# Interrogation mind-map: JPL_AUTONOMY_EDL_03

Nodes: 113 | questions: 42 | grounded claims: 36 | gaps: 35

## Questions

- **[identification]** Within-rover autonomous-drive fraction is offered as the cleanest variation, but the ground team sets that fraction by reading the terrain. Exhibit the first stage: regress autonomous fraction on the full terrain block (terrain-class FE + a-priori covariates) and report the residual share of variance. If terrain explains most of it, what quasi-random residual variation identifies the effect, and can a placebo (e.g. a within-rover, within-terrain-class split of the residual fraction that should have no productivity effect under H0) show the residual is not terrain relabeled? (raised by angrist_pischke)
- **[measurement]** The bad-controls line puts a-priori terrain class on the eligible-control side and realized slip on the mediator side, yet the ground team picks the autonomous fraction by forecasting drivability from those same a-priori terrain inputs, making the autonomy choice a function of expected productivity. Show the terrain covariates are pre-determined relative to the autonomy decision and are not themselves an outcome forecast: point to a terrain characterization timestamped before the drive plan, and show that conditioning on it does not absorb part of the autonomy effect, by reporting how the autonomous-fraction coefficient moves as the terrain block is added vs omitted. (raised by angrist_pischke)
- **[empirics]** Wild-cluster bootstrap is committed for the between-rover generation contrast, but that variation has exactly three clusters (MER, Curiosity, Perseverance), and the wild-cluster bootstrap has known severe size distortion below ~5 clusters. Run a small Monte Carlo on the candidate's own panel structure to show the actual rejection rate at nominal 5% with three between-rover clusters; if badly oversized, concede the between-rover generation coefficient is uninterpretable as a hypothesis test and only the within-rover autonomous-fraction estimate carries inferential weight. (raised by angrist_pischke)
- **[identification]** Terrain fixed effects condition out terrain suitability but leave the time-pressure selector (sols-to-deadline, conjunction blackout, end-of-campaign sprint) in the operator-chosen autonomous fraction. Name the PDS/NTRS observable that proxies mission time-pressure at the drive-sol, put it on the RHS, and show the autonomous-fraction coefficient survives; or defend exogeneity once only terrain, not schedule, is conditioned out. (raised by angrist_pischke)
- **[measurement]** Across three rovers, autonomy generation and hardware class are one-to-one collinear, so the partition of explained variance between the autonomy block and the hardware block is not unique and depends on entry order. Report the variance-share decomposition under all orderings plus a Shapley/Owen attribution, state the share range, and if it straddles 50 percent, name the identified quantity (not an order-dependent R-squared) on which the H1-over-H0 verdict rests. (raised by angrist_pischke)
- **[empirics]** Construct a true negative control: on blind-commanded segments (autonomous fraction = 0 by construction, software not driving), regress per-sol productivity on the same within-rover regressors plus a placebo 'fraction' built from a variable that cannot act through onboard autonomy (downlink-window timing, operator shift), and show that coefficient is null. Will the candidate pre-commit to this test and to the threshold at which a non-null placebo sinks H1? (raised by angrist_pischke)
- **[identification]** The within-rover autonomous-drive fraction is a continuous, non-randomly assigned (ground-team-chosen) treatment. A TOT-type parameter is identifiable under parallel trends, but reading the slope across dose levels as the marginal meters-per-sol of autonomy is the cross-dose comparison that selection-into-dose biases. Do higher-autonomous-fraction drive-sols differ systematically in pre-drive terrain ruggedness, slope, and prior-sol slip after terrain-class fixed effects and covariates? Show the balance table on residualized autonomous fraction; if imbalance survives conditioning, the slope is a dose-selection artifact. (raised by callaway_santanna)
- **[empirics]** With heterogeneous autonomy effects across terrain classes that grow within a mission, and three generations switching on at different calendar sols across rovers, a single pooled coefficient is a contaminated weighted average that can carry negative weights and flip sign. Will the candidate commit to estimating and reporting generation-by-terrain-by-mission-phase building blocks (the analogue of ATT(g,t)) aggregated with declared weights, and has the Goodman-Bacon decomposition of the between-rover TWFE specification been computed to reveal whether negative-weight comparisons (a later generation using an earlier-generation rover-segment as control) enter the headline number? (raised by callaway_santanna)
- **[identification]** The between-rover generation contrast has three full panel members and no never-treated unit (every rover eventually runs onboard autonomy), and each landing site presents non-random terrain. Identification then rests on covariate-conditional parallel trends across rovers that drove different ground, defended by a clean control group. Name the control: against which not-yet-treated rover-segments is each generation's effect identified, and show that no segment already operating under a later autonomy generation enters another generation's control comparison. Then report the pre-treatment placebo within a shared terrain class (do segments later receiving higher-generation autonomy show parallel meters-per-sol trends before the generation difference takes effect?), with wild-cluster-bootstrap inference reported jointly with a sensitivity analysis for plausible parallel-trends violations rather than a single flat-pretrend claim. (raised by callaway_santanna)
- **[identification]** Sec 4.1 uses TWFE on a continuous autonomous-drive-fraction dose but never invokes a doubly-robust / covariate-conditional design. Because the ground team chooses a higher autonomous fraction when it judges terrain suitable and time pressure high, only conditional (not unconditional) parallel trends is defensible. Re-specify as a doubly-robust ATT(g,t) modeling both the outcome and the propensity of the autonomous-fraction choice on a-priori terrain covariates, report whether the H1 coefficient is consistent under either working model alone, and show the propensity-of-autonomous-fraction surface on PDS/NTRS so the panel can judge covariate overlap across the dose range. (raised by callaway_santanna)
- **[economics]** The continuous autonomous-drive-fraction dose is the case Callaway, Goodman-Bacon & Sant'Anna (2024) warn about: treatment-on-the-treated is identified under parallel trends, but comparing productivity ACROSS dose levels does not follow, because units selecting a higher fraction differ systematically from those selecting lower. Yet H1 reads the autonomous-fraction slope as the marginal productivity of more autonomy, a cross-dose claim. From PDS/NTRS, partition drive-sols into a dose-response curve (productivity by binned autonomous fraction, within terrain class) and show whether it is identified by within-unit dose changes on the SAME terrain rather than by comparing high-dose to systematically different low-dose drives. State the stronger no-selection-into-dose assumption that licenses the cross-dose reading, or restrict H1 to treatment-on-the-treated at observed doses. (raised by callaway_santanna)
- **[empirics]** No-anticipation is a maintained assumption the design never states. The autonomous-drive fraction is not a fixed per-rover dose: the ground team ramps it UP within a mission as it accumulates trust after early autonomous drives succeed, making the dose a function of the realized outcome path, an anticipation/feedback violation distinct from the bad-controls rule already addressed. Using NTRS performance reports and the PDS sol-indexed traverse record, plot the autonomous fraction against mission sol within each rover and test whether the dose is predicted by lagged realized productivity. If the dose responds to past meters-per-sol, the within-rover slope is contaminated by reverse causation; a flat pre-dose placebo (productivity before the fraction was raised) is the check that would clear it. (raised by callaway_santanna)
- **[economics]** Will you add a measured certification-cost variable (V&V campaign, testbed re-qualification, mission-assurance review, operational risk of an in-flight autonomy upload) to the right-hand side of your software-vs-hardware cost comparison, drawn from the same flight-software-update records you cite, and state whether your central prescription survives once the institutional cost of a certified in-flight autonomy upgrade is priced against launch-mass cost? (raised by gangale)
- **[measurement]** State the bright line: give the measurable, mission-independent threshold, drawn from NTRS performance reports and not from mission self-description, that separates autonomy generation G1 from G2 from G3. If the boundary is functional (generation follows from how software was described rather than from a measured capability discontinuity), show the drive segment the classification cannot adjudicate. (raised by gangale)
- **[measurement]** Specify the temporal reference frame and conversion that makes meters-per-sol commensurable across three rovers at three landing sites with differing local solar time, mission-clock conventions, and downlink-window cadence; identify the operational error introduced if a Perseverance autonomous-drive sol and an Opportunity blind-commanded sol are pooled without correcting for the planning-cycle/downlink structure that gates how much autonomy was possible per sol. Can PDS traverse archives and NTRS operations records settle whether the sol denominator is reference-frame-clean or absorbs an autonomy-capability vs operations-cadence confound? (raised by gangale)
- **[measurement]** Specify the exact TechPort/NTRS archive field and the numeric threshold on it (a TRL step, a documented autonomous-drive-fraction band, or a stated planning-horizon/cycle-time figure) at which a rover crosses from G1 to G2 to G3, such that a blind analyst with only the archive and no mission press kit reproduces the same three generation boundaries. (raised by gangale)
- **[measurement]** Name the field in the PDS traverse-and-localization product that records available-drive-time-within-sol (or the energy-bounded drive envelope) for each drive-sol, and show that meters-per-sol is normalized to that envelope rather than to a raw sol whose usable drive window silently varies with rover, season, latitude, and downlink cadence. (raised by gangale)
- **[economics]** Across the NTRS/TechPort flight-software-update history for MER, Curiosity, and Perseverance, was any measured per-sol mobility-productivity gain actually delivered to a landed rover by post-landing autonomy upload, with before-and-after meters-per-sol on the same machine and terrain class, or is every generational step co-launched with new hardware so the retrofit window was foreclosed at launch? (raised by gangale)
- **[identification]** Pre-register the quantitative cross-platform transfer test: which off-Mars dataset (lunar VIPER/Chang'e traverse archive, terrestrial off-road benchmark, or a different-flight-processor Mars rover) you re-estimate on, and what magnitude of cross-platform coefficient stability you require, before the 'fleet-wide investment doctrine' is warranted. (raised by goswami_garretson)
- **[mechanism]** Name the observable advance commitment that would falsify the Mokyr extensibility claim itself: what record of post-landing flight-software autonomy upgrades delivering measured per-sol productivity gains on an already-deployed rover would have to be ABSENT from NTRS / mission-update record for you to reject extensibility? (raised by goswami_garretson)
- **[economics]** State the two cost curves that make the 'buy productivity with software, not mass' doctrine actionable: the marginal development-plus-verification cost of one additional unit of autonomous-drive-fraction productivity vs the marginal launch-cost of the equivalent mechanical increment, and the break-even where the software lever stops dominating. Without them, is 'invest in software' an economic doctrine or a relabeled variance-decomposition? (raised by goswami_garretson)
- **[identification]** Construct at least one budget/mission-class control from public program records (TechPort program cost, RAD750-vs-newer compute class, X-band/UHF downlink allocation per sol, or total instrument mass), enter it in the between-rover specification, and state whether the autonomy-generation coefficient survives. If autonomy generation cannot be separated from the mission-class budget it rode in on, the G1/G2/G3 contrast is a budget index wearing an autonomy label. (raised by goswami_garretson)
