[103\]](#ref-103) documented that learning rates are frequently non-constant even within a single technology and are correlated with deployment programs, reinforcing that the assumption of one stable rate is fragile. Perez [\[84\]](#ref-84) supplies the macro-level reason that learning rates cluster and shift by techno-economic paradigm, so that technologies maturing within different paradigms accumulate experience under different conditions, a further source of between-class heterogeneity. The experience-curve forecasting literature itself treats the learning rate as transferable only within a coherent technology family, never across unrelated families, which is why Nagy [\[79\]](#ref-79) and the related forecasting work estimate per-technology rates rather than a universal constant.

If the quantity to be estimated varies across the categories in the data, then pooling the categories without absorbing their differences produces an estimate that is a weighted average of distinct quantities and is biased for any one of them, so the design must absorb the category-level differences to recover a within-category parameter. The standard panel-econometric principle is that heterogeneity across known, non-sampled categories is handled by fixed effects, which absorb the time-invariant level differences across categories and identify the slope from within-category variation.

Fixed effects are the appropriate device specifically because the capability classes are the particular, named, non-sampled categories of interest rather than random draws from a larger population, which means a fixed-effects estimator is preferred to a random-effects estimator on two grounds. First, a fixed-effects estimator does not require the orthogonality assumption between the heritage variable and the class effect that a random-effects estimator imposes, and that orthogonality is implausible here because the classes with deep codified bases (Mokyr) and early investment (Arthur) are precisely the classes likely to have both lower baseline cost and more accumulated heritage, so heritage and the class effect are correlated by construction. Second, the classes are exhaustive and fixed, not a sample, so the random-effects assumption that they are exchangeable draws is false on its face. The trade-off, that fixed effects consume degrees of freedom that are scarce in a small panel, is acknowledged and is exactly why the design caps the number of capability classes at the level the sample can support and reports the realized residual degrees of freedom alongside every coefficient.

Capability-class fixed effects identify the slope from within-class variation in heritage and are therefore the right device, at high confidence, given the heterogeneity documented above. They do not resolve the path-dependence concern within a class: if a single class's heritage accumulation is itself the product of a contingent early choice, the within-class slope still reflects that history. The fixed effects remove between-class confounding, not within-class path dependence, and that residual is carried forward as a bounded threat rather than a solved problem.

A critic could argue that with so few observations per class, the within-class slope is identified off almost no variation and is hopelessly imprecise, so that the theoretically correct estimand is practically unestimable. This is the binding constraint of the study and is not waved away. The design's response, stated here and operationalized in the analysis-plan chapter, is the pre-analysis feasibility check: any capability class or decade containing only a single observation contributes nothing to the within estimator and is flagged, so that the reported slope is honest about the effective sample it is identified from, and a slope that turns out to rest on one or two classes is reported as such rather than presented as a general autonomy learning rate. The honest small-panel posture is the safeguard that keeps the theoretically correct estimand from being oversold when the data cannot support precision.

The convergence this section documents is the chapter's structural payoff: two independent theoretical anchors, Arthur on path-dependent class-specific increasing returns and Mokyr on codification-driven class-specific knowledge depth, plus the empirical heterogeneity evidence of Bhattacharya [\[14\]](#ref-14), Wei [\[103\]](#ref-103), and Perez [\[84\]](#ref-84), all point to the same estimator design. When three independent lines of reasoning converge on one specification, the specification is warranted by triangulation rather than by any single line. Confidence that within-class fixed effects are the correct identifying device is high; confidence that the resulting within-class slopes will be precisely estimated in a small panel is low and is the dominant statistical risk the later chapters confront.
## 2.7 The conceptual model and chapter synthesis

This closing section assembles the pieces developed above into the single conceptual model the empirical chapters will test, and it draws the chapter's argument together from problem to acceptable residual risk. Each link traces back to the theory built in this chapter.

The conceptual model is a structured prediction, not a result. It holds that (1) onboard autonomy qualification cost within a capability class is governed by a learning-effect mechanism (Arthur) whose observable trace is an experience curve (Wright, Section 2.1); (2) the right functional form for that curve is the log-log Wright form, on the validated forecasting evidence (Section 2.2); (3) a credible slope requires a time control to separate learning from the coincident computing-cost time trend (Wright-versus-Moore, Section 2.3); (4) the slope is expected to be negative if learning effects dominate but its magnitude is path-dependent and class-specific (Arthur, Section 2.4); (5) the slope is expected to be steeper for classes with mature, codified propositional bases and reuse infrastructure (Mokyr, Section 2.5); and (6) the correct estimand is therefore a within-class slope identified with capability-class and decade fixed effects (Section 2.6). The estimating equation \(\ln(\text{Cost}_{icd}) = \alpha + \beta \, \ln(\text{CumHeritage}_{icd}) + \gamma_c + \delta_d + \epsilon_{icd}\) is the formal compression of this six-part model, and \(\beta\) is the single quantity that the model's truth or falsity reduces to.

The chapter's argument can now be drawn together. NASA and JPL rely on an unmeasured heritage-lowers-cost assumption, and Sections 2.1 through 2.6 establish that a theory exists, learning effects traced by experience curves, that says the assumption could be true and specifies what its truth would look like, which is the precondition for treating the question as measurable rather than as a vague worry. The experience-curve literature reviewed above shows that learning rates, where they exist, are large enough to dominate portfolio economics. If an autonomy learning rate exists it is decision-relevant; if it does not, the absence is equally decision-relevant, because it removes the cost-reduction rationale for sequenced autonomy investment. The within-class log-log experience curve is not a generic regression but the specific measurable trace of Arthur's learning-effect mechanism, so a fitted slope speaks directly to the causal quantity of interest rather than to a surface correlation. The Wright log-log form is the best-validated functional form for prediction, and two-way fixed effects are the device the heterogeneity evidence and both anchors demand for separating learning from time and from intrinsic class difficulty. The principal residual risks named in this chapter, within-decade time confounding (Section 2.3), within-class path dependence (Sections 2.4 and 2.6), the gap between the heritage count and the true reusable-knowledge stock (Section 2.5), and small-panel imprecision (Section 2.6), are each bounded by a stated design response and by the design-stage posture that reports a wide interval containing zero as a failure to reject the null rather than as support for the alternative.

The chapter's contribution to the whole is to convert the heritage-lowers-cost assumption from an article of organizational faith into a structured, falsifiable conceptual model with a named mechanism, a justified functional form, a required time control, two anchor frameworks that predict the slope's sign and its heterogeneity, and a fixed estimand. The remaining chapters operationalize this model: the literature review (Chapter 3) assembles the demonstration chronology that feeds the cumulative-heritage variable, the data chapter (Chapter 4) constructs the cost and heritage measures, the design chapter (Chapter 5) develops the estimator and the threats-to-validity matrix, and the analysis plan (Chapter 6) states the estimation procedure, the pre-analysis checks, and the fixed decision rule. The theory is now in place. What it predicts is a direction and a set of moderators; what the empirical work will deliver is either a defensible negative slope that confirms the model or a defensible failure to find one that refutes it. Both outcomes are informative, and the symmetry of that informativeness is itself a product of the theoretical discipline this chapter has imposed: by stating the model as a falsifiable prediction rather than a hoped-for result, the framework guarantees that whatever the data say, the dissertation will have measured something worth knowing.



# Chapter 3. Literature Review

## 3.0 Aim and organization of the chapter

The literature already contains every observation this dissertation needs, but it has never been read as a cost-of-heritage record. That is the answer this chapter delivers, and the rest of the chapter develops it. Across five capability classes, the published autonomy-demonstration literature documents a dated, ordered sequence of flight events in which a capability is first proven, then reused, then matured, and at each step the engineering teams describe building on what came before. That sequence is the cumulative-heritage variable \(\text{CumHeritage}\) that the estimating equation requires. What the literature does not contain is the dependent variable read against that sequence: nobody has fitted the recurring qualification cost of a capability class onto its accumulated flight heritage. The experience-curve literature, reviewed in Chapter 2, supplies the method but studies manufactured-hardware unit cost; the autonomy literature reviewed here supplies the heritage chronology but treats heritage qualitatively, as an enabling narrative rather than a measurable quantity. The gap is the join. This chapter reads the autonomy literature as the empirical spine of an experience-curve study, extracts the demonstration chronology that feeds \(\text{CumHeritage}\) for each capability class, and shows that the heritage-lowers-cost claim is asserted repeatedly and measured nowhere.

The problem this chapter addresses is the distance between what the literature has recorded and what it has measured. The literature offers a rich, class-by-class demonstration record in which heritage is described in prose and the cost consequence of heritage is asserted but never estimated. What is wanted instead is a literature in which the qualification cost of each autonomy capability class can be read against its cumulative flight heritage and a learning rate can be fitted. No published study performs that join; the two relevant literatures sit side by side and unconnected. While they remain so, NASA and the Jet Propulsion Laboratory continue to sequence the Autonomous Systems and Robotics portfolio on an unmeasured assumption, with build-or-wait decisions made by assertion and no cost-anchored complement to the ordinal, non-monetary technology-readiness-level scale.

The chapter is organized to cover the five capability classes exhaustively and without overlap, because the estimating equation identifies the slope \(\beta\) from within-class, within-decade variation, and the literature must be read class by class for the chronology to be usable. Each class is treated the same way: a statement of what the demonstration record shows, the published evidence that supports it (with each major source discussed for its method, its finding, and its limitation), and the reasoning that converts that evidence into a contribution to the cumulative-heritage record. The treatment is interpretive. Every source is read for what its place in the sequence means for the argument, never merely listed. Section 3.1 covers onboard planning and scheduling, the origin class. Section 3.2 covers autonomous science target selection, the cleanest within-class heritage pair in the record. Section 3.3 covers autonomous navigation, the deepest and most continuous heritage chain. Section 3.4 covers autonomous fault detection, isolation, and recovery. Section 3.5 covers autonomous entry, descent, and landing hazard handling, the class whose maturation is most explicitly cost-and-readiness tracked. Section 3.6 treats the codified-knowledge substrate, the Core Flight System and shared autonomy frameworks, which is the Mokyr moderator that the design predicts should steepen the cost-decline curve where it is present. Section 3.7 synthesizes the whole into an explicit statement of the gap and the propositions that follow into the chapters on data and design.

A note on scope and honesty is owed at the outset, because it governs how every source below is read. This dissertation is at the design stage. The literature reviewed here establishes the heritage chronology, the capability definitions, and the qualitative scope of each demonstration. It does not, and the corpus does not, contain clean autonomy-specific recurring-engineering cost figures per demonstration. Where the published record speaks to cost, it speaks to it indirectly, through lessons-learned narratives, through technology-readiness-level transitions, and through parametric cost-model studies of adjacent space systems. The dependent variable of this study must therefore be constructed by the procedure developed in Chapter 4, not lifted ready-made from any single paper. This chapter is candid about that: it reads the literature for what it genuinely contains, the ordered heritage record, and flags where it does not contain the cost numbers the design will have to build. Confidence in the heritage chronology is high, because the demonstration sequence is documented in peer-reviewed and primary engineering sources. Confidence in any direct cost reading from this literature is low, because the cost figures are not reported in a comparable form, and that low confidence is why the study is framed as a measurement to be built rather than a result to be quoted.

## 3.1 Onboard planning and scheduling: the origin class

The claim of this section is that onboard planning and scheduling is the origin capability class of the entire study, that its heritage chain is documented with unusual completeness from prototype through first flight to operational reuse, and that the engineering teams themselves describe each step as building on the last, which is the qualitative form of the mechanism the dissertation will quantify.

The class begins before flight, with the Remote Agent prototype work at the Jet Propulsion Laboratory and NASA Ames. Pell and colleagues described the New Millennium Remote Agent architecture as an integration of constraint-based planning and scheduling, fault-tolerant multi-threaded execution, and model-based diagnosis and reconfiguration into a single onboard agent [\[82\]](#ref-82), [\[83\]](#ref-83). The method in these papers is architectural: they analyzed the spacecraft-control domain, identified its distinguishing properties of long latency, scarce communication, and the need for autonomous recovery, and designed an architecture to meet them. Their finding was that the three traditionally separate functions of planning, execution, and diagnosis could be unified under a goal-directed agent. The limitation is that these were ground prototypes; they demonstrated feasibility, not flight qualification, and they did not report the cost of carrying the prototype to flight. For this study they matter because they establish the baseline propositional content of the class, the formal planning, scheduling, and execution machinery that all later heritage reuses. In Mokyr's terms, examined in Chapter 2, this is the codified propositional base whose maturity the design predicts should govern how steeply the class's cost curve falls.

The flight event that opens the empirical record is the Remote Agent Experiment on Deep Space 1 in 1999. Bernard and colleagues reported the validation and verification of the Remote Agent as a flight experiment, the first time an artificial-intelligence agent autonomously controlled a NASA spacecraft [\[12\]](#ref-12). The companion design paper laid out how the operational rules and constraints were encoded into flight software so the spacecraft could be commanded by goals rather than by detailed command sequences [\[13\]](#ref-13). The flight-experience report by Bernard and colleagues at the 1999 Space Technology Conference documented what happened in flight, including the in-flight anomaly and its recovery [\[30\]](#ref-30). The method across these sources is a documented engineering campaign: design, encode, test, fly, and report. The finding is that goal-level autonomous control is achievable on a real deep-space platform. The most important source in this cluster for the present study is the lessons-learned record, in which the team reported that the impact of inserting system-level autonomy into a flight project was a major surprise, with the integration and verification effort far larger than anticipated [\[11\]](#ref-11). That statement is the most direct piece of qualitative evidence in the entire corpus on the recurring-engineering and organizational cost of a first-of-kind autonomy capability. It is the quantity the dependent variable is intended to capture, named in the words of the engineers who paid it. Its limitation as evidence is equally clear: it is a narrative, not a dollar figure, and it cannot by itself populate \(\ln(\text{Cost})\). It establishes that the cost was large and surprising; it does not establish how large in a form comparable across episodes.

Testing the planner is a separable and documented cost driver within the class. Smith and colleagues reported the challenges and methods in testing the Remote Agent planner, describing the verification burden specific to a planning system whose behavior is generated rather than scripted [\[88\]](#ref-88). The method is a verification-and-validation case study; the finding is that a generative planner imposes a distinctive and heavy testing cost, because the space of behaviors cannot be enumerated the way a fixed command sequence can. This is direct evidence for one component of qualification cost, the verification component, and it is also direct evidence for the Mokyr moderator: a class whose verification approaches become codified and reusable should see this component fall on later episodes, while a class that re-invents verification each time should not.

The verification-cost mechanism deserves to be named precisely, because it is the clearest causal chain the planning literature supplies for the dissertation's central claim. A goal-directed planner generates its own command sequence at run time rather than executing a sequence verified on the ground. Ground verification can therefore no longer enumerate and test every command path, so the project must instead verify the planner's model and its search behavior across a space of contingencies, a qualitatively different and larger engineering task. The observable result, reported by Smith and colleagues, is a testing campaign whose scope surprised the team. If heritage works as assumed, a later same-class episode inherits both the verification approach and the tested model fragments, so its verification cost is lower. The verification component is thus where an experience curve in this class would be most visible, and where the absence of reusable verification infrastructure would most flatten it. This is a mechanism, not a correlation, and the planning literature describes each link in it qualitatively; what no source does is measure the second-episode verification cost against the first, which is the measurement the design will attempt. The confidence that the mechanism is real is moderate-to-high, grounded in the engineers' own reporting; the confidence that its magnitude has been quantified is zero.

The within-class heritage step that the design most needs is the move from the bounded Deep Space 1 experiment to routine operational use. The supporting prototype work by Chien and colleagues on iterative-repair planning showed how to make planning responsive enough for continuous onboard operation rather than batch use, incorporating execution feedback such as early or late completion and resource over-use or under-use into a continuously repaired working plan [\[20\]](#ref-20), [\[21\]](#ref-21). The broader statement of the approach framed automated planning and scheduling as enabling a new class of goal-based autonomous spacecraft [\[19\]](#ref-19). The method in these is algorithmic and architectural; the finding is that responsiveness, the property that distinguishes an operational planner from an experimental one, is achievable by iterative repair. The limitation is that these are method papers and conference reports rather than cost studies. Their role in the spine is to document the reusable knowledge that links the first flight to the operational deployment that follows, which is the heritage the cumulative count is meant to register.

That operational deployment is the EO-1 Autonomous Sciencecraft Experiment. Chien and colleagues reported that autonomy flight software was used to improve science return on Earth Observing One, moving onboard event detection and replanning into routine Earth-observation operations after 2003 [\[24\]](#ref-24). The architecture and constraints of the agent, its autonomy, reliability, and limited-computing demands, were described in the autonomous science agent account [\[22\]](#ref-22). The operational science value was documented in two application studies: monitoring active volcanism, in which onboard detection triggered autonomous retasking [\[23\]](#ref-23), and flood detection and monitoring [\[54\]](#ref-54). The method in the EO-1 papers is operational reporting backed by science results; the finding is that onboard autonomy moved from a bounded experiment to a standing operational element that demonstrably improved the timeliness and quality of returned data. The limitation, again, is that none of these papers reports the autonomy-specific development cost in a form comparable to Deep Space 1. For the spine, EO-1 is the second observation in the planning-and-scheduling class, and the qualitative record is consistent with the heritage-lowers-barrier claim: the team reused planning and execution concepts proven on Deep Space 1. Whether that reuse lowered cost along a measurable curve is the unmeasured question.

The class extends forward into the present, which matters for the coverage window and for demonstrating that the propositional base keeps deepening. Zilberstein and colleagues developed decentralized, decomposition-based observation scheduling for a large-scale satellite constellation, formulating the problem as a distributed constraint-optimization problem with a geometric-neighborhood-decomposition heuristic to scale to millions of variables [\[113\]](#ref-113). Lenzen and colleagues reported onboard planning and scheduling autonomy within the FireBird mission [\[64\]](#ref-64), and Labreche and colleagues reported spacecraft autonomy on the OPS-SAT CubeSat using onboard machine learning [\[60\]](#ref-60). The method in these recent works ranges from constraint-optimization theory to flight demonstration on small platforms; the finding is that onboard planning has continued to mature toward larger scale and lower-cost platforms. Their limitation for this study is that constellation and CubeSat contexts may not fall within the NASA and JPL deep-space population that defines external validity, a boundary the design must police. They are included in the spine to show the class is alive and deepening, not frozen at EO-1.

Table 3.1 summarizes the planning-and-scheduling heritage chain as the spine will use it.

**Table 3.1. Onboard planning and scheduling heritage chain.**

| Episode / source | Year | Role in chain | Cost-evidence character |
|---|---|---|---|
| Remote Agent prototype [\[82\]](#ref-82), [\[83\]](#ref-83) | 1996-1997 | Propositional base, pre-flight | None (feasibility only) |
| Iterative-repair responsiveness [\[20\]](#ref-20), [\[21\]](#ref-21), [\[19\]](#ref-19) | 1998-2000 | Reusable method linking experiment to operations | Method papers, no cost |
| Remote Agent Experiment, DS1 [\[12\]](#ref-12), [\[13\]](#ref-13), [\[30\]](#ref-30) | 1999 | First flight (origin observation) | Lessons-learned narrative: large, surprising cost [\[11\]](#ref-11) |
| Planner testing burden [\[88\]](#ref-88) | 2000 | Verification cost component | Qualitative, verification-specific |
| EO-1 Autonomous Sciencecraft Experiment [\[24\]](#ref-24), [\[22\]](#ref-22), [\[23\]](#ref-23), [\[54\]](#ref-54) | 2003-2006 | Second observation, operational reuse | Operational reporting, no autonomy NRE figure |
| Constellation / CubeSat scheduling [\[113\]](#ref-113), [\[64\]](#ref-64), [\[60\]](#ref-60) | 2014-2024 | Class still deepening | Method/flight reports, external-validity boundary |

The reasoning that closes the section is this. The planning-and-scheduling class supplies a complete, dated heritage chain with explicit, in-the-engineers'-words evidence that the first-of-kind cost was large and that later episodes reused proven design patterns. That is the qualitative shape of an experience-curve relationship. What it is not is a measurement. The literature asserts heritage reuse and reports its surprise at first-of-kind cost; it never reads the second-episode cost against the first to see whether the assumed decline actually occurred or how steep it was. Confidence that the chronology is correct is high. Confidence that the literature has measured the cost consequence of that chronology is essentially zero, because no source attempts it.
## 3.2 Autonomous science target selection: the cleanest heritage pair

This section claims that autonomous science target selection, embodied in the AEGIS system, is the cleanest within-class heritage observation in the entire record, because the identical capability was reused on a second platform and the second deployment is explicitly documented as building on the first. If the heritage-lowers-cost mechanism operates anywhere in a form the design can read, it operates here.

The first deployment is AEGIS on the Mars Exploration Rover Opportunity. Estlin and colleagues reported AEGIS automated science targeting for the Opportunity rover, a system that autonomously selected science targets for the narrow-field instruments and removed the operations latency imposed by ground-in-the-loop targeting [\[34\]](#ref-34). The method is a deployed-system study with operational results; the finding is that autonomous onboard target selection both worked and materially reduced the round-trip latency that had previously bottlenecked narrow-field science. The limitation is that the paper reports operational performance, not development cost. For the spine, this is the origin observation of the science-target-selection class.

The second deployment is the heritage event the study most wants. Francis and colleagues reported AEGIS autonomous targeting for the ChemCam instrument on the Mars Science Laboratory Curiosity rover, documenting its deployment and the results of initial science-team use [\[40\]](#ref-40). What makes this pair load-bearing for the study is that it is the same AEGIS capability, ported to a different instrument on a different rover, and the paper presents it as a reuse-and-extension of the Opportunity work rather than a fresh build. The method is again a deployment study; the finding is that the ported system delivered autonomous targeting for ChemCam and was adopted into science-team operations. The limitation for cost reading is the familiar one: the paper does not report the porting cost as a dollar figure. Yet the qualitative content is the strongest single-pair evidence in the corpus for the heritage-lowers-cost mechanism, because the second deployment is documented as cheaper-by-reuse in narrative terms. The design's task is to test whether that narrative reuse corresponds to a measurable cost decline, and the AEGIS pair is where that test will be sharpest.

The instrument context that makes this class coherent is documented in the ChemCam and SuperCam instrument literature. Wiens and colleagues described the ChemCam instrument suite on the Mars Science Laboratory, including body-unit and combined-system tests [\[104\]](#ref-104), and Maurice and colleagues described the SuperCam instrument suite on the Mars 2020 Perseverance rover, the successor remote-sensing payload [\[72\]](#ref-72). The method in these is instrument-engineering description; the finding is that the narrow-field laser-spectroscopy instruments AEGIS targets form a continuous instrument lineage from Curiosity to Perseverance. Their limitation for this study is that they are instrument papers, not autonomy-cost papers; they establish that the targeting capability had a continuing instrument home, which is why the autonomous-targeting class did not terminate after Curiosity. They are spine context, not cost evidence.

Two earlier and adjacent sources establish that autonomous science behavior had antecedents and that the class connects to wider work. Woods and colleagues reported autonomous science for an ExoMars rover-like mission, demonstrating onboard science-driven autonomy in a European mission context [\[107\]](#ref-107). The Rocky 7 rover prototype reported by Hayati and colleagues demonstrated rover-based science-instrument control and autonomous navigation as a Mars sciencecraft prototype well before flight [\[48\]](#ref-48). The method in both is prototype-and-demonstration; the finding is that autonomous science behavior was an active research line with multiple independent efforts. Their limitation is that they are not within the AEGIS heritage chain proper and may sit at the boundary of the class. They are included to show the class has a research hinterland, which bears on the Mokyr moderator: a deep, codified propositional base around autonomous science increases the chance that the AEGIS reuse was genuinely cheap.

Why is the AEGIS pair the design's identification anchor even though it is a single pair? A single within-class transition carries disproportionate weight in a small panel, and that weight must be earned rather than assumed. The AEGIS-Opportunity to AEGIS-ChemCam transition has three properties that no other within-class step in the corpus combines. First, it is a same-capability reuse, not a successor capability: the autonomous-targeting function ported across is the function, not a re-conceived replacement, so the heritage link is maximal rather than partial. Second, the porting is documented contemporaneously by the deploying team as a build-on rather than a rebuild [\[40\]](#ref-40), the qualitative signature of the cost mechanism the slope is meant to capture. Third, the two deployments sit close enough in time that a decade fixed effect absorbs most of the secular computing trend between them, so the comparison is unusually clean of the maturation confound that bedevils the navigation class. These properties make AEGIS the case where a measured cost decline, if it exists, would be most credibly attributable to heritage rather than to time or to a different capability. The corresponding risk, carried forward to the design's influence diagnostics, is that a single clean pair can dominate the within-class slope; the design therefore commits in advance to reporting the science-target-selection slope with and without this pair, so that a result driven entirely by AEGIS is flagged as resting on one transition rather than presented as a class-wide regularity.

The reasoning that closes the section is that the science-target-selection class offers the design its best within-class heritage pair, AEGIS-Opportunity followed by AEGIS-ChemCam, with the second explicitly a reuse of the first. The qualitative evidence for heritage-lowers-cost is here as strong as the corpus provides. And yet no source in the class reads the ChemCam porting cost against the Opportunity development cost. The mechanism is described; the magnitude is unmeasured. Confidence in the heritage link is very high, because it is a documented same-capability reuse. Confidence that the cost consequence has been measured is zero. This is the cleanest illustration in the dissertation of the gap between an asserted and a measured learning curve.

## 3.3 Autonomous navigation: the deepest heritage chain

This section claims that autonomous surface and in-flight navigation has the deepest and most continuous heritage chain of any class in the study, running from pre-flight prototypes through Mars Exploration Rover visual odometry and AutoNav, through the enhanced navigation of Perseverance, to powered autonomous flight on Ingenuity. That depth makes the class the best-powered within-class test the panel will support, while also making it the class where confounding maturation is hardest to separate from heritage.

The class has the richest pre-flight record. Washington and colleagues reported an onboard executive for autonomous Mars rovers, anticipating the long traverses and reliable navigation that future missions would require [\[101\]](#ref-101). Goldberg and colleagues described the stereo-vision and rover-navigation software for planetary exploration that became the basis for hazard detection from passive stereo [\[45\]](#ref-45). Huntsberger reported biologically inspired autonomous rover control [\[52\]](#ref-52), and the Rocky 7 prototype again appears as an early integrated navigation demonstrator [\[48\]](#ref-48). The method across these is prototype development and algorithm design; the finding is that autonomous navigation had a deep, multi-group propositional base before it flew. The limitation is that prototypes do not report flight-qualification cost. Their role in the spine is to establish that, of all the classes, navigation entered flight with the most codified prior knowledge, which under Mokyr's framework predicts the steepest potential cost-decline curve and which the design must keep distinct from the simple fact that navigation also accumulated the most flight episodes.

The first major flight observations are the Mars Exploration Rover navigation results. Maimone and colleagues reported autonomous navigation results from the Mars Exploration Rover mission [\[68\]](#ref-68) and, in the most-cited source of the class, two years of visual odometry on the Mars Exploration Rovers, in which onboard stereo comparison gave each rover accurate knowledge of its position and let it detect and compensate for unforeseen slip [\[69\]](#ref-69). The method is a deployed-system study with quantified operational outcomes; the finding is that visual odometry worked on another world for the first time and increased science return by reducing the days needed to reach interesting terrain. The mission context is given by Crisp and colleagues' account of the Mars Exploration Rover mission [\[26\]](#ref-26) and the instrument-positioning system by Baumgartner and colleagues [\[9\]](#ref-9). The European parallel is Maurette's account of Mars rover autonomous navigation [\[71\]](#ref-71). The limitation across these is that they report navigation performance and science benefit, not autonomy development cost. For the spine, the Mars Exploration Rover work is the first dense cluster of navigation observations, and Maimone's explicit framing of slip detection as a reusable capability is qualitative heritage evidence.

The chain continues to the Mars 2020 enhanced navigation. Toupet and colleagues reported a Robot Operating System-based simulator for testing the enhanced autonomous navigation of the Mars 2020 rover, describing the new surface-mobility software developed for enhanced navigation and the testing framework built to qualify it [\[95\]](#ref-95). Verma and colleagues reported that autonomous robotics is driving Perseverance rover's progress on Mars, documenting that the AutoNav system evaluated the large majority of the distance driven in early operations [\[99\]](#ref-99). The method in Toupet is a verification-infrastructure study; the method in Verma is operational reporting. The finding is that autonomous navigation on Perseverance operated at a scale and autonomy beyond the Mars Exploration Rovers, with most driving evaluated autonomously. The limitation for cost reading is that, while Toupet's testing-framework paper speaks indirectly to the verification effort, neither reports a clean development-cost figure. For the spine, Perseverance is the most capable surface-navigation observation, and the explicit lineage from Mars Exploration Rover visual odometry to Perseverance AutoNav is documented heritage.

The class also contains the most striking single heritage extension, powered autonomous flight. Balaram and colleagues described the Ingenuity helicopter on the Perseverance rover, the first powered, autonomous flight demonstration on another planet, flown as a technology demonstration in 2021 [\[7\]](#ref-7). The method is a system-description-and-flight-results account; the finding is that autonomous powered flight was achievable on Mars, an extension of the autonomous-navigation class into a new flight regime. The limitation is that Ingenuity, as a technology demonstration, has a cost structure that may differ from operational deployments, which the design's reliability flagging must record. For the spine, Ingenuity is the frontier observation of the class and a vivid example of heritage enabling a capability that would have been unthinkable as a first-of-kind standalone development.

The class connects outward to a large terrestrial autonomous-navigation and sensing literature, relevant as propositional context but at the boundary of the population. Zhao and colleagues reported a high-accuracy autonomous navigation scheme for the Mars rover integrating inertial, visual, and celestial navigation [\[112\]](#ref-112). The wider reviews of sensor fusion for autonomous vehicles [\[110\]](#ref-110), of unmanned aerial vehicles [\[76\]](#ref-76), and of deep learning concepts and architectures [\[1\]](#ref-1) document the terrestrial methods that increasingly feed space navigation. The method in these is review and synthesis; the finding is that the propositional base for autonomous navigation is now vast and cross-domain. Their limitation for this study is sharp: terrestrial autonomy lies outside the NASA and JPL deep-space population that defines external validity, and these sources must not be allowed to inflate the heritage count. They are reviewed to make the Mokyr point that the navigation class sits on the deepest and most actively codified propositional base of any class, while being explicitly excluded from the within-class heritage tally.

Table 3.2 summarizes the navigation chain.

**Table 3.2. Autonomous navigation heritage chain.**

| Episode / source | Year | Role in chain | Cost-evidence character |
|---|---|---|---|
| Rover-navigation prototypes [\[101\]](#ref-101), [\[45\]](#ref-45), [\[52\]](#ref-52), [\[48\]](#ref-48) | 1997-2003 | Deepest pre-flight propositional base | None (prototypes) |
| MER visual odometry / AutoNav [\[68\]](#ref-68), [\[69\]](#ref-69), [\[26\]](#ref-26), [\[9\]](#ref-9), [\[71\]](#ref-71) | 2003-2007 | First dense flight cluster | Operational results, no NRE figure |
| Mars 2020 enhanced navigation [\[95\]](#ref-95), [\[99\]](#ref-99) | 2020-2023 | Most capable surface navigation | Verification-framework + operational reporting |
| Ingenuity powered flight [\[7\]](#ref-7) | 2021 | Frontier extension of class | Tech-demo cost structure flagged |
| Terrestrial/cross-domain context [\[112\]](#ref-112), [\[110\]](#ref-110), [\[76\]](#ref-76), [\[1\]](#ref-1) | 2019-2023 | Propositional context, excluded from count | Reviews, out-of-population |

One source in the class speaks directly to the verification-infrastructure cost that the heritage argument turns on, and it deserves separate emphasis. Toupet and colleagues built the Robot Operating System-based simulator specifically because the new enhanced-navigation algorithms had to be prototyped and tested before realistic flight-software testbeds existed, and they describe choosing an open, reusable robotics environment to do it [\[95\]](#ref-95). The method is the construction of a reusable test harness; the finding, read for this study, is that a substantial share of the navigation-qualification effort is verification infrastructure, and that the Mars 2020 team deliberately built that infrastructure to be reusable rather than bespoke. This is the navigation-class instance of the same verification-cost mechanism named in the planning class, and the navigation-class instance of the Mokyr codification moderator: reusable test infrastructure is exactly the codified prescriptive knowledge whose presence should lower the next episode's qualification cost. The limitation is that the paper reports the harness, not its cost or its measured saving. It is the strongest in-class hint that the heritage mechanism operates through verification reuse, and it is silent on magnitude.

The reasoning that closes the section is that navigation gives the design its deepest within-class chain and therefore its best statistical power, but also its hardest identification problem. Because the class accumulated both the most flight heritage and the most general technological maturation over the same decades, the slope estimated for navigation is the one most exposed to the confound between heritage and a secular trend in computing and software. The confound is not abstract: the same two decades that saw visual odometry give way to enhanced AutoNav also saw radiation-hardened onboard processing capacity grow by orders of magnitude, and a cheaper later episode could reflect that cheaper computing rather than reusable navigation heritage. Decade fixed effects absorb the part of that trend common to all classes, but a within-decade, navigation-specific computing trend could remain, which is why the maturation covariate and the cross-class robustness specification are not optional for this class. This is not a reason to exclude the class; it is the reason the design carries decade fixed effects and a maturation covariate. Confidence in the navigation chronology is high. Confidence that any of these sources has separated heritage from maturation in a cost measurement is zero, because none attempts a cost measurement at all.

## 3.4 Autonomous fault detection, isolation, and recovery

This section claims that autonomous fault detection, isolation, and recovery is a real and separately identifiable capability class with its own heritage chain, that it is the class whose verification cost is most plausibly large and most plausibly reusable, and that its literature is methodologically rich but thinner on operational flight episodes than the navigation and planning classes, which constrains the within-class power the panel can offer for it.

The flight origin of the class is shared with the planning class, because the Remote Agent integrated model-based diagnosis and reconfiguration alongside planning [\[83\]](#ref-83), [\[13\]](#ref-13). The fault-management function was a first-class part of that first flight, and the lessons-learned record's report of surprising integration cost applies to the fault-management component as much as to the planner [\[11\]](#ref-11). This shared origin matters for the design: an episode that fields more than one capability must be assigned to its dominant class or split, and Deep Space 1 is the canonical case where that rule bites. The fault-management heritage and the planning heritage both originate at the same flight, and the cumulative-heritage counts for the two classes must be constructed without double-counting the platform while still crediting each class with its origin.

The pre-Remote-Agent heritage of spacecraft fault tolerance is documented in the Cassini fault-tolerance work. Brown and colleagues described attitude and articulation control for the Cassini spacecraft as a fault-tolerance design, embedding autonomous failure detection, isolation, and recovery algorithms in object-oriented flight-control software [\[16\]](#ref-16). The method is a design overview; the finding is that autonomous fault tolerance was an established, if hand-crafted, flight capability before the model-based approaches arrived. The limitation is that traditional fault tolerance of this kind was bespoke per mission, which under Mokyr's framework is the flat-curve case: a capability re-implemented from scratch each time, with little codified reuse, should show little cost decline. Cassini is thus both a heritage antecedent and a theoretical contrast case for the model-based approaches that follow.

