--- marp: true theme: default paginate: true --- # EDL Heritage and Landing-Success Hazard ### Does reuse of flight-proven entry-descent-landing architecture reduce landing-failure risk? **Doctoral Defense Brief** Candidate JPL_AUTONOMY_EDL_05 COLLEGIUM 1st Battalion NORTH STAR / JPL category: Entry Descent & Landing Systems Methodological anchors: Fogel (cliometric counterfactual); Mokyr / Perez (useful knowledge) 2026-06-15 --- ## The answer, stated first This dissertation delivers the first population-level, **conditional** estimate of whether reusing a flight-proven EDL architecture lineage lowers a planetary landing's failure hazard. - The deliverable is a single coefficient `beta_1`, or a credible failure to find one. - It converts the constant "heritage is safer" intuition into a measured, falsifiable, confounding-adjusted EDL portfolio parameter. - It is not a new EDL technology, reconstruction, or architecture. It is a hypothesis test of an assertion the field makes constantly and almost never tests against the full record. - Design-stage: no coefficient is fitted here. The contribution at this stage is the pre-registered, falsifiable design itself. --- ## The contribution as one pair of hypotheses **H1 (contribution):** Conditional on target body, entry mass, and landed mass, the coefficient on the EDL-heritage-reuse index in a logistic model of landing failure is **negative and statistically distinguishable from zero** (`beta_1 < 0`). **H0 (null):** Conditional on the same controls, the coefficient is **zero** (`beta_1 = 0`). EDL architectural novelty has no effect on landing-failure probability. ``` logit Pr(failure_i = 1) = beta_0 + beta_1 * heritage_index_i + gamma' * controls_i + epsilon_i ``` The contribution lives entirely in `beta_1`. The proposition is falsifiable in the strict sense. --- ## Why a coefficient, not anecdotes - Mars: a clean near-replication chain (Phoenix to InSight; MSL to Mars 2020) looks like heritage protects. - The Moon: heritage-derived landers and first-flight commercial vehicles; several novel-architecture losses and anomalies (Vikram; SLIM's anomalous attitude) look like novelty is dangerous. - Neither reading controls for target difficulty or program depth. - A coefficient estimated against the **whole population with controls** is what a curated set of cases cannot supply. --- ## The problem frame - **Current state:** programs weight the heritage-versus-novelty EDL trade on engineering intuition, tested against curated anecdotes, never against the full population with physical and program-strength controls. - **Desired state:** a defensible, reproducible, conditional estimate of whether heritage reuse changes landing-failure probability, with the confounder bounded and uncertainty stated. - **Gap:** no such estimate exists. - **Consequence:** the trade is decided by intuition, and "heritage" may be silently rewarded as a proxy for unmeasured program strength. --- ## The gap: three literatures that never meet - **Reconstruction reports** document one mission's EDL in fine detail. They estimate no cross-mission effect. - **Forward architecture studies** reason about a hypothetical future fleet (high-mass Mars, supersonic retropropulsion, inflatable decelerators). They observe no historical effect. - **Spacecraft reliability statistics** model on-orbit longevity of operating satellites, not the discrete success or failure of a landing, and isolate no EDL-novelty regressor. No study estimates a landing-failure hazard on a constructed EDL-heritage index across Moon, Mars, and Titan while controlling for target and mass. That empty intersection is the contribution. --- ## Theoretical anchor: Fogel - An innovation's value is defined only relative to its **next-best substitute**; you must specify and quantify the world without it. - The decision-relevant quantity is not "do heritage missions succeed?" (they mostly do) but "do they succeed **more** than the same mission would have under a novel architecture at the same target and mass?" - `beta_1` is the discrete-outcome analogue of Fogel's **social saving**. Conditioning on controls builds the counterfactual into the design. - The social-savings tradition gives a prior: the unspecified-substitute error overstates value, so the naive heritage advantage should shrink once the counterfactual is built in. --- ## Theoretical anchor: Mokyr and Perez - **Propositional** knowledge (why a technique works) versus **prescriptive** knowledge (the recipe). - Heritage may protect not because hardware is old but