- **[economics]** State the two marginal cost curves (dev+V&V cost per unit of autonomy-attributable meters/sol vs launch+integration cost per unit of mass/compute-attributable meters/sol) and the real break-even in dollars-per-meter-per-sol where they cross, sourced from TechPort cost records or mission cost-estimating relationships. Absent that crossing, 'invest in software' is a within-budget variance decomposition, not an allocation rule. (raised by goswami_garretson)
- **[rival]** State the observable, advance-specified discriminator (from PDS/NTRS/TechPort, not mission self-description) that separates a genuine autonomy effect from funded-mission-class selection: e.g. a single mission where flight-software autonomy was upgraded post-landing under a flat/declining budget, delivering a measured per-sol step with mass, compute, and downlink held constant. If no such within-budget, within-platform autonomy increment exists in the NTRS record, what observation in the assembled panel could falsify the rival that autonomy is a proxy for which missions were funded to be ambitious? (raised by goswami_garretson)
- **[identification]** The additive specification (Productivity = b1*AutonomyGen + b2*Hardware + g*Terrain) imposes separability, forcing the autonomy contribution to be constant across terrain roughness, yet 5.2 predicts the autonomy effect is largest in rougher terrain. A feedback view says capability and terrain are coupled. Can the PDS drive-sol panel settle whether the autonomy-by-terrain interaction is large enough that the additive horse race is mis-specified, and on what grounds do you report the additive coefficient as 'the contribution' rather than the interaction? (raised by meadows)
- **[measurement]** Meters-per-sol is a flow rate; the binding scarcity named in Section 6 is the sol, spent inside a planning-cycle loop where today's trusted autonomy lets the ground team write sparser commands tomorrow, raising tomorrow's autonomous fraction. The candidate concedes meters-per-sol 'cannot see' that loop and would 'understate' the autonomy channel. Is the dependent variable measuring the governing stock (ground-team planning effort / trust accumulated per meter) or only an instantaneous downstream flow, and could TechPort/NTRS records build a coarse proxy for planning-cycle effort per meter? (raised by meadows)
- **[governance]** On the leverage ladder this is a parameter-level question ('buy productivity with mass or with software'), yet the Discussion locates the real lever one tier up -- at the goal the mission optimizes (sols freed for sampling/caching in MSR) and at the planning-cycle redesign a trusted-autonomy paradigm permits. If the autonomy channel's durable payoff is the goal-level reorganization of how missions are commanded rather than marginal meters on a given drive, does the between-channel coefficient inform the decision a future-rover designer faces, and what PDS/NTRS observable would adjudicate whether the high-leverage gain is the commanding-paradigm shift rather than the drive-distance scalar? (raised by meadows)
- **[economics]** State the doctrine's loss function explicitly: if 'buy productivity with software' is adopted fleet-wide and an over-trusted drive or certification gap destroys a deployed asset, WHAT stock is lost (vehicle, caching campaign, irreplaceable MSR-window sols) and over what horizon does the loss integrate? Show the loss as an integral over the binding stock, not a point estimate of forgone rate. (raised by meadows)
- **[empirics]** Will you integrate the downside loop from the same PDS/NTRS archives you use for the upside, reporting recovery-sols-lost and autonomy-attributable aborts/safe-mode entries per unit of autonomous-fraction gained alongside meters-per-sol gained, so the doctrine is scored NET rather than gross? (raised by meadows)
- **[governance]** Declare the tier of the intervention: is 'optimize for trusted autonomy fleet-wide' a bounded-downside parameter tweak or a goal-level reorganization whose tail risk is unbudgeted? Does it erode the balancing stock that makes autonomy safe (ground-team commanding skill and the conservative reserve margin), and is there a measurable operational-record proxy (rising autonomous-fraction with falling ground-intervention frequency, or thinning commanded margin over sols) that would detect the reinforcing loop drawing that stock down before catastrophe? (raised by meadows)
- **[measurement]** Name the independent observable that distinguishes a genuinely widening propositional base (Omega) in onboard perception/planning from three relabeled increments of the same trial-and-error AutoNav technique re-badged G1/G2/G3. Which variable, if flat, falsifies the Mokyr reading even while H1's autonomy-generation coefficient survives? If none exists in your sources, concede the lens is decorative. (raised by mokyr)
- **[economics]** The retrofittability claim (software gains can in principle be uploaded to a landed rover while a wheel cannot) is asserted, not estimated. What in flight-software version histories and NTRS deployment records would let you MEASURE the realized access cost of getting an autonomy increment onto a deployed rover (V&V and uplink lag, and the fraction of autonomy improvements actually retrofitted vs frozen at next launch)? If most autonomy gains shipped only at launch, the asymmetry collapses. (raised by mokyr)
- **[mechanism]** The within-rover autonomous-drive-fraction contrast measures only how much of a drive the existing technique was trusted to run (an operational trust/utilization margin), not whether the underlying knowledge base widened. What evidence separates 'the propositional base widened, making the technique extensible' (Mokyr's mechanism) from 'the same fixed technique was utilized more aggressively as ground teams gained operational confidence' (mere prescriptive habituation)? Absent that separation, why is the headline result not equally consistent with the trial-and-error stagnation case the lens is meant to rule out? (raised by mokyr)
- **[rival]** Operationalize the 'software is retrofittable' premise as a measured count from real flight-software version histories: across MER, Curiosity, and Perseverance, how many post-landing flight-software builds materially changed the AutoNav/ENav driving policy, on what sol was each uplinked, and what was the attributable change in autonomous-drive fraction or meters-per-sol? Until that count exists, the load-bearing asymmetry is asserted, not estimated. (raised by mokyr)
- **[governance]** If the post-landing retrofit count is near zero, name the discriminator in mission-assurance review records and flight-software change-control logs that separates Explanation A (autonomy is technically un-retrofittable: processor/perception/V&V-testbed cannot accommodate a better planner post-launch) from Explanation B (autonomy is institutionally un-retrofittable: governance assigns the entire downside of a bad uplink to the signer, so risk-averse authorizers freeze the policy regardless of feasibility). Who holds sign-off authority, what abort/safing-risk gate a driving-policy change must clear, and were rejected/never-proposed upgrades blocked on a technical-feasibility finding or a risk-acceptance/authority finding? (raised by mokyr)
- **[economics]** Will you add a measured certification-and-authorization cost for the retrofit path specifically (V&V re-qualification campaign, testbed re-run, mission-assurance review burden, and an expected-loss term from the residual probability that an authorized uplink degrades or loses a landed asset, drawn from actual review records and documented uplink anomalies) into the software-versus-mass ledger? Absent that term you compare the full launch-mass cost of the mechanical increment against only the development cost of the software increment, leaving the institutional access cost of the retrofit out of the ledger. (raised by mokyr)
- **[identification]** The autonomous-drive fraction is a ground-team reliance choice (Sec 4.2), so it measures calibrated reliance as much as machine competence. Specify an exogenous shifter of the autonomous fraction unrelated to per-drive hazard competence (downlink window, staffing, Earth-Mars geometry, an operations-policy autonomy-posture directive) and show the H1 coefficient survives instrumenting on it; absent that, a positive coefficient is observationally identical to the ground team learning to trust an unchanged machine. (raised by parasuraman)
- **[mechanism]** Complacency theory predicts that as autonomy becomes more reliable the ground team reallocates attention and commands longer, more sparsely specified, bolder drives (Sec 6 anticipates freed planning-cycle time), inflating meters-per-sol independently of per-drive hazard handling. Is autonomous-fraction separable in PDS/NTRS from the commanded drive-distance ceiling and route conservatism, so the coefficient reflects the machine clearing hazards rather than the team commanding bolder drives? If not, on what evidence is the gain assigned to the software channel rather than evolving operator reliance? (raised by parasuraman)
- **[measurement]** Meters-per-sol counts only successful traverse and is silent on the commission-error tail (autonomous clearance of hazards that warranted stopping, fault/safing/abort sols). A trust-calibration account predicts the coefficient shrinks once autonomy-attributable fault and recovery sols are netted out; a pure-competence account predicts it does not. Construct a competence-net measure from PDS fault/event records (meters per sol after charging autonomy-attributable faults, anomalies, aborts) and show H1 is robust to it. (raised by parasuraman)
- **[measurement]** Your autonomous-drive fraction is the ground team's revealed reliance decision, not a measure of machine competence. If it rises as a sol-indexed trust-accrual curve as false-alarm/intervention rates fall, the H1 coefficient loads operator reliance learning. Construct the within-rover time path of the autonomous fraction against cumulative sols and against the running count of autonomy-attributable aborts/faults/ground-overrides, and show the autonomy effect survives conditioning on this trust-accrual trajectory. (raised by parasuraman)
- **[mechanism]** You declare 'autonomy generation' but never decompose it into the Parasuraman-Sheridan-Wickens processing stages. AutoNav vs Enhanced Navigation differ chiefly in action-implementation timing (think-while-driving) versus a smarter analysis or decision stage. Specify which of the four stages, information acquisition, information analysis, decision/action selection, action implementation, each generation moved and to what level, and show the productivity gain attaches to the stage you claim; a single continuous 'autonomous fraction' that mixes a stage-4 timing change with stage-2/3 capability is not interpretable as 'how much the machine is doing'. (raised by parasuraman)
- **[identification]** Your model is additive and treats the human planner as a constant (meters-per-sol per drive, no operator-state term), yet the binding scarcity is the daily planning cycle and a calibrated team writes sparser, less conservative commands as workload/vigilance allow, the cry-wolf/disuse versus over-trust dynamic. Build, from PDS command logs and NTRS operations reports, a measure of commanded-drive conservatism (command sparsity, route-margin, blind-segment insertion after an anomaly) and show the autonomous-fraction coefficient is not absorbing a time-varying ground-team conservatism response, i.e. that the machine-capability effect is separable from the operators' shifting reliance posture after false alarms and aborts. (raised by parasuraman)

## Grounded claims

- **[identification]** The challenge is methodologically valid and follows directly from Angrist-Pischke design diagnostics: when treatment assignment is a deliberate function of the same covariates used for adjustment, the analyst's first obligation is to name the source of exogenous (residual) variation, and an absorbed control can leave too little identifying variation to support a causal claim. A first-stage decomposition of the autonomous fraction on the terrain block plus a within-stratum placebo is the correct exhibit to demand. However, the specific numbers this requires (residual R-squared / residual variance share, and any placebo split estimate) cannot be sourced: no retrieved corpus (AMOS, ACTA, Space Economy) contains the candidate's JPL rover panel or any rover-autonomy productivity dataset, and the AMOS SSA/SDA corpus returns zero hits on causal-inference / treatment-effect queries. The standard is grounded; the candidate's empirical result is not in evidence and must be produced by the candidate.