The model-based turn is the heritage the design predicts should bend the cost curve, if anything does. Kolcio and Fesq reported a model-based off-nominal state isolation and detection system for autonomous fault management, arguing that model-based systems provide better fault identification than traditional limit-checking and, for this study's purposes most of all, that the model can be reused and extended rather than rebuilt [\[59\]](#ref-59). The companion paper described the model-based fault detection and isolation system for increased autonomy [\[58\]](#ref-58). The method is system design with an explicit reuse argument; the finding is that model-based fault management both improves identification and offers a codified, transferable artifact. The limitation is that these are design-and-demonstration papers, not deployed-cost studies. For the spine and for the Mokyr moderator, they are the most relevant sources in the class: they describe exactly the codification-and-reuse infrastructure whose presence should steepen a class's cost-decline curve.
The class has a flight-demonstration arm on small platforms and a broad methodological hinterland. Fesq and colleagues reported extended-mission technology demonstrations using the ASTERIA spacecraft, a CubeSat that demonstrated autonomous capabilities including fault-relevant onboard functions [\[38\]](#ref-38). Wander and Forstner reported fault detection, isolation, and recovery onboard spacecraft using cognitive automation, categorizing current applications by level of autonomy [\[100\]](#ref-100). Bozzano and colleagues reported spacecraft early-design validation using formal methods, addressing the verification side of fault management [\[15\]](#ref-15). The terrestrial parallel is D'Amato and colleagues' particle-filtering approach for fault detection and isolation of unmanned-aerial-vehicle inertial sensors [\[29\]](#ref-29). The method ranges from flight demonstration to formal verification to terrestrial algorithm design; the finding is that the class has an active and methodologically deep research front. The limitation is the now-familiar pair: small-platform and terrestrial work sits at or beyond the population boundary, and none of these reports a comparable autonomy-cost figure.

The reasoning that closes the section is that fault detection, isolation, and recovery is a genuine class with a real heritage chain, anchored at Deep Space 1, with deep pre-history in bespoke fault tolerance and a clear model-based-reuse turn that is the strongest in-class instance of the Mokyr codification mechanism. The class has fewer clean operational flight episodes than navigation or planning, which means the within-class panel for fault management will be small and the slope, if estimated, will carry wide intervals. This is a power limitation to be stated honestly, not hidden. Confidence that the class is real and ordered is high. Confidence that its cost consequence has been measured is zero, and confidence that the panel can estimate a precise slope for this thin class is, by design-stage anticipation, low.

## 3.5 Autonomous entry, descent, and landing hazard handling

The claim of this section is that autonomous entry, descent, and landing hazard handling is the capability class whose maturation is most explicitly tracked against technology-readiness levels and cost in the published record, which makes it the class where the heritage-to-cost relationship is closest to the surface of the literature, even though, as everywhere else, no source closes the loop with a fitted curve.

The class is anchored by the Autonomous Landing and Hazard Avoidance Technology program, which is unusual in the corpus for stating its readiness-and-cost intent explicitly. Epp and Smith introduced the program with the goal of placing humans and cargo safely and precisely on the lunar surface with hazard-avoidance capability [\[31\]](#ref-31). Epp, Robertson, and Brady reported the program's development of an autonomous system combining guidance, navigation, and control with terrain sensing [\[32\]](#ref-32). Epp, Robertson, and Carson reported real-time hazard detection and avoidance demonstration for a planetary lander and stated the program's charter as maturing the autonomous system to technology-readiness level six [\[33\]](#ref-33). The method across the program papers is staged technology maturation with explicit readiness milestones; the finding is that hazard detection and avoidance was carried deliberately up the readiness scale through a sequence of demonstrations. The limitation is that the readiness milestones are not dollar figures, and the program's cost is not reported as autonomy-specific recurring engineering. This is the class where the literature comes closest to the dependent variable on its own terms, because the maturation was explicitly managed and reported, which is why the technology-readiness-level history will be an especially informative covariate for this class.

The sensing technology that the class depends on has its own heritage line. Amzajerdian and colleagues reported lidar sensors for autonomous landing and hazard avoidance developed under the same program, describing the imaging flash lidar, Doppler lidar, and laser altimeter that generate the terrain maps hazard avoidance needs [\[2\]](#ref-2). The method is sensor development and characterization; the finding is that the perception substrate for hazard handling matured alongside the guidance-and-control substrate. The limitation is that sensor papers report performance, not system cost. For the spine, this establishes that the class's heritage has both a perception arm and a guidance arm, which the capability-class assignment must keep coherent.

The class is genuinely international, which both enriches the heritage record and bounds external validity. Grover and colleagues described the Phoenix entry, descent, and landing system architecture, a NASA Mars-landing heritage point [\[47\]](#ref-47). Huang and colleagues reported the Tianwen-1 guidance, navigation, and control for Mars entry, descent, and landing, including actual flight results [\[51\]](#ref-51). Yu and colleagues reported autonomous hazard avoidance control for the Chang'E-3 lunar soft landing, the first Chinese soft landing on a celestial body, using onboard image data for relay hazard avoidance [\[111\]](#ref-111). Maass and colleagues reported a crater-navigation system for autonomous precision landing on the Moon developed at the German Aerospace Center [\[67\]](#ref-67). Lunghi and colleagues reported a multilayer-perceptron hazard detector for vision-based autonomous planetary landing [\[66\]](#ref-66). The method across these ranges from flight-system description to flight results to algorithm development; the finding is that autonomous landing-hazard handling is a globally pursued capability with multiple independent heritage lines. The limitation for this study is decisive on population: non-NASA, non-JPL landing systems are outside the within-class heritage that defines \(\text{CumHeritage}\) for the study's population, even though they are part of the world's autonomy record. They are reviewed to establish that the class is real and active and to make explicit the boundary the design draws around its population.

The international record is also the class's sharpest illustration of why the population boundary is a measurement decision and not a parochial one. The Tianwen-1, Chang'E-3, and German crater-navigation systems are genuine, flown or demonstrated, autonomous hazard-handling capabilities [\[51\]](#ref-51), [\[111\]](#ref-111), [\[67\]](#ref-67). Counted into \(\text{CumHeritage}\), they would inflate the heritage stock of later NASA and JPL episodes with knowledge those episodes could not necessarily reuse, because cross-agency and cross-national software and design patterns do not transfer freely. Excluded, they undercount the true world stock of hazard-handling knowledge. The design resolves this by defining the population as NASA and JPL deep-space and planetary autonomy and counting only within-population, within-class heritage, which is conservative by construction: it credits the later episode only with heritage it could plausibly have reused, and it therefore biases the estimated slope toward zero. The international record is reviewed here so that this exclusion is a stated, defensible choice rather than a silent omission, and so that the external-validity claim, that any fitted slope generalizes to the NASA and JPL population and not beyond it, is grounded in the literature rather than asserted.

The reasoning that closes the section is that entry, descent, and landing hazard handling is the class where heritage, readiness, and cost are most explicitly co-managed in the literature, through the Autonomous Landing and Hazard Avoidance Technology program's staged maturation. The program's reporting against technology-readiness milestones is the closest the corpus comes to a quasi-cost signal, because advancing a capability from one readiness level to the next is itself the expenditure of qualification effort, and the program reports that advancement in an ordered, dated form [\[31\]](#ref-31), [\[32\]](#ref-32), [\[33\]](#ref-33). That makes it the class where the technology-readiness-level covariate will do the most work in separating heritage from maturity, and the class where the design's distinction between within-population NASA and JPL heritage and the broader international record must be most carefully policed. The caution carried from Chapter 4 is that the readiness scale is ordinal and non-monetary, so a readiness transition is a maturity signal and not a cost figure; it informs the covariate, it does not become the dependent variable. Confidence in the chronology and in the readiness-tracking is high, higher than for any other class. Confidence that the cost has been read against heritage is, once more, zero.

## 3.6 The codified-knowledge substrate: the Mokyr moderator

The claim of this section is that the reusable flight-software substrate, principally the Core Flight System and the broader move toward modular, plug-and-play flight software, is the codified-knowledge infrastructure that the dissertation's theoretical framework, following Mokyr [\[114\]](#ref-114), [\[115\]](#ref-115), predicts should steepen the cost-decline curve wherever it is adopted, and that the literature documents the substrate's existence and reuse intent clearly while saying almost nothing about its cost consequence, which is itself a finding.

The substrate's origin is the Goddard Space Flight Center flight-software work. McComas described NASA Goddard's Core Flight System, the reusable flight-software framework intended to let projects compose flight software from certified, reusable components rather than rebuilding from scratch [\[73\]](#ref-73). The method is a framework description; the finding is that a deliberate, agency-level effort exists to codify and reuse flight software, which is the infrastructure Mokyr's framework identifies as the precondition for cheap reproduction and extension of a technique. The limitation, central to this study, is that the paper describes the framework and its reuse intent but does not measure the cost saved by reuse. The corpus footprint of the Core Flight System is modest, a point that stands out as an evidence gap, and that thinness is itself informative: the codified substrate whose presence should most steepen the cost curve is the part of the record least studied for its cost effect.

The substrate's intellectual lineage is the responsive-space and plug-and-play flight-software literature. Wilmot reported the implications of responsive space on flight-software architecture, arguing that rapid response requires reusable, pre-certified software modules rather than new development [\[106\]](#ref-106), and described a core plug-and-play architecture for reusable flight-software systems built explicitly to make development cost and schedule more predictable through reuse [\[105\]](#ref-105). The method is architectural; the finding is that the flight-software community framed reuse as the route to predictable cost and schedule, which is the heritage-lowers-cost claim stated about software infrastructure itself. The limitation is that these are architecture papers with cost-and-schedule aspirations rather than measured cost outcomes. They are the clearest statement in the corpus that the field believed reuse infrastructure would lower cost, and the clearest illustration that the belief was never tested as a fitted relationship.

The substrate's reach into small platforms and modular missions shows it is in active use. Latachi and colleagues reported reusable and reliable flight-control software for a fail-safe and cost-efficient CubeSat mission, explicitly pursuing reuse for cost efficiency [\[63\]](#ref-63). Gonzalez and colleagues reported an architecture-tracking approach to evaluate a modular and extensible flight software for CubeSat nanosatellites, aiming to increase reusability and reduce mission risk [\[46\]](#ref-46). The open-source simulation-framework work by Turner provided an extensible spacecraft-simulation environment under a public license to let researchers reuse and extend code [\[97\]](#ref-97). The method ranges from flight-software engineering to tool development; the finding is that modular, reusable flight software is a live and spreading practice. The limitation for this study is that the cost-efficiency claims are stated as design goals, not measured outcomes, and the platforms are often outside the deep-space population. They are reviewed to establish that the codification mechanism is real and active, so that the design's prediction of class-level heterogeneity, steeper curves where reuse infrastructure is present, has a concrete referent.

Two demonstration missions show the substrate enabling autonomy on low-cost platforms, which is the Mokyr prediction in flight. The ASTERIA CubeSat demonstrated autonomy on a platform built under a training program [\[38\]](#ref-38), the OPS-SAT CubeSat demonstrated onboard autonomy and machine learning at low cost and low bureaucracy [\[60\]](#ref-60), the Near-Earth Asteroid Scout packaged a deep-space mission into a CubeSat volume [\[65\]](#ref-65), and the Proba mission was conceived specifically to demonstrate the benefits of onboard autonomy with minimum ground involvement [\[93\]](#ref-93). The method in these is flight demonstration; the finding is that codified, reusable autonomy components increasingly let small, cheap platforms carry capabilities that once required flagship missions. The limitation is the population boundary and the absence of comparable cost figures. They are the qualitative embodiment of the moderator the design will test as effect heterogeneity across classes.

The reasoning that closes the section is that the codified-knowledge substrate is where Mokyr's framework predicts the steepest cost decline, and the literature confirms the substrate exists, is reused, and was built in the belief that reuse lowers cost, while measuring that cost effect nowhere. This is the moderator hypothesis in its rawest form: the field built the reuse infrastructure on the assumption it would pay, and never fitted the payoff. Confidence that the substrate exists and is reused is high. Confidence that its cost effect has been measured is zero, and the corpus is thinnest precisely here, which the design must treat as a known limitation on its ability to resolve the moderator.

## 3.7 Synthesis: the gap stated and the propositions that follow

The synthesis reads the five classes and the substrate as one body of evidence and states the gap precisely. Across every class, the literature delivers the same structure: a dated heritage chain, an explicit engineering narrative of building on what came before, and, at the origin of the oldest class, a direct statement that the first-of-kind cost was large and surprising. Across every class, the literature delivers none of the following: a recurring-engineering cost figure per demonstration in a comparable form, a reading of any second-episode cost against its first-episode predecessor, or a fitted relationship between cost and cumulative heritage. The recurring heritage-lowers-cost claim is made qualitatively and repeatedly. The space-autonomy review by Gao and Chien states the claim in its general form, situating the demonstrations within a longer arc and asserting that flown heritage lowers the barrier to the next demonstration [\[42\]](#ref-42). That assertion, multiplied across the class literatures reviewed above, is the conventional wisdom this dissertation converts into a testable quantity.

The synthesis must also be explicit about what the autonomy literature is and is not, relative to the experience-curve method reviewed in Chapter 2, because the contribution lives exactly in that contrast. The experience-curve literature is a literature of measurement: it fits the logarithm of unit cost on the logarithm of cumulative output, validates the functional form across many technologies, and quantifies the forecasting error of the fitted slope. The autonomy literature reviewed in this chapter is a literature of capability: it documents what was demonstrated, how it worked, and how each demonstration built on the last, but it never expresses the heritage-to-cost relationship as a fitted quantity. The two literatures are complementary in exactly the way the dissertation needs. The autonomy literature supplies the cumulative-experience axis, the ordered, dated, within-class heritage record that becomes \(\ln(\text{CumHeritage})\). The experience-curve literature supplies the method, the log-log specification, the within-family fixed-effects discipline, and the validated expectation that learning effects produce a measurable negative slope. Neither literature alone can answer the research question. The autonomy literature has the heritage but not the method; the experience-curve literature has the method but has been applied to manufactured-hardware unit cost rather than to the recurring engineering cost of qualifying a software-intensive capability for flight. The join of the two, performed on the heritage spine this chapter has assembled, is the unoccupied space the dissertation enters. Read at the level of the literature, the case for that join is straightforward. The heritage assumption is pervasive yet nowhere measured, and portfolio decisions turn on it, so the question carries practical weight rather than merely methodological interest. A within-class experience curve measures precisely the learning effect the autonomy teams describe in narrative form, and the log-log fixed-effects specification is the best-validated way to separate that learning from the confounding influences of time and scale. What can be claimed is bounded honestly: the conservative counting rules and the design-stage estimates keep the inference within the limits the evidence supports.

Table 3.3 partitions the gap cleanly across the classes, with no class omitted and none double-counted, which is the structure the estimating equation requires.

**Table 3.3. Heritage record versus cost measurement, by capability class.**

| Capability class | Heritage chain documented | First-of-kind cost evidence | Within-class reuse documented | Cost read against heritage |
|---|---|---|---|---|
| Onboard planning and scheduling | Yes, complete (DS1 to EO-1 and on) | Yes, narrative [\[11\]](#ref-11) | Yes [\[24\]](#ref-24) | No |
| Science target selection | Yes, clean pair (AEGIS x2) | Operational only | Yes, same capability [\[40\]](#ref-40) | No |
| Autonomous navigation | Yes, deepest chain | Operational only | Yes [\[69\]](#ref-69), [\[99\]](#ref-99) | No |
| Fault detection, isolation, recovery | Yes, but fewer flight episodes | Shared narrative [\[11\]](#ref-11) | Yes, model-based reuse [\[59\]](#ref-59) | No |
| EDL hazard handling | Yes, TRL-tracked | TRL milestones, not cost [\[33\]](#ref-33) | Yes, staged maturation | No |
| Codified substrate (moderator) | Yes, reuse-by-design | Reuse intent, not cost | Yes [\[105\]](#ref-105), [\[73\]](#ref-73) | No |

The final column is the gap. It is empty in every row. The literature has assembled, class by class, exactly the heritage chronology that the independent variable \(\text{CumHeritage}\) requires, and it has never measured the dependent variable against it. This is the join that no published study performs, and it is the contribution the dissertation makes: to read the autonomy literature as the empirical spine of a Wright-type experience-curve study and fit the slope \(\beta\) that the field has assumed without estimating.

From this gap, four propositions follow that carry into the chapters on data, design, and analysis. They are stated as propositions, not findings, because the study is at the design stage and no coefficient has been fitted.

The first proposition is that the demonstration literature reviewed here is sufficient to construct the cumulative-heritage variable for each capability class, because every class supplies a dated, ordered sequence of flight events with documented within-class reuse. The five heritage chains of Sections 3.1 through 3.5 supply the basis, and an experience curve requires only a dated cumulative-experience count, which these chains provide. The count is within-class and within-population and therefore conservative. The one objection it must survive is that software and design patterns cross class boundaries, which the forward-only and within-class counting rules deliberately undercount, biasing the slope toward zero and making any rejection of the flat-cost null more credible. Confidence is high that the heritage variable can be built.
The second proposition is that the literature cannot supply the dependent variable directly, because no source reports autonomy-specific recurring-engineering cost in a comparable form. The basis is the uniformly empty final column of Table 3.3: a fitted curve needs a measured cost, and the literature does not provide one. The dependent variable must therefore be constructed by the three-layer normalization developed in Chapter 4, drawing on the parametric cost-model literature rather than on any single autonomy paper. Confidence that the cost must be built rather than quoted is high. This is the load-bearing limitation of the whole design.

The third proposition is that the technology-readiness-level history is a covariate, not a cost proxy. The entry-descent-and-landing class tracks readiness milestones explicitly while leaving cost unrecorded, and the readiness scale is ordinal and non-monetary. The basis is the Autonomous Landing and Hazard Avoidance Technology program's explicit readiness charter [\[33\]](#ref-33) and the wider readiness-construct literature carried into Chapter 4. Readiness tracks maturity, which can confound heritage, so it must be controlled rather than equated with cost. This answers the claim that later, higher-readiness projects are simply cheaper, a claim the maturation covariate is designed to separate from heritage. Confidence is moderate, because readiness data are uneven across classes.

The fourth proposition is that effect heterogeneity across classes is the most policy-relevant secondary question the design can address. The codified-substrate literature documents reuse infrastructure as present in some classes and thin in others, which is the Mokyr moderator made concrete. The basis is Section 3.6, and the principle, developed in Chapter 2, holds that codification and reuse should steepen the cost-decline curve. The qualification is sharp: the corpus is thinnest exactly here, so the moderator can be tested only weakly, and a null result on heterogeneity would be uninformative rather than evidence of homogeneity. Confidence in the moderator's relevance is high; confidence that the panel can resolve it is low. That asymmetry is stated honestly so it constrains interpretation in advance.

These four propositions bridge the literature to the measurement. The literature reviewed in this chapter establishes that the heritage record exists, is ordered, is documented as reuse, and is assumed to lower cost. It establishes equally that the cost consequence of that heritage has never been measured. The chapters that follow take the heritage record assembled here and build the dependent variable, the identification strategy, and the analysis plan that will, for the first time, read autonomy qualification cost against cumulative flight heritage and return either a fitted experience-curve slope or a credible failure to find one. Either outcome answers, with a measurement, a question the literature has answered only by assertion.



# Chapter 4. Data and Measurement

## 4.0 The chapter's answer, stated first

This chapter establishes that the measurement at the center of this dissertation is constructible from real, named, and auditable sources, and that the central construct, the recurring engineering cost to qualify a class of onboard autonomy for flight, can be operationalized with explicit, reproducible rules even though no single open source reports it cleanly. The answer the chapter delivers is conditional and honest. The panel that the estimating equation \(\ln(\text{Cost}_{icd}) = \alpha + \beta \, \ln(\text{CumHeritage}_{icd}) + \gamma_c + \delta_d + \epsilon_{icd}\) requires can be assembled from four named datasets; every variable in that equation can be given an operational definition with a stated source and scale; and the dominant measurement risk, error in the dependent variable, can be made visible through a three-layer normalization procedure with a per-observation reliability flag rather than hidden inside a single point estimate. The chapter does not claim that the resulting cost figures are exact. It claims something weaker and more defensible: that the construction is transparent, that each figure carries a documented provenance and a reliability grade, and that the weakest figures can be down-weighted in a pre-registered robustness specification so that the estimated slope \(\beta\) does not rest on the least trustworthy observations.

The current state is that NASA and the Jet Propulsion Laboratory treat each flown autonomy capability as a heritage asset assumed to lower the next demonstration's cost, but the cost itself is recorded inconsistently across projects and is sometimes not disclosed at all, so the assumption has never been tested against a constructed dataset [\[42\]](#ref-42), [\[11\]](#ref-11). The desired state is a panel in which each observation is a capability-class development episode with a normalized, constant-year, scope-adjusted autonomy qualification cost, a forward-only cumulative-heritage count, a capability-class label, a decade label, and a maturation covariate, all traceable to a public record. The gap between those two states is a data-construction gap, not a modeling gap: the regression is standard, but the inputs must be built. Leaving the gap unaddressed means the experience-curve slope that the prospectus promises cannot be estimated at all, because the dependent variable would not exist in usable form. This chapter closes the construction gap at the level of design. It specifies the sources, the unit of analysis, the operationalization of every variable, the data-quality and validation procedures, and the access and ethics posture, and it carries the prospectus data section forward by elaborating each element into a procedure that a reviewer who did not build the panel could reproduce and check.

Two scope commitments govern the chapter and are stated once here so they are not relitigated below. First, the chapter is design-stage. Every cost figure, heritage count, and panel dimension discussed as a worked example illustrates the procedure rather than reporting an executed measurement; no observation's normalized cost is reported as a fitted or final value, because the panel assembly and cost normalization are not complete. Second, the architecture layer is out of scope. This is an econometric measurement study, and no real DoDAF or BEA capability, system, or data-service exchange is the subject of the dissertation, so the chapter describes datasets and variables in the vocabulary of empirical economics rather than forcing an architecture chain onto a cost panel.

## 4.1 The four named datasets

The panel is assembled from four named sources, each contributing a distinct and non-substitutable element. No single source contains the full panel; the contribution of the chapter is in part the specification of how the four are joined. This section treats each source in depth along five dimensions the prospectus named but did not fully develop: provenance, access path, coverage, unit of analysis as the source records it, and known biases that the source carries into the panel.

### 4.1.1 NASA TechPort project records and technology-readiness-level histories

**Provenance.** TechPort is NASA's authoritative system of record for technology development projects. The agency maintains it as the inventory and reporting backbone for technology investments across the mission directorates, and it records project descriptions, organizational ownership, start and end dates, technology-readiness-level entry and exit estimates, and a taxonomy classification that maps each project into NASA's technology area structure. Because it is an internal system of record exposed through a public interface, its provenance is institutional rather than scholarly: the projects themselves author the entries and the agency curates them, which makes them authoritative for what NASA considers a project and how NASA classifies it, but dependent on the projects' own reporting discipline for completeness.

**Access path.** The data are reached through the TechPort public application programming interface and the bulk data export available from the TechPort public portal. The bulk export is preferred for panel construction because it permits a reproducible snapshot: the entire project inventory can be downloaded at a dated version, retained, and re-queried, which is the standard a reproducible cost study requires. The application programming interface is used for targeted field retrieval and for confirming that a project's recorded technology-readiness-level history matches the snapshot.

**Coverage.** TechPort covers NASA-funded technology projects, a broader set than the flight-demonstrated autonomy episodes this study analyzes. That breadth is useful, because it supplies the denominator against which the autonomy episodes are identified, and it supplies the maturation covariate and project-timing fields for episodes that are otherwise documented only in the technical literature. The coverage is also the source's first limitation: TechPort records technology projects, and a flown autonomy capability may have been developed inside a flight project that TechPort does not classify as a technology project, in which case the episode is present in the published literature but thin or absent in TechPort. The construction procedure therefore treats TechPort as the project-inventory and covariate source, not as the sole arbiter of which episodes exist.

**Unit as recorded.** TechPort's native unit is the technology project, identified by a project record with a unique identifier, a start and end date, an owning organization, and a technology-readiness-level trajectory. This is close to, but not identical with, the study's unit of analysis, the capability-class development episode. The mapping from TechPort projects to episodes is a construction step, not a lookup, and Section 4.2 specifies it.

**Known biases.** Three biases enter the panel through TechPort. First, self-reported technology-readiness levels are applied inconsistently across projects and drift in meaning as the scale is used outside its original setting, so the maturation covariate carries measurement error that the study must treat as such rather than as a clean number [\[53\]](#ref-53), [\[92\]](#ref-92). Second, project-timing fields reflect administrative start and end dates that may not coincide with the engineering start of the autonomy development effort, which matters because the forward-only heritage rule keys on development-start timing. Third, coverage is denser for recent projects than for the earliest demonstrations, because the system of record was populated more completely over time, which biases the panel toward better documentation in later decades and is one reason the decade fixed effects are necessary.

### 4.1.2 NTRS autonomy demonstration reports

**Provenance.** The NASA Technical Reports Server is the agency's repository of technical and lessons-learned documentation. For this study it is the primary-source archive that supplies the engineering record of the named autonomy demonstrations: the Remote Agent Experiment on Deep Space 1 [\[12\]](#ref-12), [\[88\]](#ref-88), the Deep Space 1 autonomy-infusion lessons-learned record [\[11\]](#ref-11), the EO-1 Autonomous Sciencecraft Experiment, AEGIS, and Ingenuity. Its provenance is the strongest available for the qualitative content the study needs, because these are the project teams' own technical reports rather than secondary summaries.

**Access path.** NTRS is reached through its citations search application programming interface, which returns document metadata and, where the document is releasable, a downloadable full text. The construction procedure records, for each retrieved report, the NTRS citation identifier so that every qualitative claim about an episode's scope or cost can be traced to a specific document, which is the auditability standard the chapter commits to.

**Coverage.** NTRS coverage for the named demonstrations is rich on the engineering and lessons-learned side and sparse on the cost side. The lessons-learned record from Deep Space 1, for instance, documents the organizational difficulty of inserting system-level autonomy into a flight project and reports that the impact of that insertion surprised the project, which is direct primary evidence on the non-recurring engineering and organizational cost of a first-of-kind autonomy capability [\[11\]](#ref-11). That is precisely the quantity the dependent variable is intended to capture, and NTRS supplies the qualitative scope from which a normalized figure is built, but it rarely supplies a clean dollar figure for the autonomy portion alone.

**Unit as recorded.** NTRS records documents, not episodes. An episode is reconstructed from one or several NTRS documents plus the peer-reviewed literature. The construction procedure therefore treats NTRS as the source of the heritage chronology, the capability definitions, and the scope information used to size each episode, and it cross-references each episode to the document set that establishes it.

**Known biases.** NTRS carries a documentation-survivorship bias: demonstrations that were written up thoroughly are better represented than those that were not, and successful demonstrations are more likely to have generated celebratory technical reports than quietly cancelled efforts. This biases the panel toward documented successes and against the abandoned attempts whose cost would, if recorded, inform the true cost of qualifying a capability. The study cannot recover the undocumented attempts; it acknowledges the bias and confines its claim to the population of documented flight demonstrations.

### 4.1.3 NICM-class parametric development-cost estimates
**Provenance.** The NASA Instrument Cost Model and the broader family of single-variable and multivariable parametric cost models for space systems are the agency-standard estimators of development cost from measurable scope drivers [\[77\]](#ref-77), [\[89\]](#ref-89), [\[90\]](#ref-90), [\[96\]](#ref-96), [\[39\]](#ref-39). Their provenance is methodological. They are calibrated regressions of historical development cost on physical and programmatic scope variables, published and maintained as cost-estimating tools. The NICM family estimates instrument development cost; the single-variable space-telescope cost models of Stahl and colleagues exemplify the parametric approach the study adopts as its normalization basis [\[89\]](#ref-89), and the survey of space-telescope cost models and of early-phase hardware cost-estimation methods situates that approach within the wider parametric-estimation literature [\[90\]](#ref-90), [\[96\]](#ref-96).

**Access path.** The published parametric models supply the functional forms and the calibrated coefficients. Where a NICM-class tool is available to the analyst, it supplies estimates directly; where only the published model is available, the estimate is reconstructed from the published form. The construction procedure records, for each imputed figure, which model and which scope inputs produced it, so that the imputation is reproducible.

**Coverage.** Parametric cost models cover the development cost of space hardware and instruments well, because that is the population on which they were calibrated. They cover the non-recurring engineering cost of a software-intensive autonomy capability less directly, because autonomy qualification is not the population the NICM family was built on. This is the source's central limitation for this study and the reason the dependent-variable construction is the most delicate step in the entire design. The parametric model is used here not to estimate autonomy cost as though autonomy were an instrument, but to provide a consistent, scope-anchored normalization basis and an imputation of last resort when the autonomy-specific figure cannot be extracted from project documentation. Section 4.3 develops this point.

**Unit as recorded.** Parametric models estimate a cost per defined scope unit, typically an instrument or a subsystem characterized by mass, power, data rate, or analogous drivers. The study uses this per-scope basis to normalize development cost so that the residual variation reflects autonomy qualification effort rather than platform or instrument scale.

**Known biases.** Parametric cost models carry the bias of their calibration population. Applied outside that population, their estimates are extrapolations whose error is larger than the in-sample fit suggests. The technology-readiness-level-conditioned cost-estimation literature documents that cost and maturity interact in ways a single parametric form may not capture, which is why the study treats the maturation covariate separately rather than folding it into the cost estimate [\[17\]](#ref-17), [\[57\]](#ref-57), [\[8\]](#ref-8). The reliability flag described in Section 4.3 prevents these extrapolated figures from being treated as equal in trustworthiness to figures extracted directly from project records.

### 4.1.4 Published mission and autonomy literature

**Provenance.** The peer-reviewed and conference literature on the named demonstrations supplies capability scope, demonstration dates, and the heritage links between episodes. Its provenance is scholarly and, for the central demonstrations, authored by the engineering teams themselves. The EO-1 Autonomous Sciencecraft Experiment papers [\[24\]](#ref-24), the AEGIS deployments on Opportunity and on the Curiosity ChemCam instrument [\[34\]](#ref-34), [\[40\]](#ref-40), the Perseverance autonomous-navigation record [\[99\]](#ref-99), and the Ingenuity powered-flight demonstration [\[7\]](#ref-7) are primary scholarly sources for the chronology and scope that feed the heritage variable.

**Access path.** These sources are reached through their digital object identifiers and through the standard scholarly databases. Each is retained with its identifier so that every scope and dating decision is traceable.

**Coverage.** The literature covers the demonstrations' scientific and engineering content thoroughly and their cost sparsely, mirroring NTRS. The space-autonomy review by Gao and Chien situates the demonstrations in a longer arc and states the recurring heritage-lowers-cost claim that this study converts into a testable quantity [\[42\]](#ref-42), which makes the review a coverage backbone for the chronology even though it reports no cost figures.

**Unit as recorded.** The literature records demonstrations and capabilities, which the study maps to episodes. The mapping is the same construction step described for NTRS, and the two sources are used jointly to fix each episode's date, scope, and capability class.

**Known biases.** The literature shares NTRS's survivorship bias toward documented successes and adds a publication-incentive bias: novel first-of-kind demonstrations attract more publication than incremental reuses of an existing capability, which can make the heritage link between a first deployment and a second deployment thinner in the record than the engineering reality. The AEGIS sequence is the partial exception and the cleanest within-class heritage pair, because the second deployment on ChemCam is documented as building on the first on Opportunity [\[34\]](#ref-34), [\[40\]](#ref-40). That documentation is what makes the pair the central illustrative case for the heritage mechanism.

### 4.1.5 How the four sources join

The join is keyed on the episode. TechPort supplies the project inventory, the maturation covariate, and administrative timing; NTRS and the published literature supply the heritage chronology, the capability definitions, and the qualitative scope used to size each episode; the parametric cost models supply the normalization basis and the imputation of last resort for the dependent variable. A claim about any single episode is therefore typically supported by more than one source, and the construction procedure records the source set for each episode so that the panel is auditable observation by observation. Confidence in this join is moderate at the design stage. The join logic is fully specified and the sources are real and accessible, but the join has not been executed across the full episode set, so the realized completeness of the merged panel is not yet known. What would raise confidence is the completed merge with a documented count of episodes for which all four sources resolve. What would lower it is discovery, during assembly, that a material fraction of episodes resolve in the literature but not in TechPort, which would weaken the maturation covariate and is flagged as a watch item.

## 4.2 Unit of analysis and the capability-class taxonomy

The unit of analysis is the capability-class development episode, defined exactly as in the prospectus: a single mission or technology demonstration that fields one defined onboard autonomy capability, assigned to exactly one primary capability class, with multi-capability episodes split or flagged to their dominant class. This section elaborates the taxonomy and the assignment rule into an operational procedure, because capability-class assignment is one of the four named data limitations and because the within-class fixed-effects estimator depends on the assignment being defensible.

### 4.2.1 The five capability classes

The taxonomy partitions onboard autonomy into five coherent functional classes, carried unchanged throughout:

1. **Onboard planning and scheduling.** The capability to generate, repair, and execute activity plans onboard rather than uplinking a fully specified command sequence. Its origin observation is the Remote Agent Experiment, which integrated onboard planning, execution, and fault recovery [\[12\]](#ref-12), and its operational transition is the EO-1 Autonomous Sciencecraft Experiment, which moved onboard detection and replanning into routine Earth observation [\[24\]](#ref-24).

2. **Autonomous science target selection.** The capability to choose science targets onboard from sensor data without ground-in-the-loop targeting. Its origin and within-class heritage pair is the AEGIS deployment on Opportunity and then on the Curiosity ChemCam instrument [\[34\]](#ref-34), [\[40\]](#ref-40).

3. **Autonomous surface or in-flight navigation.** The capability to evaluate terrain or flight conditions and select a safe path onboard. Its most recent and most capable observations are Perseverance autonomous navigation at scale and Ingenuity autonomous powered flight [\[99\]](#ref-99), [\[7\]](#ref-7), with the Mars 2020 enhanced-navigation work documented in the supporting record [\[95\]](#ref-95).

4. **Autonomous fault detection, isolation, and recovery.** The capability to detect off-nominal states and recover onboard, originating in the Remote Agent's model-based fault recovery and advanced in model-based off-nominal state isolation and detection work [\[59\]](#ref-59).

5. **Autonomous entry-descent-and-landing hazard handling.** The capability to detect and avoid landing hazards autonomously during the brief, communication-isolated descent phase, developed through the Autonomous Landing and Hazard Avoidance Technology program and related lidar-based hazard-detection work [\[31\]](#ref-31), [\[32\]](#ref-32), [\[2\]](#ref-2), [\[33\]](#ref-33), with the entry-descent-and-landing architecture context supplied by the Phoenix and Tianwen-1 landing-system records [\[47\]](#ref-47), [\[51\]](#ref-51).

### 4.2.2 The assignment rule

Each episode is assigned to exactly one primary capability class by the dominant-capability rule: the class that the episode's primary documented objective fields. Where an episode fields more than one capability, two treatments are available and the choice is recorded per episode. The first splits the episode into class-specific sub-episodes when the documentation separates the development effort by capability; the second assigns the episode to its dominant class and sets a multi-capability flag when the effort cannot be separated. The Remote Agent Experiment is the canonical multi-capability case, because it fielded onboard planning, execution, and fault recovery together [\[12\]](#ref-12), [\[83\]](#ref-83). The construction procedure assigns it to onboard planning and scheduling as its dominant documented objective and flags it as also originating the fault-detection-and-recovery class, with the flag recorded so that a robustness check can test sensitivity to the assignment.