because flight converts prescriptive recipes into propositional understanding; failure modes become understood, not merely avoided. - The NTRS reconstruction reports are the literal artifact of that conversion, which grounds the **reconstruction-depth weighting** of the index. - Prediction: protective effect concentrates in deeply reconstructed elements; novelty grounded in strong analysis (supersonic retropropulsion) carries less excess risk than ungrounded novelty. This is the **novelty decomposition**. --- ## How the two anchors jointly fix the design - **Fogel fixes what to estimate and how to identify it:** a conditional, discrete-outcome logistic hazard on the heritage index with physical and program-strength controls; `beta_1` is the counterfactual contrast. - **Mokyr / Perez fix what the estimate means and where it should concentrate:** reconstruction-depth weighting and the grounded-versus-ungrounded novelty decomposition. - Neither anchor alone yields the full design. Every term in the equation is required by one anchor or the other. --- ## Data: four named real sources - **NTRS EDL reconstruction reports** define each architecture's elements and measured in-flight behavior. - **Global record of Moon, Mars, Titan landing attempts** is the population frame; one row per attempt. - **TechPort EDL-technology TRL records** date when each EDL element first became flight-proven, and in what regime. - **GAO program-history reports** supply cost, schedule, and organizational context for the program-strength control. TechPort and GAO are a **data substrate** (accessed as data via public APIs and archives), not bibliography. The NASA Technology Taxonomy supplies the TRL-classification framework. --- ## Unit, outcome, and the heritage index - **Unit:** the individual landing attempt at the Moon, Mars, or Titan. Orbital insertions, flybys, and Earth-return entries excluded. - **Outcome `failure`:** binary; loss of primary surface function = failure; survivable-but-degraded = success. Boundary cases (SLIM) recoded both ways. - **`heritage_index` in [0,1]:** criticality-weighted mean over six elements (aeroshell/TPS, entry guidance, supersonic deceleration, terminal descent/propulsion, terminal guidance/hazard avoidance, touchdown), each scored **against the regime in which it was proven**, optionally weighted by reconstruction depth. - Constructed from documents, not expert opinion. Reproducible from named sources. --- ## Controls and the central confounder - **Physical controls:** target body (Moon / Mars / Titan), entry mass, landed mass. Mass is not a nuisance covariate; it mechanically determines how much novelty an attempt is forced to carry. - **`program_strength`:** a deliberately coarse ordinal index from GAO cost-and-schedule data plus organizational EDL flight experience. - The central confounder is that well-funded, experienced programs both reuse heritage and execute better. The program-strength control is the main-specification guard against it. --- ## Estimator and identification - **Primary estimator:** Firth-penalized (bias-reduced) logistic regression. The frame is small with rare-event structure and quasi-separation is live; ordinary maximum likelihood is unsafe. - **Inference:** exact or permutation-based on `beta_1`, stratified by target body; asymptotic Wald intervals are untrustworthy at this size. - **Identification:** `beta_1` is the within-stratum heritage-versus-novel contrast at matched target, mass, and program strength. This is the Fogelian counterfactual made operational. - Reverse causation (novelty forced onto harder missions) attenuates `beta_1` toward zero, so a rejection of H0 is **conservative**. --- ## Threats to validity - **Internal:** confounding by program strength (with/without-control bounding); reverse causation (biases against H1); coding endogeneity (blind, pre-registered, inter-coder coding). - **External:** Mars-and-Moon dominance; the single Titan row probes but cannot establish generalization; crewed-class masses out of frame. - **Construct:** the heritage index is a proxy; defended by element-wise document-based, regime-aware coding and inter-coder reliability. - **Statistical-conclusion:** small sample; Firth penalization, exact inference, a single pre-registered specification, and a reported minimum detectable effect. --- ## The pre-registered analysis plan (seven steps) 1. Assemble and **freeze** the population frame. 2. Code the six EDL elements; second coder on a random subsample. 3. Code outcomes **blind** to the heritage coding. 