    - Angrist & Pischke, 'The Credibility Revolution in Empirical Economics', J. Economic Perspectives 24(2) | https://doi.org/10.1257/jep.24.2.3 | grade A
    - Angrist & Pischke, Mostly Harmless Econometrics: An Empiricist's Companion | https://doi.org/10.1515/9781400829828 | grade A
- **[measurement]** The bad-controls concern is correctly framed: a control that is a function of the outcome (or of expected outcome) reintroduces bias, and a pre-determined control must be timestamped before treatment to qualify as such. Angrist-Pischke explicitly flag conditioning on post-treatment or outcome-anticipating variables as the failure mode, and the coefficient-stability exhibit (movement of the treatment coefficient as the suspect block is added vs omitted) is the right diagnostic to demand. The methodological standard and the demand are fully grounded. The data-provenance fact the question asks for (an actual timestamp on the terrain characterization relative to the drive plan) and the coefficient-stability numbers are candidate-/mission-data facts that no retrieved corpus supplies; they cannot be asserted from evidence here.
    - Angrist & Pischke, Mastering 'Metrics: The Path from Cause to Effect (as recorded in the angrist_pischke design-diagnostics dossier) | https://doi.org/10.1515/9781400829828 | grade A
- **[empirics]** The premise is well supported by the cluster-robust inference methods literature: valid cluster-robust and wild-cluster-bootstrap inference degrades sharply when the number of clusters is very small, and the wild bootstrap with few clusters is a recognized special case requiring dedicated treatment (Cameron-Gelbach-Miller; MacKinnon-Webb). With G=3 between-rover clusters, the default wild-cluster-bootstrap p-values are not credible, and the standard remedy (a design-specific Monte Carlo to report the true rejection rate, then retreating to the within-rover estimand if size is distorted) is the appropriate demand. The methodological challenge is grounded. The candidate's own Monte Carlo rejection rate at nominal 5% is not in any retrieved source and must be produced by the candidate; it cannot be asserted here.
    - Cameron, Gelbach & Miller, 'Bootstrap-Based Improvements for Inference with Clustered Errors' | https://doi.org/10.2139/ssrn.956890 | grade A
    - MacKinnon & Webb, 'The wild bootstrap for few (treated) clusters', The Econometrics Journal 20(2) | https://doi.org/10.1111/ectj.12107 | grade A
- **[identification]** Methodologically the objection is the canonical omitted-variable / selection-on-observables problem: if the autonomous fraction is chosen on BOTH terrain AND a schedule selector, and the schedule selector is itself correlated with meters-per-sol, then conditioning on terrain alone does not deliver conditional independence. The selection-on-observables (CIA) defense requires that ALL selectors that also move the outcome be on the RHS; a defensible CIA claim must name and include the time-pressure variable, not assume it away. This is the inferential standard the candidate must meet, per Angrist & Pischke.
    - Angrist & Pischke, Mastering 'Metrics: The Path from Cause to Effect (2014); Mostly Harmless Econometrics (2009) | https://doi.org/10.1515/9781400829828 | grade A
- **[measurement]** Methodologically the objection is sound on its face: under perfect (or near-perfect) collinearity between two regressor blocks, incremental-R-squared / sequential-sums-of-squares attribution is order-dependent and the between-block partition is not identified; a horse-race verdict that rests on 'which block absorbs the larger incremental share' is therefore an artifact of covariate entry order, not an identified quantity. This is a direct instance of the design-econometrics principle that explained-variance shares are not causal estimands and that a credible verdict must rest on an identified contrast (an experiment-like comparison), not on a partitioned R-squared.
    - Angrist & Pischke, Mostly Harmless Econometrics: An Empiricist's Companion (2009) | https://doi.org/10.1515/9781400829828 | grade A
- **[empirics]** Methodologically the demand is the standard placebo / negative-control falsification: a credible within-rover design should produce a null coefficient on a treatment-like variable on a margin where the hypothesized channel provably cannot operate; a non-null placebo indicates the within-rover contrast is capturing whatever else co-moves with the team's decision to drive autonomously (selection into autonomy), not the software channel. Pre-committing the test and a rejection threshold is the appropriate falsification discipline, consistent with the candidate's existing falsification checks being insufficient because they are all internal to H1 rather than negative controls.
    - Angrist & Pischke, Mostly Harmless Econometrics (2009), ch. on regression and bad controls; Mastering 'Metrics (2014) | https://doi.org/10.1515/9781400829828 | grade A
- **[identification]** This is the exact pathology Callaway, Goodman-Bacon & Sant'Anna (2024) formalize for continuous treatment: treatment-on-the-treated parameters are identified under a parallel-trends assumption analogous to the binary case, but COMPARING parameters across dose levels does NOT follow from parallel trends, because units selecting into higher doses can differ systematically. So the candidate's preferred within-rover slope (marginal m/sol of autonomy) is precisely the dose-comparison estimand that selection-into-dose contaminates. The settling condition is a balance test: regress the autonomous fraction on terrain-class fixed effects + a priori terrain covariates (ruggedness, slope, prior-sol slip), then check whether the residualized dose is balanced against those covariates. If imbalance survives conditioning, the within-rover slope is a dose-selection artifact, not the autonomy effect, and at most a covariate-conditional TOT at observed doses is defensible (double robustness via Sant'Anna & Zhao buys consistency if either the dose-assignment or outcome model is correct). Candidate-domain corpora (AMOS, ACTA, Space Economy) return zero on continuous-treatment DiD, so no rover-specific precedent exists to lean on.
    - Callaway, Goodman-Bacon & Sant'Anna, 'Difference-in-Differences with a Continuous Treatment' (2024 working paper) | https://doi.org/10.2139/ssrn.4716682 | grade B
    - Sant'Anna & Zhao, 'Doubly robust difference-in-differences estimators,' Journal of Econometrics (2020) | https://doi.org/10.1016/j.jeconom.2020.06.003 | grade A
    - callaway_santanna reviewer-brain dossier (Hall of Shoulders), AMOS/ACTA zero-hit finding on causal-DiD terms | D:/Claude_Code/brain/collegium/hall_of_shoulders/brains/callaway_santanna/ | grade C
- **[empirics]** Goodman-Bacon (2021) proves the TWFE estimator equals a weighted average of all 2x2 DiD comparisons in the data, and under treatment-effect heterogeneity with staggered timing, comparisons that use already-treated units as controls can receive negative weights, making the pooled coefficient an uninterpretable, possibly sign-flipped blend. The candidate's own stated conditions (heterogeneity by terrain, within-mission growth, staggered generation switch-on) are exactly when this occurs. Callaway & Sant'Anna (2021) supply the disciplined replacement: define disaggregated ATT(g,t) building blocks, restrict controls to never-treated or not-yet-treated units so a unit's own past treatment never contaminates another's estimate, and aggregate transparently with researcher-chosen, defensible weights rather than regression-imposed opaque weights. The settling artifacts are (a) a disaggregated generation-by-terrain-by-mission-phase table and (b) a Goodman-Bacon decomposition exhibiting the weight on each comparison type and flagging any negative-weight 'forbidden comparison.' Acknowledging the literature without reporting the disaggregated estimand is insufficient.
    - Goodman-Bacon, 'Difference-in-differences with variation in treatment timing,' Journal of Econometrics (2021) | https://doi.org/10.1016/j.jeconom.2021.03.014 | grade A
    - Callaway & Sant'Anna, 'Difference-in-Differences with multiple time periods,' Journal of Econometrics (2021) | https://doi.org/10.1016/j.jeconom.2020.12.001 | grade A
- **[identification]** Callaway & Sant'Anna (2021) require the candidate to name the clean control group explicitly: with no never-treated unit, each generation's effect must be identified off not-yet-treated rover-segments as of each period, and the estimator must demonstrate that no already-treated (later-generation) segment enters any control comparison, since that is the contamination channel the method exists to close. Because the load-bearing assumption is covariate-conditional parallel trends (terrain, slope, slip differ non-randomly across landing sites), double robustness (Sant'Anna & Zhao 2020) gives consistency if either the generation-assignment propensity model or the meters-per-sol outcome model is correct. The pre-treatment placebo is ATT(g,t) on pre-period leads within a shared terrain class. Critically, Roth (2022) shows a flat pre-trend is NECESSARY BUT NOT SUFFICIENT: pre-trends tests have low power, and conditioning the analysis on having passed one can itself worsen bias. Therefore the candidate must report a Roth-style sensitivity analysis bounding the estimate under plausible parallel-trends violations (per the synthesis in Roth, Sant'Anna, Bilinski & Poe 2023), paired with wild-cluster-bootstrap inference appropriate to the small number of rover clusters, rather than a single point estimate plus a flat-pretrend plot.