This assignment is a construct-validity decision, and the chapter states its limitation plainly. The class boundaries are defensible but contestable, particularly between planning-and-scheduling and fault-detection-and-recovery, which the Remote Agent fielded as an integrated system, and between science-target-selection and navigation, which share onboard perception machinery. The study's response is threefold: the assignment rule is fixed before estimation so it is not tuned to produce a result; every multi-capability episode carries a flag; and the analysis plan includes a robustness specification in which contested assignments are reclassified to test whether the slope depends on a boundary call. Confidence in the taxonomy is moderate. The five classes are functionally coherent and grounded in the demonstration record, but the small number of episodes per class means a single contested assignment can move a within-class estimate, which is why the influence diagnostics in the analysis plan are not optional.
### 4.2.3 The measurement table

Table 4.1 gives the full operationalization of every variable in the estimating equation. It is the chapter's central deliverable. It converts the abstract equation into a set of measurable quantities, each with an operational definition, a source, and a scale, and it is the object against which a reviewer checks that the model's inputs are real and constructible.

**Table 4.1. Operationalization of every variable in the estimating equation \(\ln(\text{Cost}_{icd}) = \alpha + \beta \, \ln(\text{CumHeritage}_{icd}) + \gamma_c + \delta_d + \epsilon_{icd}\).**

| Construct | Operational definition | Source(s) | Scale / units | Enters model as |
|---|---|---|---|---|
| Dependent variable: autonomy qualification cost (\(\text{Cost}_{icd}\)) | Recurring non-recurring-engineering cost to qualify the capability for flight, normalized by capability scope on a NICM-class basis, expressed in constant-year dollars; built by the three-layer procedure of Section 4.3 with a per-observation reliability flag | Project documentation (NTRS, published literature) for direct extraction; NICM-class parametric models for imputation [\[77\]](#ref-77), [\[89\]](#ref-89), [\[90\]](#ref-90) | Continuous, constant-year US dollars per unit scope | \(\ln(\text{Cost}_{icd})\) |
| Independent variable: cumulative flight-demonstrated heritage (\(\text{CumHeritage}_{icd}\)) | Count of prior flight demonstrations in the same capability class that reached flight operation before the episode's development-start date; forward-only counting rule; first-in-class set to 1 before logging | NTRS chronology + published demonstration dates [\[12\]](#ref-12), [\[24\]](#ref-24), [\[34\]](#ref-34), [\[40\]](#ref-40), [\[99\]](#ref-99), [\[7\]](#ref-7) | Integer count, >= 1 | \(\ln(\text{CumHeritage}_{icd})\) |
| Capability-class fixed effects (\(\gamma_c\)) | Indicator for the episode's primary capability class (five classes); absorbs time-invariant baseline-cost and difficulty differences across classes | Taxonomy assignment (Section 4.2) from NTRS + literature | Categorical, 5 levels | Indicator set \(\gamma_c\) |
| Decade fixed effects (\(\delta_d\)) | Indicator for the decade of development start; absorbs economy-wide and agency-wide time trends in cost, tooling, and computing | TechPort development-start date; literature where TechPort is thin | Categorical, by decade | Indicator set \(\delta_d\) |
| Maturation covariate (robustness) | Technology-readiness level at development start | TechPort technology-readiness-level history [\[70\]](#ref-70), [\[81\]](#ref-81) | Ordinal, 1-9 | Covariate in robustness specification only |
| Reliability flag (auxiliary) | Which normalization layer produced the cost figure (direct / parametric-imputed / deflated-only) and an associated trust grade | Construction log (Section 4.3) | Ordinal, 3 layers + grade | Weight in inverse-imputation-error robustness specification |

The maturation covariate is deliberately confined to the robustness specification and never enters the baseline, for a construct reason the chapter states plainly. The technology-readiness-level scale is ordinal and non-monetary, it is applied inconsistently and loses precision as it travels outside its original setting, and it tracks maturity rather than the cost of advancing maturity [\[70\]](#ref-70), [\[81\]](#ref-81), [\[53\]](#ref-53). Treating it as a cost proxy would import its measurement error into the dependent variable. Treating it as a separate robustness covariate lets the study check whether the heritage effect is merely the simpler effect of later projects starting at higher maturity, without contaminating the baseline estimate. This is the disciplined use of the scale that the calibration and shortcomings literature recommends [\[92\]](#ref-92), [\[81\]](#ref-81).

## 4.3 The dependent variable: three-layer cost normalization and reliability flags

Constructing the dependent variable is the most delicate step in the design, and this section elaborates the prospectus's three-layer procedure into an operational, auditable method. The section argues the following.

The autonomy qualification cost can be operationalized as a normalized, constant-year, scope-adjusted figure whose measurement error is explicit and bounded rather than hidden, such that the estimated slope \(\beta\) can be made robust to the least reliable observations.

Public development-cost figures for autonomy capabilities are reported inconsistently. Some are full project costs that bundle platform, instrument, and autonomy; some are subsystem costs; some are not disclosed [\[11\]](#ref-11), [\[42\]](#ref-42). A defensible cost study cannot treat these heterogeneous figures as interchangeable, so the construction proceeds in three explicit layers, each recorded per observation.

The standard a cost study must meet is that a reviewer who did not build the panel can check it. A layered procedure that records, for each figure, which layer produced it and how reliable that layer is, makes the measurement auditable and lets the weakest figures be down-weighted. That is what makes the slope defensible against the objection that it rests on imputed numbers.

Parametric cost estimation is the agency-standard method for development cost when direct figures are unavailable, and its use here follows the established space-cost-modeling literature [\[77\]](#ref-77), [\[89\]](#ref-89), [\[90\]](#ref-90), [\[96\]](#ref-96), [\[39\]](#ref-39). The technology-readiness-conditioned cost-estimation literature establishes that cost and maturity interact and that the reliability of an estimate degrades when the estimator is applied outside its calibration population, which is the backing for the reliability flag [\[17\]](#ref-17), [\[57\]](#ref-57), [\[8\]](#ref-8), [\[18\]](#ref-18), [\[10\]](#ref-10), [\[87\]](#ref-87).

The procedure does not eliminate measurement error in the dependent variable. It makes the error explicit, graded, and down-weightable. The residual error remains and is the dominant construct-validity threat the study carries forward.

If, on assembly, the autonomy-specific cost cannot be extracted for the great majority of episodes, so that the panel is dominated by layer-two parametric imputations, the dependent variable would rest mostly on extrapolations of estimators calibrated on hardware rather than software-intensive autonomy, and the measurement's credibility would fall accordingly. The study's pre-commitment is to report the layer composition of the assembled panel so that this objection can be evaluated by the reader rather than concealed, and to treat a panel dominated by layer-two figures as yielding a weaker, not a stronger, conclusion.

### 4.3.1 Layer one: direct extraction of the autonomy non-recurring engineering portion

The first layer extracts, where the documentation permits, the autonomy-specific non-recurring engineering portion of development cost, separating it from platform and instrument cost. The Deep Space 1 lessons-learned record is the archetype: it documents the organizational and engineering cost of inserting system-level autonomy into a flight project as a distinct and surprising burden, which is direct qualitative evidence on the autonomy portion even where a single dollar figure is not stated [\[11\]](#ref-11). Where a project report states an autonomy-subsystem cost or a separable autonomy work-breakdown element, that figure is taken at layer one. Layer-one figures carry the highest reliability grade because they measure the construct most directly.

### 4.3.2 Layer two: parametric imputation of the autonomy development effort

Where the autonomy portion is not separately reported, the second layer imputes the autonomy development effort from the documented scope using a NICM-class parametric estimate [\[77\]](#ref-77), [\[89\]](#ref-89). The imputation is anchored on the scope drivers the documentation provides, for instance the functional breadth of the onboard capability and the verification burden it imposed, and it uses the parametric form as a consistent estimator across episodes rather than as a precise per-episode truth. The space-telescope single-variable cost models exemplify the parametric approach adopted as the normalization basis [\[89\]](#ref-89), and the survey of space-cost models and of early-phase hardware cost-estimation methods supplies the broader methodological grounding [\[90\]](#ref-90), [\[96\]](#ref-96), [\[39\]](#ref-39). Layer-two figures carry a lower reliability grade than layer-one figures, recorded in the flag, because they are model-based extrapolations into a population the model was not calibrated on.

### 4.3.3 Layer three: deflation to constant-year dollars

The third layer expresses every figure, whether extracted at layer one or imputed at layer two, in constant-year dollars using a standard deflator, so that figures from different decades are comparable and the decade fixed effects absorb time trends rather than nominal-price drift. The deflator basis is recorded so the constant-year conversion is reproducible. A figure that required only deflation, because it was already a clean autonomy figure, retains its layer-one grade; the deflation step does not by itself degrade reliability, it only standardizes the unit.

### 4.3.4 The reliability flag and its use in inference

For each observation, the construction log records which layer produced the figure and an associated reliability grade, and this flag is then used in the third pre-registered robustness specification, which weights observations by the inverse of their imputation error so that observations resting on the weakest imputation receive less weight. This is the operational device that protects the slope from being driven by the least trustworthy figures. The flag also feeds the layer-composition report committed to above. The confidence attached to the dependent-variable construction is moderate at the design stage: the procedure is fully specified and the parametric basis is real and standard, but the realized reliability of the panel depends on how many episodes resolve at layer one versus layer two, which is not known until assembly. A panel in which a substantial share of episodes carry layer-one figures would raise confidence; a panel dominated by layer-two imputations would lower it, which the study has pre-committed to disclose.

## 4.4 The independent variable: the forward-only cumulative-heritage counting rule

The independent variable, cumulative flight-demonstrated heritage, is the experience stock whose logarithm carries the slope \(\beta\), and its construction is governed by a single fixed convention carried from the prospectus: a prior demonstration counts toward the heritage stock of a later episode only if it reached flight operation before the later episode's development-start date.

The forward-only counting rule produces a cumulative-heritage measure that cannot mechanically manufacture a heritage-cost association and that is conservative against the study's own hypothesis.

Heritage that arrives after a development has begun cannot have lowered that development's cost, so counting it would credit an episode with experience it could not have used. The forward-only rule excludes such heritage by construction.
A measure that can only undercount, never overcount, the usable experience stock biases the estimated slope toward zero, which means any rejection of the flat-cost null is achieved against a conservative measure and is therefore more credible, not less.

The experience-curve tradition requires that the cumulative-experience variable be the stock available at the time the cost was incurred, and the Wright-versus-Moore distinction in the forecasting literature warns that a cumulative-experience measure entangled with calendar time will confound learning with a coincident time trend [\[79\]](#ref-79), [\[35\]](#ref-35). Keying the count on development-start timing and absorbing calendar time with decade fixed effects is the discipline that literature prescribes.

The rule is coarse in one respect it cannot fix: it counts within-class demonstrations only, and it counts demonstrations rather than the finer-grained reusable-knowledge stock that actually transfers between episodes. It is a proxy for the experience stock, not the stock itself.

Heritage is not always within-agency or within-class: software components and design patterns cross capability classes, so the within-class count understates the true reusable stock, which biases against finding an effect. The study addresses this through a robustness specification that broadens heritage to include cross-class software-component reuse, and it treats the within-class count as the conservative baseline.

### 4.4.1 Operational steps

The heritage count for each episode is built in three operational steps. First, the dated demonstration record is assembled from NTRS and the published literature, fixing the flight-operation date of every demonstration in each capability class [\[12\]](#ref-12), [\[24\]](#ref-24), [\[34\]](#ref-34), [\[40\]](#ref-40), [\[99\]](#ref-99), [\[7\]](#ref-7). Second, for each episode, the count of prior same-class demonstrations whose flight-operation date precedes the episode's development-start date is computed. Third, the first observation in each class, which has no prior heritage, is set to a cumulative count of one before the logarithm is taken, so that the first demonstration is the curve origin and contributes a defined \(\ln(\text{CumHeritage})\) of zero. The AEGIS pair is the clearest worked example: the ChemCam deployment's development began after the Opportunity deployment reached flight operation, so ChemCam carries a within-class heritage count of one prior demonstration while Opportunity, as first-in-class, is set to the origin [\[34\]](#ref-34), [\[40\]](#ref-40). The pair illustrates how the rule operates; it is not presented as a fitted result.

### 4.4.2 The counting log

Every heritage count is recorded in a counting log that states, per episode, the prior demonstrations counted, their flight-operation dates, the episode's development-start date, and the resulting count. The log is retained as an appendix so that the heritage variable is reproducible and so that a reviewer can check any single count against the dated record. This is the same auditability standard applied to the dependent variable. Confidence in the heritage construction is high relative to the dependent variable, because the inputs, demonstration dates, are far better documented in the open record than costs are, and the counting rule is mechanical once the dates are fixed. The residual risk lies in the development-start dates, which the rule keys on and which TechPort records administratively rather than as engineering-start dates; the log makes any contestable date visible.

## 4.5 Fixed effects, the maturation covariate, and the parametric cost-model basis

This section operationalizes the remaining terms of the estimating equation and states how each is constructed and validated.

### 4.5.1 Capability-class fixed effects

The capability-class fixed effects \(\gamma_c\) are indicator variables for the five classes of Section 4.2. They absorb time-invariant differences in baseline cost and intrinsic difficulty across classes, so that the slope is not contaminated by the simple fact that some classes are more expensive than others for reasons unrelated to heritage. The construction is mechanical once the taxonomy assignment is fixed. Their validation is the feasibility check committed to in the analysis plan: any class containing only a single observation contributes nothing to the within estimator and is reported as such, because the slope is then identified off a smaller effective sample than the raw episode count suggests.

### 4.5.2 Decade fixed effects

The decade fixed effects \(\delta_d\) are indicator variables for the decade of development start. They absorb economy-wide and agency-wide time trends in cost, tooling, and computing that affect all classes, and they separate genuine within-class learning from a coincident secular decline in software-development cost. Their construction depends on the development-start date, taken from TechPort where available and from the literature where TechPort is thin. The Wright-versus-Moore concern is the reason these fixed effects are not optional: without a time control, an experience curve and a time trend are observationally similar where experience grows with time, and the slope would not be cleanly identified [\[79\]](#ref-79), [\[35\]](#ref-35).

### 4.5.3 The maturation covariate

The maturation covariate is the technology-readiness level at development start, drawn from the TechPort technology-readiness-level history and entered only in a robustness specification, never in the baseline, for the construct reason given in Section 4.2.3. Its validation is twofold. First, the study treats the scale as an ordinal, error-bearing covariate rather than a cardinal cost proxy, consistent with the documented shortcomings of the scale [\[81\]](#ref-81), [\[70\]](#ref-70), [\[53\]](#ref-53). Second, the calibration literature, which attempts to anchor technology-readiness levels to mission data, is the backing for using the scale as a maturity check while acknowledging that its meaning is imperfectly calibrated [\[92\]](#ref-92). The covariate's role is diagnostic: if adding it collapses the slope, the raw heritage association was partly the effect of later projects starting at higher maturity, which the study would report as a weakening of the contribution rather than concealing.

### 4.5.4 The parametric cost-model basis as a cross-check

The parametric cost models serve a second function beyond imputation: they cross-check the layer-one figures. Where a direct autonomy figure is extracted at layer one and a NICM-class parametric estimate can also be formed from the documented scope, the two are compared, and a large discrepancy is recorded as a data-quality flag that triggers review of the layer-one extraction. This is a validation-against-a-second-method procedure that the broader technology-cost literature supports as a way to bound estimation error when a single estimator is untrustworthy outside its calibration set [\[90\]](#ref-90), [\[96\]](#ref-96), [\[17\]](#ref-17), [\[57\]](#ref-57). The cross-check does not assume the parametric estimate is correct; it uses agreement or disagreement between two independent constructions as a signal about the reliability of each.

## 4.6 Coverage window, validation against known values, ethics and access, and the four material limitations

### 4.6.1 Coverage window

The intended coverage runs from the Deep Space 1 Remote Agent Experiment in 1999 through the Mars 2020 demonstrations and their immediate successors, including the SuperCam-era and OPS-SAT-style onboard-learning context where it bears on capability scope [\[72\]](#ref-72), [\[60\]](#ref-60), and the entry-descent-and-landing hazard-handling lineage from the Autonomous Landing and Hazard Avoidance Technology program through later lander hazard-detection demonstrations [\[31\]](#ref-31), [\[32\]](#ref-32), [\[2\]](#ref-2), [\[33\]](#ref-33), [\[66\]](#ref-66), [\[51\]](#ref-51), [\[47\]](#ref-47). This window spans the period in which onboard autonomy moved from single experiments to routine mission elements, which is precisely the period over which an experience curve, if it exists, should be detectable. A narrower window would risk too few episodes per class for any within-class estimate; a wider window backward would reach a period with too little onboard autonomy to populate the classes.

### 4.6.2 Validation against known values

Three validation procedures check the constructed variables against independently known quantities. First, the demonstration dates that drive the heritage count are validated against multiple primary sources per demonstration, so that a date used in a count is corroborated rather than taken from a single report; the Remote Agent flight-experience and design records, for instance, corroborate the Deep Space 1 timing [\[12\]](#ref-12), [\[13\]](#ref-13), [\[83\]](#ref-83). Second, the dependent-variable figures are validated by the layer-one-versus-parametric cross-check of Section 4.5.4, which flags any direct figure that diverges sharply from a scope-anchored estimate for re-examination. Third, the maturation covariate is checked against the calibration literature's anchoring of technology-readiness levels to mission data, so that a recorded level that is implausible given the episode's documented maturity is flagged [\[92\]](#ref-92), [\[81\]](#ref-81). None of these procedures certifies the variables as exact; each provides an independent check that surfaces the most serious construction errors. The confidence this validation supports is moderate: it catches gross errors and documents the basis for each figure, but it cannot remove the irreducible measurement error in costs that are not cleanly reported in the open record.

### 4.6.3 Ethics and access

The data are drawn entirely from public sources: the TechPort public application programming interface and bulk export, the NTRS citations interface, the published scholarly literature, and published parametric cost models. No proprietary, classified, or controlled-unclassified information is used, and no MITRE-internal or working-note material is cited or relied upon, consistent with the program's standard that internal working documents are context only and not citable in the published work. There are no human subjects, so no human-subjects review is implicated. The access posture is reproducibility-first: every source is retained at a dated version, every figure carries a recorded provenance, and the assembled panel, the imputation log, and the counting log are retained so the measurement can be checked by a reviewer who did not build it. This access and retention commitment is the ethical core of a cost study whose dependent variable is partly imputed: the imputation is defensible only if it is fully visible.
### 4.6.4 The four material data limitations

The prospectus named four material limitations. The chapter carries them forward with the construction detail that makes each one concrete and bounds the strength of any eventual conclusion.

First, the population of flight-demonstrated autonomy episodes is small, on the order of tens rather than thousands, which limits statistical power and constrains the number of fixed effects the panel can support. The construction response is a feasibility check that reports any singleton class or decade and an honest accounting of the effective, within-transformed sample rather than the raw count.

Second, development-cost figures are heterogeneous in definition and sometimes not public, which forces NICM-class imputation and introduces measurement error in the dependent variable. The construction response is the three-layer procedure, the reliability flag, the inverse-imputation-error weighting specification, and a committed disclosure of the panel's layer composition.

Third, capability-class assignment requires judgment and can be contested at the boundaries. The construction response is the fixed dominant-capability rule, the multi-capability flag, and the reclassification robustness check.

Fourth, heritage is not always within-agency or within-class, because software components and design patterns cross capability classes. The within-class heritage count does not capture this, which biases against finding an effect. The construction response is the cross-class-reuse robustness specification and the framing of the within-class count as the conservative baseline.

These four limitations are not defects to be hidden. They are the boundary conditions of the measurement, and stating them with their construction responses is what lets the chapter claim that the panel is auditable rather than exact. The chapter's argument can therefore be drawn together. The dependent-variable measurement problem is genuine, because costs are reported inconsistently and sometimes not at all, and it is consequential, because cost is the variable the entire contribution turns on. The three-layer normalization meets the problem at its root by making the heterogeneity explicit and gradeable rather than averaging over it silently, which is preferable to taking heterogeneous figures at face value and letting the least reliable of them drive the slope undetected. The error that remains in imputed costs is tolerable precisely because it is disclosed, graded, down-weightable, and bounded by validation against an independent estimator. That argument is what the chapter contributes to the dissertation as a whole, and it is what the research-design and analysis-plan chapters build on when they specify how the slope is estimated from these constructed variables and how its confidence interval is read against the fixed decision rule.

### 4.6.5 What the chapter establishes and what it defers

This chapter establishes the data foundation: four named, real, accessible sources; a unit of analysis with an operational assignment rule; a full measurement table operationalizing every variable in the estimating equation; a three-layer, reliability-flagged construction of the delicate dependent variable; a conservative forward-only heritage count with a retained log; and a validation, ethics, and access posture built for reproducibility. It defers, by design and not by oversight, the execution: the panel is not yet assembled, the cost normalization is not yet complete across all episodes, and so no normalized cost, no heritage-count distribution, and no panel dimension is reported here as a final value. The worked examples, the AEGIS heritage pair and the Deep Space 1 cost-burden record, illustrate the procedures and are labeled as illustrations of method, not as measurements. The research-design chapter that follows takes the constructed variables specified here and develops the estimator, the identification strategy, and the threats-to-validity matrix that turn this data foundation into a falsifiable test of whether each successive autonomous-operations flight demonstration lowers the cost-to-field of the next.



# Chapter 5. Research Design and Identification

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

This chapter delivers the inferential machinery that converts the theoretical model of Chapter 2 and the measured panel of Chapter 4 into a falsifiable test, and it states the design in one sentence at the outset. The design is this: the experience-curve slope of onboard-autonomy qualification cost is estimated by ordinary least squares on a log-log specification with two-way fixed effects for capability class and decade, identified off within-class, within-decade variation in cumulative flight-demonstrated heritage, hardened by three pre-registered robustness specifications and a four-way threats-to-validity matrix, and protected by a small-sample inferential procedure and an influence diagnostic, so that the single coefficient on which the contribution turns is read against a fixed decision rule rather than constructed after the fact. The estimating equation that this design serves is fixed and carried unchanged from the prospectus and from Chapter 2:

\[ \ln(\text{Cost}_{icd}) = \alpha + \beta \, \ln(\text{CumHeritage}_{icd}) + \gamma_c + \delta_d + \epsilon_{icd} \qquad\qquad (1) \]

for episode \(i\) in capability class \(c\) and decade \(d\), where \(\text{Cost}_{icd}\) is normalized development cost, \(\text{CumHeritage}_{icd}\) is cumulative within-class flight-demonstrated heritage, \(\gamma_c\) are capability-class fixed effects, \(\delta_d\) are decade fixed effects, and \(\epsilon_{icd}\) is the error term. The single parameter \(\beta\) is the experience-curve slope, and the implied learning rate is \(1 - 2^{\beta}\). The two hypotheses turn on \(\beta\): H0 holds that \(\beta\) is statistically indistinguishable from zero, so that per-episode autonomy development cost is flat with respect to cumulative prior flight demonstrations within a capability class; H1 holds that \(\beta\) is negative and statistically significant after controlling for capability class and decade. Everything in this chapter exists to make the estimate of \(\beta\) defensible, to name in advance every way the estimate could mislead, and to bound what may and may not be concluded from it.

The inferential problem this chapter addresses can be set out plainly. The theoretical framework supplies a functional form and the data chapter supplies a constructed panel, but neither by itself answers the questions that decide whether the eventual estimate of \(\beta\) means anything: which estimator is correct for a panel of tens of heterogeneous engineering episodes, what variation the slope is actually identified off once class and time are absorbed, which confounders survive that absorption, and how inference should be conducted when degrees of freedom are scarce. What is needed is a complete identification argument in which the estimator is justified against its rivals, the source of identifying variation is named explicitly, every threat to internal, external, construct, and statistical-conclusion validity is paired with a stated design response, the robustness battery is fixed before estimation, and the inferential procedure is calibrated to the small panel rather than borrowed uncritically from the data-rich experience-curve literature. Without that argument, an estimate of \(\beta\), even a precisely computed one, would be uninterpretable, because the reader could not tell whether a negative slope reflected learning, a residual time trend, a scale artifact, or a single influential first-of-kind episode. To leave the question open is to produce a number with no warrant, which is the failure mode the contribution is designed to avoid: the whole value of the study is that it replaces an assertion with a measured, defended quantity, and an undefended measurement is no improvement on the assertion it replaces.

This chapter closes the gap at the level of inference. It does five things, in order. First, it justifies the estimator: ordinary least squares on the log-log two-way fixed-effects form, and it argues explicitly why a non-linear or fully Bayesian estimator is not preferred in a panel this small (Section 5.1). Second, it states the identification strategy: what \(\beta\) is identified off, what each fixed-effect set removes, and what the maturation covariate adds (Section 5.2). Third, it specifies the three pre-registered robustness specifications, namely the technology-readiness-level covariate, the cross-class reuse measure, and the inverse-imputation-error weighting, and explains what each one buys (Section 5.3). Fourth, it develops the full four-way threats-to-validity matrix, internal, external, construct, and statistical-conclusion, with a design response paired to each threat and a confidence statement attached (Section 5.4). Fifth, it treats external validity and the path-dependence caution that the Arthur anchor supplies, bounding the generality of any estimated slope (Section 5.5). The chapter then draws the design's argument together and states what would have to be true for the identification to fail. The chapter constructs no DoDAF, BEA, or capability-architecture chain; this is an econometric measurement study, and architecture vocabulary is omitted rather than forced.

A note on register and epistemic discipline is owed before the argument begins, and it is the same note that governs every chapter of this design-stage dissertation. No coefficient has been fitted on the full dataset. Every statement about what \(\beta\) will be, what its confidence interval will contain, or what minimum effect the design can detect is conditional and explicitly labeled as expected or illustrative. The design developed here fixes a procedure and predicts its properties; it does not assert a result. Where a causal claim is made, the mechanism that carries the cause through to the effect it predicts is named, and where only correlation is available the chapter says so and lowers the stated confidence. Confidence is reported as low, moderate, high, or very high, and each level is tied to the evidence that would raise or lower it. The discipline is not decorative: in a study whose entire contribution is the credibility of a single estimate, the honesty of the design-stage posture is itself part of what makes the result believable.

## 5.1 The estimator: ordinary least squares on the log-log two-way-fixed-effects form

The first claim of this chapter is the foundational design commitment: the correct estimator for this study is ordinary least squares applied to the log-log experience-curve specification with capability-class and decade fixed effects, and the alternatives that a reader might expect, a non-linear power-law fit or a hierarchical Bayesian model, are not preferred here for reasons specific to the small panel. The argument runs as follows.

The experience-curve slope \(\beta\) should be estimated by ordinary least squares on the linearized log-log form with two-way fixed effects, rather than by non-linear least squares on the untransformed power law or by a Bayesian hierarchical estimator, because the linear fixed-effects estimator delivers a single interpretable slope, conserves the degrees of freedom that are the binding constraint in this panel, and inherits the predictive validation that the experience-curve literature has established for this functional form.

The basis is threefold. First, the log-log linearization is the form whose out-of-sample predictive performance has been measured and found superior across a large and heterogeneous library of technologies. Nagy, Farmer, Bui, and Trancik [\[79\]](#ref-79) compared candidate functional forms and found the Wright power law, taken in its logarithmic linear form, performs at least as well as and generally better than the alternatives; Farmer and Lafond [\[35\]](#ref-35) then quantified the forecast-error distribution of that form, and Lafond and colleagues [\[61\]](#ref-61) extended it to full distributional forecasts. Estimating the same form by ordinary least squares on logged variables is the canonical way that this validated relationship is fit, and the energy-technology learning-rate literature, which is the most mature body of applied experience-curve estimation, fits learning rates by exactly this log-log ordinary-least-squares procedure. McDonald and Schrattenholzer [\[74\]](#ref-74) estimate learning rates for energy technologies this way; Neij and colleagues [\[80\]](#ref-80) do the same for power-production technologies; and the review by Rubin, Azevedo, Jaramillo, and Yeh [\[86\]](#ref-86) documents that the log-log regression of log cost on log cumulative output is the standard estimator across the field. Second, the linear form produces a single slope coefficient \(\beta\) whose interpretation, the percentage cost change per doubling of cumulative experience through \(1 - 2^{\beta}\), is exactly the quantity the hypotheses are written about, so the estimator returns the estimand without an intervening transformation. Third, the fixed effects enter as indicator variables for the specific, non-sampled capability classes and decades that are the categories of interest, which absorbs class-specific and time-specific level differences without imposing the orthogonality condition between the heritage regressor and the absorbed effects that a random-effects estimator would require.

When a functional form has been competitively validated for prediction, when its standard applied estimator is well established in the adjacent field that fits it most often, and when the quantity of interest is the single coefficient that estimator returns, the careful analyst should adopt that estimator rather than a more elaborate one whose additional structure is not warranted by the data. The governing principle is estimator parsimony under data scarcity: in a small panel, every additional estimated parameter is paid for in degrees of freedom and in the variance of the coefficient that matters, so the estimator should be no more complex than the inferential target requires.

That preference is supported by the explicit demonstration in the experience-curve estimation literature that the log-log ordinary-least-squares slope is both the conventional and the validated way to recover a learning rate, and that elaborations are introduced only when the data support them. Rubin and colleagues [\[86\]](#ref-86) survey the range of learning-rate estimates and the methods that produce them and document that the simple log-log regression is the workhorse; Yeh and Rubin [\[109\]](#ref-109) then catalogue the uncertainties in those experience-curve estimates and show that the dominant sources of uncertainty are data quality and specification choice, not the choice between ordinary least squares and a more complex fitting routine. This evidence matters because it locates the real inferential risk where it belongs, in the data and the specification rather than in the estimator, and so justifies spending the study's analytic budget on cost normalization, identification, and robustness rather than on estimator sophistication.

The preference for ordinary least squares holds with high confidence conditional on the panel being small and the inferential target being a single within-class slope. One caveat bounds this choice: if the panel were data-rich, a hierarchical Bayesian model that partially pooled learning rates across capability classes would be defensible and possibly preferable, because it would borrow strength across classes while still allowing class-specific slopes, and the energy-technology literature does use such richer structures where the data permit. The claim here is not that ordinary least squares is universally superior; it is that ordinary least squares is the correct choice for a panel of tens of observations in which the cost of every additional estimated parameter is acute. This qualifier is what licenses the later commitment to report residual degrees of freedom alongside every coefficient.
The estimator choice could be wrong if the true cost-experience relationship were strongly non-linear in logs, in which case the linear slope would average over curvature and mislead, or if the error structure violated the homoscedastic, independent-error assumption underlying naive ordinary-least-squares inference in a way that a Bayesian or generalized model would handle better. Both objections are taken seriously. The first is addressed by the diagnostic plots of log cost against log heritage by class specified in the analysis plan, which would reveal gross curvature, and by the influence diagnostics of Chapter 6, which would reveal whether a single episode is bending the line. The second is addressed not by changing the estimator but by changing the inference: the design pre-commits to small-sample-robust standard errors rather than to naive ordinary-least-squares standard errors, so that the point estimator and the inferential procedure are chosen separately, each for its own reason.

The estimator decision here is evidence-based and parsimony-driven rather than conventional by default. Read the cluster of method sources for what their convergence means rather than as a list: Nagy [\[79\]](#ref-79), the validation work of Farmer and Lafond [\[35\]](#ref-35) and Lafond [\[61\]](#ref-61), and the applied estimation practice documented by McDonald and Schrattenholzer [\[74\]](#ref-74), Neij [\[80\]](#ref-80), Rubin [\[86\]](#ref-86), and Yeh and Rubin [\[109\]](#ref-109) converge on a single message. The log-log ordinary-least-squares slope is the validated, standard, and interpretable estimator for a learning rate, and the binding risks in such estimation are data and specification rather than the fitting routine. That convergence is the justification a small-panel study needs to adopt the simple estimator deliberately rather than apologetically. Confidence that ordinary least squares on the log-log two-way fixed-effects form is the right estimator for this panel is therefore high. It is explicitly conditioned on the small-sample setting and would be revised toward a partially pooled estimator only if the panel grew large enough to support class-specific slopes.

One further design point on the estimator belongs here, because it determines how the two fixed-effect sets are entered. The capability-class and decade effects are entered as additive indicator variables rather than as an interacted class-by-decade saturation, a deliberate degrees-of-freedom decision. A fully saturated class-by-decade design would absorb so much variation that the within-cell heritage variation left to identify \(\beta\) could vanish in a panel this small, and in the limit a cell containing a single episode contributes nothing to the within estimator. The additive two-way structure absorbs the dominant confounders, class-level baseline difficulty and economy-wide time trends, while preserving the within-class, within-decade comparisons that carry the identifying information. The analysis plan of Chapter 6 makes the feasibility of even this additive structure an explicit pre-analysis check, reporting the realized effective sample once singleton cells are accounted for, so that the estimator's degrees-of-freedom cost is disclosed rather than assumed away.

## 5.2 Identification: what \(\beta\) is identified off, and the role of each fixed-effect set

The second claim is the identification claim, and it is the heart of the chapter, because the credibility of the entire contribution rests on the answer to a single question: when the slope is read off this regression, what comparison is actually generating it? The argument is stated and defended below.

The slope \(\beta\) is identified off within-capability-class, within-decade variation in cumulative flight-demonstrated heritage, and only off that variation. The capability-class fixed effects remove between-class differences in baseline cost and intrinsic difficulty, the decade fixed effects remove economy-wide and agency-wide time trends common to all classes, and what remains to identify the slope is the comparison of episodes in the same capability class facing different accumulated heritage within the same decade.

The basis is the structure of the estimating equation and the documented confounders the fixed effects are built to remove. The first confounder is between-class heterogeneity. Capability classes differ in intrinsic difficulty: autonomous entry-descent-and-landing hazard handling, with its hard real-time constraints and its irreversibility, is not the same engineering proposition as onboard planning and scheduling, and the heterogeneity of learning rates across settings is itself documented. Bhattacharya and colleagues [\[14\]](#ref-14) show, on cost-quantity data from India, that learning rates are diverse across settings rather than constant, direct evidence that pooling heterogeneous classes without class effects would conflate a difficult class's high cost with a low-heritage state and bias the slope. The capability-class fixed effects \(\gamma_c\) remove exactly this by absorbing each class's level. The second confounder is the secular time trend. The cost of all software-intensive engineering fell over the study window for reasons common to every class, principally the steep decline in computing cost and the maturation of software tooling, and the experience-curve literature is explicit that a credible learning estimate must separate genuine learning, a function of cumulative output, from a coincident calendar-time trend. Nagy and colleagues [\[79\]](#ref-79) establish that where output grows exponentially the Wright cumulative-output form and the Moore calendar-time form become observationally similar, the formal reason a time control is mandatory. The decade fixed effects \(\delta_d\) absorb this common time trend. After both sets are absorbed, the only remaining variation in \(\text{CumHeritage}\) is within class and within decade, so that is the variation that identifies \(\beta\).

A coefficient in a two-way fixed-effects regression is identified off the variation in the regressor that survives the absorption of both fixed-effect sets, and that surviving variation is interpretable as the causal effect of interest if and only if it is plausibly unconfounded by anything correlated with it that the fixed effects did not remove. The standard within-estimator logic applies: the fixed effects convert a comparison across classes and across decades, contaminated by class difficulty and time trends, into a comparison within class and within decade, contaminated only by whatever still moves with heritage inside a class-decade cell.