4. Run pre-estimation diagnostics (reliability, separation, collinearity, events-per-variable). 5. Fit the primary Firth-penalized logistic with the three physical controls and program strength. 6. Report `beta_1`, its exact/permutation interval, and the IQR-implied failure-probability change. 7. Run the full robustness battery and the minimum-detectable-effect / power assessment. The order is load-bearing: each step commits a quantity later steps cannot renegotiate. --- ## The fixed decision rule Reject H0 in favor of H1 **if and only if**, in the primary specification: 1. `beta_1` is below zero; **and** 2. its exact or permutation 95% interval excludes zero; **and** 3. the sign is stable across the entire robustness set (cloglog link; drop low-documentation rows; recode boundary partial-successes both ways; with and without program strength; Mokyr novelty decomposition). An interval including zero fails to reject H0. A `beta_1` negative without the control but collapsing with it is reported as **confounded**. Inseparability of heritage from program strength **falsifies** the independent-value claim. --- ## Expected results (design-stage, not executed) **No coefficient is fitted on the full population. Every number below is a reporting-format illustration, not an estimate.** - Under H1: a negative `beta_1` whose exact interval excludes zero. - Illustrative reporting form only: "a move from the lowest to the highest heritage-index quartile is associated with a drop in modeled landing-failure probability from roughly forty to roughly fifteen percent at fixed Mars mass." This is a placeholder showing the output shape. - Mokyr decomposition (sub-implication): if the effect is knowledge-based, excess risk loads on **ungrounded** novelty, not on analytically matured novelty such as supersonic retropropulsion. --- ## Confidence and uncertainty - Confidence in the **design** is high: the estimator, controls, dataset, and falsification conditions are fixed in advance and strict. - Confidence in any **particular empirical pattern** is, by construction, suspended; no coefficient has been fitted. - The binding constraint is **power**: a frame of several dozen attempts may not distinguish a portfolio-material effect from zero. - The minimum detectable effect is reported alongside every non-rejection, so a null is never read as more decisive than power supports. A credibly bounded non-answer is an acceptable, honest deliverable. --- ## What a non-confirming result would mean - **Clean null:** heritage carries no independent protective effect once physics and program strength are matched. Redirect investment to engineering reserves and verification rigor (the Fogelian conclusion). - **Confounded:** the apparent heritage effect dissolves with the program-strength control. The trade should be reframed around program strength, which is doing the real work. This is the most probable single outcome on prior reasoning. - **Contrary (`beta_1 > 0`):** heritage flown outside its proven regime, or inherited software defects, import rather than retire risk. - **Inseparable:** heritage and program strength are non-identified; the intuition is shown untestable on the present record. All four are reported with equal prominence. --- ## How the argument holds together **Central goal:** a single pre-registered, conditional, confounding-adjusted `beta_1` is the correct, falsifiable, decision-relevant way to test whether heritage reuse lowers a landing's failure hazard, sound whichever way the coefficient falls. | Element of the argument | Developed in | |-------------------------|--------------| | The problem is real | Ch1, Ch3 (Vikram, SLIM, the heritage spectrum) | | The problem is material | Ch1, Ch3, Ch7 (recurring portfolio trade; mass scaling) | | The design addresses the mechanism | Ch2, Ch4, Ch5 (Fogel counterfactual; conditional logistic) | | It beats the alternatives | Ch5 (Firth dominates ML; cloglog link check) | | Residual risk is acceptable | Ch4, Ch5, Ch6 (pre-registration; exact inference; sensitivity) | --- ## Assumptions and residual risk - **Assumptions:** the record is a codeable population of binary trials; heritage attaches to elements-in-regimes; the coarse program-strength index bounds the confounder; conditional independence holds after conditioning; Firth + exact inference is valid on the small frame. - **Scope:** observational, design-stage measurement study; architecture-traceability vocabulary is out of scope. `beta_1` is, in plain prose, an input a program weighs against the cost of qualifying