    - Callaway & Sant'Anna, 'Difference-in-Differences with multiple time periods,' Journal of Econometrics (2021) | https://doi.org/10.1016/j.jeconom.2020.12.001 | grade A
    - Roth, 'Pretest with Caution: Event-Study Estimates after Testing for Parallel Trends,' American Economic Review: Insights (2022) | https://doi.org/10.1257/aeri.20210236 | grade A
    - Roth, Sant'Anna, Bilinski & Poe, 'What's trending in difference-in-differences? A synthesis of the recent econometrics literature,' Journal of Econometrics (2023) | https://doi.org/10.1016/j.jeconom.2023.03.008 | grade A
    - Sant'Anna & Zhao, 'Doubly robust difference-in-differences estimators,' Journal of Econometrics (2020) | https://doi.org/10.1016/j.jeconom.2020.06.003 | grade A
- **[identification]** The methodological re-specification is correct and answerable: Sant'Anna & Zhao (2020) give a doubly-robust ATT estimator that is consistent if EITHER the propensity-score model OR the outcome-regression working model is correctly specified (not necessarily both), and is the engine under the did/DRDID software. When dose selection depends on terrain, identification needs covariate-CONDITIONAL parallel trends, and double robustness buys consistency if either the propensity model (which terrain covariates drive the ground team's fraction choice) or the outcome model is right. So the candidate SHOULD re-specify the bare TWFE-on-a-dose as a conditional doubly-robust ATT(g,t) and report which working model carries identification. The candidate-specific deliverable the question also demands, the estimated propensity-of-autonomous-fraction surface on the actual PDS/NTRS panel and the resulting covariate-overlap diagnostic, is an empirical artifact not present in any retrieved source and must be produced by the candidate; see gap g1.
    - Sant'Anna & Zhao, Doubly robust difference-in-differences estimators, Journal of Econometrics (2020) | https://doi.org/10.1016/j.jeconom.2020.06.003 | grade A
    - callaway_santanna dossier (Hall of Shoulders brain), conditional parallel trends + double robustness prescription for terrain-driven dose selection | D:/Claude_Code/brain/collegium/hall_of_shoulders/brains/callaway_santanna/ | grade C
- **[economics]** The critique is exactly correct and grounded: Callaway, Goodman-Bacon & Sant'Anna (2024) show that for a continuous treatment, treatment-on-the-treated-type parameters are identified under an analogous parallel-trends assumption, but COMPARING these parameters across dose levels is challenging because parallel trends does not rule out selection bias, units that select into higher doses differ systematically. Therefore H1's reading of the autonomous-fraction slope as marginal productivity of more autonomy is a forbidden cross-dose comparison UNLESS the candidate either (a) identifies the dose-response from within-unit dose changes on the same terrain class, or (b) states the stronger no-selection-into-dose (selection-on-observables-into-intensity) assumption, or (c) retreats H1 to treatment-on-the-treated at observed doses. That triage is the defensible response. The within-terrain binned dose-response curve from the actual PDS/NTRS drive-sol records is a candidate-specific empirical product not contained in any retrieved source; see gap g2.
    - Callaway, Goodman-Bacon & Sant'Anna, Difference-in-Differences with a Continuous Treatment (2024 working paper) | https://doi.org/10.2139/ssrn.4716682 | grade B
    - callaway_santanna dossier (Hall of Shoulders brain), cross-dose comparison needs more than parallel trends; selection-into-dose generates bias | D:/Claude_Code/brain/collegium/hall_of_shoulders/brains/callaway_santanna/ | grade C
- **[empirics]** The no-anticipation / dynamic-feedback concern is a real and standard requirement, and the proposed checks are the right ones. Callaway & Sant'Anna identification rests on BOTH parallel trends (unconditional or covariate-conditional) AND a no-anticipation assumption, and pre-treatment ATT(g,t) placebo estimates are the prescribed diagnostic for it. A trust-ramped dose driven by realized success makes the fraction a function of the outcome path, which is exactly an anticipation/feedback channel; regressing the autonomous fraction on lagged meters-per-sol tests it, and a flat pre-dose placebo (productivity before the fraction was raised) is the clearing check. The candidate must ALSO heed Roth (2022): pre-trends/placebo tests can have low power, and conditioning the analysis on having passed them can itself worsen bias, so a flat pre-dose placebo is necessary but not sufficient and should be paired with sensitivity to plausible violations. The actual within-rover dose-vs-sol plot and the lagged-productivity regression on PDS/NTRS data are candidate-specific empirical artifacts absent from all retrieved sources; see gap g3.
    - Callaway & Sant'Anna, Difference-in-differences with multiple time periods, Journal of Econometrics (2021) | https://doi.org/10.1016/j.jeconom.2020.12.001 | grade A
    - Roth, Pretest with Caution: Event-Study Estimates after Testing for Parallel Trends, AER: Insights (2022) | https://doi.org/10.1257/aeri.20210236 | grade A
    - callaway_santanna dossier (Hall of Shoulders brain), no-anticipation + pre-treatment ATT(g,t) placebo; flat pre-trend necessary not sufficient | D:/Claude_Code/brain/collegium/hall_of_shoulders/brains/callaway_santanna/ | grade C
- **[economics]** The institutional cost of certifying an in-flight rover autonomy upgrade is real and measurable from the records already in scope, so it can be added as a right-hand-side variable rather than treated as free. The MER mobility flight-software record ('Driving Ambition') documents the surface-test and validation regime behind each mobility-software capability; the directed-vs-autonomous tradeoffs record measures the time/throughput tradeoff of running autonomy versus blind-commanded driving; the global-planner 'software integration and surface test' record documents the testbed/surface-test burden of fielding an in-flight capability upgrade; and the Verma et al. 2025 Enhanced Autonomous Navigation record documents the calendar/verification effort of a certified in-flight upgrade to Perseverance. These supply a measured certification-cost term. However, whether the candidate's CENTRAL PRESCRIPTION survives once that term is priced cannot be settled from the literature: it depends on the candidate's own cost model and its parameter values, which are internal to the dissertation and not in retrieval. The existence and measurability of the cost is grounded; the survival of the prescription is a gap the candidate must close with his own model.
    - Biesiadecki & Maimone, 'The Mars Exploration Rover Surface Mobility Flight Software: Driving Ambition,' IEEE Aerospace Conf. 2006 | https://doi.org/10.1109/aero.2006.1655723 | grade A
    - Biesiadecki et al., 'Tradeoffs Between Directed and Autonomous Driving on the Mars Exploration Rovers,' Int. J. Robotics Research, 2007 | https://doi.org/10.1177/0278364907073777 | grade A
    - 'Global planning on the Mars Exploration Rovers: Software integration and surface test,' J. Field Robotics, 2009 | https://doi.org/10.1002/rob.20287 | grade A
    - Verma et al., 'Enhanced Autonomous Navigation on the Perseverance Mars Rover,' IEEE Trans. Field Robotics, 2025 | https://doi.org/10.1109/tfr.2025.3636366 | grade A
- **[measurement]** A mission-independent capability axis on which a generation discontinuity could be defined exists in the literature and is measured per-sol: drive distance per sol and autonomous-vs-directed drive fraction are reported across MER, Curiosity, and Perseverance. That much is grounded. But retrieval returns NO published source that defines a G1/G2/G3 bright line as a measured discontinuity on that axis; the three-level generation label is the candidate's own construct and is not externally validated by any retrieved performance report. The operationalizable substrate (measured meters/sol and autonomy fraction) is real; the bright line itself is unsupported by retrieval and is therefore the burden of proof the candidate has not yet met. The measurement demand is half-answerable (the axis exists) and half-gap (no source adjudicates the boundary).
    - Biesiadecki et al., 'Tradeoffs Between Directed and Autonomous Driving on the Mars Exploration Rovers,' Int. J. Robotics Research, 2007 | https://doi.org/10.1177/0278364907073777 | grade A
    - Rankin et al., 'Driving Curiosity: Mars Rover Mobility Trends During the First Seven Years,' IEEE Aerospace Conf. 2020 (journal version J. Field Robotics 2021, doi 10.1002/rob.22011) | https://doi.org/10.1109/aero47225.2020.9172469 | grade A
    - Verma et al., 'Enhanced Autonomous Navigation on the Perseverance Mars Rover,' IEEE Trans. Field Robotics, 2025 | https://doi.org/10.1109/tfr.2025.3636366 | grade A
- **[measurement]** The objection is well-founded and the confound is real, not hypothetical, and it is settleable from the cited archives. The sol is a per-rover LOCAL solar reference frame; Gangale's own Darian-calendar work establishes that off-Earth timekeeping is designed infrastructure, not a neutral universal denominator, so a meters-per-sol value carries a rover-specific time frame that must be declared and converted before pooling. Independently, the rover-operations literature documents that per-sol productivity is gated by the uplink/downlink and tactical-planning-cycle structure: the relation between downlink products and the next uplink command set, and autonomous-planner concepts (IRONCAP) built specifically to relax the ground-in-the-loop per-sol bottleneck, both show that how much autonomy was even possible per sol is co-determined by operations cadence. Pooling a Perseverance autonomous-drive sol with an Opportunity blind-commanded sol without correcting for differing planning-cycle/downlink-window structure therefore risks attributing to the autonomy treatment a per-sol productivity difference that is partly an operations-cadence artifact. PDS traverse archives (per-sol drive geometry/distance) and NTRS operations records (planning-cycle and downlink-window cadence per mission) jointly contain the variables needed to test whether the sol denominator is reference-frame-clean or confounded; the candidate must declare the temporal frame, state the conversion, and add the cadence covariate.
    - Gangale, 'The Architecture of Time, Part 2: The Darian System for Mars,' SAE Technical Paper 2006-01-2249 | https://doi.org/10.4271/2006-01-2249 | grade B
    - 'Relating downlink products to uplink commands in Mars rover operations,' IEEE Aerospace Conf. 2002 | https://doi.org/10.1109/aero.2002.1035301 | grade A
    - 'Innovative Rover Operations Concepts - Autonomous Planner (IRONCAP): Supporting Rover Operations,' AIAA SpaceOps 2012 (10.2514/6.2012-1294460; journal 2013 doi 10.2514/5.9781624102080.0557.0572) | https://doi.org/10.2514/6.2012-1294460 | grade B
    - Biesiadecki et al., 'Tradeoffs Between Directed and Autonomous Driving on the Mars Exploration Rovers,' Int. J. Robotics Research, 2007 | https://doi.org/10.1177/0278364907073777 | grade A
- **[identification]** The dissertation already names the replication platforms (lunar rover programs, future different-processor Mars rovers, terrestrial off-road autonomy) and the estimator to re-run, and explicitly disavows generalization 'without re-estimation'; transfer is labeled a low-confidence conjecture, not a result. The corpora confirm lunar autonomy is an active research frontier (Chen et al. 2025, the candidate's own ref [33]) but supply NO lunar traverse-productivity archive with a published meters-per-sol record analogous to the Mars PDS products, and NO source sets a numeric coefficient-stability threshold. So the candidate has a named-platform / named-estimator transfer plan but no pre-registered quantitative threshold, and the concrete off-Mars estimation dataset does not yet exist in the retrieved record.