This within-transformation is how the energy-technology learning-rate literature isolates learning from confounds, and the discipline of pairing a cumulative-experience regressor with a time control is treated there as a requirement rather than a refinement. Wei, Smith, and Sohn [\[103\]](#ref-103) show that retrospective experience-curve learning rates are not constant and correlate with deployment programs, a concrete demonstration that an uncontrolled experience curve mixes learning with program-specific and time-varying effects, which therefore backs the necessity of the decade control as the device that strips the common time-varying component out before the slope is read. The convergence of the cumulative-output-versus-time identification problem in Nagy [\[79\]](#ref-79) with the non-constant-rate finding in Wei [\[103\]](#ref-103) is the backing for treating the decade fixed effects not as a nuisance adjustment but as a load-bearing element of identification.

Identification holds at moderate confidence, and the qualifier names exactly why it is not high. The within-class, within-decade variation that identifies \(\beta\) is real but thin: a panel of tens of episodes spread across roughly five capability classes and three decades leaves few comparisons inside any single class-decade cell, and in the limiting case where a cell holds one episode that cell contributes nothing to the within estimator. The slope is therefore identified off a smaller effective sample than the raw episode count suggests, and confidence in identification is bounded by that thinness rather than by any flaw in the logic. This caveat is addressed by the Chapter 6 pre-analysis feasibility check, which reports the effective sample after singleton cells are removed, so that the reader sees how much identifying variation actually exists before interpreting the slope.

Identification fails if something correlated with cumulative heritage within a class and within a decade, and not removed by either fixed-effect set, also drives cost. Three such possibilities are real. The first is a within-decade computing or tooling trend: decade effects remove the common decade-level shift but not a finer-grained time trend operating inside a decade, so if heritage accumulates monotonically within a decade it could pick up the residual within-decade cost decline. The second is reverse selection: if cheaper-to-qualify autonomy capabilities are attempted only after heritage in their class already exists, then the heritage-cost association is partly a selection artifact in which low cost causes the heritage state rather than the reverse. The third is the maturity confound: later episodes tend to start at higher technology-readiness levels, and if higher starting maturity independently lowers qualification cost, then a slope attributed to heritage could be a slope on maturity. Each objection is taken seriously, and each is the reason for a specific element of the robustness battery in Section 5.3: the cross-class reuse measure and the influence diagnostics speak to the residual-trend and selection concerns, and the technology-readiness-level covariate speaks directly to the maturity confound by entering starting maturity as a separate regressor so that the heritage slope is identified net of it.

Identification in this study is not assumed but earned by the fixed-effect structure, and its credibility is deliberately bounded by the thinness of the surviving variation. The role of each fixed-effect set is specific and non-interchangeable: \(\gamma_c\) is the answer to between-class difficulty heterogeneity documented by Bhattacharya [\[14\]](#ref-14), and \(\delta_d\) is the answer to the cumulative-output-versus-calendar-time confound formalized by Nagy [\[79\]](#ref-79) and made concrete by Wei [\[103\]](#ref-103). The maturation covariate is not part of the baseline identification but is the designed check that the heritage slope is not a relabeled maturity slope. Confidence in identification is therefore moderate, held back from high by the small effective sample and by the residual within-decade-trend and selection concerns. It would rise if the panel grew, if a finer time control than the decade became feasible without exhausting degrees of freedom, or if the maturity and cross-class robustness specifications left the slope intact.

## 5.3 The three pre-registered robustness specifications

The third claim is that the design controls its own multiplicity risk by fixing, before estimation, exactly three robustness specifications, each targeting a named threat from Section 5.2, rather than searching over specifications until one is congenial. I state and defend the claim.

Three robustness specifications, and only these three, are pre-registered: a specification adding the technology-readiness-level maturation covariate, a specification broadening the heritage measure to include cross-class software-component reuse, and a specification weighting observations by the inverse of the imputation error. Each targets a distinct, pre-named threat, and fixing the set in advance is what converts robustness checking from a vulnerability into a strength.

The basis is that each robustness specification answers one of the three objections raised against identification in Section 5.2, and that pre-registering the set forecloses the specification search that would otherwise inflate the false-positive rate. The first specification adds the technology-readiness-level at development start as a covariate, drawn from the TechPort maturation history described in Chapter 4. Its purpose is to separate the heritage effect from the simpler effect of starting at higher maturity. The technology-readiness-level construct is the standard maturity tracker, introduced and reviewed by Mankins [\[70\]](#ref-70), and its known limitations, principally that it is ordinal and non-monetary and so cannot serve as a cost proxy, are documented by Olechowski and colleagues [\[81\]](#ref-81). Those limitations are why the technology-readiness-level enters as a robustness covariate rather than as a structural part of the baseline: it is a maturation signal, not a cost, and its job is to absorb the maturity confound so the heritage slope is read net of starting maturity. The wide adoption of the technology-readiness-level as a comparable maturation measure across domains, documented by Hedér [\[53\]](#ref-53) and used in cost-and-readiness assessments such as those of Bukar and colleagues [\[18\]](#ref-18) and Khan and colleagues [\[57\]](#ref-57), supports treating the starting technology-readiness-level as a meaningful, comparable covariate across episodes. The second specification broadens \(\text{CumHeritage}\) to count cross-class software-component reuse, not only within-class flight demonstrations. Its purpose is to attenuate the downward bias the baseline within-class count carries, because software components and design patterns cross capability classes and the within-class count cannot see that shared heritage; if the true heritage stock includes cross-class reuse, the within-class baseline understates heritage and biases the slope toward zero, so broadening the measure tests whether a steeper slope emerges when shared components are counted. The third specification weights each observation by the inverse of its cost-imputation error, using the three-layer reliability flag constructed in Chapter 4, so that episodes whose cost figures rest on the weakest NICM-class imputation receive less weight. Its purpose is to prevent the least reliable cost figures from driving the slope, addressing the construct-validity concern in the dependent variable directly through the estimator's weighting rather than through a post-hoc judgment.

Pre-registering a fixed, small set of robustness specifications, each motivated by a named threat, is the correct way to test the durability of an estimate without inflating the false-positive rate, because it removes the analyst's freedom to search over specifications and report the favorable one. The inferential principle is that the credibility of a robustness battery comes from its being committed in advance and tied to threats, not from its size or from the analyst's discretion in assembling it.

This rests on the documented finding in the experience-curve literature that specification choice, not estimator choice, is a dominant source of uncertainty in learning-rate estimates. Yeh and Rubin [\[109\]](#ref-109) catalogue the uncertainties in experience-curve estimates and show that how the analyst specifies the relationship materially moves the estimated rate, direct backing for treating specification discipline as a first-order design concern; Rubin and colleagues [\[86\]](#ref-86) reinforce this by showing the wide spread of published learning rates that specification and data differences produce. If specification freedom is a leading source of uncertainty, then constraining it by pre-registration is a leading remedy, and that is the backing for fixing the three specifications in advance rather than choosing them after seeing the baseline result.

The robustness battery is held to give high confidence that a surviving slope is not a specification artifact, conditional on the three specifications being the right three. One caveat is that the battery is not exhaustive: there are specification choices it does not vary, for example the boundary definitions of the capability classes themselves and the precise deflator used to express costs in constant-year dollars. The design does not claim that no specification could overturn the result; it claims that the three pre-registered specifications target the three most consequential named threats and that their pre-commitment removes the most dangerous degree of analyst freedom. This caveat is addressed by the reproducibility commitment: the assembled panel, the imputation log, and the estimation code are retained, so that a reviewer can run a specification the design did not pre-register and see for themselves.

The battery could mislead if all three specifications shared a common blind spot, so that a confounder invisible to within-class counting, to cross-class counting, and to technology-readiness-level adjustment alike drove the slope in every specification. The clearest candidate is the residual within-decade computing-cost trend, which none of the three specifications directly removes, because all three hold the decade fixed-effect structure fixed and vary other things. This objection is acknowledged honestly and is the reason the discussion of any fitted result will weigh the within-decade-trend rival explicitly rather than treating survival across the three specifications as proof of a learning effect; survival across the battery raises confidence but does not, on its own, exclude a confounder common to all three.

The robustness specifications are a designed defense, each mapped one-to-one to a named threat, and their value is in their pre-commitment. The decision rule that uses them is fixed and stated in Chapter 6: H0 is rejected for H1 if and only if \(\beta\) is negative and its confidence interval excludes zero in the baseline specification and in at least two of the three robustness specifications. Requiring survival in two of three, rather than in all three or in any one, is a deliberate middle commitment: it tolerates the failure of a single specification, which a thin panel could produce by chance, without accepting a result that holds only in the baseline. Confidence that a slope surviving this rule reflects a real within-class learning effect rather than a specification artifact is high; it is bounded below high only by the acknowledged possibility of a confounder common to all three specifications, principally the residual within-decade trend.

## 5.4 Threats to validity and the design responses

The fourth claim organizes every remaining inferential risk into the four canonical validity categories and pairs each threat with a stated design response, so that no threat is left unnamed and none is left unanswered. I treat the four categories in turn, each paired with its evidence, its design response, and a confidence statement, and I am explicit where a response only mitigates rather than eliminates a threat.

**Internal validity.** The primary internal threat is omitted-variable confounding between cumulative heritage and the secular decline in computing and software cost, and the decade fixed effects mitigate but do not eliminate it. Computing and tooling cost fell steeply and monotonically across the study window, cumulative heritage also rose monotonically with time, and two monotone series are mechanically correlated; the decade fixed effects remove the between-decade portion of this co-movement, but a within-decade computing trend can survive, because absorbing a decade level does not absorb a trend operating inside the decade. The design response is layered: the decade fixed effects remove the dominant common time component, the cross-class reuse robustness specification tests whether the slope is carried by genuine reusable heritage rather than by a generic time trend, and the discussion commits to weighing the residual-trend rival explicitly against any fitted slope. A second internal threat is reverse pathways, in which cheaper episodes are attempted only after heritage exists, so that the heritage-cost association is partly a selection artifact; the design response is the maturation covariate, which separates the effect of starting at higher maturity from the effect of accumulated heritage, together with the forward-only counting rule established in Chapter 4, which credits an episode only with heritage that reached flight before that episode's development start and so cannot manufacture a mechanical correlation from heritage that arrived too late to have lowered cost. The forward-only rule is conservative by construction: it tends to undercount heritage and therefore biases the slope toward zero, which makes any rejection of the flat-cost null more credible rather than less, because the rule works against finding the effect H1 predicts. Internal validity is held at moderate confidence: the design removes the dominant confounders and biases conservatively, but the residual within-decade trend cannot be fully excluded with a decade-level control, and confidence would rise only with a finer time control that the degrees-of-freedom budget does not currently permit.

**External validity.** The claim is that any estimated slope generalizes to NASA and Jet Propulsion Laboratory onboard autonomy for deep-space and planetary missions within the demonstration window, and no further. The basis is that the panel is drawn entirely from that population, the autonomy demonstrations that anchor it, from the Deep Space 1 Remote Agent through the Mars 2020 demonstrations, are NASA and Jet Propulsion Laboratory missions, and the cost-normalization basis is the NICM-class parametric family built for NASA development cost. The design response is to state the scope honestly and to refrain from claiming a universal learning rate: the fixed-effects estimator recovers an average within-class slope for this population, not a constant that transfers to commercial autonomy, to terrestrial autonomy, or to mission classes outside the window. The space-autonomy review by Gao and Chien [\[42\]](#ref-42) situates the anchored demonstrations within the NASA and Jet Propulsion Laboratory arc and supports treating them as a coherent population rather than a sample of a broader autonomy universe, which is what supports the bounded external-validity claim. External validity is held at high confidence within its stated scope and is deliberately not extended beyond it; the scope limitation is a feature of honest design, not a deficiency to be apologized for, because a learning rate measured on this population answers the NASA and Jet Propulsion Laboratory portfolio question the contribution is aimed at.
**Construct validity.** The claim is that the two central constructs, autonomy qualification cost and reusable-knowledge heritage, are measured imperfectly by the available records, and that the design makes the imperfection explicit and auditable rather than pretending it away. The basis is detailed in Chapter 4 and recalled here for the threat analysis. Public development-cost figures bundle platform, instrument, and autonomy inconsistently, some are not disclosed, and the three-layer normalization is itself a construct choice that introduces measurement error into the dependent variable: extract the autonomy non-recurring engineering where possible, impute via a NICM-class parametric estimate where not, and deflate to constant-year dollars. The parametric-cost basis is the single-variable cost-model family for space systems exemplified by Stahl and colleagues [\[89\]](#ref-89) and surveyed in the companion cost-model review [\[90\]](#ref-90); using it to impute the autonomy portion is a defensible but not error-free construct decision. On the independent side, the within-class heritage count is a coarse proxy for the true reusable-knowledge stock that the Mokyr framework of Chapter 2 identifies as the real driver, because reusable knowledge crosses class boundaries and lives partly in shared flight-software substrates that a within-class count cannot see. The design response is the inverse-imputation-error weighting, which down-weights the observations whose cost construct is weakest, and the cross-class reuse specification, which partly closes the gap between the heritage count and the reusable-knowledge stock by counting shared components. Construct validity is held at moderate confidence: the constructs are imperfect, the imperfection is made explicit and partly corrected by the weighting and cross-class specifications, and confidence would rise materially only if project-level autonomy cost figures became separable from platform cost in the source records, which is a data-construction task for the build phase rather than a thing more citations could fix.

**Statistical-conclusion validity.** The claim is that the dominant statistical threat is the small sample, that with tens of observations and two sets of fixed effects degrees of freedom are scarce and standard errors are wide, and that the design responds by changing the inferential procedure rather than by pretending the sample is large. The reasoning is arithmetic. A panel on the order of tens of episodes, once roughly five capability-class effects and three decade effects are estimated, leaves few residual degrees of freedom, and naive ordinary-least-squares inference computed against a large-sample reference distribution would understate the true uncertainty. The design response has three parts. First, the design pre-commits to small-sample-robust inference rather than to naive standard errors, so that the inferential procedure is matched to the panel size; the point estimator and the inferential procedure are chosen separately, the former for interpretability and parsimony in Section 5.1 and the latter for honesty under small samples here. Second, the design pre-commits to reporting confidence intervals rather than only point estimates and significance stars, and to treating a wide interval that contains zero as a failure to reject H0 rather than as evidence for H1, which is the correct reading of an imprecise estimate. Third, multiplicity is controlled by the pre-registration of exactly three robustness specifications in Section 5.3, rather than by searching over specifications, so that the family of tests is fixed in advance and the false-positive rate is not inflated by the search. The documented finding that data quality and specification, not estimator choice, dominate experience-curve uncertainty, established by Yeh and Rubin [\[109\]](#ref-109) and Rubin and colleagues [\[86\]](#ref-86), is the backing for locating the statistical-conclusion defense in inference and pre-registration rather than in a more elaborate estimator. Statistical-conclusion validity is held at moderate confidence: the design cannot manufacture power the small sample does not contain, but it can and does ensure that the uncertainty is reported honestly and that a precise-looking but fragile result is not produced by specification search. Confidence in the conclusion, whichever way it falls, rises with the width of the reported interval being taken at face value; the design's contribution to statistical-conclusion validity is to guarantee that face value is honest.

A power and minimum-detectable-effect consideration belongs to statistical-conclusion validity and is stated here at the design stage as an expectation, not a computed result. Because the panel is small and the identifying variation is within class and within decade, the design has limited power to detect a shallow slope, and there exists a minimum-detectable-effect below which a true negative \(\beta\) would not be distinguishable from zero at conventional confidence. The honest design-stage posture is that this minimum-detectable-effect is expected to be non-trivial: a learning rate that is real but mild could fail to clear the detection threshold in a panel of tens of within-class comparisons, in which case the study would report a failure to reject H0 that reflects limited power rather than a genuinely flat cost. The design responds to this not by overstating power but by pre-committing to report the achieved confidence interval, to interpret a wide, zero-spanning interval as inconclusive rather than as support for the null, and to report the effective within-cell sample so the reader can judge the power directly. The illustrative magnitude of the detectable slope, like every other quantitative expectation in this dissertation, is deferred to Chapter 6 and is labeled there as non-empirical; no minimum-detectable-effect number is asserted here as a computed quantity, because the panel that would anchor it is not yet assembled.

## 5.5 External validity and the path-dependence caution

The fifth claim closes the chapter by developing the deepest external-validity bound, the one supplied by the Arthur anchor, and explaining why it constrains how any estimated slope may be interpreted as a general planning parameter. I state and defend it.

Even a cleanly identified, robust, negative slope \(\beta\) must be interpreted as a path-contingent average for this population and this historical sequence, not as a universal learning constant, because the increasing-returns mechanism that an experience curve measures is itself path-dependent and non-ergodic, so the realized slope reflects which capability classes received early investment as much as it reflects any intrinsic learnability.

The basis is Arthur's theory of increasing returns and its explicit warning about path dependence. Arthur [\[3\]](#ref-3) identifies learning effects as one of the mechanisms that make a technology more attractive the more it is adopted, and learning effects are precisely what an experience curve measures: each use of a technique lowers the cost of the next, generating positive feedback. That is the mechanism the slope captures, and it is the reason H1 is the theoretically expected direction. Arthur [\[4\]](#ref-4) also establishes that increasing-returns systems are path-dependent: early choices lock in, and the realized state of the system reflects the historical sequence of choices rather than only the technical fundamentals. Arthur [\[5\]](#ref-5) formalizes the non-ergodicity of such systems, in which the outcome depends on the path taken and not merely on the underlying technology, which means that the cost path actually observed in the autonomy record is one realization of many that could have occurred under different early-investment choices. The path-dependence concept has been developed and made testable in the broader literature, and the conceptual clarifications it has received, including the distinction between contingency and self-reinforcement, sharpen what the caution does and does not claim.

When the mechanism a measurement captures is known to be path-dependent and non-ergodic, the measured value must be interpreted as conditional on the realized historical path, not as a context-free constant, because a different path would in principle have produced a different value through the same mechanism. The methodological principle here is that an estimate of a path-dependent quantity inherits the path-dependence: the slope is a property of this sequence of autonomy investments, and its transportability to a counterfactual sequence is not guaranteed by its having been cleanly identified within the observed one.

Path dependence in technology trajectories is not a speculative caution but a documented and recurring phenomenon, and the increasing-returns framework that predicts the learning effect is the same framework that predicts its path-contingency, so the two cannot be separated. The non-constant-rate finding of Wei, Smith, and Sohn [\[103\]](#ref-103), in which retrospective learning rates vary and correlate with the specific deployment programs that produced them, is direct empirical backing for the claim that a learning rate is a property of a realized program path rather than a transportable constant. The convergence of Arthur's theoretical non-ergodicity [\[5\]](#ref-5) with the empirical non-constancy documented by Wei [\[103\]](#ref-103) is the backing for treating the path-dependence caution as a binding interpretive constraint, not an optional hedge.

The path-dependence caution is held at high confidence as a constraint on interpretation and at lower confidence as a quantitative bound, because the design can name the contingency but cannot measure how much of any realized slope is path-specific versus intrinsic. This caveat does not invalidate the measurement; it bounds its generalization. A robust negative slope still answers the NASA and Jet Propulsion Laboratory portfolio question for the realized portfolio, which is the population the contribution targets, and the build-or-wait decision it informs is itself a decision about the realized path, so a path-contingent slope is the right input for that path-contingent decision. This caveat is addressed by treating effect heterogeneity across capability classes, the Mokyr-predicted secondary finding developed in Chapter 7, as the partial empirical handle on path-contingency: classes whose codified-knowledge substrate is mature should show steeper, more transportable slopes, and the cross-class comparison is the design's way of distinguishing the more intrinsic from the more path-specific component of learning, to the limited extent a small panel allows.

The caution would be over-stated if autonomy learning effects were in fact governed by stable technical fundamentals that transferred across paths, so that the realized slope was close to a true constant and the non-ergodicity were small in magnitude. This is possible and is taken seriously, because the experience-curve literature does find broadly stable learning rates within coherent technology families even as it finds heterogeneity across families. The design's response is not to adjudicate this in advance but to report the slope with its cross-class heterogeneity, so that the data speak to how path-specific the autonomy learning rate is: tight, similar slopes across classes would suggest more transportable fundamentals, and dispersed, class-specific slopes would suggest stronger path-contingency. The objection is therefore converted into an empirical question the design can partially answer rather than a stance the design must assert.

The interpretive reading of this section is that the Arthur anchor performs two distinct jobs for the design, and both are load-bearing rather than decorative. It supplies the mechanism, learning effects as positive feedback, that gives the slope its causal content and makes H1 the expected direction, and it supplies the binding caution, path dependence and non-ergodicity, that bounds the slope's external validity to the realized portfolio path. Holding both at once is the disciplined position: the slope is worth measuring because the learning-effect mechanism is real, and the slope must be read as path-contingent because the same mechanism is path-dependent. Confidence in the path-dependence caution as an interpretive constraint is high; confidence in any quantitative claim about how path-contingent a specific estimated slope is would be low and is deferred to the cross-class heterogeneity finding rather than asserted here.

## 5.6 Summary and implications of the design

This chapter has built the inferential half of the dissertation's argument, and it closes by stating that half explicitly, so that the design's claim on the reader's belief is visible as a structured argument rather than a sequence of separate decisions. What is asserted here is the design-specific portion of that argument, which the later chapters integrate with the theoretical and empirical portions.

That the underlying question is genuine is carried by the other chapters and taken as established: NASA and the Jet Propulsion Laboratory rely on an unmeasured heritage-lowers-cost assumption, documented qualitatively by Gao and Chien [\[42\]](#ref-42) and by the demonstration record. That the question is consequential follows from the portfolio-decision stakes and the documented gap between the ordinal technology-readiness-level scale and a cost-anchored learning rate, established by Mankins [\[70\]](#ref-70) and Olechowski [\[81\]](#ref-81). Sections 5.1 and 5.2 of this chapter show that the measurement reaches the causal mechanism rather than a generic correlation: the within-class log-log fixed-effects estimator measures the increasing-returns learning effect that Arthur [\[3\]](#ref-3) identifies, off variation the fixed effects have stripped of the leading class and time confounds. Section 5.1 establishes why the design is preferable to its rivals: ordinary least squares on the validated log-log form, established by Nagy [\[79\]](#ref-79) and the applied learning-rate literature [\[74\]](#ref-74), [\[80\]](#ref-80), [\[86\]](#ref-86), conserves the degrees of freedom that a small panel makes precious, and the two-way fixed-effects structure separates learning from the time and scale confounds that a bare experience curve would conflate, as Wei [\[103\]](#ref-103) shows happens when the time component is left uncontrolled. Sections 5.3, 5.4, and 5.5 together keep the remaining risk within tolerable bounds: the small-sample, measurement-error, and path-dependence risks are held in check by the pre-registered robustness battery, by the inference and influence diagnostics matched to the panel size, and by the honest design-stage framing that interprets a wide interval as inconclusive rather than as support for the null. That remaining risk is tolerable not because it is eliminated, which a small panel of imperfectly measured costs forbids, but because every component of it is named in advance and paired with a stated, pre-committed response, so that the eventual estimate, whichever way it falls, is read against a procedure the analyst could not have bent to a desired answer.

The design developed in this chapter therefore delivers what the contribution requires: a defensible procedure for estimating the single coefficient \(\beta\), an explicit account of the variation that identifies it, a fixed set of robustness checks tied to named threats, a complete and honestly bounded threats-to-validity analysis, and an interpretation discipline that holds the slope to be a path-contingent average for the NASA and Jet Propulsion Laboratory autonomy portfolio rather than a universal constant. What the design does not yet contain is a number, and that absence is deliberate and stated: the panel is not assembled, no coefficient is fitted, and the analysis plan that turns this design into an executed estimate, with its five-step procedure, its two mandatory pre-analysis checks, its fixed decision rule, and its explicitly illustrative expected-results block, is the subject of Chapter 6. The identification argument is complete; the estimation remains to be run.



# Chapter 6. Analysis Plan and Expected Results

## 6.0 Aim of the chapter

This chapter fixes, in advance of any estimation, exactly how the experience-curve slope will be computed, exactly how that slope will be turned into an accept-or-reject verdict on the null hypothesis, and exactly what each possible outcome would mean. The thesis of the chapter is a single sentence: because the autonomy panel is small and the dependent variable is constructed rather than directly observed, the credibility of the eventual measurement depends far more on the discipline of a pre-registered analysis plan than on the cleverness of the estimator, and so the plan is written now, while no coefficient has yet been fitted, so that interpretation is constrained by rules set before the data could influence them. Everything that follows develops that thesis. The five-step estimation procedure is specified so that a reviewer could reproduce it; two mandatory pre-analysis checks are committed to so that a single high-leverage episode or a degenerate fixed-effects cell cannot silently drive the result; the decision rule on the null is stated as a fixed function of the estimated slope and its confidence interval, carried unchanged from the prospectus; the expected-results discussion is presented in conditional, explicitly illustrative form so that no reader can mistake a design-stage expectation for an executed finding; and the falsification condition is stated symmetrically, so that the flat-cost null is as reportable a result as the declining-cost alternative.

The reason this matters is mechanical, not rhetorical. The experience-curve estimation literature is concentrated and consistent on one point: the central empirical risk is not that the wrong functional form is chosen, because the Wright power law in log-log form is the best-validated form available [\[79\]](#ref-79), but that an apparent cost decline is read as learning when it is in fact a confound, an artifact of a few influential observations, or the product of an analyst's freedom to choose among specifications after seeing the data [\[61\]](#ref-61), [\[35\]](#ref-35). A design-stage analysis plan is the instrument that removes the third of those risks by construction and bounds the first two by committing to diagnostics in advance. This chapter is therefore not a description of results. It is the pre-commitment that gives any future result its evidentiary weight.

A note on status is owed at the outset and is repeated where it matters. The full panel has not been assembled, the three-layer cost normalization is not complete, and no coefficient has been fitted on the full dataset. Every number that appears in Section 6.4 is illustrative and is labelled as such; it exists only to show the shape of a result and how it would be read against the hypotheses, and it is not an estimate. The result tables in Section 6.5 are specified but deliberately left unpopulated. This is the honest posture of a proposal, and the chapter keeps it throughout.

## 6.1 Problem frame for this chapter

The current state is that NASA and the Jet Propulsion Laboratory reason about the heritage value of flown autonomy qualitatively. A capability that has flown is treated as a heritage asset assumed to lower the cost of the next demonstration in its class, and that assumption drives portfolio sequencing and the technology-readiness-level accounting that tracks maturation but not its cost [\[70\]](#ref-70), [\[81\]](#ref-81). The desired state is a falsifiable, cost-anchored learning rate per autonomy capability class: a number, with an honest interval, that says how much cheaper the next same-class demonstration becomes per doubling of cumulative flight-demonstrated heritage. The gap, at the level this chapter operates, is procedural. Chapters 4 and 5 specified the data and the estimator; what remains is the precise analytic recipe that turns the constructed panel into that number and into a verdict, written so that the recipe cannot be quietly bent once the data are in hand. The consequence of leaving that recipe implicit would be severe in a study of this kind: with tens of observations and two sets of fixed effects, the analyst's degrees of freedom in choosing transformations, exclusions, and specifications are large relative to the data, and an undisciplined search across them could manufacture a significant slope from noise [\[61\]](#ref-61). The remedy is to fix the recipe in advance. That is the work of this chapter.

This framing carries forward the argument established in the earlier chapters. The heritage-lowers-cost assumption is relied upon yet unmeasured [\[42\]](#ref-42), and it matters because portfolio decisions turn on it while the only standing maturity instrument, the technology-readiness-level scale, is ordinal and non-monetary [\[81\]](#ref-81). A within-class log-log experience curve is the direct measurement of the learning effect that the heritage argument invokes [\[3\]](#ref-3), [\[79\]](#ref-79), and ordinary least squares on the log-log two-way fixed-effects form is the best-validated specification, with the fixed effects separating learning from time and scale confounds [\[79\]](#ref-79), [\[94\]](#ref-94). The small-sample, measurement-error, and path-dependence threats that remain are held within tolerable bounds by the pre-registered robustness specifications and the diagnostics this chapter commits to, and the design-stage framing makes no claim it has not yet earned [\[14\]](#ref-14), [\[5\]](#ref-5). The plan below is where the fourth and fifth links of that argument are operationalized.
This chapter does not invoke any architecture-traceability layer. By the study's explicit scope, it is an econometric measurement rather than an analysis of a real system, capability, or data-service exchange, so no capability-to-system-to-data-exchange chain is constructed and no DoDAF or BEA vocabulary is forced onto the argument. The single permitted conceptual link, that the fitted slope is an input to a portfolio decision, is reserved for the discussion chapter and is stated there in plain prose, not as an architecture table.

## 6.2 The estimation procedure

The estimation proceeds in five ordered steps. Each step is specified at the level of detail a second analyst would need to reproduce it, because reproducibility is part of the contribution of a cost study: a measurement a reviewer cannot re-derive is an assertion, not a measurement.

### 6.2.1 Step one: build the episode inventory

The first step assembles the episode inventory. The inventory is the population of capability-class development episodes defined in Chapter 4: each row is a single mission or technology demonstration that fielded one defined onboard autonomy capability, assigned to exactly one primary capability class, with multi-capability episodes split or flagged to their dominant class. The inventory is built from two of the four named sources. NASA TechPort supplies the project descriptions, start and end dates, organizational ownership, technology-readiness-level entry and exit estimates, and taxonomy classification, and it is the system of record that anchors project timing. The NASA Technical Reports Server supplies the primary engineering and lessons-learned documentation for the named demonstrations and the capability definitions used to assign each episode to a class. Step one outputs a dated, class-assigned episode list with one row per episode and a provenance note recording which source established each field. The development-start date is recorded with particular care, because that date is the cut point for the heritage count constructed in step two; where TechPort and NTRS disagree on timing, the rule is to prefer the TechPort project record as the system of record and to flag the discrepancy.

The claim that this step can be executed reliably rests on the grounds that both sources are structured and publicly accessible and that the named demonstrations are documented in the primary literature with dated, attributable detail [\[42\]](#ref-42). The connecting logic is that a panel whose timing and class assignment are sourced to a system of record and to dated primary reports is auditable by a reviewer who did not build it. One caveat must be stated plainly: capability-class assignment requires judgment at the boundaries, because some episodes field capabilities that straddle two classes, and the dominant-class flag is a judgment call that a different analyst might make differently. The design accepts that contested boundary assignments are recorded and reported, not hidden, so that the sensitivity of the result to any single contested assignment can be examined. Confidence in the reliability of step one is high for the dated chronology and moderate for the class assignment, and what would raise the latter is independent coding of the boundary cases by a second reader with a measured agreement rate.

### 6.2.2 Step two: construct cumulative within-class heritage

The second step constructs the independent variable. For each episode, the cumulative within-class flight-demonstrated heritage is the count of prior flight demonstrations in the same capability class that reached flight operation before that episode's development-start date. The counting rule is forward-only by design: a prior demonstration counts toward a later episode's heritage stock only if it reached flight operation before the later episode's development began, because heritage that arrives after development has started cannot have lowered that development's cost. The first observation in each class is the curve origin, and its cumulative count is set to one before the natural logarithm is taken, so that the logarithm is defined and the first demonstration anchors the class curve at the origin rather than being dropped. Step two outputs, for each episode, an integer heritage count and its natural logarithm, together with a counting log that records, for every nonzero count, which prior episodes were credited and the dates that justified the credit.

The forward-only rule is a deliberate methodological choice with a known directional consequence, and naming that consequence is part of the justification for the choice. The rule is conservative: by refusing to credit an episode with heritage it could not have used, it tends to undercount heritage and therefore biases the estimated slope toward zero. The mechanism is direct. If some later episode genuinely benefited from a near-contemporaneous demonstration whose flight operation post-dated the later episode's development start by a small margin, the forward-only rule withholds that credit and attenuates the measured association between heritage and cost. The result is that any rejection of the flat-cost null obtained under this rule is more credible than it would be under a permissive rule, because the rule works against finding an effect. This is the desired direction for a conservative test: a design should err toward failing to confirm its own hypothesis, not toward confirming it. Confidence that the rule is conservative in direction is high, because the reasoning is a property of the counting convention and not of the data; the data will determine only the magnitude of the attenuation, not its sign.

### 6.2.3 Step three: assemble normalized development cost

The third step assembles the dependent variable, and it is the most delicate step in the procedure, as Chapter 4 established at length. Normalized development cost is the recurring engineering cost to qualify the capability for flight, expressed in constant-year dollars and normalized by capability scope on a NICM-class basis, and it is assembled through the three-layer procedure fixed in Chapter 4. The first layer extracts, where the documentation permits, the autonomy-specific non-recurring engineering portion of a project's cost, separating it from platform and instrument cost. The second layer, where the autonomy portion is not separately reported, applies a NICM-class parametric estimate of the autonomy development effort given the documented scope, using the parametric cost-model approach exemplified for space instruments by Stahl and colleagues [\[89\]](#ref-89). The third layer expresses every figure in constant-year dollars using a standard deflator. For each observation the procedure records which layer produced the figure and an associated reliability flag, and that flag feeds the inverse-imputation-error weighting in one of the pre-registered robustness specifications so that observations resting on the weakest imputation receive less weight. Step three outputs, for each episode, a normalized cost in constant-year dollars, its natural logarithm, the layer that produced it, and the reliability flag.

What this step claims is bounded honestly: the procedure does not eliminate measurement error in the dependent variable, it makes the error explicit and auditable. Public development-cost figures are reported inconsistently, some as full project costs that bundle platform, instrument, and autonomy, some as subsystem costs, and some not at all, so that a single uniform extraction is impossible. A layered procedure with a recorded provenance flag per observation converts an uncontrolled measurement problem into a controlled one, because the analyst and the reviewer both know which figures rest on direct reporting and which rest on imputation, and the weighting specification lets the result be examined with the weakest figures down-weighted. NICM-class imputation is itself a construct choice, and the parametric basis carries its own error [\[89\]](#ref-89). The design accepts, rather than dismisses, that the dependent variable is the dominant construct-validity threat to the whole study and is treated as such in the threats analysis of Chapter 5; the analysis plan's contribution is not to remove that threat but to make its footprint visible in the result through the reliability flag and the weighted specification. Confidence that the dependent variable is auditable is high; confidence that it is free of measurement error is, correctly, low, and the plan never claims otherwise.

### 6.2.4 Step four: fit the baseline and robustness specifications

The fourth step fits the models. The baseline is the canonical estimating equation carried unchanged from the prospectus. For episode i in capability class c and decade d,

\[ \ln(\text{Cost}_{icd}) = \alpha + \beta \, \ln(\text{CumHeritage}_{icd}) + \gamma_c + \delta_d + \epsilon_{icd} \qquad\qquad (1) \]

where \(\text{Cost}_{icd}\) is normalized development cost, \(\text{CumHeritage}_{icd}\) is cumulative within-class flight-demonstrated heritage, \(\gamma_c\) are capability-class fixed effects, \(\delta_d\) are decade fixed effects, and \(\epsilon_{icd}\) is the error term. The coefficient \(\beta\) is the experience-curve slope and is the single parameter on which the hypotheses turn; the implied learning rate is one minus two raised to the power \(\beta\), reported for interpretability. The estimator is ordinary least squares with the two sets of fixed effects entered as indicator variables. The choice of ordinary least squares on the log-log form, rather than a non-linear or Bayesian estimator, follows the experience-curve literature directly: the log-log linearization is the form whose predictive performance has been validated across many technologies [\[79\]](#ref-79), [\[35\]](#ref-35), it yields a single interpretable slope, and it keeps the degrees-of-freedom cost low, which matters acutely in a panel of tens of observations. The fixed effects are entered as indicators rather than as a random-effects structure because the capability classes and decades are the specific, non-sampled categories of interest, not draws from a larger population, and because a fixed-effects estimator does not impose the orthogonality assumption between the heritage variable and the class or decade effects that a random-effects estimator would require.