a novel element. No architecture table is populated. - **Residual risk (accepted):** limited power; an unobserved confounder uncorrelated with program strength; documentation asymmetry; analyst-coded construct risk. Each is named and bounded, not eliminated. --- ## Mechanism, not correlation ``` Reuse of a flight-proven EDL lineage (driver) -> flight exposed, reconstructed, and codified the failure modes into propositional knowledge, and the element runs inside its proven regime (mechanism) -> fewer in-flight-only-discoverable failure modes during the irreversible EDL event -> a measurable beta_1 < 0 in the conditional hazard (operational consequence) -> an evidence-based qualification bar for novelty (strategic implication) ``` `beta_1` is identified off a conditional within-stratum contrast. Where heritage proves inseparable from program strength, the design downgrades to correlation and says so. --- ## Stakeholder implications - **JPL EDL architects:** under H1, a novel element's excess risk acquires a price and a budgetable qualification investment; under the null, a heritage label cannot substitute for verifying an element runs inside its proven regime. - **NASA portfolio decision-makers:** under H1, reconstruction depth is an investment with measurable downstream return; under confounding, dollars go to engineering reserves and verification rigor. - **Commercial lander providers:** a number to negotiate against under H1; a caution against penalizing missing lineage when program strength is the operative factor under confounding. --- ## Contribution, restated - One conditional coefficient, or one credible failure to find one, both decision-relevant. The disjunction is the deliverable. - A permanent bridge between three literatures that have never been joined: reconstruction lineage, the novelty taxonomy, and the discrete-outcome estimation precedent, fused into one hazard model. - Fogel makes `beta_1` a measured counterfactual contrast; Mokyr tells the field whether the operative mechanism is hardware age or codified knowledge. - The instrument outlives the first estimate: a frozen frame, a documented rubric, a pre-registered estimator, and a reproducible coding protocol that grow with the record. --- ## From design to execution - Freeze the frame from the four named sources. - Code the index (second coder on a subsample); code outcomes blind; build the program-strength index. - Fit the primary model; run the full robustness set; report the minimum detectable effect. - Extensions: Earth-entry robustness, outer-planet / ocean-world external validity, and a more granular propositional-knowledge coding of novel-element maturation. --- ## Defense questions anticipated - How do you separate heritage from program strength when the two are nearly collinear? - Is a sample of a few dozen attempts adequate, and what is the minimum detectable effect? - How are partial successes (SLIM's anomalous attitude) coded, and how sensitive are results to that rule? - Does the single Titan point do real work, or is the claim effectively Mars-and-Moon only? - Could novelty be endogenous to mission difficulty, biasing the coefficient toward zero? --- ## References - Full list: 149 verified entries in the dissertation back matter (92 grade-A peer-reviewed, 57 grade-B report/preprint; 99 with a registered DOI, all with a resolvable URL). - Methodological anchors: Fogel, *Railroads and American Economic Growth* (1964) [121]; Leunig, "Social Savings" (2010) [133]; Perez, "Technological revolutions and techno-economic paradigms" (2010) [139], applied alongside Mokyr's *The Gifts of Athena* (2002) and *A Culture of Growth* (2017). - EDL spine: MSL overview [88]; Mars 2020 overview [74]; Phoenix performance [48]; InSight reconstruction [82]; Tianwen-1 [144]; SLIM [109]; Vikram loss [64]. - Estimator: Puhr et al. (2017) [52]; Heinze logistf [68]; Shen and Gao [8]; Grile and Bettinger reliability [127]. --- ## Closing The dissertation does not need the coefficient to be negative to be worth defending. It needs the coefficient **estimated against the whole population with controls, pre-registered, and falsifiable**, so that whichever branch the data select, NASA and JPL gain a defensible EDL portfolio parameter where before they had only an intuition tested against anecdotes. The deliverable at this stage is the pre-registered, falsifiable design. The next step is executing the frozen plan on the assembled population and reporting the full robustness set, with null and contrary outcomes given equal prominence.