    - JPL_AUTONOMY_EDL_03 dissertation.md, Sec 5.5.2 External validity + Sec 7.7 (lines 825-827, 1156-1160, 1205, 1219) | D:/Claude_Code/brain/collegium/candidates/dissertations/JPL_AUTONOMY_EDL_03/dissertation.md | grade B
    - Chen, Jackson, Allard, Beltrame, 'Path planning algorithm for a South Pole lunar rover mission', Acta Astronautica 237 (2025) | https://doi.org/10.1016/j.actaastro.2025.07.059 | grade A
- **[mechanism]** The candidate already identifies the disciplining observation: Perseverance's Enhanced Navigation is cited as the deployed proof-of-concept that the computational lever is real and deliverable after landing, raising autonomous-drive fraction and distance-per-sol over earlier generations (Verma et al. 2025), and extensibility is explicitly labeled 'a hypothesis for future replication, not a result.' The candidate thus stakes extensibility on the existence of a measured post-landing autonomy-upgrade productivity gain. What the dissertation does NOT do is state the converse falsifier as a pre-registered absence-condition over the NTRS record, so the mechanism is anchored to one confirming instance rather than disciplined by a named would-falsify observation.
    - JPL_AUTONOMY_EDL_03 dissertation.md, Sec 1.x significance + 2.4 Mokyr lens + 5.5.2 (lines 184, 196, 827) | D:/Claude_Code/brain/collegium/candidates/dissertations/JPL_AUTONOMY_EDL_03/dissertation.md | grade B
- **[economics]** The dissertation concedes on its own terms that it estimates a productivity coefficient, not a cost-per-unit-productivity on either channel, and computes no break-even; the 'cost paid in development and verification vs at launch' contrast is asserted qualitatively, not quantified. The Space Economy corpus supplies a real launch-cost anchor for the mass side (cost-per-kg-to-LEO curves; a bullish $33/kg figure) but NO software development-plus-verification cost-per-unit-productivity figure, so the two curves cannot be paired and no break-even can be derived from the retrieved evidence. On the retrieved record the 'invest in software' conclusion is a variance-decomposition with an economic label, not a costed allocation rule.
    - JPL_AUTONOMY_EDL_03 dissertation.md, Sec 1.x decision relevance (lines 184, 1138) + limitations (line 1227) | D:/Claude_Code/brain/collegium/candidates/dissertations/JPL_AUTONOMY_EDL_03/dissertation.md | grade B
    - OCEA, 'Forecasting the Space Economy' (Space_Economy_Papers SD08), bullish launch cost ~$33/kg | D:/Claude_Code/Space_Economy_Papers (source_key SD08) | grade C
    - 'The Role of Space in Driving Sustainability, Security, and Development' (Space_Economy_Papers SD13), sample launch cost-per-kg curves | D:/Claude_Code/Space_Economy_Papers (source_key SD13) | grade C
- **[identification]** On the panelist's own terms the between-rover G1/G2/G3 contrast does NOT survive a mission-class/budget control, and the dissertation already concedes the mechanism that makes it fail. The design conditions only on hardware covariates (wheel diameter, mass class, actuator class, drive energy) which are admittedly collinear with autonomy generation by construction because both are rover-level attributes a rover fixed effect fully absorbs (dissertation Sec 2.5, lines 303,370); TechPort is used only as an independent TRL maturity index and robustness check (Sec 4.5, line 244), NOT as a dollar program-cost, downlink-per-sol, or instrument-mass budget control, so no budget/mission-class regressor exists in the specification. NASA's own Mars 2020 mobility record confirms the confound physically: Perseverance was designed to 'replicate Curiosity except where new mission objectives mandated a change,' and the key upgrades were bundled together (more tractive/robust tires, new engineering cameras, a NEW COMPUTER DEDICATED FOR IMAGE PROCESSING, and a more efficient AutoNav), i.e. the autonomy generation rode in with compute and mass on a richer mission. The dissertation itself flags the latent version of this objection: 'a finding that the flight processor, not the software, gated the generation step would partly merge the two channels the model seeks to separate' (line 392). Therefore the candidate's honest posture is already that the between-rover coefficient is confounded and the within-rover autonomous-fraction contrast, not the G1/G2/G3 contrast, carries the causal weight (lines 166, 230, 309). What the candidate has NOT done, and what no retrieved source supplies, is enter a MEASURED budget control (TechPort program cost in dollars, X-band/UHF downlink allocation per sol, total instrument mass) and run the survival test numerically.
    - JPL_AUTONOMY_EDL_03 dissertation.md, Sec 2.5 collinearity of rover-level attributes + Sec 4.5 TechPort as TRL robustness check + line 392 flight-processor-gating caveat (lines 244,303,370,392) | D:/Claude_Code/brain/collegium/candidates/dissertations/JPL_AUTONOMY_EDL_03/dissertation.md | grade B
    - 'Planning for a Martian Road Trip - The Mars2020 Mobility Systems Design' (NASA NTRS 20230005676): Perseverance design replicates Curiosity except where new objectives mandate change; key upgrades bundle new tires, new engineering cameras, a new computer dedicated to image processing, and a more efficient AutoNav | https://ntrs.nasa.gov/citations/20230005676 | grade C
    - Goswami & Garretson four-driver decomposition / preference-indicator pre-registration (F2,F4) - capability is endogenous to elite preference and program funding (thinker dossier; Scramble for the Skies) | https://doi.org/10.5040/9781978730243.ch-005 | grade A
- **[rival]** The within-budget, within-platform autonomy increment the panelist demands DOES exist in the NTRS record, and it is precisely the variation the dissertation makes primary, which is the correct answer to the funded-mission-class-selection rival. Two discriminators are advance-specifiable from public archives with the budget, mass, compute and downlink held constant by construction: (1) WITHIN-MISSION, post-landing autonomy increments on a fixed platform - the Mars Exploration Rover record documents directed-versus-autonomous driving as a tradeoff that shifted as onboard planning capability deepened on the SAME deployed rover (NTRS 'Tradeoffs between directed and autonomous driving on the MER', 20060042880), and Perseverance's Enhanced Navigation is a post-landing flight-software capability that runs on the fixed flight processor and raises autonomous-drive fraction and meters/sol without any change to mass, tires, or downlink (dissertation Sec 1.x significance, line 184, citing Verma et al. 2025 [112]). (2) The dissertation's PRIMARY identifying variation is exactly this within-budget contrast: the within-rover autonomous-drive-fraction comparison holds the entire mission class, budget, mass, compute, and downlink fixed at the level of the individual machine and asks only whether the share of the drive the software executed moves productivity (lines 162,166,291,357). The named FALSIFIER for the budget-endogeneity rival is the dissertation's own first pre-specified check, restated in budget terms: if the within-rover autonomous-fraction coefficient is statistically indistinguishable from zero after terrain conditioning - where the budget is constant by construction - then there is no autonomy effect where funded-mission-class is held fixed, and the apparent G1/G2/G3 rise is shown to be the funded-mission-class index the panelist posits (lines 176,309). That within-rover null is the observation in the assembled panel that falsifies a real autonomy effect and leaves the budget-proxy rival standing.
    - 'Tradeoffs between directed and autonomous driving on the Mars Exploration Rover' (NASA NTRS 20060042880) - same-platform directed-vs-autonomous driving contrast on the deployed MER rovers | https://ntrs.nasa.gov/citations/20060042880 | grade C
    - JPL_AUTONOMY_EDL_03 dissertation.md - within-rover autonomous-drive-fraction contrast as primary identification holding the machine (hence budget/mass/compute/downlink) fixed, plus the pre-specified within-rover null falsification check (lines 162,166,176,184,291,309,357) | D:/Claude_Code/brain/collegium/candidates/dissertations/JPL_AUTONOMY_EDL_03/dissertation.md | grade B
    - Goswami & Garretson preference-indicator pre-registration (F4): name in advance the observable and the falsifier (thinker dossier; Scramble for the Skies ch.4-6) | https://doi.org/10.5040/9781978730243.ch-004 | grade A
- **[identification]** The thinker's frameworks decide the modeling half of this: a system is elements interconnected so as to produce behavior over time; feedback and delay routinely defeat the additive, separable intuition. An additive main-effects model that holds the autonomy term constant across terrain is exactly the 'correlating snapshots' error the framework warns against when a coupled mechanism is at work; if 5.2's own mechanism (autonomy reshapes which terrain the rover dares to enter, changing the marginal value of autonomy) is real, then the interaction IS the mechanism and the additive coefficient is a structurally biased summary of a coupled relationship, not 'the contribution.' So the candidate cannot have it both ways: a coupled mechanism in 5.2 falsifies the separability his headline coefficient assumes. What the retrieved corpora canNOT settle is the empirical adjudication question -- whether the specific PDS drive-sol record has the cross-terrain-class variation and sample to estimate that interaction term with enough power to reject the additive model. No retrieved source characterizes PDS drive-sol statistical power, so that half is a gap, not an answer.
    - Thinking in Systems: A Primer (Donella H. Meadows) -- state lives in stocks, feedback loops and delays drive counterintuitive behavior; bounded-rational actors respond to limited information and cannot mentally simulate the coupled system | https://us.macmillan.com/books/9781603580557/thinkinginsystems | grade A
    - Leverage Points: Places to Intervene in a System (Meadows 1999, reprinted) -- structure of stocks/flows and feedback-loop structure rank above parameter numbers; small structural couplings produce large, durable behavior change | 10.4324/9781849773386-15 | grade A
- **[measurement]** The thinker's stock-and-flow discipline answers the diagnostic half cleanly: state lives in stocks (accumulations); stocks change only through flows (rates). Meters-per-sol is a flow; the accumulating planning-cycle trust / commanding-effort the candidate himself names as the loop driver is the stock. A model that correlates the instantaneous flow while the governing variable is the unmeasured accumulating stock makes its growth conclusions unfalsifiable -- the framework's standing charge to any analyst who 'reports launch rates and deployment counts' instead of identifying the governing stock and integrating its inflows/outflows. The candidate's own concession (meters-per-sol 'cannot see' the planning loop and 'understates' autonomy) is therefore a self-admission that the chosen scalar is a low-leverage flow proxy for the real stock. What the retrieved corpora do NOT establish is whether TechPort or NTRS in fact contain a constructable per-meter planning-effort proxy -- no retrieved source indexes TechPort/NTRS commanding-effort fields, so the existence of that alternative measure is a gap, not a grounded affirmative.