After the baseline, the three pre-registered robustness specifications are fit, and they are fixed in advance precisely so that they are not chosen after seeing the baseline result. Each is paired with the specific threat to validity it is designed to probe, so that a robustness specification is a directed test of a named alternative explanation rather than a generic sensitivity check.

The first robustness specification adds the technology-readiness-level maturation covariate, the technology-readiness-level at development start from TechPort, to separate the heritage effect from the simpler effect of starting at a higher maturity. The threat it probes is that later episodes tend to begin development at a higher maturity, and that the higher maturity, not the within-class heritage stock, is what lowers their cost; if the heritage slope survives the inclusion of the maturation covariate, the heritage effect is not merely a relabelling of a maturity effect. The interpretive rule is stated in advance: if the slope is materially unchanged when the covariate enters, the heritage reading is reinforced, and if the slope collapses toward zero, the apparent learning was maturity in disguise. The technology-readiness-level scale is used here strictly as a maturation control and not as a cost proxy, consistent with its documented ordinal and non-monetary character [\[70\]](#ref-70), [\[81\]](#ref-81).

The second robustness specification broadens the heritage measure to include cross-class software-component reuse, which should attenuate any downward bias from shared components that the within-class count misses. The threat it probes is the opposite of the first: that the within-class count understates the true reusable-knowledge stock because heritage flows across capability classes through shared flight-software components and design patterns, biasing the within-class slope toward zero. The energy-technology learning-rate literature is the closest external analogue for how a broadened experience measure shifts an estimated rate, and it shows that the choice of experience stock materially moves the learning rate, with rates differing depending on whether cumulative production, cumulative capacity, or a broader experience proxy is used [\[80\]](#ref-80), [\[74\]](#ref-74), [\[86\]](#ref-86). The interpretive rule is that if the broadened measure recovers a steeper negative slope than the within-class measure, the contrast localizes where reusable knowledge actually accumulates, which is a finding in its own right about the codification substrate.

The third robustness specification weights observations by the inverse of the imputation error so that episodes with the least reliable cost figures contribute least, using the reliability flag from step three. The threat it probes is that the result is driven by the most heavily imputed, least reliable cost figures; if the slope survives down-weighting those figures, the result does not rest on the weakest data. The interpretive rule is that a slope robust to inverse-error weighting is one whose sign is not an artifact of the NICM-class imputation, while a slope that depends on the heavily imputed observations is reported as contingent on the imputation procedure.

Inference is small-sample-robust throughout. Because the number of fixed-effect clusters is small, conventional cluster-robust standard errors are known to over-reject the null and to understate the true sampling variability, and the plan therefore commits in advance to small-sample-appropriate inference and to reporting confidence intervals rather than bare p-values, consistent with the uncertainty-quantification discipline that the experience-curve forecasting literature treats as essential [\[109\]](#ref-109), [\[61\]](#ref-61). The mechanism behind the over-rejection is well understood: with few clusters the cluster-robust variance estimator is itself estimated with large error, so the resulting test statistic does not follow its assumed reference distribution and rejects too often, which in this setting would mean declaring a learning effect present when the data do not support it. Reporting an interval rather than a point verdict, and characterizing that interval honestly when it is wide, is the design's protection against that failure. Step four outputs, for each of the four specifications, an estimate of beta, its small-sample-robust confidence interval, the implied learning rate, the residual degrees of freedom, and the standard regression diagnostics.

### 6.2.5 Step five: report the verdict and the diagnostics

The fifth step reports. The plan commits to reporting, for the baseline and each robustness specification, the estimated beta, its confidence interval, the implied learning rate, and the accept-or-reject decision on the null under the fixed rule of Section 6.3, together with diagnostic plots of log cost against log heritage by capability class so that the shape of the relationship is visible and not only summarized by a coefficient. The estimation code, the assembled panel, and the imputation log are retained and released so that the measurement is reproducible and auditable, which is the standard a defensible cost study must meet and which the energy-technology experience-curve reviews treat as a precondition for trusting a published learning rate [\[86\]](#ref-86), [\[109\]](#ref-109). Step five outputs the populated result tables specified in Section 6.5 and the accompanying diagnostic figures. At the design stage those tables are specified and left unpopulated by design; populating them is the work of the build phase, not of this proposal.

The five steps are stated as a fixed sequence because their order matters. Heritage counts cannot be constructed before the dated inventory exists; cost cannot be normalized before the scope of each episode is fixed by the inventory; the models cannot be fit before both variables are assembled; and the verdict cannot be read before the diagnostics confirm the slope is not an artifact of a degenerate cell or a single episode. The next section specifies the two checks that gate the move from step four to step five.

### 6.2.6 Why the procedure is fixed rather than exploratory
The procedure above is written as a fixed recipe rather than as a menu of analytic options to be selected during estimation, and that choice is itself a methodological commitment with a defensible justification. In a large panel, an analyst can afford to explore. Split-sample validation, held-out forecasting, and post-hoc specification search are disciplined when there are enough observations to keep an exploratory and a confirmatory sample separate. In a panel of tens of observations there is no such luxury. Every observation must enter the single confirmatory fit, which means the analyst's choices among transformations, exclusions, and specifications cannot be validated against held-out data and must instead be constrained by pre-commitment. The experience-curve forecasting literature shows that the predictive reliability of a learning rate degrades sharply when the estimation choices are made after seeing the series, because the analyst's freedom is large relative to the information content of a short series [\[61\]](#ref-61), [\[35\]](#ref-35). The mechanism is the familiar one of researcher degrees of freedom: with many defensible analytic paths and a noisy short series, at least one path will, by chance, produce an apparently clean negative slope, and an analyst who reports that path without having committed to it in advance reports noise as signal. The fixed recipe removes that path-selection freedom by construction. A fixed recipe also forecloses the discovery of structure the design did not anticipate; the design accepts this cost deliberately, because in a small panel the risk of manufacturing a false positive through flexibility outweighs the value of the discoveries flexibility might enable, and any structure the recipe misses can be pursued in a clearly labelled, exploratory follow-up that does not contaminate the confirmatory verdict.

## 6.3 Pre-analysis checks and the fixed decision rule

### 6.3.1 Why two checks gate the verdict

The claim of this section is that in a small panel the estimated slope must pass two feasibility-and-robustness checks before it is interpreted, and that committing to these checks in advance is what separates a credible small-N measurement from a fragile one. The basis is the structural facts of the design. With tens of observations and two sets of fixed effects, degrees of freedom are scarce, and two failure modes are foreseeable rather than hypothetical. The first failure mode is a degenerate fixed-effects cell. The second is a single high-leverage episode driving the slope. Both belong as gating checks rather than after-the-fact caveats because the analyst can detect both mechanically before reading the coefficient, and detecting them in advance prevents the slope from being reported as if it rested on more evidence than it does. This rests on the experience-curve literature's repeated finding that learning-rate estimates are sensitive to the data window and to influential points, so an estimate reported without an influence check is incompletely characterized [\[109\]](#ref-109), [\[61\]](#ref-61). These checks bound but do not remove small-sample fragility; a panel of tens of observations is intrinsically limited and no diagnostic manufactures statistical power that the data do not contain. The honest response to limited power is to report it transparently, not to suppress it, which is why both checks feed directly into how the slope is described rather than into whether it is described.

### 6.3.2 The fixed-effects feasibility check

The first check is fixed-effects feasibility. Before the slope is interpreted, the plan checks whether any capability class or any decade contains only a single observation. A class or decade with a single observation contributes nothing to the within estimator, because the within transformation removes all variation in a singleton cell, and so the effective sample that identifies the slope is smaller than the raw episode count suggests. The check records, for each capability class and each decade, the number of episodes it contains, and it reports the effective sample after singleton cells are accounted for. The affected observations are flagged so the reader knows the slope is identified off a smaller effective sample than the headline count implies. This is not a discretionary diagnostic to be run if the result looks suspect. It runs unconditionally, and its output is reported alongside the coefficient whatever the coefficient turns out to be. The mechanism by which it protects the inference is direct: it prevents the residual degrees of freedom from being overstated, which would otherwise narrow the reported confidence interval spuriously and make a noisy slope look more certain than it is. This check is necessary because singleton cells are a near-certainty given the small, unevenly distributed autonomy demonstration record across classes and decades.

### 6.3.3 The small-panel influence diagnostic

The second check is the influence diagnostic. Because the panel is small, a single episode can exert outsized leverage on the fitted slope; an unusually expensive first-of-kind demonstration, for example the Remote Agent origin observation whose lessons-learned record documents that the impact of inserting system-level autonomy into a flight project was a major surprise, is exactly the kind of point that could anchor a curve by itself. The plan therefore reports the slope estimated with and without each potentially influential observation, identified by standard leverage and influence measures, and it treats a slope that depends on a single episode as not robust. The rule is stated in advance: if removing any one observation moves the slope from the rejection region to the non-rejection region, or vice versa, the slope is reported as fragile and the conclusion is qualified accordingly. This converts a known small-sample vulnerability into a reported property of the result rather than an unexamined risk. Influence diagnostics are the standard small-sample safeguard, and the experience-curve forecasting literature treats sensitivity to individual data points as a first-order source of learning-rate uncertainty [\[109\]](#ref-109), [\[61\]](#ref-61). With very few observations even a robust slope rests on thin evidence, and the diagnostic certifies only that no single point is decisive, not that the slope is precise. Both checks are stated here so they constrain interpretation in advance rather than being selected after the result is seen, which is the entire point of pre-registering them.

The influence diagnostic interacts with the origin-observation convention in a way that must be anticipated, because the first-in-class episodes are simultaneously the most influential and the most structurally necessary points in the panel. The forward-only counting rule places each class's first demonstration at the curve origin with its cumulative heritage set to one, which means the first-in-class episodes carry the smallest value of the independent variable and, if first-of-kind demonstrations are systematically the most expensive, also the largest value of the dependent variable. A point with an extreme value on both axes is precisely the configuration that maximizes leverage. The design does not respond by discarding these points, because they are the anchors that define the experience curve and removing them would amount to estimating a learning rate without an origin. Instead it reports the slope's sensitivity to each origin observation explicitly and interprets a slope that depends entirely on the origin points with caution. The mechanism worth naming is that an experience curve fit through a high-cost origin and a cluster of lower-cost successors will register a negative slope whether the cost decline is genuine learning or merely a single expensive first attempt followed by ordinary successors; the influence diagnostic distinguishes a curve supported by a graded decline across multiple heritage levels from one driven by the gap between a single expensive origin and everything after it. Reporting that distinction is the difference between a defensible learning rate and an artifact of the origin convention, and the plan commits to drawing it.

A further consequence of the small panel deserves statement here, because it bounds what the influence diagnostic can deliver. With a panel of tens of observations distributed unevenly across capability classes and decades, the number of points that are individually influential may be a substantial fraction of the sample rather than a handful of outliers. The plan does not treat a high count of influential points as a reason to suppress the result; it treats it as information about the fragility of the measurement and reports it as such. If the diagnostic reveals that the verdict survives the removal of any single point but not the joint removal of the two or three most influential, that pattern is reported, because a verdict robust to one deletion but not to a small joint deletion is weaker than one robust to both, and the reader is entitled to that distinction. This is the honest face of a small-N design: the diagnostics do not rescue the sample from its size, they make the limitations of that size legible.

### 6.3.4 The fixed decision rule

The decision rule is fixed before estimation and is carried unchanged from the prospectus. The null hypothesis is rejected in favor of the alternative if and only if the estimated beta is negative and its confidence interval excludes zero in the baseline specification and in at least two of the three pre-registered robustness specifications. If beta is not distinguishable from zero, or is positive, the null is not rejected and the flat-cost conclusion stands. This rule is a function only of the sign of the slope and whether its interval excludes zero across a fixed majority of specifications; it contains no free parameter that could be tuned after the fact. The requirement that the result hold in the baseline and in at least two of three robustness specifications is the multiple-specification discipline the design substitutes for an undisciplined search: rather than reporting whichever of many specifications gives the strongest result, the plan pre-commits to a conjunction across a fixed, small set, so a result driven by one fragile specification cannot clear the bar [\[61\]](#ref-61), [\[14\]](#ref-14). The decision rule also interacts with the two pre-analysis checks. A slope that satisfies the numerical rule but fails the influence diagnostic, in the sense that its rejection of the null depends on a single observation, is reported as a rejection that is not robust, and the verdict is qualified rather than asserted. The rule and the checks together are the mechanism by which the design controls the analyst's degrees of freedom; they are the operational core of the claim that the residual risk of a spurious finding is acceptable.

## 6.4 Illustrative, explicitly non-empirical expected results

### 6.4.1 The status of every number in this section

Before any expectation is stated, the status of this section is fixed without ambiguity. The full panel has not been assembled, the cost normalization is not complete, and no coefficient has been fitted on the full dataset. Every figure that appears below is illustrative. None is derived from a fitted model, and none should be read as an estimate. The figures exist only to make concrete the shape of a result that would reject the null and the shape of a result that would not, so a reader can see in advance how the eventual coefficient will be interpreted against the hypotheses. This labelling is not a disclaimer added for caution; it is a structural commitment of the design-stage posture, and it is repeated in this section because this is the section a reader is most likely to mistake for a findings report. There is no findings report in this dissertation. There is a plan, and an illustration of how a finding would be read.

### 6.4.2 The shape of a result that would reject the null

Consider first the shape of a result consistent with the alternative hypothesis. If onboard autonomy qualification cost followed an experience curve with a learning rate of the order commonly observed for software-intensive and technology-development cost series, the estimated slope beta would be a moderate negative number, the implied per-doubling cost reduction would be a meaningful fraction, and the confidence interval would exclude zero in the baseline and in a majority of the robustness specifications. To make the reading concrete and nothing more, suppose, purely illustratively, that a fitted baseline returned a slope whose implied per-doubling cost reduction fell somewhere in the low tens of percent. Such a value would be unremarkable against the broad range of learning rates documented across technologies, where the energy-technology reviews report learning rates spread widely around a central tendency near the high teens to low twenties of percent per doubling and emphasize that the spread across technologies is large [\[86\]](#ref-86), [\[74\]](#ref-74), [\[80\]](#ref-80). The interpretation of such an illustrative value would be that, in the autonomy classes for which heritage accumulated within the window, each doubling of cumulative flight-demonstrated heritage was associated with a cost reduction in that range after capability class and decade were controlled. This direction is the theoretically expected signature if learning effects dominate, because a negative log-log slope is what Arthur's learning-effect mechanism predicts: each use of a technique lowers the cost of the next use, generating the positive feedback that an experience curve measures [\[3\]](#ref-3). The expected-sign reasoning is therefore not a guess but a mechanism. Cumulative within-class heritage produces codified, reusable design patterns, verification approaches, and flight-software components; those reduce the recurring engineering needed to qualify the next same-class capability; and that reduction registers as a negative slope. Stating the mechanism is what licenses the expected sign; without it, a negative slope would be a bare correlation and the confidence attached to it would be lower.

The expected heterogeneity of the slope is a second, sharper prediction, and it follows from the Mokyr lens carried through the dissertation. Mokyr's distinction between propositional and prescriptive useful knowledge predicts that the cost-decline curve will be steeper for capability classes whose underlying propositional base is mature and whose heritage knowledge is openly codified and reusable, for example through a shared flight-software substrate, and flatter where each project re-implements from scratch. The illustrative expectation is therefore not a single slope but a family of class-specific slopes, with the steeper members of the family being the classes that sit on a codified, reusable substrate. This expectation is testable as effect heterogeneity across capability classes and is the most policy-relevant secondary pattern the design can produce, but at the design stage it remains an expectation, not a measurement, and it is labelled as such. Confidence that the direction of any detected effect would be negative is moderate, resting on the convergence of the experience-curve mechanism and the autonomy heritage-reuse argument; confidence in any specific magnitude is low by construction, because no magnitude has been estimated, and the illustrative figure is offered only to anchor the reader's expectation of scale, not to forecast a value.

A second feature of the expected-sign reasoning deserves to be made explicit, because it bears on how a detected slope would be read. The expected magnitude is not anchored only to the central tendency of cross-technology learning rates; it is also bounded above by the conservative counting rule established in Section 6.2.2. Because the forward-only rule withholds heritage credit from near-contemporaneous demonstrations, the measured slope is expected to be shallower in magnitude than the true within-class learning rate, whatever that rate is. The two illustrative considerations therefore point in a consistent direction: the central tendency of the cross-technology evidence suggests a moderate negative slope is plausible if learning operates [\[86\]](#ref-86), [\[74\]](#ref-74), and the conservative counting rule suggests that whatever slope is measured will, if anything, understate the underlying effect. This matters for interpretation because it means a measured slope near zero is genuinely ambiguous between a true absence of learning and a true learning effect attenuated by the counting rule and the small sample, whereas a measured slope clearly bounded away from zero is, given the conservative rule, strong evidence that learning operates. The plan reports the slope in a way that keeps this asymmetry visible, so the reader does not over-read a near-zero estimate as a confident null.

A third illustrative consideration concerns what the implied learning rate would mean in operational terms if a moderate negative slope were obtained, because the point of the measurement is to be usable, not only to be significant. The implied learning rate, one minus two raised to the power beta, translates a slope into a per-doubling cost reduction that a program office could reason with. Illustratively, and only illustratively, a learning rate in the low tens of percent would mean that fielding a target autonomy capability becomes affordable for a cost-capped mission only after the cumulative within-class heritage has roughly doubled some specific number of times, and the slope would tell the portfolio how many intermediate demonstrations that doubling requires. The energy-technology reviews are again the relevant analogue for the practice of converting a fitted slope into a planning-grade learning rate and for the caution that such a rate carries wide uncertainty when the underlying series is short [\[109\]](#ref-109), [\[80\]](#ref-80). At the design stage this translation is shown only to demonstrate that the measurement, if obtained, would be decision-relevant in the specific build-or-wait sense the prospectus identified; no learning rate has been estimated and none is forecast here.

### 6.4.3 The shape of a result that would not reject the null

Consider next the shape of a result consistent with the null hypothesis, which the plan treats as an equally reportable outcome. The analysis may return a slope indistinguishable from zero, or a positive slope, or a negative slope whose confidence interval contains zero, or a negative slope that fails the influence diagnostic because it depends on a single episode. Any of these would fail to reject the flat-cost null. The substantive reading of such an outcome would be that within-class heritage, as measured by the forward-only cumulative count, does not on its own lower qualification cost along a detectable experience curve once capability class and decade are controlled. This reading would not be a failure of the study; it would be a finding. It would indicate that the heritage-lowers-cost assumption, which NASA and JPL currently treat as self-evident, is not supported as an experience-curve relationship in the open record, and it would shift the portfolio rationale for autonomy investment away from cost reduction and toward capability value. The mechanism by which the null could obtain is also nameable in advance: heritage may flow across capability classes through shared software components and design patterns that the within-class count cannot see, so the true reusable-knowledge stock is poorly proxied by the within-class heritage variable, and the measured within-class slope is attenuated toward zero even where genuine learning occurs. The cross-class robustness specification is the design's attempt to detect exactly this pattern; if the broadened heritage measure recovers a negative slope that the within-class measure misses, that contrast is itself an informative result about where reusable knowledge actually lives.

The symmetry between the two readings is the design's central virtue and it is worth stating plainly. Both outcomes are decision-relevant and both are reportable. The reasoning rests on the structure of the hypotheses: a negative significant slope yields a planning parameter for portfolio sequencing, while a flat slope reframes the portfolio rationale around capability value rather than cost decline. A study designed so that only one outcome is publishable is a study whose result is partly predetermined by the publication incentive, whereas a study whose two outcomes are both informative is one whose finding is driven by the data [\[14\]](#ref-14). A null result in a small, low-power panel is weaker evidence for the absence of an effect than a rejection is for its presence, because failure to detect can reflect insufficient power rather than a true zero; the plan therefore commits to reporting the confidence interval and the effective sample so a wide interval containing zero is read as inconclusive rather than as positive evidence of a flat cost, and treats a wide interval that contains zero as a failure to reject rather than as evidence for the alternative. No claim is made here about which outcome will occur. The plan is built to make either one credible.

## 6.5 Falsification, the specified result tables, and reproducibility

### 6.5.1 What a falsifying result looks like
The contribution of the dissertation is falsifiable, and this section states the falsification condition exactly. The contribution is falsified if, on the assembled panel, the estimated beta is zero or positive, or its confidence interval contains zero across the baseline and the robustness specifications under the fixed decision rule of Section 6.3. In that case the heritage-lowers-cost assumption is not supported as an experience-curve relationship, and the study reports the flat-cost null as its finding. The contribution is also weakened, though not strictly falsified, in a specific and pre-named way. If an apparently negative slope in the baseline collapses toward zero once the technology-readiness-level maturation covariate or the cross-class heritage measure is added, that collapse indicates that the raw association reflected maturity or shared components rather than within-class flight heritage, and the slope is reported as not robust to the named confound. Naming these two distinct ways the contribution can fail, outright falsification and confound-driven attenuation, in advance is what makes the test honest: the design specifies the conditions under which it will declare its own central claim unsupported, and those conditions are not adjustable after the data are seen.

### 6.5.2 The specified but unpopulated result tables

The result tables are specified here and left unpopulated by design, because no coefficient has been fitted. Three tables are specified. The first is the main results table: one row per specification (baseline; baseline plus the technology-readiness-level covariate; baseline with cross-class heritage; baseline with inverse-imputation-error weighting), with columns for the estimated slope beta, its small-sample-robust confidence interval, the implied learning rate, the residual degrees of freedom, and the effective sample after the fixed-effects feasibility check. The second is the influence-diagnostic table: one row per potentially influential episode, with the slope estimated with and without that episode and a flag indicating whether its removal changes the verdict. The third is the heterogeneity table: one row per capability class, reporting the class-specific slope where the within-class effective sample supports estimating one and a not-estimable flag where it does not, so that the Mokyr-predicted codification heterogeneity can be read where the data permit and its absence acknowledged where they do not. Each table's column headers, units, and notes are fixed now; the cells are blank at the design stage and are populated only in the build phase. Presenting the tables as specified-but-blank is a deliberate honesty device. It shows the reader exactly what will be reported, prevents the eventual results from being reshaped to flatter a hypothesis, and makes unmistakable that nothing in this dissertation is an executed estimate.

### 6.5.3 Reproducibility and the retained imputation log

Reproducibility is the final commitment of the analysis plan, and it is a substantive part of the contribution rather than an administrative afterthought. The plan commits to retaining and releasing the estimation code, the assembled panel, the forward-only cumulative-heritage counting log, and the three-layer cost-normalization imputation log with its per-observation reliability flags. The imputation log matters most of these. Because the dependent variable is the study's dominant construct-validity threat, a log that records, for every observation, which normalization layer produced its cost figure and how reliable that figure is judged to be allows a reviewer to re-run the analysis under different weighting choices and see whether the verdict survives [\[89\]](#ref-89). The case for treating reproducibility as load-bearing is drawn directly from the energy-technology experience-curve reviews, which find that published learning rates vary substantially with the analyst's data choices and conclude that a learning rate is interpretable only when its construction is fully documented and checkable [\[86\]](#ref-86), [\[109\]](#ref-109), [\[80\]](#ref-80). One caveat is that reproducibility certifies the procedure, not the truth of the underlying cost figures: a fully reproducible analysis built on imperfectly measured costs inherits that measurement error, and the plan does not pretend otherwise. The design accepts that the appropriate response to an imperfectly measured dependent variable is to make the imperfection auditable and to report the result's sensitivity to it, which the retained logs and the weighted robustness specification together accomplish.

The reproducibility commitment extends beyond releasing artifacts to specifying, in advance, the precise analytic choices that a re-analyst would otherwise have to guess. Three such choices are fixed here so that they cannot become hidden researcher degrees of freedom. The first is the deflator: every cost figure is expressed in constant-year dollars using a single named deflator applied uniformly, and the base year and the deflator series are recorded so that the constant-year transformation is exactly reproducible and so that a reviewer who prefers a different deflator can re-run the analysis and compare. The second is the treatment of the origin observation: the convention that each class's first demonstration takes a cumulative heritage of one before logging is recorded as a stated rule with its rationale, so that a re-analyst understands why the origin points are not dropped and can test the sensitivity of the slope to that convention directly. The third is the boundary coding of contested capability-class assignments: the dominant-class flag for every multi-capability episode is recorded with the reasoning behind it, so that the sensitivity of the verdict to any single contested assignment is examinable rather than buried. Fixing these three choices in the released log is what allows the central claim of the study, the fitted slope, to be checked rather than merely believed, and it is the operational meaning of the requirement that the residual risk be acceptable: a measurement whose every consequential analytic choice is recorded and testable carries a residual risk a reviewer can bound, whereas a measurement whose choices are opaque carries a residual risk no reviewer can assess. The energy-technology reviews make exactly this point when they trace the dispersion of published learning rates partly to undocumented analyst choices and call for the construction of each rate to be made transparent [\[86\]](#ref-86), [\[109\]](#ref-109). The plan adopts that standard as a binding commitment rather than an aspiration.

### 6.5.4 How this chapter advances the argument

This chapter has done the work its place in the argument requires. It has specified a five-step estimation procedure detailed enough to reproduce; committed in advance to two pre-analysis checks that convert the foreseeable small-sample failure modes into reported properties of the result; fixed a decision rule on the null that is a function only of the sign of the slope and whether its interval excludes zero across a pre-registered majority of specifications; illustrated, in explicitly non-empirical and clearly labelled form, the shape of a result that would reject the null and the shape of one that would not, with the expected sign grounded in a named mechanism rather than asserted as a correlation; and stated the falsification condition symmetrically, so that the flat-cost null is as reportable as the declining-cost alternative. In doing so it has operationalized the two final links of the argument: the design addresses the causal learning-effect mechanism through a within-class log-log curve [\[3\]](#ref-3), [\[79\]](#ref-79), and the residual risk of a spurious or fragile finding is held to an acceptable level by the pre-registered robustness conjunction, the influence diagnostic, the fixed-effects feasibility check, and the retained reproducibility artifacts [\[61\]](#ref-61), [\[14\]](#ref-14), [\[109\]](#ref-109). What the chapter has deliberately not done is report a result, because there is none to report; the panel is unbuilt, the cost normalization is incomplete, and every number above is an illustration. The contribution remains a single falsifiable measurement to be made under rules now fixed: a fitted experience-curve slope of autonomy qualification cost on cumulative flight-demonstrated heritage, or a credible failure to find one, each of which is decision-relevant for the Autonomous Systems and Robotics portfolio [\[42\]](#ref-42), [\[81\]](#ref-81). The discipline of writing the plan before the estimate is what will let that eventual number, whatever its sign, be believed.



# Chapter 7. Discussion

## 7.0 The chapter's central claim

Whether the experience-curve slope \(\beta\) comes back negative or comes back flat, the completed study converts a NASA and Jet Propulsion Laboratory planning question from an assertion into a measured quantity, and the measured quantity is decision-relevant in either case. That is the claim this chapter develops. The two outcomes are not a success branch and a failure branch. They are two readings of the same instrument, each of which changes what a program office can defensibly say about the cost of fielding the next onboard-autonomy capability. If H1 holds, the fitted slope is a usable sequencing parameter, and the heritage-lowers-cost argument that justifies much autonomy investment is vindicated and quantified. If H0 holds, the same argument is shown to rest on something other than a measurable cost decline, and the portfolio rationale must be rebuilt on capability value rather than on declining qualification cost. The chapter takes each branch in turn, then engages the rival explanations that could counterfeit a learning curve, then states what the design can and cannot say about classes it did not observe and worlds outside NASA and JPL deep-space autonomy.

The discussion is bounded by the honest posture carried from the prospectus. No coefficient has been fitted on the full dataset. Every numerical reading in this chapter is illustrative and explicitly labeled as such; the conditional mood ("if the panel yields a slope of the form...", "were the interval to exclude zero...") is used deliberately and is not throat-clearing. A design-stage discussion states in advance how each possible result will be interpreted, so that the interpretation is a commitment rather than a rationalization assembled after the number is seen. That pre-commitment is itself part of the contribution, because it is what makes the eventual measurement falsifiable in the strong sense: the reader knows now what a confirming result and a disconfirming result will each be taken to mean.

### 7.0.1 The problem this chapter addresses

The current state is that NASA and JPL hold a working belief, documented across the autonomy demonstration record, that each flown autonomy capability is a heritage asset that lowers the cost of the next same-class demonstration [\[42\]](#ref-42), [\[11\]](#ref-11), [\[24\]](#ref-24), [\[34\]](#ref-34), [\[40\]](#ref-40). The belief is operationally consequential: it underwrites portfolio sequencing in the Autonomous Systems and Robotics line, and it is the implicit justification for treating an intermediate, lower-stakes demonstration as an investment that buys down the cost of a later, more ambitious one. The desired state is a discussion that says, precisely, what follows for that belief once a fitted slope exists, under both possible values of the slope, with the rival explanations weighed and the limits of generalization stated. The gap is that a measurement, however clean, does not interpret itself; an experience-curve coefficient on a small panel of spacecraft-autonomy episodes can be read naively as a learning rate when it is in fact a time trend, a scale effect, or a selection artifact, and a flat coefficient can be over-read as proof that heritage is worthless when it may only mean that the within-class heritage count is too coarse a proxy for the reusable-knowledge stock that actually matters. Leave that interpretive gap unfilled and the eventual number, whatever it is, becomes available for misuse in both directions: a negative slope cited as license to spend, a flat slope cited as license to stop investing, neither reading disciplined by the threats the design was built to address. This chapter fills the gap by fixing the interpretation in advance and by carrying the argument all the way to the decision endpoint.

## 7.1 If H1 holds: the slope as a planning parameter

If the assembled panel returns a \(\beta\) that is negative and whose confidence interval excludes zero in the baseline specification and in at least two of the three pre-registered robustness specifications, then the fitted slope is a usable planning parameter for autonomy portfolio sequencing, and specifically it makes the build-or-wait decision answerable by calculation rather than by assertion.

The estimating equation is \(\ln(\text{Cost}_{icd}) = \alpha + \beta \, \ln(\text{CumHeritage}_{icd}) + \gamma_c + \delta_d + \epsilon_{icd}\), and \(\beta\) is by construction the elasticity of normalized autonomy qualification cost with respect to cumulative within-class flight-demonstrated heritage. A negative \(\beta\) means that each proportional increase in accumulated heritage is associated with a proportional decline in the cost of qualifying the next same-class capability, after capability-class and decade effects are absorbed. The quantity \(1 - 2^{\beta}\) translates that elasticity into the more legible learning rate: the fraction by which qualification cost falls each time the cumulative heritage stock doubles. This is the standard Wright reading of an experience-curve slope, and it is the reading the experience-curve literature has validated as a forecasting object across many technologies [\[79\]](#ref-79), [\[35\]](#ref-35), [\[94\]](#ref-94).

A slope is a planning parameter precisely when it converts a future cost into a function of a present investment decision. The build-or-wait problem is the canonical case. Suppose a future mission needs an autonomy capability in class \(c\) whose current qualification cost, at the heritage stock that exists today, exceeds the mission's cost cap. The program office faces a choice: build the capability now at the unaffordable price, or wait and let an intermediate, lower-stakes demonstration first accumulate one more unit of heritage in class \(c\), lowering the later qualification cost. A fitted \(\beta\) answers how much that intermediate demonstration would buy down the later cost, because the cost at heritage stock \(h+1\) relative to the cost at stock \(h\) is \(\left(\frac{h+1}{h}\right)^{\beta}\) in the model's units, holding class and decade fixed. With a measured \(\beta\), the office can compute roughly how many intermediate demonstrations are required to bring the target capability inside the cap, and can compare the summed cost of those intermediate demonstrations against the cost saved on the target, which is exactly the comparison a sequencing decision requires. Without \(\beta\), that comparison is made by assertion, and the prospectus is explicit that this is how the decision is made today.

The principle that a learning rate functions as a sequencing input is not novel to this study; it is the operating logic of the broader technology-cost-forecasting program. Nagy and colleagues establish that the Wright power law predicts technological cost with quantifiable accuracy [\[79\]](#ref-79); Farmer and Lafond characterize the error distribution of such forecasts so that a planner can attach a confidence band to a projected cost rather than a point [\[35\]](#ref-35); Lafond and colleagues extend this to full distributional forecasts, which is what a portfolio office needs when it is reasoning about a cost cap that the projected cost must fall below with some stated probability [\[61\]](#ref-61). The portfolio-selection literature supplies the decision frame into which such a parameter feeds: project portfolio selection is an explicit optimization over candidate investments under resource constraints, and a cost-decline parameter is exactly the kind of forward cost input that decision-support methods for portfolio selection consume [\[43\]](#ref-43). The launch-cadence evidence supplies a concrete, recent, space-sector instance of the same learning-effect mechanism operating on operational cost, where each flight lowers the cost and raises the automation of the next [\[91\]](#ref-91), [\[37\]](#ref-37); that evidence is taken up in detail in Section 7.5 as the closest available analogue.

The strength of this claim is conditional and bounded. It holds only if \(\beta\) survives the decision rule, and even then the planning parameter it yields is an average within-class slope over the observation window, not a guaranteed forward rate. Confidence in the planning use is at most **moderate** at the design stage, for three reasons that no fitted number can remove. First, the panel is small, on the order of tens of episodes, so even a slope that clears the decision rule will carry a wide interval, and the planning use must propagate that width rather than treat the point estimate as exact. Second, the slope is an average that may mask the heterogeneity Section 7.4 argues is the more policy-relevant finding; a single pooled rate applied to a class with an atypical learning rate would mislead. Third, and most fundamentally, the slope describes a realized historical path and is not a law; Arthur's path-dependence caution, developed in Section 7.5, means the rate that held over the observed sequence is not guaranteed to hold forward if the portfolio's investment pattern changes [\[4\]](#ref-4), [\[5\]](#ref-5). The planning parameter is therefore a disciplined prior for the build-or-wait decision, to be updated, not a deterministic cost oracle.

The claim would be defeated, even with a qualifying \(\beta\), if the slope were shown to be an artifact rather than a learning effect. That is the burden of Section 7.3. If the apparent decline collapses once the maturation covariate or the cross-class reuse measure is added, then \(\beta\) was capturing maturity or shared components rather than within-class flight heritage, and it cannot be used as a heritage-investment planning parameter even though it cleared the baseline. The decision rule's requirement that \(\beta\) survive in at least two of three robustness specifications is the structural protection against this objection, and the planning use is licensed only after that protection holds.