    - Thinking in Systems: A Primer (Donella H. Meadows) -- state lives in stocks (accumulations); stocks change only through flows (rates); confusing the rate for the accumulation produces counterintuitive, unfalsifiable conclusions | https://us.macmillan.com/books/9781603580557/thinkinginsystems | grade A
    - Density-based evolutionary model of the space debris environment in LEO -- treats the population as an accumulating stock whose net growth, not instantaneous rate, governs long-run behavior; the canonical demonstration that the stock, not the flow snapshot, is the variable to budget | 10.1016/j.actaastro.2024.03.008 | grade A
- **[governance]** The leverage-points hierarchy ranks parameters/numbers (subsidies, standards, the mass-vs-software trade) at the bottom and the system's goal and the paradigm out of which goals arise at the top; people reliably push the obvious low-leverage parameters, often in the wrong direction, while durable change requires moving up to rules, goals, and paradigm. By the candidate's own Discussion the autonomy channel's durable payoff is goal/paradigm-level (reorganizing how missions are commanded so sols are freed for the MSR sampling-and-caching goal), which the framework ranks far above the mass-vs-software parameter the between-channel coefficient estimates. So optimizing the visible meters-per-sol parameter risks exactly the framework's predicted failure: telling the designer the answer at the tier where the lever is weakest, while the high-leverage reorganization sits where the chosen scalar is blind. The framework therefore endorses the question's worry and instructs the candidate to state the actual goal his architecture optimizes and show whether his lever is upstream or downstream of it. What the retrieved corpora do NOT supply is the specific PDS or NTRS observable that would empirically distinguish a commanding-paradigm shift from a drive-distance gain -- no retrieved source names such an archival indicator, so that adjudication is a gap.
    - Leverage Points: Places to Intervene in a System (Meadows 1999, reprinted) -- twelve places ranked least to most effective: 12 constants/parameters/numbers lowest; goals and the paradigm out of which goals arise are the highest-leverage tiers; people reliably push low-leverage parameters in the wrong direction | 10.4324/9781849773386-15 | grade A
    - Thinking in Systems: A Primer (Donella H. Meadows) -- a system optimized for 'maximize deployment now' behaves differently from one optimized for 'sustain capacity'; ask what an architecture is FOR, and whether the intervention is upstream or downstream of the true goal | https://us.macmillan.com/books/9781603580557/thinkinginsystems | grade A
- **[economics]** The candidate's upside framing commits the rate-for-stock error Meadows names as fatal to growth conclusions: system state lives in STOCKS (accumulations), which change only through FLOWS (rates), and behavior is governed by the accumulation of flows into stocks, not by the instantaneous rate managers watch. Meters-per-sol is a flow; the binding scarcity is a stock (one non-replaceable vehicle plus a finite, non-renewable stock of mission-sols inside an MSR launch window). A correct loss function is therefore an integral, not a point estimate: L = integral over the remaining mission horizon of the value-density of sols actually delivered, with the catastrophic branch driving that integrand to zero at the abort sol and holding it at zero for every subsequent sol in the window (the asset is a stock that does not refill, so the loss compounds as forgone caching across the entire residual horizon). Because the vehicle stock is one and irreplaceable, the downside is bounded below not by a recoverable rate dip but by the total residual integral, which is precisely the 'size of buffer / structure of stocks' tier (Meadows leverage points 10-11) that the candidate's rate-only Discussion never budgets.
    - Donella H. Meadows, Thinking in Systems: A Primer (via meadows hall-of-shoulders brain dossier + OpenAlex record) | https://us.macmillan.com/books/9781603580557/thinkinginsystems | grade A
    - Donella H. Meadows, 'Leverage Points: Places to Intervene in a System' (1999) | 10.4324/9781849773386-15 | grade A
- **[empirics]** The demand is methodologically correct and the downside loop is real and archivally observable in principle: the operational record of rover autonomy carries both the gain (visual-odometry / AutoNav increasing safe drive distance) and the cost (autonomy is a recognized verification-and-validation hard problem whose faults can propagate when a sparser, more-trusted command set reduces ground insight). A doctrine estimated on the gross upside alone cannot establish the SIGN of the net effect, which is exactly Meadows's objection that a lever scored only on its visible flow is unfalsifiable.
    - Maimone, Cheng, et al., 'Two years of Visual Odometry on the Mars Exploration Rovers,' Journal of Field Robotics (2007) | 10.1002/rob.20184 | grade A
    - Brat & Jonsson, 'Challenges in verification and validation of autonomous systems for space exploration,' IJCNN (2005) | 10.1109/ijcnn.2005.1556387 | grade A
- **[governance]** By Meadows's own ranking the doctrine is NOT a bounded parameter tweak: 'optimize for trusted autonomy fleet-wide' changes the GOAL of the operations system (leverage tier 3, goals; near the high-gain end above buffers and balancing loops), and a high-leverage lever pushed in the wrong direction is worse than no intervention because the lever's gain amplifies the error. The mechanism of harm is balancing-loop erosion: ground-commanding skill and the conservative reserve margin are stabilizing stocks (a negative/balancing feedback that returns the system to safety); a reinforcing loop that rewards higher autonomous-fraction can draw those stocks down (skill atrophies when unused, margin is spent to buy rate), weakening the very balancing feedback that keeps autonomy safe. This is detectable in the operational record exactly as Meadows prescribes: name the governing stock, write its inflow and outflow, and watch the stock rather than the headline rate. The wrong-direction signature is a rising autonomous-fraction co-occurring with a falling ground-intervention frequency and a thinning commanded margin over sols, the operational proxy for the balancing stock being drained, which is testable on the same telemetry that yields the meters-per-sol upside.
    - Donella H. Meadows, 'Leverage Points: Places to Intervene in a System' (1999) | 10.4324/9781849773386-15 | grade A
    - meadows hall-of-shoulders brain dossier, summarizing Meadows 'Leverage Points' | 10.4324/9781849773386-15 | grade A
    - Donella H. Meadows, Thinking in Systems: A Primer (via meadows dossier) | https://us.macmillan.com/books/9781603580557/thinkinginsystems | grade A
- **[measurement]** An independent propositional-base observable DOES exist and is measurable separately from the autonomy-generation coefficient: the predictive accuracy of the rover's onboard SLIP/SINKAGE and stereo-hazard-geometry MODELS against measured terra-mechanical ground truth. Mokyr's apparatus requires that a prescriptive technique (lambda) rest on a widening propositional base (Omega) that makes it 'extensible and self-correcting,' whereas trial-without-theory stagnates (Gifts of Athena, 2002). The directly testable Omega-variable is whether onboard slip prediction migrated from no model, to image-learned slip prediction (Angelova et al., 'Slip Prediction Using Visual Information,' RSS 2006), to mechanics-based wheel slip/sinkage estimation (Iagnemma-lineage methods, 2010), a deepening theory of WHY the terrain behaves as it does. If across G1/G2/G3 the onboard slip/sinkage prediction error and stereo-hazard model coverage are FLAT (same fixed heuristics merely re-badged), the Mokyr reading is falsified even if the autonomous-drive-fraction coefficient is large and significant. The candidate's design does not measure this variable; it lets the autonomy-generation coefficient stand in for Omega. Absent that slip-model-accuracy / planning-completeness series in PDS/NTRS, the propositional/prescriptive lens is a flattering narrative bolted onto a regression, not a tested proposition.
    - Hall of Shoulders mokyr dossier (Mokyr, The Gifts of Athena, Princeton 2002), review lens item 1 'Propositional vs. prescriptive test' | local:brain/collegium/hall_of_shoulders/brains/mokyr/mokyr.db | grade A
    - Angelova, Matthies, Helmick, Perona, 'Slip Prediction Using Visual Information,' Robotics: Science and Systems | https://doi.org/10.15607/rss.2006.ii.014 | grade A
    - 'Methods for Wheel Slip and Sinkage Estimation in Mobile Robots' (2010) | https://doi.org/10.5772/9279 | grade B
- **[economics]** The realized access cost IS measurable and the historical record supports Mokyr's stickiness prediction over the design's frictionless 'in principle' claim. Two observables exist: (a) the dated flight-software version history per landed asset, each onboard-autonomy upgrade is a discrete, V&V-gated, uplinked software version, and the realized access cost is the calendar lag between development and the version becoming operational on the surface; the autonomy-survey literature documents that onboard autonomy on landed assets advances in such discrete, validated software releases rather than continuous frictionless upload (Gao & Chien, 'Autonomy for Space Robots: Past, Present, and Future,' 2021). (b) The launch-cohort attribution: the major step changes in rover onboard autonomy track ROVER GENERATIONS (MSL Curiosity, then Mars 2020 Perseverance), i.e. they shipped predominantly AT LAUNCH with new flight hardware, with mid-mission uplinks being the smaller, V&V-bounded increment (MSL eight-year mission overview, 2022; Mars 2020 next-generation imaging system, 2020). Mokyr's access-cost thesis predicts exactly this contested, tacit, slow diffusion: the aerospace STI record shows the mechanisms by which technical information diffuses from producers to users are poorly understood and far stickier than 'in principle' transfer (NASA/DOD Aerospace Knowledge Diffusion Research Project, NTRS 19960052732), and the Spinoff/tech-transfer apparatus exists precisely because access costs are high (KSC Tech Transfer, NTRS 20120014076). The candidate must produce the per-asset version-history lag distribution and the retrofitted-vs-frozen-at-launch fraction; if, as the generational pattern suggests, most autonomy gains shipped at launch, the software/hardware asymmetry that makes the result decision-relevant collapses and the autonomy lever is as frozen-at-launch as the wheel.
    - Gao & Chien (eds./authors), 'Autonomy for Space Robots: Past, Present, and Future,' Current Robotics Reports | https://doi.org/10.1007/s43154-021-00057-2 | grade A
    - NASA/DOD Aerospace Knowledge Diffusion Research Project: US Scientific and Technical Information Policy (NTRS) | https://ntrs.nasa.gov/citations/19960052732 | grade B
    - 'The Mars 2020 Engineering Cameras and Microphone on the Perseverance Rover' (2020); 'Mission Overview ... Mars Science Laboratory Curiosity Rover After Eight Years of Surface Operations' (2022) | https://doi.org/10.1007/s11214-020-00765-9 | grade A
    - KSC Tech Transfer News (NTRS) | https://ntrs.nasa.gov/citations/20120014076 | grade C
- **[mechanism]** PARTIAL ANSWER WITH CONCEDED GAP. The separating evidence Mokyr requires would be a knowledge-base-widening series that is logically independent of utilization frequency: (i) the onboard slip/sinkage and stereo-hazard MODEL accuracy series (Omega deepening, per Angelova 2006 and slip/sinkage estimation 2010), and (ii) the planning-completeness / decision-horizon of the onboard planner, whether each generation makes the NEXT increment cheaper and diagnosable, the signature of an extensible propositional base (mokyr dossier; Gifts of Athena 2002). The autonomy literature confirms these are conceptually distinct axes: 'adaptive and intelligent' navigation (model-driven extensibility) is surveyed separately from raw autonomous traverse utilization ('Adaptive and intelligent navigation of autonomous planetary rovers, a survey,' 2017). HOWEVER, the candidate's panel as described supplies only the within-rover autonomous-drive-fraction (a utilization/trust margin), which by construction can be large and significant with a fixed technique merely switched on more often where terrain permits. Because the design provides no model-accuracy or planning-completeness series to stand beside the fraction, the Mokyrian widening mechanism is UNIDENTIFIED, and the headline coefficient is observationally equivalent to the 'same fixed technique utilized more aggressively as ground-team confidence grew' (prescriptive habituation), i.e. it does not rule out the trial-and-error stagnation rival the lens claims to exclude. Mokyr's own propositional-vs-prescriptive and Cardwell reversibility tests demand that separation before the extensibility reading can be granted.
    - Hall of Shoulders mokyr dossier, review lens items 1-2 (Mokyr, Gifts of Athena 2002; Cardwell's Law / reversibility test) | local:brain/collegium/hall_of_shoulders/brains/mokyr/mokyr.db | grade A
    - 'Adaptive and intelligent navigation of autonomous planetary rovers, A survey' (2017) | https://doi.org/10.1109/ahs.2017.8046384 | grade B
    - Angelova et al., 'Slip Prediction Using Visual Information,' RSS 2006 | https://doi.org/10.15607/rss.2006.ii.014 | grade A
- **[economics]** Partial grounding only: the certification-and-authorization cost term Q3 demands is real and measurable in principle. A 2025 method paper formalizes surface-autonomy certification as a structured verification-and-validation campaign with non-trivial assurance burden, establishing that a retrofit-path certification cost is a legitimate, non-zero ledger entry rather than an assumed-near-zero quantity. This supports the OBJECTION that the candidate's ledger omits an access cost, but it does NOT supply the Mars rover-specific magnitude (the actual MER/Curiosity/Perseverance re-qualification cost, review burden, or uplink-anomaly expected-loss), which was not retrievable this turn.
    - Mendoza et al., 'A Method for Lunar Surface Autonomy Certification: Application to a Construction Pathfinder Mission,' Aerospace 12(12):1115, 2025 (via OpenAlex + Crossref) | https://doi.org/10.3390/aerospace12121115 | grade A
    - mokyr dossier, Hall of Shoulders brain (Cardwell's Law / access-cost framing: incumbents who bear the downside of allowing the next innovation freeze the channel; sustained progress requires keeping the channel contestable) | D:/Claude_Code/brain/collegium/hall_of_shoulders/brains/mokyr/ | grade C
- **[identification]** The challenge is theoretically valid: the autonomous-drive fraction, when set by ground-team trust, is a calibrated-reliance variable subject to misuse (over-reliance) and disuse (under-reliance), so a positive H1 coefficient is genuinely confounded by operators learning to trust an unchanged machine and is not, on its own, evidence of machine competence. Use/misuse/disuse/abuse is the governing frame for this confound. However, retrieval located NO source establishing that the candidate's PDS command logs or NTRS Sol-by-Sol AutoNav reports contain an exogenous shifter (downlink-window, staffing, Earth-Mars geometry, or documented autonomy-posture directive) usable as an instrument, nor any demonstration that H1 survives such instrumenting. The identification fix is unsupported by available evidence.
    - Parasuraman, R. & Riley, V. (1997), 'Humans and Automation: Use, Misuse, Disuse, Abuse,' Human Factors 39(2):230-253 | https://doi.org/10.1518/001872097778543886 | grade A
- **[mechanism]** The mechanism challenge is theoretically valid: high automation reliability predictably degrades operator monitoring of the automated function (automation-induced complacency), an attentional-reallocation effect that is robust and dosage-dependent. This supports the prediction that a trusted, reliable AutoNav frees planning-cycle attention and induces the ground team to command longer/bolder drives, inflating meters-per-sol through the commanding channel rather than the hazard-clearing channel. Therefore the productivity gain cannot be attributed to the software channel unless the commanded drive-distance ceiling and route conservatism are separately observable and held fixed. Retrieval found NO source showing those commanded-route variables are separable in the candidate's PDS/NTRS data, so the assignment of the gain to the software channel is, on available evidence, unestablished.
    - Parasuraman, R. & Manzey, D.H. (2010), 'Complacency and Bias in Human Use of Automation: An Attentional Integration,' Human Factors 52(3):381-410 | https://doi.org/10.1177/0018720810376055 | grade A
    - Parasuraman, R., Molloy, R. & Singh, I.L. (1993), 'Performance Consequences of Automation-Induced Complacency,' Int. J. of Aviation Psychology 3(1):1-23 | https://doi.org/10.1207/s15327108ijap0301_1 | grade A
- **[measurement]** The measurement challenge is theoretically valid: automation reliance produces both errors of omission (failing to act on conditions the automation did not flag) and errors of commission (acting on the automation against contrary evidence). A meters-per-sol outcome that counts only successful traverse omits the commission/fault tail, so the H1 coefficient can be an artifact of counting successes rather than the net competence of the human-machine system; netting autonomy-attributable fault, safing, and abort sols is the correct robustness test and a trust-calibration account predicts the coefficient would shrink. Retrieval found NO source demonstrating that such a competence-net measure was constructed from the candidate's PDS fault/event records or that H1 is robust to it; the empirical resolution is unsupported.
    - Parasuraman, R. & Manzey, D.H. (2010), 'Complacency and Bias in Human Use of Automation: An Attentional Integration,' Human Factors 52(3):381-410 | https://doi.org/10.1177/0018720810376055 | grade A
- **[measurement]** The framework underwriting this challenge is sound and well-established: 'use' of automation is the operator's voluntary activation/disengagement decision, governed by trust, workload, and risk, and is distinct from the automation's true reliability; over-reliance is 'misuse' and under-reliance 'disuse', so a use-fraction is a calibrated-reliance variable, not a competence measure (Parasuraman & Riley 1997). The candidate must therefore treat the autonomous-drive fraction as a reliance variable and net out a trust-accrual trajectory before reading the coefficient as machine capability. HOWEVER, the specific empirical object the question demands, the dated within-rover time path of the autonomous fraction against cumulative sols and against the running count of autonomy-attributable aborts/faults/ground-overrides, is NOT recoverable from any source retrieved this turn; AMOS, ACTA, and the Space-Economy corpus returned zero hits and NTRS returned only one unrelated 2008 visual-target-tracking paper, so no retrieved source supplies that reliance trajectory.
    - Parasuraman & Riley, 'Humans and Automation: Use, Misuse, Disuse, Abuse', Human Factors (1997) | https://doi.org/10.1518/001872097778543886 | grade A
- **[mechanism]** The type-and-level demand is grounded and binding: automation is not a switch but applies to four distinct stages of human information processing, (1) information acquisition, (2) information analysis, (3) decision and action selection, (4) action implementation, each automatable to a different level, so a single autonomy scalar that collapses these stages is mis-specified and the candidate must state which stage each generation moved and to what level (Parasuraman, Sheridan & Wickens 2000). The same source establishes the out-of-the-loop performance cost that makes the stage distinction consequential for any productivity claim. HOWEVER, the engineering fact needed to settle whether AutoNav vs Enhanced Navigation is a stage-4 timing change versus a stage-2/3 capability change, the NTRS/PDS description of what each rover-autonomy generation actually automates, was NOT returned by any retrieval this turn (AMOS=0, ACTA=0, NTRS yielded no AutoNav engineering description), so the stage attribution of the productivity gain cannot be asserted from a source.
    - Parasuraman, Sheridan & Wickens, 'A model for types and levels of human interaction with automation', IEEE Trans. SMC-A (2000) | https://doi.org/10.1109/3468.844354 | grade A
- **[identification]** The identification concern is theoretically well-founded: operator reliance is dynamic, not constant. Vigilance and workload are time-varying variables a static design ignores at its peril, and a candidate arguing automation enables higher cadence at constant safety must show reliance was calibrated and the operator-workload curve was measured, not assumed flat (Parasuraman 2003, neuroergonomics). The alerting-threshold choice itself drives operators toward disuse via the cry-wolf failure when false-alarm rate is high (Parasuraman & Riley 1997), so a ground-team conservatism response co-moving with the autonomous fraction is exactly the omitted dynamic confounder the disuse/misuse framework predicts. The additive, constant-ideal-human specification therefore omits the reliance-response channel and the candidate must demonstrate separability. HOWEVER, whether such a conservatism measure can in fact be built from the candidate's PDS command logs and NTRS operations reports, and whether the autonomous-fraction coefficient survives its inclusion, is NOT established by any source retrieved this turn; no command-log or operations-report dataset surfaced (AMOS=0, ACTA=0, NTRS=1 unrelated), so the empirical separability result cannot be asserted.