### 7.1.1 What the H1 reading does and does not authorize

State sharply what a qualifying slope would and would not license, because the misuse risk is asymmetric and runs toward over-claiming. A qualifying \(\beta\) would authorize a program office to treat heritage accumulation as a quantified cost-buy-down mechanism within a capability class, to rank candidate intermediate demonstrations by how much later cost each buys down per dollar spent, and to defend a wait-and-accumulate sequencing choice with a number and an interval rather than with a narrative. It would also supply the cost-anchored complement to the technology-readiness-level scale that the prospectus identifies as missing: the TRL scale tracks maturity but is ordinal and non-monetary [\[70\]](#ref-70), [\[81\]](#ref-81), so a program that knows a capability is at TRL 6 still does not know, from the TRL alone, the dollars required to advance it; a fitted \(\beta\) attaches a cost trajectory to the maturation path that TRL describes only qualitatively.

A qualifying \(\beta\) would not authorize extrapolation of the rate to capability classes outside the panel, to non-NASA or commercial autonomy, or to heritage stocks far beyond the observed range, because the experience-curve tradition treats a learning rate as transferable only within a coherent technology family and over the range where it was estimated [\[79\]](#ref-79), [\[14\]](#ref-14). Nor would it authorize treating the slope as a fixed agency-wide constant; the within-class fixed-effects design estimates an average within-class slope and explicitly does not claim a universal learning rate. The discipline here is the same the experience-curve literature imposes on itself: the curve is a within-family forecasting object with a stated error, not a physical law, and the policy reading must inherit that modesty.
## 7.2 If H0 holds: the rationale shifts from cost reduction to capability value

If the assembled panel returns a \(\beta\) that is statistically indistinguishable from zero, or positive, or whose confidence interval contains zero across the baseline and robustness specifications, then the flat-cost null stands. The consequence is not that autonomy investment is unjustified but that its justification must shift from heritage-driven cost reduction to the value of the capability itself.

Under the fixed decision rule, a slope that does not clear the negative-and-significant threshold in the baseline and in at least two robustness specifications leaves H0 unrejected. The substantive content of an unrejected H0 is that, on the measured panel, accumulated within-class flight heritage as operationalized here does not lower the recurring cost of qualifying the next same-class capability in a way the data can detect. That is a finding about the heritage-cost relationship, not about whether autonomy is worth flying.

An argument that justifies an investment by appeal to a cost decline is only as strong as the cost decline is real. If the cost decline is absent or undetectable, the argument does not merely weaken at the margin. Its load-bearing premise is removed, and the investment must be carried by a different premise or not at all. The heritage-reuse argument, as documented across the demonstration record, is exactly an appeal to cost decline: the next mission can field the capability more cheaply because the design patterns, verification approaches, and flight-software components are reusable [\[40\]](#ref-40). A flat \(\beta\) says that, whatever reuse is occurring, it does not show up as a measurable per-episode cost decline within the class. The honest response is to stop resting the portfolio case on that premise and to rest it instead on the premise that remains intact: that the capability itself raises mission science and operational return. That is the original reason autonomy is flown under long communication delays and constrained ground contact, and it holds independent of whether qualification cost falls.

That autonomy delivers capability value independent of any cost-learning effect is the better-documented half of the record. The EO-1 Autonomous Sciencecraft Experiment improved the quality and timeliness of returned science by moving detection and replanning onboard, used operationally to detect and monitor active volcanism, floods, and cryospheric change [\[24\]](#ref-24), [\[24\]](#ref-24). AEGIS reduced the operations latency imposed by ground-in-the-loop targeting of narrow-field instruments, first on Opportunity and then on ChemCam [\[34\]](#ref-34), [\[12\]](#ref-12). Perseverance flew autonomous navigation across the large majority of its early traverse, and Ingenuity demonstrated powered autonomous flight as a technology demonstration [\[99\]](#ref-99), [\[7\]](#ref-7). None of these value claims depends on a cost-learning slope; each asserts what the capability enables. A flat \(\beta\) would leave every one of them standing. The space-autonomy review situates these within a longer arc and states the value case directly, separate from the cost-heritage claim [\[42\]](#ref-42).

Confidence that an unrejected H0 should be read as "shift the rationale" rather than as "heritage genuinely does not lower cost" is **low to moderate**, and the asymmetry matters. A null result on a small, measurement-noisy panel is consistent with two very different worlds: one in which heritage truly does not lower autonomy qualification cost, and one in which it does but the instrument is too blunt to see it. The prospectus is explicit that the within-class heritage count is a coarse measure of the true reusable-knowledge stock that Mokyr's framework identifies as the real driver, and that software components and design patterns cross capability classes in ways the within-class count cannot capture, biasing the estimate toward zero. A flat slope is therefore weaker evidence against the heritage mechanism than a negative slope is for it, because the measurement error and the conservative forward-only counting rule both push toward zero. The H0 reading must say this plainly: the finding is that the cost decline is not detectable as specified, which is a reason to stop asserting it as a planning fact, not a proof that it is absent.

The H0 reading would be overturned if a later study with a larger panel, a finer heritage measure that credits cross-class component reuse, or a cleaner cost-normalization recovered a detectable negative slope. The present design anticipates this by pre-registering the cross-class reuse robustness specification, precisely so that a slope hidden by within-class-only counting has a chance to appear. If even that broadened specification returns zero, the null is more credible, but it remains an interval statement about one panel, not a universal claim. The honest framing converts the objection into a research agenda rather than treating the null as final.

### 7.2.1 The constructive content of a null

A null result is often treated as a non-finding. In this design it is a finding with operational content, and stating that content is part of the contribution's symmetric value. If the flat-cost null holds, three things follow that a program office can act on. First, the agency should not price a future autonomy capability by assuming that prior demonstrations have already bought down its qualification cost along a curve; the budget for that capability should be set from its own scope, not discounted for heritage that the data cannot show is lowering cost. Second, the case for an intermediate, lower-stakes demonstration can no longer be made on cost-buy-down grounds and must rest, if at all, on the claim that the intermediate demonstration retires technical or programmatic risk or delivers its own mission value; this is a real reframing of the build-or-wait decision, not its abolition. Third, the null sharpens the question that the experience-curve framework leaves open and that Section 7.4 takes up: if heritage does not lower cost on average across classes, perhaps it lowers cost in some classes and not others, and the policy-relevant object is then the heterogeneity rather than the average. An unrejected H0 thus redirects rather than ends the inquiry, and a discussion that pre-commits to that redirection is more useful than one that treats the null as a dead end.

## 7.3 Rival explanations and the responses to them

A negative \(\beta\), even one that clears the decision rule, is not self-evidently a learning curve. Three rival mechanisms could each produce a downward-sloping relationship between accumulated heritage and qualification cost without any genuine within-class learning effect being present. This section names each rival, states the mechanism by which it would counterfeit a learning curve, states the design response, and states honestly the residual risk the design response leaves unaddressed. The discipline is that the claim that \(\beta\) measures learning is only as strong as the case that it does not measure these three things instead.

### 7.3.1 Rival one: the secular computing-and-software cost trend

**The rival mechanism.** Over the observation window, the cost of computing and of software development fell for the whole economy and the whole agency, for reasons unrelated to autonomy-specific learning: cheaper processors, better development tooling, more capable compilers and test infrastructure. Because later autonomy episodes both have more accumulated heritage and enjoy cheaper computing, a naive regression of cost on heritage would attribute the general cost decline to heritage and report a spurious negative slope. Here calendar time is the true cause: an economy-wide decline in input costs lowers the cost of later episodes, and a naive estimate reads that fall as a \(\beta\) that looks like learning but is really a time trend.

**The design response.** This is the confound the decade fixed effects \(\delta_d\) are built to absorb. The Wright-versus-Moore distinction in the experience-curve literature is the formal statement of this threat: where output grows with time, a cost-versus-cumulative-output relationship and a cost-versus-time relationship are observationally similar, and only an explicit time control separates genuine cumulative-experience learning from a coincident time trend [\[79\]](#ref-79). The design adopts that discipline directly: after the decade effects are absorbed, \(\beta\) is identified only from within-decade variation in cumulative heritage, so a cost decline that is uniform across all classes within a decade is removed before \(\beta\) is estimated.

**The residual risk.** Decade fixed effects remove a between-decade time trend but cannot remove a within-decade computing trend that happens to track heritage accumulation inside a single decade. If, within a decade, the classes that accumulated the most heritage were also the ones that benefited most from a mid-decade tooling improvement, some of that tooling effect would remain in \(\beta\). The residual risk is therefore real but bounded, and the design downgrades confidence accordingly: a qualifying \(\beta\) is read as consistent with a learning effect, not as proof of one, until the maturation covariate and cross-class specifications corroborate it. Confidence that \(\beta\) is not merely a within-decade computing trend is **moderate**.

### 7.3.2 Rival two: scale effects masquerading as heritage

**The rival mechanism.** Larger, better-funded missions may field autonomy more efficiently per unit of capability simply because they have the engineering depth, the integration infrastructure, and the schedule slack to do so, independent of any heritage. If later autonomy episodes tend to be embedded in larger missions, a regression of total or even partially normalized cost on heritage would pick up the scale efficiency and report it as a heritage effect. Here mission scale is the true cause: static scale economies in engineering and integration make per-capability cost lower on larger later missions, and a naive estimate reads that as a \(\beta\) that is really a scale elasticity wearing a learning curve's clothes.

**The design response.** This is the confound the three-layer cost normalization is built to remove. The dependent variable is not a raw project cost; it is the recurring autonomy qualification cost normalized by capability scope on a NASA-Instrument-Cost-Model-class parametric basis [\[89\]](#ref-89). Normalization by scope is designed to strip out the part of cost that moves with instrument or platform scale, so that the residual reflects the autonomy qualification effort rather than the size of the mission carrying it. The parametric cost-model literature that supplies this normalization is the same tradition that single-variable space-system cost models exemplify, where a cost is expressed as a function of a scope driver so that residual variation can be attributed to factors other than scale [\[89\]](#ref-89). The accuracy of parametric estimation tools in this family for NASA space missions has been subject to blind validation: Friz and colleagues [\[117\]](#ref-117) conducted a blind validation study of the SEER-H parametric cost estimation tool across NASA missions and found that parametric models of this class produce estimates within a defensible error band when applied to comparable programs, which provides external evidence for the reliability of the imputation layer this design employs.

**The residual risk.** Normalization reduces but does not eliminate the scale confound, because the scope driver used to normalize is itself imperfect and because some scale efficiencies operate on the autonomy effort directly rather than on the platform it rides. The prospectus is candid that the dependent variable is the most delicate construct and is auditable, not error-free, with each observation flagged by which of the three imputation layers produced it and weighted in the robustness specification accordingly. The residual scale risk is what the inverse-imputation-error weighting specification partly addresses, by down-weighting the observations whose cost figures rest on the weakest scope normalization. Confidence that \(\beta\) is not a disguised scale elasticity is **moderate**, contingent on the normalization passing its own audit.

### 7.3.3 Rival three: selection on cost

**The rival mechanism.** Suppose cheap-to-qualify autonomy demonstrations are attempted only after heritage already exists, while the first-of-kind demonstrations, attempted before any heritage, are the expensive ones. Then the association between high heritage and low cost would be partly a selection artifact: heritage does not lower the cost of a fixed capability, but the agency self-selects into cheap demonstrations once heritage is available. Here the true cause is programmatic selection: by choosing which demonstrations to attempt conditional on existing heritage, the agency makes high-heritage episodes cheap, and a naive estimate reads that as a \(\beta\) that reflects what the agency chose to attempt rather than what heritage did to cost.

**The design response.** Two design features address this. The maturation covariate, the technology-readiness-level at development start, is included as a robustness covariate precisely to test whether the apparent heritage effect is really the simpler effect of later projects starting at higher maturity, which is one channel through which cost-based selection would operate [\[70\]](#ref-70), [\[81\]](#ref-81). The cross-class reuse robustness specification addresses a second channel by broadening the heritage measure to include software-component reuse that crosses capability classes. If selection were driving the result, broadening the heritage definition would not strengthen a genuine learning signal; if learning is real, the broadened measure should attenuate the downward bias from shared components and the slope should be at least as strong. The forward-only counting rule is a third, structural protection: heritage counts toward an episode only if it reached flight operation before that episode's development start. This prevents the mechanical correlation in which later episodes are credited with heritage they could not have used, and it biases the estimated slope toward zero, making any rejection of the flat null more credible rather than less.

**The residual risk.** Selection that operates through a channel not captured by maturity or by cross-class reuse, for example an unobserved programmatic judgment that a demonstration is "ready" for reasons orthogonal to TRL, would remain in \(\beta\). There is no instrument in the design that fully closes this channel, and honesty requires saying so. The path-dependence caution in Section 7.5 is the theoretical statement of the same concern: the realized cost path reflects which classes were chosen for early investment, not only the intrinsic learnability of those classes [\[4\]](#ref-4), [\[5\]](#ref-5). Confidence that \(\beta\) is free of selection is **low to moderate**, and this is the rival the design controls least completely; a qualifying slope is therefore reported with this limitation foregrounded rather than buried.

### 7.3.4 Rival four: imputation bias confounded with heritage

**The rival mechanism.** The dependent variable for most observations in the panel is not a directly reported autonomy qualification cost but a figure imputed through one of three layers of decreasing directness: a TechPort-reported project cost from which the autonomy fraction is estimated, a parametric NICM-class estimate scaled to the capability scope, or a cross-mission analogical imputation. If the observations that carry the most accumulated heritage are systematically also the observations whose cost figures come from cleaner, more directly reported sources, the observed cost decline with heritage could reflect measurement favorability rather than genuine learning. The mechanism is that cheaper-to-qualify autonomy capabilities, being closer to established engineering practice, may be both easier to cost-track directly and more heritage-rich; a naive regression then confounds improved measurement with improved cost performance, and the estimated slope understates the true measurement uncertainty or overstates the true learning signal.
**The design response.** The design's primary protection is the reliability-flag stratification already embedded in the data structure: each observation is tagged by which imputation layer produced its cost figure, and the third robustness specification weights observations by the inverse of the imputation error, so that poorly measured cost figures contribute less to the estimated slope. A pre-committed slope-robustness check compares \(\beta\) across the three reliability strata, high-reliability (layer-one direct reporting), medium-reliability (NICM-class parametric), and low-reliability (analogical imputation). If imputation bias were driving the result, \(\beta\) would differ across strata, because the direct-reporting observations carry the heritage-richest, best-measured items. A stable slope across strata weakens the imputation-bias explanation: a learning signal that appears in both the well-measured and the poorly measured subsets is less plausibly an artifact of measurement quality. This check is pre-registered rather than post-hoc.

**The residual risk.** The stratification check can detect imputation bias that correlates with heritage but not bias that is uniform across strata. Suppose the parametric tools used in layer two systematically under-estimate cost for later, more heritage-rich episodes, for reasons unrelated to heritage. That bias would appear in the medium-reliability stratum across all heritage levels and would not be flagged by the cross-stratum comparison. The blind validation evidence for SEER-H-class tools [\[117\]](#ref-117) gives some assurance that the parametric layer is not systematically biased in its application to NASA missions, but that evidence is aggregate and does not address heritage-correlated patterns within the autonomy panel specifically. Confidence that \(\beta\) is free of imputation bias confounded with heritage is **moderate**, conditional on the cross-stratum slope-robustness check passing; it would fall to **low** if the slope proved materially unstable across reliability strata.

### 7.3.5 Why the four responses are jointly, not individually, the case

No single robustness specification defeats all four rivals, and the chapter does not claim otherwise. The strength of the design is the joint structure: the computing-trend rival is addressed by decade effects, the scale rival by scope normalization and inverse-error weighting, the selection rival by the maturation covariate, the cross-class measure, and the forward-only rule, and the imputation-bias rival by the reliability-flag stratification and the pre-committed cross-stratum slope check. The decision rule's requirement that \(\beta\) survive in the baseline and in at least two of three robustness specifications is how the design confronts the rivals jointly. A slope that is a computing trend would not survive the maturation and cross-class specifications unchanged; a slope that is a scale effect would move under inverse-error weighting; a slope that is selection would attenuate when the heritage definition is broadened; a slope that is imputation bias would fracture across reliability strata. A \(\beta\) that holds across these is not proven to be learning, but it has survived the four best alternative explanations the design can mount, and that is the standard a measurement study can meet. Where it cannot fully close a channel, as with orthogonal selection or stratum-uniform parametric bias, the residual risk is named and the confidence is downgraded rather than the risk concealed. The chapter therefore leaves the measurement standing on a within-class log-log curve that targets the learning mechanism, that has outlasted its four most serious rivals jointly, and whose remaining uncertainty is acknowledged and bounded rather than claimed away.

## 7.4 Effect heterogeneity across classes: the Mokyr codification moderator

The most policy-relevant secondary finding the design can produce is not the average slope but the heterogeneity of the slope across capability classes. Mokyr's distinction between propositional and prescriptive knowledge supplies the prediction: the cost-decline curve will be steeper in classes whose underlying knowledge is mature, codified, and reusable, and flatter in classes where each project re-implements from scratch.

The estimating equation absorbs capability-class baseline differences in \(\gamma_c\), but it estimates a single pooled within-class slope \(\beta\). The design can be extended, within the same framework, to recover class-specific slopes where the per-class sample permits, and the heterogeneity of those slopes is an observable. Mokyr's framework predicts the pattern of that heterogeneity: a technique becomes cheap to reproduce and extend only when it rests on a wide base of propositional understanding, and techniques discovered by trial without underlying theory tend to stagnate rather than improve [\[115\]](#ref-115). Mokyr further argues that access cost to knowledge governs diffusion, which predicts that the slope will be steeper where heritage knowledge is openly codified and reusable and flatter where it is locked in bespoke artifacts [\[115\]](#ref-115).

A moderator that the theory predicts and the data can test is more valuable to a portfolio than an average, because it tells the program office not merely whether heritage lowers cost on average but where to expect it to lower cost and where not to. If the codification moderator holds, the office can prioritize building shared, codified infrastructure in the classes where the propositional base is mature, expecting a steep return, and can budget conservatively in the classes where each demonstration is still bespoke, expecting little cost transfer. This is a sharper instrument than a single learning rate, and it maps directly onto the build-or-wait decision class by class.

The codified-knowledge substrate that Mokyr's framework identifies as the steepening factor is documented in the autonomy record as the reusable flight-software base on which the heritage-reuse argument depends. The Core Flight System and shared flight-software architecture work is the concrete instance: a core, reusable, plug-and-play flight-software architecture is the kind of codified prescriptive knowledge that should, on Mokyr's account, lower the access cost to heritage and steepen the within-class slope [\[105\]](#ref-105), [\[63\]](#ref-63). Architecture-tracking and modular-extensible flight-software studies for small spacecraft document the reuse economics of such substrates [\[46\]](#ref-46), and the NASA technology taxonomy that organizes the autonomy capability classes is itself a codification artifact that structures how heritage is tracked across the portfolio [\[75\]](#ref-75). These sources converge on a single proposition: reuse infrastructure exists in some classes more than others, and where it exists it is the mechanism by which heritage becomes cheap to redeploy. That convergence is what licenses treating codification as a testable moderator rather than a decorative reference to Mokyr.

Confidence in recovering the codification moderator is **low** at the design stage, and lower than confidence in the average slope, for a hard sample-size reason. The full panel is small, on the order of tens of episodes; partitioning it into class-specific slopes spreads that already-scarce data across five capability classes, and any class with only one or two episodes contributes nothing to a within-class slope. The fixed-effects feasibility check pre-registered in the analysis plan will report exactly which classes can and cannot support a slope, and the heterogeneity finding is available only for the classes that survive that check. The codification moderator is therefore the design's most ambitious secondary aim and its least certain. It should be reported as suggestive heterogeneity where the data permit, not as a fitted moderation coefficient with its own confidence interval, unless a class is unusually well-populated.

The heterogeneity reading would be defeated if the apparent cross-class differences in slope were driven by differences in cost-measurement quality across classes rather than by differences in codification. If the classes with the most codified substrate also happen to be the classes whose costs are best documented and least imputed, then a steeper apparent slope in those classes could reflect cleaner data rather than steeper learning. The reliability flags and the inverse-imputation-error weighting are the partial protection: the heterogeneity is more credible if it survives in the weighted specification, where poorly measured observations are down-weighted. Where the data cannot separate codification from measurement quality, the chapter must say so and report the heterogeneity as a hypothesis the design generates rather than a result it establishes.

### 7.4.1 Why heterogeneity is the finding the field should want

In a study built around a single falsifiable slope, the temptation is to treat the average \(\beta\) as the whole prize and the heterogeneity as an afterthought. The opposite is closer to the truth for the portfolio. An average learning rate, applied uniformly across classes that in fact learn at very different rates, would over-invest in the slow-learning classes and under-invest in the fast-learning ones, because the build-or-wait calculation would use a rate that is wrong for both. The heterogeneity, even if recovered only coarsely, corrects this: it tells the office that the heritage-investment dollar buys more cost-decline in the codified classes and less in the bespoke ones, which is the actionable shape of the result. The Mokyr lens earns its place in the design precisely here. It does not merely predict that heritage lowers cost, which is Arthur's mechanism; it predicts where heritage lowers cost most, and it ties that prediction to an observable, manipulable feature of the portfolio, namely whether the agency invests in codified, reusable substrate. That makes the codification moderator the bridge from measurement to action, and it is the reason the chapter treats it as the most policy-relevant secondary finding rather than a robustness footnote.

## 7.5 Path dependence and the external-validity bound; the launch-cadence analogue

Whatever slope the panel returns, it is the slope of a realized historical path and not a universal constant. Arthur's theory of increasing returns and path dependence both supplies the learning-effect mechanism the curve measures and bounds how far any single estimated rate can be carried beyond the sequence that produced it.

Arthur's framework identifies learning effects as one of the mechanisms that make a technology more attractive the more it is adopted, and learning effects are exactly what an experience curve measures: each use of a technique lowers the cost of the next, generating positive feedback [\[3\]](#ref-3), [\[4\]](#ref-4). This is why Arthur's lens predicts H1 as the theoretically expected direction. But the same framework holds that increasing-returns systems are path-dependent and non-ergodic: early choices lock in, and the realized cost path reflects historical sequence as much as technical fundamentals [\[4\]](#ref-4), [\[5\]](#ref-5). The two claims are one theory. The learning effect that produces a measurable slope is also what makes the slope contingent on which classes received early investment.

A learning rate is externally valid only over the family and the historical regime in which it was estimated, because the rate is a property of the realized path of adoption, not of the technology in the abstract. If the classes that accumulated the most heritage in the observation window did so because they were chosen for early investment, then the measured average slope encodes that choice. A different agency, or the same agency in a different funding regime, that invested in a different sequence of classes would realize a different path and could realize a different rate. The reason for treating the slope as bounded is therefore the non-ergodicity of the system: the past path does not pin down the future path, so the estimated rate is a description of what happened, usable as a disciplined prior, not a forecast guaranteed to hold under a changed investment pattern.

The clearest contemporary space-sector instance of Arthur's learning-effect mechanism operating on cost is the launch-cadence evidence, and it is worth developing as the analogue rather than merely citing it. The launch-market analysis documents that higher launch cadence lowers the cost and raises the automation of constellation operations, with each flight reducing the cost of the next, which is the learning-effect mechanism in its purest operational form: a repeated activity whose per-instance cost falls with cumulative repetition [\[91\]](#ref-91), [\[37\]](#ref-37). The analogue is instructive in two directions. It is encouraging for H1, because it shows the mechanism is alive and measurable in a closely related space context, which raises the prior that an autonomy experience curve exists. But it is also a caution, because the launch case is one where cadence, the cumulative-repetition driver, grew dramatically and the cost decline was correspondingly visible, whereas the autonomy demonstration record has far fewer episodes and a far slower cadence, so the same mechanism, even if present, would be harder to detect with the same confidence. The launch analogue thus both motivates the hypothesis and calibrates expectations about statistical power: the mechanism is real in the neighborhood, but the autonomy panel is thin where the launch panel is thick.

Confidence that the estimated slope generalizes beyond NASA and JPL deep-space and planetary autonomy is **low**, and this is a deliberate, not an apologetic, statement of scope. The findings, if any, generalize to NASA and JPL onboard autonomy for deep-space and planetary missions over the observation window. They may not generalize to commercial autonomy, to terrestrial autonomy, or to mission classes outside the demonstration window, and the fixed-effects design estimates an average within-class slope rather than a universal learning rate. The path-dependence caution is the theoretical reason this scope statement is binding rather than conventional: because the rate is path-contingent, extrapolating it to a different path is not a small extension but a category error.

The external-validity bound would be loosened, though not removed, if the heterogeneity analysis in Section 7.4 showed that the per-class slopes were stable and theoretically interpretable, because a slope that varies across classes in the way Mokyr predicts is more plausibly a real learning property of each class than an accident of investment sequence. Conversely, if the per-class slopes were erratic and uninterpretable, that would strengthen the path-dependence reading, in which the realized rates are artifacts of historical sequence. The two anchor lenses thus do work against each other in a productive way: Arthur's path dependence is the threat to external validity, and Mokyr's codification moderator is the test of whether the slope is a stable knowledge property or a contingent path artifact. The chapter does not resolve this tension by assertion; it states that the heterogeneity analysis is the evidence that would move confidence one way or the other, and that at the design stage neither reading is established.

### 7.5.1 The conceptual objective-to-decision endpoint

Architecture vocabulary is out of scope for this econometric study, and this chapter honors that by not populating any capability, system, or data-service-exchange chain. The single permitted conceptual statement is this, in plain prose. The strategic objective is an affordable, well-sequenced autonomy portfolio in the Autonomous Systems and Robotics line. The fitted slope \(\beta\), under the H1 branch, is an input to a portfolio decision: it feeds the build-or-wait determination for a target capability against its cost cap, and, through the heterogeneity analysis, it informs where shared codified substrate is worth building. Under the H0 branch, the same measurement informs the decision differently, by removing cost-buy-down as a defensible premise and redirecting the justification to capability value. In both branches the endpoint is the same kind of object: a measured quantity that enters a sequencing decision the agency currently makes by assertion. That is the whole architecture-relevant content of the contribution, and it is stated here as a sentence, not a chain.

## 7.6 Synthesis: the argument carried to the decision

The discussion can now be drawn together as a completed argument, each part interpreted under both branches of the disjunction. NASA and JPL rely on an unmeasured heritage-lowers-cost assumption documented across the autonomy demonstration record [\[42\]](#ref-42), [\[11\]](#ref-11), [\[24\]](#ref-24), [\[34\]](#ref-34), [\[40\]](#ref-40), and neither branch of the result leaves that reliance unexamined. The assumption carries real weight: autonomy investment and portfolio sequencing turn on it, the sequencing decision is a constrained portfolio-selection problem that consumes forward cost inputs [\[43\]](#ref-43), and the TRL scale tracks maturity but not cost [\[70\]](#ref-70), [\[81\]](#ref-81), [\[99\]](#ref-99), [\[7\]](#ref-7), so the measurement supplies something the existing maturity instrument cannot. A within-class log-log experience curve measures Arthur's learning-effect mechanism directly [\[3\]](#ref-3), [\[79\]](#ref-79), and the discussion has shown how a qualifying slope is read as that mechanism and a flat slope as its non-detection. Section 7.3 weighed the computing-trend, scale, and selection rivals and showed that the joint robustness structure defeats them where it can and names the residual where it cannot. What uncertainty remains is bounded honestly: the small panel, the measurement error in the dependent variable, and the path dependence are addressed by pre-registered robustness, influence diagnostics, honest design-stage framing, and the explicit external-validity bound, and where a risk cannot be closed it is stated and the confidence downgraded rather than the risk denied [\[81\]](#ref-81), [\[5\]](#ref-5), [\[84\]](#ref-84).

The two branches of the disjunction are the chapter's central interpretive commitment, and they are symmetric in value. If H1 holds, the contribution is a planning parameter and a vindicated, quantified heritage-reuse argument, qualified by the heterogeneity that Mokyr predicts and bounded by the path dependence that Arthur warns of. If H0 holds, the contribution is a disciplined removal of a load-bearing but unverified premise, a redirection of the portfolio rationale to capability value, and a sharpened question about where, if anywhere, heritage lowers cost. Neither branch is a non-result. The design's worth is that it forces the agency's assumption to declare itself as a number with an interval, and a number with an interval, whatever its sign, is more useful to a program office than an assumption that is never tested because it is never stated. That is the thesis with which the chapter opened, now carried through both outcomes, both anchor lenses, all three rivals, and the explicit bound on how far the eventual measurement may travel.
# Chapter 8. Conclusion

This dissertation delivers a single, falsifiable instrument: the first fitted measurement of the experience-curve slope of onboard-autonomy qualification cost on cumulative flight-demonstrated heritage. This concluding chapter argues that the instrument is worth building and defending even before it is run, because it converts a load-bearing NASA and JPL assumption from assertion into a planning parameter, and because the value of the design does not depend on which way the result turns. Whether the assembled panel yields a negative slope \(\beta\) that rejects the flat-cost null, or a slope indistinguishable from zero that fails to reject it, the agency gains a defensible number where it previously had a habit of belief. That symmetry, established in the prospectus and carried unchanged through every preceding chapter, is the foundation on which this conclusion rests.

The chapter does four things in sequence. It restates the contribution and isolates what survives even if the alternative hypothesis is not confirmed. It states the limitations honestly, distinguishing the limitations the design retires from those it can only bound. It specifies a concrete future-research program, with attention to the path from a fully developed design to an executed measurement on the full data. It then closes the argument and the dissertation.

## 8.1 The Contribution Restated: One Slope, or One Credible Failure to Find One

The problem this dissertation addresses can be stated in four parts. The current state is that NASA and the Jet Propulsion Laboratory treat each flown autonomy capability as a heritage asset assumed to lower the cost of fielding the next demonstration in the same capability class, and they sequence portfolio investments on that assumption without ever expressing it as a measurable quantity. The desired state is a cost-anchored, falsifiable learning rate, estimated per autonomy capability class, that a program office can read as a number rather than infer from a narrative. The gap is that no published study fits a Wright-type log-log model of capability-class autonomy qualification cost on cumulative flight-demonstrated heritage; the experience-curve literature concerns hardware unit cost, and the space-autonomy literature describes heritage only qualitatively [\[94\]](#ref-94), [\[42\]](#ref-42). Leaving the gap open means that build-or-wait and sequencing decisions in the Autonomous Systems and Robotics portfolio remain made by assertion, with no quantitative complement to the technology-readiness-level scale, which is ordinal and non-monetary by construction [\[81\]](#ref-81).

The contribution that answers this problem is deliberately narrow. It is the measurement itself: a fitted slope \(\beta\), its confidence interval, the implied learning rate \(1 - 2^{\beta}\), and an explicit accept-or-reject decision on the null. The canonical estimating equation, carried unchanged from the prospectus and used in every chapter, is

\[ \ln(\text{Cost}_{icd}) = \alpha + \beta \, \ln(\text{CumHeritage}_{icd}) + \gamma_c + \delta_d + \epsilon_{icd} \qquad\qquad (1) \]

for episode \(i\) in capability class \(c\) and decade \(d\), where \(\text{Cost}\) is normalized recurring engineering cost to qualify the capability for flight, \(\text{CumHeritage}\) is the cumulative within-class count of prior flight demonstrations that reached flight operation before the episode's development-start date, \(\gamma_c\) are capability-class fixed effects, and \(\delta_d\) are decade fixed effects. The two hypotheses are stated here in their fixed form so the conclusion cannot be read as having softened them. H0 holds that per-episode autonomy development cost is flat with respect to cumulative prior flight demonstrations within a capability class, so that \(\beta\) is statistically indistinguishable from zero. H1 holds that per-episode cost declines along a log-log experience curve as cumulative flight-demonstrated heritage accumulates, so that \(\beta\) is negative and statistically significant after controlling for capability class and decade.

What survives a non-confirming result is worth stating carefully. The contribution of this dissertation does not turn on the direction of the finding, because the instrument and not the slope is the deliverable. That property rests on the construction itself: a four-source panel from NASA TechPort, NTRS autonomy demonstration reports, NICM-class parametric cost estimates, and the published mission literature; a defined unit of analysis; a three-layer cost-normalization procedure with per-observation reliability flags; a forward-only heritage-counting rule; a two-way fixed-effects estimator; and a decision rule fixed before estimation. The logic is the standard logic of falsifiable measurement. A test that is informative only when it confirms a prior is not a test but an advocacy exercise, whereas a test whose two outcomes are both decision-relevant earns its keep regardless of which outcome occurs. The experience-curve econometric tradition supports this stance: it treats the learning rate as the object of measurement and accepts that some technologies show shallow or absent learning, so that a null is a legitimate scientific finding rather than a failure of the method [\[79\]](#ref-79), [\[94\]](#ref-94). What is robust here is the contribution, not any particular policy recommendation. A null result narrows the policy options as surely as a negative slope opens them, but it does not leave the agency without guidance. One objection deserves a direct answer: a reader might judge a design-stage instrument with no executed result to be no contribution at all. The answer, developed across Chapters 4 through 6, is that the delicate and contestable parts of this study are precisely the construction choices, the cost-normalization layers, the heritage-counting convention, the fixed-effects feasibility logic, and the influence diagnostics, and these are fully specified and auditable here. The number the panel will eventually return is the mechanical output of choices already on the record.

What survives if H1 is not confirmed is considerable. The capability-class taxonomy survives as a reusable organizing frame for the autonomy demonstration record. The forward-only heritage-counting rule survives as a defensible convention that biases against finding an effect and so makes any rejection of the null more credible. The three-layer cost normalization survives as an auditable procedure for converting heterogeneous, inconsistently reported development figures into a comparable dependent variable. The pre-registered decision rule survives as a commitment device that prevents the result from being narrated into significance after the fact. And a credible failure to reject H0 is itself a finding with a clear operational reading: it would say that within-class flight heritage, as measured, does not lower qualification cost on its own, and that the portfolio rationale for autonomy investment must rest on capability value rather than on an assumed cost-decline curve. The agency would then know not to budget future demonstrations as if each one buys down the next, a materially different planning posture from the one it holds by default.

## 8.2 Both Outcomes Are Decision-Relevant: The Symmetric Value of the Design

The single most important property of this design is that it is symmetric in its value. The disjunction between H1 and H0 is not a gamble in which one branch is a success and the other a wasted effort. It is a fork in which each branch hands the Autonomous Systems and Robotics portfolio a different, equally usable, planning conclusion.

Consider the branch in which the panel yields a negative slope whose confidence interval excludes zero in the baseline specification and in at least two of the three pre-registered robustness specifications, satisfying the fixed decision rule. The causal reading of that result names a mechanism rather than asserting a bare correlation. Cumulative flight heritage within a capability class is the cause, working through the accumulation of codified, reusable design patterns, verification-and-validation approaches, and flight-software components that each demonstration leaves behind. That reusable stock lowers the recurring non-recurring-engineering cost to qualify the next same-class capability, and the lowered cost registers as the measurable negative log-log slope \(\beta\). The slope then becomes a planning parameter the portfolio can use directly in a build-or-wait decision: when a future mission needs a capability currently too expensive to qualify within its cost cap, a known slope tells the portfolio roughly how many intermediate demonstrations would buy the target capability down to affordability. This is the mechanism W. Brian Arthur identifies, in which each use of a technique lowers the cost of the next and generates positive feedback [\[3\]](#ref-3). The slope is the empirical measure of exactly that learning effect.

Now consider the branch in which the panel yields a slope indistinguishable from zero, or positive, so that the decision rule is not satisfied and H0 is not rejected. This is not a null in the dismissive sense; it carries its own operational consequence. If within-class heritage does not lower qualification cost along a measurable curve, then the heritage-reuse argument that quietly justifies much autonomy investment is weaker than assumed, and the portfolio must justify each demonstration on the value of the capability it delivers rather than on a promised discount to its successors. The mechanism account here is a refusal to infer a mechanism from a correlation the data do not support: where the slope is flat, the design says so and downgrades confidence in the cost-decline story accordingly. That refusal is a service to the agency, because budgeting an intermediate demonstration as a cost-reduction investment, when no such reduction is measurable, would misallocate scarce portfolio funds.

The confidence attached to this symmetry claim is high, and the basis for that confidence is structural rather than empirical. The symmetry does not depend on any particular value of \(\beta\). It depends only on the decision rule being fixed before estimation, on both outcomes mapping to distinct and stated planning postures, and on the dependent variable being a true per-episode cost rather than a total that grows mechanically with scale. All three conditions are established in the design and do not await the data. What would lower this confidence is a discovery during panel assembly that the effective sample, after the fixed-effects feasibility check removes singleton cells, is too small to distinguish a moderate negative slope from zero at any usable confidence level. In that case the study would return neither a clean rejection nor a clean failure to reject, but an indeterminate interval, and the honest report would be that the question cannot be answered at the available power. The analysis plan anticipates this by pre-committing to report the realized residual degrees of freedom alongside every coefficient and to treat a wide interval that contains zero as a failure to reject rather than as evidence for H1. The symmetry of value is therefore protected by the same machinery that protects the inference.

## 8.3 The Anchors Sharpened the Test

The two methodological anchors did analytical work in this dissertation rather than ornamental work, and this section restates how. Arthur and Mokyr each supplied a testable prediction that the design operationalizes, and neither was imported as decoration.

W. Brian Arthur's theory of increasing returns identifies learning effects as one of the mechanisms that make a technology more attractive the more it is used, and it is the learning effect that an experience curve measures [\[3\]](#ref-3). Arthur's contribution to the test is twofold. First, he supplies the theoretically expected direction: if learning effects dominate, a measurable negative slope is the expected signature, which is what H1 asserts. Second, and more usefully for a critical design, he supplies a caution the design must absorb. Arthur's increasing-returns systems are path-dependent and non-ergodic, so early investment choices lock in and the realized cost path reflects historical sequence rather than only technical fundamentals [\[5\]](#ref-5). The structure of this caution is worth making explicit. Even a clean negative slope cannot be read as a universal, transferable learning rate, because the capability classes that received early investment accumulated their own increasing returns, so the measured average slope encodes which classes happened to be funded first. Arthur's path-dependence result, which holds that the realized path in a non-ergodic system is contingent on its history, is what carries that conclusion. It bounds the external validity of any single estimated slope without invalidating it as a within-sample measurement. The design forecloses the temptation to publish a single pooled learning rate as a constant of nature; the fixed-effects specification, which estimates a within-class slope rather than one universal slope, is the formal expression of Arthur's caution built into the estimator.

Joel Mokyr's distinction between propositional knowledge, the knowledge of why something works, and prescriptive knowledge, the knowledge of how to do it, supplied the dissertation's most policy-relevant secondary prediction. Mokyr argues that a technique becomes cheap to reproduce and extend only when it rests on a wide base of propositional understanding, and that techniques discovered by trial without underlying theory tend to stagnate rather than improve. Translated into this study, Mokyr predicts that the experience-curve slope will be heterogeneous across capability classes and steeper where the underlying knowledge is mature, codified, and reusable, for example through the Core Flight System substrate and shared autonomy frameworks, and flatter where each project re-implements from first principles. This prediction is not a flourish; it is a testable moderator. The design can examine effect heterogeneity across capability classes and ask whether the classes resting on a codified, openly reusable knowledge base show steeper cost decline than the classes that do not. That heterogeneity, if it appears, is the most actionable finding the design can produce, because it would tell the agency that investment in reuse infrastructure, not only in capabilities, is what bends the cost curve. Mokyr's reversibility caution, that progress is not guaranteed to persist and that access cost to knowledge governs diffusion, also enters as a hedge against assuming that a steep slope, once observed, will hold indefinitely.

The confidence in these anchor-derived predictions is moderate, calibrated to the design-stage evidence grade. Arthur's learning-effect mechanism is well established in the increasing-returns literature, and the launch-cadence record provides a concrete space example in which each flight lowers the cost of the next [\[91\]](#ref-91), [\[37\]](#ref-37); that raises confidence in the predicted direction. The Mokyr moderator is more speculative because the codified-reuse substrate evidence is thinner in the open literature than the demonstration spine, as noted earlier. What would raise confidence in the moderator is a focused sweep on Core Flight System reuse economics during the build phase; what would lower it is a finding that the codification variable cannot be measured cleanly enough to enter the heterogeneity analysis. The anchors therefore sharpened the test by converting two general theories into one directional hypothesis and one moderator hypothesis, both of which the estimator can address, and both stated with the epistemic modesty the design stage requires.

There is a further contribution from the anchors that is methodological rather than substantive, and it is worth isolating because it survives any outcome of the estimation. The pairing of Arthur and Mokyr resolves an ambiguity that the bare experience-curve literature leaves open. That literature, in its hardware-unit-cost form, treats the learning rate as a single number attached to a technology and is largely silent on why the number takes the value it does or whether it should be expected to differ across related technologies [\[94\]](#ref-94), [\[79\]](#ref-79). Arthur supplies the mechanism that makes the number meaningful, the learning effect operating through positive feedback, and Mokyr supplies the moderator that predicts when the number should be large rather than small, the depth and codification of the underlying propositional base. Together they convert the experience curve from a curve-fit into a mechanism-bearing measurement, which is the standard this dissertation holds itself to when it insists that every causal claim trace its cause through to an observed consequence rather than rest on a fitted correlation. A reader who rejects the autonomy application entirely can still take from the design this approach to putting an experience curve on a mechanistic footing, and that approach is reusable for any software-intensive capability whose qualification cost is the object of interest.

## 8.4 From Design to Execution: The Remaining Build Steps and Reproducibility Commitments

The path from this fully specified design to an executed measurement is concrete, ordered, and bounded, and the reproducibility commitments are part of the deliverable rather than an afterthought. The work that remains is not a research question to be discovered; it is a construction task whose steps are already named.

The first remaining step is panel assembly. The episode inventory must be built from NASA TechPort project records and NTRS autonomy demonstration reports, with each episode assigned to exactly one primary capability class and its development-start date recorded. TechPort supplies the project inventory, the taxonomy mapping, the maturation covariate, and project timing; NTRS supplies the heritage chronology and the capability definitions for the named demonstrations, including Remote Agent on Deep Space 1, the EO-1 Autonomous Sciencecraft Experiment, AEGIS on Opportunity and on the Curiosity ChemCam instrument, and the Mars 2020 autonomous-navigation and Ingenuity demonstrations [\[34\]](#ref-34), [\[40\]](#ref-40), [\[7\]](#ref-7), [\[99\]](#ref-99). The AEGIS sequence in particular remains the cleanest within-class heritage pair in the record and the central illustrative case for the mechanism the study tests.
The second remaining step is cost normalization, the most delicate and most load-bearing construction task. Public development-cost figures for autonomy capabilities are reported inconsistently: some are full project costs that bundle platform, instrument, and autonomy; some are subsystem costs; some are not disclosed at all. The three-layer procedure must be executed in full. The first layer extracts, where possible, the autonomy-specific non-recurring-engineering portion from project documentation, separating it from platform and instrument cost; the Deep Space 1 lessons-learned record, which reports that inserting system-level autonomy into a flight project surprised the project team, is direct primary evidence on this quantity for the first-of-kind case [\[11\]](#ref-11). The second layer applies, where the autonomy portion is not separately reported, a NICM-class parametric estimate of the autonomy development effort given the documented scope, using the parametric cost-model approach exemplified by single-variable parametric models for space systems [\[89\]](#ref-89). The third layer expresses every figure in constant-year dollars and records, per observation, which layer produced the figure and an associated reliability flag, so that the weakest imputations can be down-weighted in the inverse-imputation-error robustness specification. This procedure does not eliminate measurement error in the dependent variable. It makes the error explicit and auditable, which is the standard a defensible cost study must meet.

The third remaining step is heritage counting under the forward-only rule, in which a prior demonstration counts toward a later episode's heritage stock only if it reached flight operation before the later episode's development-start date. The convention is conservative by design: it tends to undercount heritage and so biases the estimated slope toward zero, making any rejection of the flat-cost null more credible.

The fourth and fifth steps are estimation and reporting. The two-way fixed-effects log-log regression and the three pre-registered robustness specifications (the TRL maturation covariate, the cross-class reuse measure, and the inverse-imputation-error weighting) are fit, and \(\beta\), its confidence interval, the implied learning rate, and the accept-or-reject decision are reported together with diagnostic plots of log cost against log heritage by class. Two pre-analysis checks gate the interpretation. The fixed-effects feasibility check reports any capability class or decade containing only a single observation, because such cells contribute nothing to the within estimator and the slope is then identified off a smaller effective sample than the raw count suggests. The influence diagnostic reports the slope with and without each potentially high-leverage episode, because in a panel of tens of observations a single unusually expensive first-of-kind demonstration could drive the result; a slope that depends on one episode is treated as not robust. These checks are fixed in advance so that they constrain interpretation rather than being selected after the result is seen.

Several limitations of the executed study survive careful construction and must be carried forward honestly. The population of flight-demonstrated autonomy episodes is small, on the order of tens, which limits statistical power and constrains the number of fixed effects the panel can support; this is the dominant statistical-conclusion threat, and it is intrinsic to the subject rather than a defect of the design. Development-cost figures are heterogeneous in definition and sometimes undisclosed, which forces the NICM-class imputation that the design makes auditable but cannot make error-free; no quantity of additional citation closes this gap, because it is a data-construction reality rather than a literature gap. Capability-class assignment requires judgment and can be contested at the boundaries. Heritage crosses capability classes through shared software components and design patterns that the within-class count does not capture, which biases against finding an effect and which the cross-class robustness specification only partly addresses. The findings, if any, generalize to NASA and JPL onboard autonomy for deep-space and planetary missions within the demonstration window, and not necessarily to commercial autonomy, terrestrial autonomy, or mission classes outside that window. These limitations bound the strength of any conclusion, and the report will state them alongside the result rather than after it.

Beyond the steps that complete the present study, the design opens a future-research program with three distinct lines, each of which extends the instrument rather than merely repeating it. The first line extends the panel beyond the original demonstration window as new autonomy demonstrations reach flight. Every successive same-class demonstration adds an observation at a higher value of cumulative heritage, which is exactly the variation the slope is identified off, so the estimate sharpens over time without any change to the specification. The design is therefore a living instrument: re-running it as the record grows converts a presently underpowered measurement into a progressively more precise one, and the influence diagnostic becomes less binding as the number of observations rises and no single episode can dominate the fit. The second line is the heterogeneity analysis that operationalizes the Mokyr moderator. Once the panel supports it, the study can ask whether the slope differs across capability classes in the direction the codification hypothesis predicts, and whether the presence of a shared, codified flight-software substrate is associated with the steeper classes. This is the most policy-relevant secondary product the design can yield, because it would distinguish investment in capabilities from investment in the reuse infrastructure that bends the cost curve. The third line is the cross-class reuse measure, which the baseline within-class count deliberately omits and which the robustness specification only begins to capture. A future build that traces software-component and design-pattern reuse across capability classes would test directly whether heritage that crosses class boundaries lowers cost, and would quantify the downward bias that the within-class count imposes on the baseline slope. None of these lines requires a new estimator; each requires more data and more careful provenance tracing, which places them squarely in the build phase rather than in a new theoretical departure.

The reproducibility commitments are explicit and binding. The estimation code, the assembled panel, the imputation log, and the heritage-counting log will be retained so that a reviewer who did not build the study can check it. Pre-registering the baseline specification, the three robustness specifications, and the fixed decision rule means the analysis cannot be retrofitted to its result. Multiple-specification testing is controlled by pre-registering the robustness specifications rather than searching over them. A focused follow-up sweep on small-sample and few-clusters fixed-effects inference, for example wild-cluster bootstrap and small-N robust standard errors, is the highest-value methodological strengthening before the estimation is run, because the econometric-method facet is the thinnest part of the evidence base and the inference must withstand the scarcity of degrees of freedom [\[61\]](#ref-61), [\[14\]](#ref-14). A parallel sweep on Core Flight System reuse economics is the highest-value substantive strengthening, because the codified-reuse substrate evidence that underwrites the Mokyr moderator is thinner than the demonstration spine and the moderator analysis is only as strong as the codification variable it rests on. With these steps and commitments, the design becomes an execution, and the execution becomes a number that the agency can audit.

## 8.5 Closing

This dissertation began from a discrepancy between how confidently NASA and JPL act on the belief that flown autonomy lowers the cost of the next demonstration and how little that belief has been measured. The argument carried across the dissertation answers that discrepancy, and it is worth restating in closing. The agency relies on an unmeasured heritage-lowers-cost assumption that is visible across the autonomy demonstration record [\[42\]](#ref-42), [\[11\]](#ref-11), and the reliance is consequential because portfolio sequencing and build-or-wait decisions turn on it while the technology-readiness-level scale tracks maturity without tracking the cost of advancing it [\[81\]](#ref-81). A within-class log-log experience curve measures Arthur's learning effect directly rather than gesturing at it [\[5\]](#ref-5), [\[79\]](#ref-79), and ordinary least squares on the log-log two-way fixed-effects form is the best-validated way to do so: it produces a single interpretable slope and lets the fixed effects separate genuine learning from time and scale confounds while keeping the degrees-of-freedom cost low in a small panel [\[79\]](#ref-79), [\[94\]](#ref-94). The uncertainty that remains is bounded honestly. Small sample size, measurement error in the dependent variable, and path dependence are held in check by the pre-registered robustness specifications, the influence and feasibility diagnostics, the conservative forward-only heritage rule, and the design-stage framing that labels every expected result as illustrative and refuses to report a fitted coefficient as an executed finding.

What the completed study will deliver is therefore a defensible number, or a defensible failure to find one, on a question the agency currently answers by assumption. If the slope is negative and significant, NASA and JPL gain a planning parameter that quantifies how much heritage investment buys down the cost of the next demonstration and that supports the build-or-wait decision with arithmetic instead of advocacy. If the slope is indistinguishable from zero, they gain the equally useful knowledge that the cost-decline story does not hold as an experience-curve relationship and that the portfolio rationale must rest on the value of the capability rather than on a discount to its successors. The anchors sharpen rather than soften this conclusion: Arthur names the mechanism the curve measures and warns that its realized path is contingent on history, and Mokyr predicts that the curve will bend most where the underlying knowledge is codified and reusable, which points the agency toward investment in reuse infrastructure as the lever on the slope. The contribution is the measurement, the design is complete, and the construction steps are named. What remains is the work of execution, offered in the same spirit of service that the autonomy enterprise has shown across a generation of flights: to replace a trusted belief with a measured one, so that those who decide how the next mission is funded may decide with evidence in hand. The instrument is ready to be run.
# References

The reference list below is compiled directly from the verified project bibliography (`research/corpus.jsonl`, 117 entries). Entries are ordered alphabetically by first-author surname and rendered in a single consistent style. Every digital object identifier and every uniform resource locator is rendered as a clickable link, per the standing hyperlink convention. Grey literature and unrefereed technical reports are flagged with a grade marker so that the reader can weight them accordingly; consistent with the study's grading convention, NASA Technical Reports Server records, conference proceedings, and policy reports carry a Grade B marker, and unrefereed preprints or repository documents carry a Grade C marker. The seed references that anchor the contribution and the two methodological lenses (Arthur and the experience-curve econometrics) are all Grade A peer-reviewed sources and are not marked. No MITRE-internal or working-note source appears in this list; the bibliography is composed entirely of externally resolvable scholarly and agency documents.

<span id="ref-1"></span>[1] Laith Alzubaidi, Jinglan Zhang, Amjad J. Humaidi, Ayad Q. Al-Dujaili, Ye Duan, Omran Al-Shamma, "Review of deep learning: concepts, CNN architectures, challenges, applications, future directions," *Journal of Big Data*, 2021. doi: [10.1186/s40537-021-00444-8](https://doi.org/10.1186/s40537-021-00444-8).
<span id="ref-2"></span>[2] Farzin Amzajerdian, Diego F. Pierrottet, Larry B. Petway, Glenn D. Hines, Vincent E. Roback, Robert A. Reisse, "Lidar Sensors for Autonomous Landing and Hazard Avoidance," *AIAA SPACE 2013 Conference and Exposition*, 2013. doi: [10.2514/6.2013-5312](https://doi.org/10.2514/6.2013-5312). [Grade B]
<span id="ref-3"></span>[3] W. B. Arthur, "Competing Technologies, Increasing Returns, and Lock-In by Historical Events," *The Economic Journal*, 1989. doi: [10.2307/2234208](https://doi.org/10.2307/2234208).
<span id="ref-4"></span>[4] W. B. Arthur, "Increasing Returns and Path Dependence in the Economy," *University of Michigan Press*, 1994. doi: [10.3998/mpub.10029](https://doi.org/10.3998/mpub.10029).
<span id="ref-5"></span>[5] W. B. Arthur, "Foundations of complexity economics," *Nature Reviews Physics*, 2021. doi: [10.1038/s42254-020-00273-3](https://doi.org/10.1038/s42254-020-00273-3).
<span id="ref-6"></span>[6] Robert U. Ayres, "Technological Progress: A Proposed Measure," *Technological Forecasting and Social Change*, 1998. doi: [10.1016/s0040-1625(98)00029-8](https://doi.org/10.1016/s0040-1625(98)00029-8).
<span id="ref-7"></span>[7] J. Balaram, M. Aung, M. P. Golombek, "The Ingenuity Helicopter on the Perseverance Rover," *Space Science Reviews*, 2021. doi: [10.1007/s11214-021-00815-w](https://doi.org/10.1007/s11214-021-00815-w).
<span id="ref-8"></span>[8] P. Baron, S. M. Cornet, E. D. Collins, G. DeAngelis, G. D. Del Cul, Yu. S. Fedorov, et al., "A review of separation processes proposed for advanced fuel cycles based on technology readiness level assessments," *Progress in Nuclear Energy*, 2019. doi: [10.1016/j.pnucene.2019.103091](https://doi.org/10.1016/j.pnucene.2019.103091).
<span id="ref-9"></span>[9] Eric Baumgartner, R. G. Bonitz, Joseph Melko, Lori Shiraishi, Patrick C. Leger, "The Mars Exploration Rover instrument positioning system," *2005 IEEE Aerospace Conference*, 2005. doi: [10.1109/aero.2005.1559295](https://doi.org/10.1109/aero.2005.1559295). [Grade B]
<span id="ref-10"></span>[10] Ramon Filipe Beims, Cândida Luiza Simonato, Vinicyus Rodolfo Wiggers, "Technology readiness level assessment of pyrolysis of triglyceride biomass to fuels and chemicals," *Renewable and Sustainable Energy Reviews*, 2019. doi: [10.1016/j.rser.2019.06.017](https://doi.org/10.1016/j.rser.2019.06.017).
<span id="ref-11"></span>[11] D. E. Bernard, et al., "Infusion of Autonomy Technology into Space Missions: DS1 Lessons Learned," *NASA Technical Reports Server*, 1997. [Online]. Available: [https://ntrs.nasa.gov/citations/20210003565](https://ntrs.nasa.gov/citations/20210003565). [Grade B]
<span id="ref-12"></span>[12] D. E. Bernard, et al., "Validation and Verification of the Remote Agent for Spacecraft Autonomy," *NASA Technical Reports Server*, 1999. [Online]. Available: [https://ntrs.nasa.gov/citations/20210003369](https://ntrs.nasa.gov/citations/20210003369). [Grade B]
<span id="ref-13"></span>[13] Douglas E. Bernard, Gregory A. Dorais, Charles Fry, Edward B. Gamble, Bob Kanefsky, James Kurien, et al., "Design of the Remote Agent experiment for spacecraft autonomy," *1998 IEEE Aerospace Conference*, 1998. doi: [10.1109/aero.1998.687914](https://doi.org/10.1109/aero.1998.687914). [Grade B]
<span id="ref-14"></span>[14] S. Bhattacharya, "The cost-quantity relations and the diverse patterns of learning by doing: Evidence from India," *Research Policy*, 2017. doi: [10.1016/j.respol.2017.09.005](https://doi.org/10.1016/j.respol.2017.09.005).
<span id="ref-15"></span>[15] Marco Bozzano, Alessandro Cimatti, Joost-Pieter Katoen, Panagiotis Katsaros, Konstantinos Mokos, Viet Yen Nguyen, et al., "Spacecraft early design validation using formal methods," *Reliability Engineering & System Safety*, 2014. doi: [10.1016/j.ress.2014.07.003](https://doi.org/10.1016/j.ress.2014.07.003).
<span id="ref-16"></span>[16] G. M. Brown, Douglas E. Bernard, Robert Rasmussen, "Attitude and articulation control for the Cassini spacecraft: a fault tolerance overview," *14th Digital Avionics Systems Conference*, 1995. doi: [10.1109/dasc.1995.482828](https://doi.org/10.1109/dasc.1995.482828). [Grade B]
<span id="ref-17"></span>[17] Georg A. Buchner, Arno Zimmermann, Arian E. Hohgräve, Reinhard Schomäcker, "Techno-economic Assessment Framework for the Chemical Industry Based on Technology Readiness Levels," *Industrial & Engineering Chemistry Research*, 2018. doi: [10.1021/acs.iecr.8b01248](https://doi.org/10.1021/acs.iecr.8b01248).
<span id="ref-18"></span>[18] Ahmed M. Bukar, Muhammad Asif, "Technology readiness level assessment of carbon capture and storage technologies," *Renewable and Sustainable Energy Reviews*, 2024. doi: [10.1016/j.rser.2024.114578](https://doi.org/10.1016/j.rser.2024.114578).
<span id="ref-19"></span>[19] Steve Chien, Ben Smith, G. Rabideau, Nicola Muscettola, Kanna Rajan, "Automated planning and scheduling for goal-based autonomous spacecraft," *IEEE Intelligent Systems and their Applications*, 1998. doi: [10.1109/5254.722362](https://doi.org/10.1109/5254.722362).
<span id="ref-20"></span>[20] Steve Chien, Russell Knight, Andre Stechert, Rob Sherwood, Gregg Rabideau, "Using Iterative Repair to Increase the Responsiveness of Planning and Scheduling for Autonomous Spacecraft," *NASA Technical Reports Server*, 1998. [Online]. Available: [https://ntrs.nasa.gov/api/citations/20000052453/downloads/20000052453.pdf](https://ntrs.nasa.gov/api/citations/20000052453/downloads/20000052453.pdf). [Grade C]
<span id="ref-21"></span>[21] Steve Chien, Russell Knight, Andre Stechert, Rob Sherwood, Gregg Rabideau, "Using Iterative Repair to Improve the Responsiveness of Planning and Scheduling," *Proc. AIPS*, 2000. [Online]. Available: [http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.96.1796](http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.96.1796). [Grade B]
<span id="ref-22"></span>[22] Steve Chien, Rob Sherwood, Daniel Tran, Benjamin Cichy, Gregg Rabideau, Rebecca Castano, et al., "The EO-1 Autonomous Science Agent Architecture," *Proc. International Workshop on Planning and Scheduling for Space*, Darmstadt, Germany, 2004. [Online]. Available: [https://ntrs.nasa.gov/citations/20060043730](https://ntrs.nasa.gov/citations/20060043730). [Grade B]
<span id="ref-23"></span>[23] S. Chien, et al., "Monitoring active volcanism with the Autonomous Sciencecraft Experiment on EO-1," *Remote Sensing of Environment*, 2005. doi: [10.1016/j.rse.2005.08.007](https://doi.org/10.1016/j.rse.2005.08.007).
<span id="ref-24"></span>[24] S. Chien, et al., "Using Autonomy Flight Software to Improve Science Return on Earth Observing One," *Journal of Aerospace Computing, Information, and Communication*, 2005. doi: [10.2514/1.12923](https://doi.org/10.2514/1.12923).
<span id="ref-25"></span>[25] R. Cowan, Philip Gunby, "Sprayed to Death: Path Dependence, Lock-in and Pest Control Strategies," *The Economic Journal*, 1996. doi: [10.2307/2235561](https://doi.org/10.2307/2235561).
<span id="ref-26"></span>[26] J. A. Crisp, Mark Adler, J. Matijevic, S. W. Squyres, R. E. Arvidson, D. M. Kass, "Mars Exploration Rover mission," *Journal of Geophysical Research: Planets*, 2003. doi: [10.1029/2002je002038](https://doi.org/10.1029/2002je002038).
<span id="ref-27"></span>[27] Paul A. David, "Path dependence, its critics, and the quest for 'historical economics'," *Edward Elgar Publishing eBooks*, 2001. doi: [10.4337/9781781950227.00006](https://doi.org/10.4337/9781781950227.00006).
<span id="ref-28"></span>[28] Leonhard Dobusch, Elke Schüssler, "Theorizing path dependence: a review of positive feedback mechanisms in technology markets, regional clusters, and organizations," *Industrial and Corporate Change*, 2012. doi: [10.1093/icc/dts029](https://doi.org/10.1093/icc/dts029).
<span id="ref-29"></span>[29] Egidio D'Amato, V. Nardi, Immacolata Notaro, Valerio Scordamaglia, "A Particle Filtering Approach for Fault Detection and Isolation of UAV IMU Sensors: Design, Implementation and Sensitivity Analysis," *Sensors*, 2021. doi: [10.3390/s21093066](https://doi.org/10.3390/s21093066).
<span id="ref-30"></span>[30] Douglas E. Bernard, Edward B. Gamble, Guy K. Man, Gregory A. Dorais, Bob Kanefsky, et al., "Spacecraft autonomy flight experience: The DS1 Remote Agent Experiment," *Space Technology Conference and Exposition*, 1999. doi: [10.2514/6.1999-4512](https://doi.org/10.2514/6.1999-4512). [Grade B]
<span id="ref-31"></span>[31] Chirold Epp, Thomas B. Smith, "Autonomous Precision Landing and Hazard Detection and Avoidance Technology (ALHAT)," *2007 IEEE Aerospace Conference*, 2007. doi: [10.1109/aero.2007.352724](https://doi.org/10.1109/aero.2007.352724). [Grade B]
<span id="ref-32"></span>[32] Chirold D. Epp, Edward A. Robertson, Tye Brady, "Autonomous Landing and Hazard Avoidance Technology (ALHAT)," *2008 IEEE Aerospace Conference*, 2008. doi: [10.1109/aero.2008.4526297](https://doi.org/10.1109/aero.2008.4526297). [Grade B]
<span id="ref-33"></span>[33] Chirold Epp, Ed Robertson, John M. Carson, "Real-Time Hazard Detection and Avoidance Demonstration for a Planetary Lander," *AIAA SPACE 2014 Conference and Exposition*, 2014. doi: [10.2514/6.2014-4312](https://doi.org/10.2514/6.2014-4312). [Grade B]
<span id="ref-34"></span>[34] T. A. Estlin, et al., "AEGIS Automated Science Targeting for the MER Opportunity Rover," *ACM Transactions on Intelligent Systems and Technology*, 2012. doi: [10.1145/2168752.2168764](https://doi.org/10.1145/2168752.2168764).
<span id="ref-35"></span>[35] J. D. Farmer, F. Lafond, "How predictable is technological progress?," *Research Policy*, 2016. doi: [10.1016/j.respol.2015.11.001](https://doi.org/10.1016/j.respol.2015.11.001).
<span id="ref-36"></span>[36] Christopher J. Farrell, "A theory of technological progress," *Technological Forecasting and Social Change*, 1993. doi: [10.1016/0040-1625(93)90025-3](https://doi.org/10.1016/0040-1625(93)90025-3).
<span id="ref-37"></span>[37] International Astronautical Federation, "Impact of Launch Cadence on the Automation and Economics of Constellation Operations," *IAC Proceedings*, 2024. doi: [10.52202/078367-0036](https://doi.org/10.52202/078367-0036). [Grade B]
<span id="ref-38"></span>[38] Lorraine Fesq, P. Beauchamp, Amanda Donner, Rob Bocchino, Brian Kennedy, Faiz Mirza, et al., "Extended Mission Technology Demonstrations Using the ASTERIA Spacecraft," *2019 IEEE Aerospace Conference*, 2019. doi: [10.1109/aero.2019.8742020](https://doi.org/10.1109/aero.2019.8742020). [Grade B]
<span id="ref-39"></span>[39] Veronica Foreman, Jacqueline Le Moigne, Olivier de Weck, "A Survey of Cost Estimating Methodologies for Distributed Spacecraft Missions," *AIAA SPACE 2016*, 2016. doi: [10.2514/6.2016-5245](https://doi.org/10.2514/6.2016-5245). [Grade B]
<span id="ref-40"></span>[40] R. Francis, et al., "AEGIS autonomous targeting for ChemCam on Mars Science Laboratory: Deployment and results of initial science team use," *Science Robotics*, 2017. doi: [10.1126/scirobotics.aan4582](https://doi.org/10.1126/scirobotics.aan4582).
<span id="ref-41"></span>[41] Alan R. Fusfeld, "The technological progress function: A new technique for forecasting," *Technological Forecasting*, 1970. doi: [10.1016/0099-3964(70)90031-1](https://doi.org/10.1016/0099-3964(70)90031-1).
<span id="ref-42"></span>[42] Y. Gao, S. Chien, "Autonomy for Space Robots: Past, Present, and Future," *Current Robotics Reports*, 2021. doi: [10.1007/s43154-021-00057-2](https://doi.org/10.1007/s43154-021-00057-2).
<span id="ref-43"></span>[43] Fereidoun Ghasemzadeh, Norm Archer, "Project portfolio selection through decision support," *Decision Support Systems*, 2000. doi: [10.1016/s0167-9236(00)00065-8](https://doi.org/10.1016/s0167-9236(00)00065-8).
<span id="ref-44"></span>[44] Gunther Glenk, Rebecca Meier, Stefan Reichelstein, "Cost Dynamics of Clean Energy Technologies," *Schmalenbach Journal of Business Research*, 2021. doi: [10.1007/s41471-021-00114-8](https://doi.org/10.1007/s41471-021-00114-8).
<span id="ref-45"></span>[45] S. Goldberg, Mark Maimone, Larry Matthies, "Stereo vision and rover navigation software for planetary exploration," *Proceedings of the IEEE Aerospace Conference*, 2002. doi: [10.1109/aero.2002.1035370](https://doi.org/10.1109/aero.2002.1035370). [Grade B]
<span id="ref-46"></span>[46] Carlos E. Gonzalez, Camilo Rojas, Alexandre Bergel, Marcos Díaz, "An Architecture-Tracking Approach to Evaluate a Modular and Extensible Flight Software for CubeSat Nanosatellites," *IEEE Access*, 2019. doi: [10.1109/access.2019.2927931](https://doi.org/10.1109/access.2019.2927931).
<span id="ref-47"></span>[47] Myron R. Grover, Benjamin Cichy, Prasun N. Desai, "Overview of the Phoenix Entry, Descent, and Landing System Architecture," *Journal of Spacecraft and Rockets*, 2011. doi: [10.2514/1.46548](https://doi.org/10.2514/1.46548).
<span id="ref-48"></span>[48] S. Hayati, R. Volpe, Paul Backes, J. Balaram, Richard V. Welch, R. Ivlev, et al., "The Rocky 7 rover: a Mars sciencecraft prototype," *Proceedings of the IEEE International Conference on Robotics and Automation*, 1997. doi: [10.1109/robot.1997.619330](https://doi.org/10.1109/robot.1997.619330). [Grade B]
<span id="ref-49"></span>[49] Eric von Hippel, Marcie J. Tyre, "How learning by doing is done: problem identification in novel process equipment," *Research Policy*, 1995. doi: [10.1016/0048-7333(93)00747-h](https://doi.org/10.1016/0048-7333(93)00747-h).
<span id="ref-50"></span>[50] Kerstin Hötte, "How to accelerate green technology diffusion? Directed technological change in the presence of coevolving absorptive capacity," *Energy Economics*, 2020. doi: [10.1016/j.eneco.2019.104565](https://doi.org/10.1016/j.eneco.2019.104565).
<span id="ref-51"></span>[51] Xiang-Yu Huang, Maodeng Li, Xiaolei Wang, Jinchang Hu, Yu Zhao, Minwen Guo, et al., "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).
<span id="ref-52"></span>[52] Terry Huntsberger, "Biologically Inspired Autonomous Rover Control," *Autonomous Robots*, 2001. doi: [10.1023/a:1012467829785](https://doi.org/10.1023/a:1012467829785).
<span id="ref-53"></span>[53] Mihály Héder, "From NASA to EU: the evolution of the TRL scale in Public Sector Innovation," *SZTAKI Publication Repository*, 2017. [Online]. Available: [http://eprints.sztaki.hu/9204/](http://eprints.sztaki.hu/9204/). [Grade B]
<span id="ref-54"></span>[54] Felipe Ip, J. M. Dohm, Victor R. Baker, T. Doggett, A. G. Davies, Rebecca Castaño, et al., "Flood detection and monitoring with the Autonomous Sciencecraft Experiment onboard EO-1," *Remote Sensing of Environment*, 2006. doi: [10.1016/j.rse.2005.12.018](https://doi.org/10.1016/j.rse.2005.12.018).
<span id="ref-55"></span>[55] Aymen Kayal, "Measuring the Pace of Technological Progress," *Technological Forecasting and Social Change*, 1999. doi: [10.1016/s0040-1625(98)00030-4](https://doi.org/10.1016/s0040-1625(98)00030-4).
<span id="ref-56"></span>[56] Martín Kenney, "Technology, entrepreneurship and path dependence: industrial clustering in Silicon Valley and Route 128," *Industrial and Corporate Change*, 1999. doi: [10.1093/icc/8.1.67](https://doi.org/10.1093/icc/8.1.67).
<span id="ref-57"></span>[57] Saleem Nawaz Khan, Yang Zhu, Weiguo Dong, Ming Zhao, "Cost and technology readiness level assessment of emerging technologies, new perspectives, and future research directions in H2 production," *Sustainable Energy & Fuels*, 2022. doi: [10.1039/d2se00988a](https://doi.org/10.1039/d2se00988a).
<span id="ref-58"></span>[58] Ksenia O. Kolcio, "Model-Based Fault Detection and Isolation System for Increased Autonomy," *AIAA SPACE 2016*, 2016. doi: [10.2514/6.2016-5225](https://doi.org/10.2514/6.2016-5225). [Grade B]
<span id="ref-59"></span>[59] Ksenia Kolcio, Lorraine Fesq, "Model-based off-nominal state isolation and detection system for autonomous fault management," *2016 IEEE Aerospace Conference*, 2016. doi: [10.1109/aero.2016.7500793](https://doi.org/10.1109/aero.2016.7500793). [Grade B]
<span id="ref-60"></span>[60] Georges Labrèche, David J. Evans, Dominik Marszk, Tom Mladenov, Vasundhara Shiradhonkar, Tanguy Soto, et al., "OPS-SAT Spacecraft Autonomy with TensorFlow Lite, Unsupervised Learning, and Online Machine Learning," *2022 IEEE Aerospace Conference*, 2022. doi: [10.1109/aero53065.2022.9843402](https://doi.org/10.1109/aero53065.2022.9843402). [Grade B]
<span id="ref-61"></span>[61] F. Lafond, et al., "How well do experience curves predict technological progress? A method for making distributional forecasts," *Technological Forecasting and Social Change*, 2018. doi: [10.1016/j.techfore.2017.11.001](https://doi.org/10.1016/j.techfore.2017.11.001).
<span id="ref-62"></span>[62] B. Lane, J. Reed, B. Shaffer, S. Samuelsen, "Forecasting renewable hydrogen production technology shares under cost uncertainty," *International Journal of Hydrogen Energy*, 2021. doi: [10.1016/j.ijhydene.2021.06.012](https://doi.org/10.1016/j.ijhydene.2021.06.012).
<span id="ref-63"></span>[63] Ibtissam Latachi, Tajjeeddine Rachidi, Mohammed Karim, Ahmed Hanafi, "Reusable and Reliable Flight-Control Software for a Fail-Safe and Cost-Efficient CubeSat Mission: Design and Implementation," *Aerospace*, 2020. doi: [10.3390/aerospace7100146](https://doi.org/10.3390/aerospace7100146).
<span id="ref-64"></span>[64] C. Lenzen, M. Wörle, T. Göttfert, Falk Mrowka, M. Wickler, "Onboard Planning and Scheduling Autonomy within the Scope of the FireBird Mission," *SpaceOps 2014 Conference*, 2014. doi: [10.2514/6.2014-1759](https://doi.org/10.2514/6.2014-1759). [Grade B]
<span id="ref-65"></span>[65] Tiffany Russell Lockett, Julie Castillo-Rogez, Les Johnson, Joe Matus, Jack Lightholder, Anne Marinan, et al., "Near-Earth Asteroid Scout Flight Mission," *IEEE Aerospace and Electronic Systems Magazine*, 2020. doi: [10.1109/maes.2019.2958729](https://doi.org/10.1109/maes.2019.2958729).
<span id="ref-66"></span>[66] Paolo Lunghi, Marco Ciarambino, Michèle Lavagna, "A multilayer perceptron hazard detector for vision-based autonomous planetary landing," *Advances in Space Research*, 2016. doi: [10.1016/j.asr.2016.04.012](https://doi.org/10.1016/j.asr.2016.04.012).
<span id="ref-67"></span>[67] Bolko Maass, Svenja Woicke, Willem M. Oliveira, Bronislovas Razgus, Hans Krüger, "Crater Navigation System for Autonomous Precision Landing on the Moon," *Journal of Guidance, Control, and Dynamics*, 2020. doi: [10.2514/1.g004850](https://doi.org/10.2514/1.g004850).
<span id="ref-68"></span>[68] Mark Maimone, Andrew Johnson, Yang Cheng, Reg G. Willson, Larry Matthies, "Autonomous Navigation Results from the Mars Exploration Rover (MER) Mission," *Springer Tracts in Advanced Robotics*, 2006. doi: [10.1007/11552246_1](https://doi.org/10.1007/11552246_1).
<span id="ref-69"></span>[69] Mark Maimone, Yang Cheng, Larry Matthies, "Two years of Visual Odometry on the Mars Exploration Rovers," *Journal of Field Robotics*, 2007. doi: [10.1002/rob.20184](https://doi.org/10.1002/rob.20184).
<span id="ref-70"></span>[70] J. C. Mankins, "Technology readiness assessments: A retrospective," *Acta Astronautica*, 2009. doi: [10.1016/j.actaastro.2009.03.058](https://doi.org/10.1016/j.actaastro.2009.03.058).
<span id="ref-71"></span>[71] M. Maurette, "Mars Rover Autonomous Navigation," *Autonomous Robots*, 2003. doi: [10.1023/a:1022283719900](https://doi.org/10.1023/a:1022283719900).
<span id="ref-72"></span>[72] S. Maurice, R. C. Wiens, P. Bernardi, Ph. Caïs, S. Robinson, T. Nelson, et al., "The SuperCam Instrument Suite on the Mars 2020 Rover: Science Objectives and Mast-Unit Description," *Space Science Reviews*, 2021. doi: [10.1007/s11214-021-00807-w](https://doi.org/10.1007/s11214-021-00807-w).
<span id="ref-73"></span>[73] David McComas, "NASA/GSFC's Flight Software Core Flight System," *NASA STI Repository*, 2013. [Online]. Available: [http://hdl.handle.net/2060/20130013412](http://hdl.handle.net/2060/20130013412). [Grade B]
<span id="ref-74"></span>[74] Alan McDonald, Leo Schrattenholzer, "Learning rates for energy technologies," *Energy Policy*, 2001. doi: [10.1016/s0301-4215(00)00122-1](https://doi.org/10.1016/s0301-4215(00)00122-1).
<span id="ref-75"></span>[75] David Miranda, "2020 NASA Technology Taxonomy," *NASA STI Repository*, 2019. [Online]. Available: [http://hdl.handle.net/2060/20190032038](http://hdl.handle.net/2060/20190032038). [Grade B]
<span id="ref-76"></span>[76] Syed Agha Hassnain Mohsan, Nawaf Qasem Hamood Othman, Yanlong Li, Mohammed H. Alsharif, Muhammad Asghar Khan, "Unmanned aerial vehicles (UAVs): practical aspects, applications, open challenges, security issues, and future trends," *Intelligent Service Robotics*, 2023. doi: [10.1007/s11370-022-00452-4](https://doi.org/10.1007/s11370-022-00452-4).
<span id="ref-77"></span>[77] J. Mrozinski, G. Fox, Hamid Habib-Agahi, "NASA Instrument Cost/Schedule Model," *2011 IEEE Aerospace Conference*, 2011. doi: [10.1109/aero.2011.5747633](https://doi.org/10.1109/aero.2011.5747633). [Grade B]
<span id="ref-78"></span>[78] Kris Myny, Steve Smout, Maarten Rockelé, Ajay Bhoolokam, Tung Huei Ke, Soeren Steudel, et al., "A thin-film microprocessor with inkjet print-programmable memory," *Scientific Reports*, 2014. doi: [10.1038/srep07398](https://doi.org/10.1038/srep07398).
<span id="ref-79"></span>[79] B. Nagy, J. D. Farmer, Q. M. Bui, J. E. Trancik, "Statistical Basis for Predicting Technological Progress," *PLOS ONE*, 2013. doi: [10.1371/journal.pone.0052669](https://doi.org/10.1371/journal.pone.0052669).
<span id="ref-80"></span>[80] Lena Neij, "Cost development of future technologies for power generation: A study based on experience curves and complementary bottom-up assessments," *Energy Policy*, 2008. doi: [10.1016/j.enpol.2008.02.029](https://doi.org/10.1016/j.enpol.2008.02.029).
<span id="ref-81"></span>[81] A. L. Olechowski, S. D. Eppinger, N. Joglekar, K. Tomaschek, "Technology readiness levels: Shortcomings and improvement opportunities," *Systems Engineering*, 2020. doi: [10.1002/sys.21533](https://doi.org/10.1002/sys.21533).
<span id="ref-82"></span>[82] Barney Pell, Douglas E. Bernard, Steve Chien, Erann Gat, Nicola Muscettola, P. Pandurang Nayak, et al., "Remote agent prototype for spacecraft autonomy," *Proceedings of SPIE*, 1996. doi: [10.1117/12.255150](https://doi.org/10.1117/12.255150). [Grade B]
<span id="ref-83"></span>[83] Barney Pell, Douglas E. Bernard, Steve Chien, Erann Gat, Nicola Muscettola, P. Pandurang Nayak, et al., "An autonomous spacecraft agent prototype," *Proceedings of the First International Conference on Autonomous Agents*, 1997. doi: [10.1145/267658.267724](https://doi.org/10.1145/267658.267724).
<span id="ref-84"></span>[84] C. Perez, "Technological revolutions and techno-economic paradigms," *Cambridge Journal of Economics*, 2009. doi: [10.1093/cje/bep051](https://doi.org/10.1093/cje/bep051).
<span id="ref-85"></span>[85] Paul Pierson, "Increasing Returns, Path Dependence, and the Study of Politics," *American Political Science Review*, 2000. doi: [10.2307/2586011](https://doi.org/10.2307/2586011).
<span id="ref-86"></span>[86] Edward S. Rubin, Inês M. L. Azevedo, Paulina Jaramillo, Sonia Yeh, "A review of learning rates for electricity supply technologies," *Energy Policy*, 2015. doi: [10.1016/j.enpol.2015.06.011](https://doi.org/10.1016/j.enpol.2015.06.011).
<span id="ref-87"></span>[87] Justyna Rybicka, Ashutosh Tiwari, Gary A. Leeke, "Technology readiness level assessment of composites recycling technologies," *Journal of Cleaner Production*, 2015. doi: [10.1016/j.jclepro.2015.08.104](https://doi.org/10.1016/j.jclepro.2015.08.104).
<span id="ref-88"></span>[88] B. D. Smith, et al., "Challenges and Methods in Testing the Remote Agent Planner," *NASA Technical Reports Server*, 2000. [Online]. Available: [https://ntrs.nasa.gov/citations/20210001679](https://ntrs.nasa.gov/citations/20210001679). [Grade B]
<span id="ref-89"></span>[89] H. P. Stahl, T. Henrichs, A. Luedtke, M. West, "Single-variable parametric cost models for space telescopes," *Optical Engineering*, 2010. doi: [10.1117/1.3456582](https://doi.org/10.1117/1.3456582).
<span id="ref-90"></span>[90] H. Philip Stahl, "Survey of cost models for space telescopes," *Optical Engineering*, 2010. doi: [10.1117/1.3430603](https://doi.org/10.1117/1.3430603).
<span id="ref-91"></span>[91] Center for Security and Emerging Technology, "Shaping the U.S. Space Launch Market," *CSET*, 2024. doi: [10.51593/20240017](https://doi.org/10.51593/20240017). [Grade B]
<span id="ref-92"></span>[92] R. J. Terrile, Fred G. Doumani, Gary Y. Ho, Byron Jackson, "Calibrating the Technology Readiness Level (TRL) scale using NASA mission data," *2015 IEEE Aerospace Conference*, 2015. doi: [10.1109/aero.2015.7119313](https://doi.org/10.1109/aero.2015.7119313). [Grade B]
<span id="ref-93"></span>[93] F. Teston, Richard Creasey, Jo Bermyn, D. Bernaerts, Karim Mellab, "Proba: ESA's Autonomy and Technology Demonstration Mission," *Proc. AIAA/USU Conference on Small Satellites*, 1999. [Online]. Available: [https://digitalcommons.usu.edu/cgi/viewcontent.cgi?article=2152&context=smallsat](https://digitalcommons.usu.edu/cgi/viewcontent.cgi?article=2152&context=smallsat). [Grade B]
<span id="ref-94"></span>[94] P. Thompson, "The Relationship between Unit Cost and Cumulative Quantity and the Evidence for Organizational Learning-by-Doing," *Journal of Economic Perspectives*, 2012. doi: [10.1257/jep.26.3.203](https://doi.org/10.1257/jep.26.3.203).
<span id="ref-95"></span>[95] Olivier Toupet, Tyler Del Sesto, Masahiro Ono, Steven Myint, Joshua Vander Hook, Michael McHenry, et al., "A ROS-based Simulator for Testing the Enhanced Autonomous Navigation of the Mars 2020 Rover," *2020 IEEE Aerospace Conference*, 2020. doi: [10.1109/aero47225.2020.9172345](https://doi.org/10.1109/aero47225.2020.9172345). [Grade B]
<span id="ref-96"></span>[96] Olga Trivailo, Martin Sippel, Y. Ahmet Şekercioğlu, "Review of hardware cost estimation methods, models and tools applied to early phases of space mission planning," *Progress in Aerospace Sciences*, 2012. doi: [10.1016/j.paerosci.2012.02.001](https://doi.org/10.1016/j.paerosci.2012.02.001).
<span id="ref-97"></span>[97] Andrew Turner, "An Open-Source, Extensible Spacecraft Simulation and Modeling Environment Framework," *VTechWorks (Virginia Tech)*, 2003. [Online]. Available: [http://hdl.handle.net/10919/34727](http://hdl.handle.net/10919/34727). [Grade B]
<span id="ref-98"></span>[98] Jean-Philippe Vergne, Rodolphe Durand, "The missing link between the theory and empirics of path dependence: Conceptual clarification, testability issue, and methodological implications," *Journal of Management Studies*, 2010. doi: [10.1111/j.1467-6486.2009.00913.x](https://doi.org/10.1111/j.1467-6486.2009.00913.x).
<span id="ref-99"></span>[99] V. Verma, et al., "Autonomous robotics is driving Perseverance rover's progress on Mars," *Science Robotics*, 2023. doi: [10.1126/scirobotics.adi3099](https://doi.org/10.1126/scirobotics.adi3099).
<span id="ref-100"></span>[100] Alexandra Wander, Roger Förstner, "Innovative fault detection, isolation and recovery on-board spacecraft: Study and implementation using cognitive automation," *2013 Conference on Control and Fault-Tolerant Systems (SysTol)*, 2013. doi: [10.1109/systol.2013.6693950](https://doi.org/10.1109/systol.2013.6693950). [Grade B]
<span id="ref-101"></span>[101] R. Washington, Keith Golden, John Bresina, David E. Smith, C. Anderson, Trey Smith, "Autonomous rovers for Mars exploration," *1999 IEEE Aerospace Conference*, 1999. doi: [10.1109/aero.1999.794236](https://doi.org/10.1109/aero.1999.794236). [Grade B]
<span id="ref-102"></span>[102] Rupert Way, F. Lafond, F. Lillo, Valentyn Panchenko, J. D. Farmer, "Wright meets Markowitz: How standard portfolio theory changes when assets are technologies following experience curves," *Journal of Economic Dynamics and Control*, 2017. doi: [10.2139/ssrn.2965695](https://doi.org/10.2139/ssrn.2965695).
<span id="ref-103"></span>[103] M. Wei, S. Smith, M. Sohn, "Non-constant learning rates in retrospective experience curve analyses and their correlation to deployment programs," *Energy Policy*, 2017. doi: [10.1016/j.enpol.2017.04.035](https://doi.org/10.1016/j.enpol.2017.04.035).
<span id="ref-104"></span>[104] R. C. Wiens, S. Maurice, B. L. Barraclough, M. Saccoccio, Walter Barkley, J. F. Bell, et al., "The ChemCam Instrument Suite on the Mars Science Laboratory (MSL) Rover: Body Unit and Combined System Tests," *Space Science Reviews*, 2012. doi: [10.1007/s11214-012-9902-4](https://doi.org/10.1007/s11214-012-9902-4).
<span id="ref-105"></span>[105] Jonathan Wilmot, "A Core Plug and Play Architecture for Reusable Flight Software Systems," *2nd IEEE International Conference on Space Mission Challenges for Information Technology*, 2006. doi: [10.1109/smc-it.2006.7](https://doi.org/10.1109/smc-it.2006.7). [Grade B]
<span id="ref-106"></span>[106] Jonathan Wilmot, "Implications of Responsive Space on the Flight Software Architecture," *NASA Technical Reports Server*, 2006. [Online]. Available: [http://hdl.handle.net/2060/20060026205](http://hdl.handle.net/2060/20060026205). [Grade C]
<span id="ref-107"></span>[107] Mark Woods, Andy Shaw, Dave Barnes, D. M. Price, Derek Long, D. Pullan, "Autonomous science for an ExoMars Rover-like mission," *Journal of Field Robotics*, 2009. doi: [10.1002/rob.20289](https://doi.org/10.1002/rob.20289).
<span id="ref-108"></span>[108] Adrian Yao, Sally M. Benson, William C. Chueh, "Critically assessing sodium-ion technology roadmaps and scenarios for techno-economic competitiveness against lithium-ion batteries," *Nature Energy*, 2025. doi: [10.1038/s41560-024-01701-9](https://doi.org/10.1038/s41560-024-01701-9).
<span id="ref-109"></span>[109] Sonia Yeh, Edward S. Rubin, "A review of uncertainties in technology experience curves," *Energy Economics*, 2011. doi: [10.1016/j.eneco.2011.11.006](https://doi.org/10.1016/j.eneco.2011.11.006).
<span id="ref-110"></span>[110] De Jong Yeong, Gustavo Velasco-Hernandez, John M. Barry, J. L. Walsh, "Sensor and Sensor Fusion Technology in Autonomous Vehicles: A Review," *Sensors*, 2021. doi: [10.3390/s21062140](https://doi.org/10.3390/s21062140).
<span id="ref-111"></span>[111] Jie Yu, Honghua Zhang, Ming Cheng, Jun Liang, Yu Zhao, Ji Li, "Autonomous hazard avoidance control for Chang'E-3 soft landing," *Scientia Sinica Technologica*, 2014. doi: [10.1360/092014-51](https://doi.org/10.1360/092014-51).
<span id="ref-112"></span>[112] Yunan Zhao, Xinlong Wang, Qunsheng Li, Dun Wang, Yuanwen Cai, "A high-accuracy autonomous navigation scheme for the Mars rover," *Acta Astronautica*, 2019. doi: [10.1016/j.actaastro.2018.10.036](https://doi.org/10.1016/j.actaastro.2018.10.036).
<span id="ref-113"></span>[113] Itai Zilberstein, Ananya Rao, M. Salis, Steve A. Chien, "Decentralized, Decomposition-Based Observation Scheduling for a Large-Scale Satellite Constellation," *Proceedings of the International Conference on Automated Planning and Scheduling*, 2024. doi: [10.1609/icaps.v34i1.31535](https://doi.org/10.1609/icaps.v34i1.31535).

<span id="ref-114"></span>[114] Joel Mokyr, *The Lever of Riches: Technological Creativity and Economic Progress*. New York: Oxford University Press, 1990. doi: [10.1093/acprof:oso/9780195074772.001.0001](https://doi.org/10.1093/acprof:oso/9780195074772.001.0001).

<span id="ref-115"></span>[115] Joel Mokyr, *The Gifts of Athena: Historical Origins of the Knowledge Economy*. Princeton: Princeton University Press, 2002. doi: [10.1515/9781400829439](https://doi.org/10.1515/9781400829439).

<span id="ref-116"></span>[116] T. P. Wright, "Factors Affecting the Cost of Airplanes," *Journal of the Aeronautical Sciences*, vol. 3, no. 4, pp. 122-128, 1936. doi: [10.2514/8.155](https://doi.org/10.2514/8.155).
<span id="ref-117"></span>[117] P. D. Friz, S. Hosder, B. B. Leser, and B. C. Towle, "Blind validation study of parametric cost estimation tool SEER-H for NASA space missions," *Acta Astronautica*, vol. 166, pp. 358-368, 2020. doi: [10.1016/j.actaastro.2019.09.030](https://doi.org/10.1016/j.actaastro.2019.09.030).



# Appendices

## Appendix A. Variable and Data Dictionary, with the Capability-Class Taxonomy

This appendix is the authoritative dictionary for every symbol in the estimating equation and every field in the constructed panel. The estimating equation is fixed and is reproduced here without alteration:

\[ \ln(\text{Cost}_{icd}) = \alpha + \beta \, \ln(\text{CumHeritage}_{icd}) + \gamma_c + \delta_d + \epsilon_{icd} \qquad\qquad (1) \]

for episode \(i\) in capability class \(c\) and decade \(d\). The dictionary entries below pin each term to a precise definition, a unit, a source field, and an admissible range, so that the panel can be assembled without recourse to the body chapters.

**Episode (`i`), the unit of analysis.** A single mission or technology demonstration that fields one defined onboard autonomy capability and is assigned to exactly one primary capability class. The identifier is a composite of the mission name and the capability tag (for example, "EO-1 / onboard-planning"). When a single mission fields more than one autonomy capability, it generates more than one episode record, and a `multi_capability_flag` marks each split so that the influence diagnostics in the analysis plan can test whether splitting drives the slope. The source of record for episode existence and timing is the NASA TechPort project inventory, cross-checked against the NTRS demonstration reports listed in the reference list.

**Dependent variable, \(\text{Cost}\).** The recurring non-recurring-engineering cost to qualify the named capability for flight, expressed in constant-year United States dollars after normalization by capability scope on a NICM-class basis. The variable is strictly positive and enters the model as \(\ln(\text{Cost})\). Its construction is the most delicate step in the study and is given a full derivation in Appendix B. Each \(\text{Cost}\) value carries a companion field, `cost_layer` (one of extract, impute, deflate-only), recording which normalization layer produced it, and a `reliability_flag` (high, medium, low) used to weight the inverse-imputation-error robustness specification. The admissible range is open above zero; values are recorded to the nearest thousand constant-year dollars to avoid false precision given the imputation error.

**Independent variable, \(\text{CumHeritage}\).** For each episode, the integer count of prior flight demonstrations within the same capability class that reached flight operation strictly before the episode's development-start date. The forward-only counting rule is binding: a demonstration that reached flight after the focal episode's development start contributes nothing to that episode's heritage, because heritage that arrives late cannot have lowered the cost already being incurred. The first episode in each class takes \(\text{CumHeritage} = 1\) before the logarithm is applied, fixing the curve origin. The variable enters as \(\ln(\text{CumHeritage})\). Appendix C tabulates the counting log episode by episode.

**Capability-class fixed effects, \(\gamma_c\).** A set of indicator variables, one per capability class, absorbing time-invariant differences in baseline cost and intrinsic difficulty across classes. The five initial classes form the taxonomy below. Class membership is a categorical field, `capability_class`, populated from the TechPort taxonomy mapping and confirmed against the NTRS capability definitions.

**Decade fixed effects, \(\delta_d\).** A set of indicator variables, one per decade of development start (1990s, 2000s, 2010s, 2020s), absorbing economy-wide and agency-wide trends in computing cost, tooling, and software practice that affect all classes simultaneously. The decade is derived deterministically from the `dev_start_date` field.

**Maturation covariate, \(\text{TRL\_start}\).** The technology-readiness level at development start, taken from the TechPort technology-readiness history, included only in the robustness specification that separates a heritage effect from the simpler effect of starting at higher maturity. It is an ordinal field on the standard one-through-nine scale and is never treated as a cost proxy, consistent with the documented ordinal and non-monetary character of the scale.

The capability-class taxonomy, which defines `c`, is as follows. Class 1, onboard planning and scheduling, covers goal-directed onboard plan generation, execution, and replanning (the Remote Agent and EO-1 Autonomous Sciencecraft lineage). Class 2, autonomous science target selection, covers onboard selection of science targets for narrow-field instruments (the AEGIS lineage from Opportunity to ChemCam). Class 3, autonomous surface or in-flight navigation, covers visual-odometry traverse autonomy and powered autonomous flight (the Mars Exploration Rover AutoNav through Perseverance and Ingenuity lineage). Class 4, autonomous fault detection, isolation, and recovery, covers model-based onboard fault management. Class 5, autonomous entry-descent-and-landing hazard handling, covers terrain-relative hazard detection and avoidance during descent. Each episode is assigned to exactly one of these five classes; boundary cases are resolved to the dominant capability and flagged, and the flag is carried into the analysis so the reader can see which assignments are contestable.

## Appendix B. Derivation of the Three-Layer Cost Normalization

The dependent variable cannot be lifted from public records as a clean number, because development-cost figures for autonomy capabilities are reported inconsistently: some are full project costs that bundle platform, instrument, and autonomy together; some are subsystem costs; and many are not disclosed. This appendix derives the three-layer normalization that converts whatever is available for each episode into a comparable per-episode autonomy qualification cost, and it states the bias direction of each layer so the measurement error is auditable rather than hidden.

Let \(\text{C\_raw}\) be the most autonomy-specific cost figure obtainable for an episode, \(S\) be a documented scope measure (capability size proxy drawn from the demonstration report), and \(P(t)\) be a standard deflator mapping nominal dollars in year \(t\) to constant-year dollars. The normalized dependent variable is built in three ordered layers. Layer one, extraction: where the autonomy-specific non-recurring-engineering portion is separately reported, set \(\text{C\_norm} = P(t) \cdot \text{C\_auto}\), where \(\text{C\_auto}\) is the extracted autonomy portion net of platform and instrument cost. This layer carries the least measurement error and earns `reliability_flag = high`. Layer two, imputation: where the autonomy portion is not separately reported, estimate it parametrically from the documented scope using a NICM-class single-variable parametric relationship of the form \(\hat{\text{C}}_{\text{auto}} = f(S)\), calibrated on the parametric cost-model literature for space development, then deflate: \(\text{C\_norm} = P(t) \cdot \hat{\text{C}}_{\text{auto}}\). This layer is load-bearing precisely because direct autonomy-cost figures are scarce in the open record, and it earns `reliability_flag = medium` or `low` depending on how far the episode's scope sits from the calibration support. Layer three, deflation-only: where a credible autonomy-specific nominal figure exists but no scope normalization is needed, apply \(\text{C\_norm} = P(t) \cdot \text{C\_raw}\) and record `cost_layer = deflate-only`.

Two properties of this derivation matter for inference. First, the normalization does not eliminate measurement error in \(\text{Cost}\); it makes the error explicit, attaches a reliability flag to every value, and routes the weakest values into a down-weighting robustness specification rather than discarding them. Second, the imputation layer introduces a known risk: if \(f(S)\) systematically compresses the spread of autonomy cost across scope, it biases \(\beta\) toward zero, which is conservative for a study whose alternative hypothesis is a negative slope. The retained imputation log, preserved with the panel, records \(\text{C\_raw}\), \(S\), the layer applied, the deflator vintage, and the resulting \(\text{C\_norm}\) for every episode, so that a reviewer can re-run the normalization and reproduce the dependent variable exactly.

The deflator term \(P(t)\) is itself a documented choice rather than a free parameter, because the conclusion that a cost fell over a multi-decade window is only meaningful relative to a stated price base. The study fixes a single constant-year base, applies one published economy-wide or agency-appropriate deflator series to every nominal figure, and records the deflator vintage in the imputation log so that a reviewer who prefers a different base can re-deflate without re-extracting. This matters for identification: because decade fixed effects already absorb the part of cost movement common to all classes within a decade, any residual sensitivity of the slope to the deflator choice would surface as a within-decade artifact, and the analysis plan therefore re-estimates the slope under an alternative deflator as an informal robustness check separate from the three pre-registered specifications. The scope measure \(S\) is likewise drawn only from documented quantities in the demonstration reports (for example, the number of distinct autonomy functions integrated, or a coarse software-size proxy where reported), never from a quantity that is itself a function of cost, so that the normalization does not circularly divide cost by a cost-derived denominator. Where two scope proxies are available for the same episode, the more conservative proxy (the one that yields the smaller implied cost decline) is used, holding the study to the same toward-zero bias discipline that governs the heritage count.

## Appendix C. The Forward-Only Cumulative-Heritage Counting Log

This appendix specifies the deterministic procedure that produces \(\text{CumHeritage}\) for every episode, and it provides the log template that records the result. The procedure has four steps. Step one: sort all episodes within a capability class by `flight_operation_date`, the date the demonstration reached flight operation. Step two: for each focal episode, read its `dev_start_date`. Step three: count the number of same-class episodes whose `flight_operation_date` is strictly earlier than the focal episode's `dev_start_date`; this integer is the raw heritage stock. Step four: if the raw stock is zero (the focal episode is first in class), set \(\text{CumHeritage} = 1\) so the logarithm is defined and the first demonstration anchors the curve origin; otherwise set \(\text{CumHeritage}\) to the raw stock. The rule is forward-only by construction, which prevents the mechanical correlation in which a later, more expensive episode is wrongly credited with heritage it could not have used, and it is conservative because it tends to undercount heritage and thereby biases the slope toward zero. The log template carries, per episode, the fields `episode_id`, `capability_class`, `dev_start_date`, `flight_operation_date`, `raw_heritage_count`, \(\text{CumHeritage}\), and a free-text `heritage_provenance` note naming each prior demonstration counted, so the count is checkable against the dated demonstration record in the reference list.

## Appendix D. Pre-Registration of Specifications and the Fixed Decision Rule

This appendix fixes, in advance of any estimation, the baseline specification, the three robustness specifications, and the decision rule, so that the analysis cannot be tuned to a desired result after the panel is seen. The baseline specification is the two-way fixed-effects log-log regression in Appendix A, estimating \(\beta\) off within-class, within-decade variation in cumulative heritage. Robustness specification one adds the \(\text{TRL\_start}\) maturation covariate, separating a genuine heritage effect from the simpler effect of starting at higher maturity. Robustness specification two broadens heritage to include cross-class software-component reuse, which should attenuate any downward bias arising from shared components and tests whether the within-class count understates the true reusable-knowledge stock. Robustness specification three weights observations by the inverse of the imputation error, down-weighting episodes whose cost rests on the weakest layer-two imputation. The decision rule is fixed: reject the null in favor of the alternative if and only if \(\beta\) is negative and its confidence interval excludes zero in the baseline specification and in at least two of the three robustness specifications; otherwise the flat-cost null stands. Two pre-analysis checks are mandatory before \(\beta\) is interpreted: a fixed-effects feasibility check that flags any class or decade cell containing a single observation, since such cells contribute nothing to the within estimator, and a small-panel influence diagnostic that re-estimates \(\beta\) with each potentially influential episode removed and treats any slope that depends on a single episode as not robust. Inference uses small-sample-robust standard errors, and a wide interval that contains zero is reported as a failure to reject the null rather than as evidence for the alternative.

## Appendix E. Supplementary Tables and the Triaged Literature

This appendix holds the supplementary material that supports the body without interrupting it. It carries three tabular artifacts whose schemas are fixed here and whose values are populated in the build phase. The episode inventory table records, per episode, the dictionary fields from Appendix A together with the cost-layer and heritage-provenance fields from Appendices B and C, and it is the single object from which the regression matrix is assembled. The capability-class summary table aggregates, per class, the episode count, the decade span, the realized within-class heritage range, and the count of single-observation cells flagged by the feasibility check, so that a reader can see at a glance how much identifying variation each class supplies and how thin the effective sample is. The specification-results template reserves the rows for \(\beta\), its confidence interval, the implied learning rate (\(1 - 2^{\beta}\), reported for interpretability), and the accept-or-reject decision under the baseline and the three robustness specifications; every cell is left explicitly empty at the design stage, because no coefficient has been fitted to the full dataset and none will be reported until the panel and the cost normalization are complete. This appendix also holds the extended literature table, which carries the demonstration-record and codified-reuse sources triaged out of the main literature review so that the body cites the strongest spine while the full evidentiary base remains visible and auditable here. Marking these tables as templates rather than results is the honest design-stage posture the whole dissertation maintains: the structure is complete and reproducible, the values are a build-phase deliverable, and the distinction between the two is stated plainly so that no illustrative figure is ever mistaken for an executed finding.

**Table E.1. Episode inventory schema (design-stage template; no values populated).**

| Field | Type | Source | Notes |
|-------|------|--------|-------|
| `episode_id` | string | TechPort + NTRS | mission name + capability tag |
| `capability_class` | categorical (1-5) | TechPort taxonomy | dominant class; boundary cases flagged |
| `multi_capability_flag` | boolean | analyst | true when a mission split into multiple episodes |
| `dev_start_date` | date | TechPort | drives decade and heritage cutoff |
| `flight_operation_date` | date | NTRS / literature | drives heritage counting |
| \(\text{Cost}\) (\(\text{C\_norm}\)) | numeric (constant-year USD) | Appendix B | strictly positive; enters as \(\ln(\text{Cost})\) |
| `cost_layer` | categorical | Appendix B | extract / impute / deflate-only |
| `reliability_flag` | categorical | Appendix B | high / medium / low |
| \(\text{CumHeritage}\) | integer (>= 1) | Appendix C | enters as \(\ln(\text{CumHeritage})\) |
| \(\text{TRL\_start}\) | ordinal (1-9) | TechPort | robustness covariate only |
| `decade` | categorical | derived | from `dev_start_date` |

**Table E.2. Capability-class summary schema (design-stage template).**

| Field | Type | Purpose |
|-------|------|---------|
| `capability_class` | categorical (1-5) | row key |
| `n_episodes` | integer | identifying-variation budget per class |
| `decade_span` | range | which decade fixed effects the class touches |
| \(\text{heritage\_range}\) | range | realized \(\text{CumHeritage}\) spread within class |
| `n_singleton_cells` | integer | cells flagged by the feasibility check |

**Table E.3. Specification-results template (all cells empty at design stage by construction).**

| Specification | \(\beta\) | 95% CI | Implied LR (\(1 - 2^{\beta}\)) | H0 decision |
|---------------|--------|--------|---------------------------|-------------|
| Baseline (two-way FE) | (empty) | (empty) | (empty) | (empty) |
| R1: + \(\text{TRL\_start}\) | (empty) | (empty) | (empty) | (empty) |
| R2: cross-class reuse | (empty) | (empty) | (empty) | (empty) |
| R3: inverse-error weights | (empty) | (empty) | (empty) | (empty) |

The empty cells in Table E.3 are not an omission; they are the literal embodiment of the design-stage guardrail. A completed dissertation will fill exactly these four rows and report the accept-or-reject decision the body has pre-committed to, and the symmetry of the design means the table is informative whichever way it resolves: a populated baseline with a negative \(\beta\) whose interval excludes zero, confirmed in at least two robustness rows, rejects the flat-cost null and yields a planning parameter, while a \(\beta\) indistinguishable from zero leaves the null standing and reframes the portfolio rationale around capability value. Neither outcome can be read off these appendices today, and the dissertation is explicit that it does not pretend otherwise.

## Appendix F. Instrument and Data-Source Query Details

This appendix records the operational query details for the four named data sources, so that the panel assembly can be reproduced rather than merely described. It is the instrument section of a study whose instruments are public application programming interfaces and document repositories rather than physical sensors.

The first source is NASA TechPort, the agency system of record for technology projects, accessed through its public application programming interface and bulk export at the TechPort public portal. The retrieval procedure filters the project inventory to the Autonomous Systems and Robotics taxonomy area, extracts for each candidate project the start and end dates, the owning organization, the technology-readiness entry and exit estimates, and the taxonomy classification, and writes these into the episode inventory fields of Table E.1. Because TechPort coverage is uneven across the early part of the window, every TechPort-derived episode is cross-checked against an independent NTRS or peer-reviewed record before it is admitted, and an admission note records the corroborating citation.

The second source is the NASA Technical Reports Server, accessed through its citations search application programming interface. The query set targets the named demonstrations that form the empirical spine: Remote Agent on Deep Space 1 (records 20210003369, 20210001679, 20210003565 in the reference list), the EO-1 Autonomous Sciencecraft Experiment, the AEGIS deployments, and the Ingenuity flight demonstration. NTRS supplies the heritage chronology (`flight_operation_date`), the capability definitions that fix class membership, and the qualitative scope information that sizes each episode for the Appendix B normalization. The third source is the NICM-class parametric cost-model literature, which is not queried through an interface but is applied as the calibrated function `f(S)` of Appendix B; the governing references are the single-variable parametric space-cost models and the NICM instrument cost and schedule model in the reference list. The fourth source is the published mission and autonomy literature catalogued in the reference list, which supplies capability scope, demonstration dates, and the documented heritage links between episodes. For every source, the query strings, the retrieval date, and the count of admitted records are logged alongside the panel, so the assembly is auditable to the same standard as the cost normalization and the heritage counting.