    - Parasuraman, 'Neuroergonomics: Research and practice', Theoretical Issues in Ergonomics Science (2003) | https://doi.org/10.1080/14639220210199753 | grade A
    - Parasuraman & Riley, 'Humans and Automation: Use, Misuse, Disuse, Abuse', Human Factors (1997) | https://doi.org/10.1518/001872097778543886 | grade A

## Gaps

- **[identification]** No retrieved source contains the first-stage decomposition (residual variance share of the autonomous fraction after the terrain block) or any within-rover/within-terrain-class placebo split for this candidate. The AMOS (2,346 SSA/SDA papers), ACTA, and Space Economy corpora return zero hits on rover-autonomy productivity panels and on causal-inference/treatment-effect methods. The quantitative answer (how much identifying variation survives terrain absorption, and whether a placebo clears) is a candidate-supplied exhibit that is absent from evidence; it cannot be asserted or invented. (raised by angrist_pischke)
- **[measurement]** No retrieved source establishes the provenance/timestamp of the candidate's a-priori terrain characterization relative to the drive-plan decision, nor reports the autonomous-fraction coefficient with vs without the terrain block. These are mission-telemetry and candidate-regression facts absent from AMOS/ACTA/Space Economy and from NTRS (which returned zero relevant rows for rover-autonomy productivity). Whether the terrain covariate is genuinely pre-treatment or an outcome-anticipating bad control cannot be settled from retrieved evidence. (raised by angrist_pischke)
- **[empirics]** No retrieved source reports a Monte Carlo on the candidate's three-between-rover-cluster panel, so the actual rejection rate of the proposed wild-cluster bootstrap at nominal 5% is unknown from evidence. Whether the between-rover generation coefficient survives as an interpretable hypothesis test, or must be abandoned in favor of the within-rover estimand, is a candidate-supplied simulation result absent here and cannot be invented. (raised by angrist_pischke)
- **[identification]** Whether a specific PDS/NTRS observable (sols-to-deadline, planning-cycle cadence, tactical-vs-strategic plan flag) actually exists in the record at drive-sol resolution, and whether the autonomous-fraction coefficient empirically survives its inclusion, cannot be settled: no retrieved source contains the candidate's dissertation regressions or any Mars-rover-productivity econometric study (AMOS/ACTA/Space-Economy corpora returned zero hits; dossier independently confirms AMOS returns zero causal-inference/treatment-effect hits). The empirical resolution is unverified. (raised by angrist_pischke)
- **[measurement]** The actual variance-share decomposition under all orderings, the Shapley/Owen attribution on the assembled three-rover panel, and the numeric share range (and whether it straddles 50 percent) cannot be reported: no retrieved source contains the candidate's panel, the per-block R-squared values, or any computed decomposition. The numeric resolution is unverified. (raised by angrist_pischke)
- **[empirics]** Whether the placebo coefficient is in fact null on the candidate's blind-commanded segments, and what the empirical threshold should be, cannot be settled: no retrieved source contains the candidate's blind-commanded subsample, the placebo construction, or any executed negative-control result. The empirical resolution is unverified. (raised by angrist_pischke)
- **[identification]** No retrieval this turn contains the candidate's estimated propensity-of-autonomous-fraction surface, the terrain covariates that drive the ground team's dose choice, or the covariate-overlap diagnostic across the dose range on the PDS/NTRS panel. The general method is grounded (c1); these candidate-specific empirical artifacts are not in any source and cannot be asserted. The candidate must estimate and exhibit the propensity surface and demonstrate overlap before the cross-dose reading can stand. (raised by callaway_santanna)
- **[economics]** No retrieval this turn contains the candidate's binned within-terrain-class dose-response curve, the partition of PDS/NTRS drive-sols by autonomous-fraction bin, or evidence on whether the slope is identified by within-unit same-terrain dose changes versus across-unit comparisons. The forbidden-cross-dose theorem is grounded (c2); the empirical partition that would (or would not) rescue H1 is absent from sources and must be built by the candidate. (raised by callaway_santanna)
- **[empirics]** No retrieval this turn contains the candidate's within-rover autonomous-fraction-versus-sol trajectories, a regression of the dose on lagged realized productivity, or a pre-dose placebo result on the PDS/NTRS traverse record. The no-anticipation requirement and the placebo/lagged-regression and sensitivity recipe are grounded (c3); whether THIS candidate's dose actually responds to lagged meters-per-sol is an empirical question no source settles and the candidate must run. (raised by callaway_santanna)
- **[measurement]** No retrieved source supplies a published, measured, mission-independent threshold that adjudicates the G1/G2/G3 generation boundary as a capability discontinuity rather than a mission-described relabeling. The three-level generation taxonomy is the candidate's own construct; retrieval grounds the measurement substrate (meters/sol, autonomy fraction) but not the boundary. The candidate must either (a) define the bright line as a measured discontinuity on the reported metrics, or (b) concede the generation variable is functionally (mission-description) defined and therefore confounded with branding and at risk of the rise-and-stall failure of activity-defined classifications. Unresolved on present evidence. (raised by gangale)
- **[measurement]** REFUSED on the empirical core. No retrieved source supplies an exact TechPort/NTRS field name with a numeric threshold (TRL step, drive-fraction band, or cycle-time figure) that fixes the G1/G2/G3 boundaries archive-derivably. The gangale corpus supplies only the LENS, not the fact: Gangale's 'Functional Approach: Its Rise and Stall' (DOI 10.1163/9789004366022_013) and How High the Sky? (DOI 10.1163/9789004366022) establish that an activity-defined construct that cannot deliver a measurable bright-line threshold rises and then stalls in practice, which is exactly the failure mode the candidate's autonomy-generation treatment risks if G2-vs-G3 reduces to a program calling its planner 'Enhanced Navigation.' Crossref confirms such relabeling-prone artifacts exist (e.g., 'Enhanced Autonomous Navigation on the Perseverance Mars Rover,' DOI 10.1109/tfr.2025.3636366; 'AutoNav Mark3,' DOI 10.2514/6.2006-6708) but none expose a TechPort field + numeric cut a blind analyst could apply. The candidate must convert the treatment into an archive-thresholded boundary or concede it is a paywalled label. (raised by gangale)
- **[measurement]** REFUSED on the empirical core. No retrieved source names the PDS traverse/localization field that bounds usable-drive-time or energy envelope per drive-sol, nor demonstrates meters-per-sol is normalized to it. The gangale corpus supplies only the reference-frame-rigor LENS: Gangale's review lens explicitly demands that any operation spanning regimes 'specify the time and coordinate reference frame and the conversion between them' (gangale dossier, grade A), and his Mars-timekeeping work, 'The Architecture of Time, Part 2: The Darian System for Mars' (DOI 10.4271/2006-01-2249) and 'Part 3: Project Management in Two-Dimensional Time' (DOI 10.2514/6.2007-6073), treats a sol as a designed temporal reference frame, not a uniform unit of opportunity. This grounds the CRITIQUE that a sol's usable drive window varies by rover/season/latitude/downlink, but does NOT supply the missing PDS field or normalization proof. The candidate must name the field and show the denominator is the energy-bounded envelope, or concede the autonomy coefficient is partly absorbing an unconverted power-and-timekeeping frame. (raised by gangale)
- **[economics]** REFUSED on the empirical core. No retrieved source documents a measured before-and-after per-sol mobility gain delivered to a LANDED rover by post-landing software upload on a fixed machine and terrain class. Crossref/OpenAlex surface adjacent generational-autonomy artifacts ('Enhanced Autonomous Navigation on the Perseverance Mars Rover,' DOI 10.1109/tfr.2025.3636366; 'AutoNav Mark3,' DOI 10.2514/6.2006-6708; 'Terrain Aware Traverse Planning for Mars Rovers,' DOI 10.33915/etd.7956) and an MSL mission overview (DOI 10.1002/2014je004622), but none isolates a retrofit-window per-sol delta on a single landed rover. The gangale corpus supplies the pre-emption-vs-reaction LENS: Gangale's standing argument is to 'design the operationalizable artifact... before the scarcity hardens into entrenched first-mover positions' (gangale dossier, grade A), which sharpens the timing critique that if no landed rover in the panel ever bought productivity by software after touchdown, the upload option is a design-time fork that must be committed before launch, not a surface-time patch, mistiming the central investment recommendation. The candidate must produce the same-machine before/after evidence or concede the recommendation is mistimed. (raised by gangale)
- **[identification]** No retrieved source provides a pre-registered numeric cross-platform coefficient-stability threshold, nor a concrete off-Mars traverse-productivity dataset (lunar VIPER/Chang'e meters-per-sol archive or a different-processor Mars-rover productivity series) on which the autonomy coefficient could be re-estimated. The panelist's demanded quantitative transfer indicator cannot be supplied from evidence; absent it, the 'investment doctrine' remains an articulated vision plus a Mars-fleet retrospective, exactly as the panelist's F3 lens charges. (raised by goswami_garretson)
- **[mechanism]** No retrieved source (and not the dissertation) states the pre-registered NTRS-absence test that would falsify extensibility. Retrieval could not independently corroborate the Verma et al. 2025 Enhanced Navigation per-sol productivity figure (the ACTA/AMOS corpora returned zero on 'Enhanced Navigation Perseverance'); only the candidate's own citation attests it. Without a named would-falsify observation, the extensibility claim risks the unfalsifiable-ideational-residual failure the panelist's F4 lens targets. (raised by goswami_garretson)
- **[economics]** No retrieved source provides the marginal development-plus-verification cost per unit of autonomous-drive-fraction productivity, nor the marginal launch-cost of the equivalent mechanical increment in the same units, nor a break-even point. The cost side of each channel is unquantified in both the dissertation and the corpora, so the economic-doctrine claim cannot be grounded; this is the F1/F5 break-even objection sustained. (raised by goswami_garretson)
- **[identification]** No retrieved source supplies the measured budget/mission-class control the panelist demands in usable units: TechPort program cost in dollars per mission, X-band/UHF downlink allocation per sol, and total instrument-payload mass per rover were not found in the corpora or via NTRS/OpenAlex this turn, and the candidate did not construct any of them as a regressor. So the survival test of the autonomy-generation coefficient against a budget index cannot be EXECUTED from the evidence; the qualitative verdict (it would not survive, by the dissertation's own collinearity admission and the NTRS bundling record) stands, but the quantitative 'enter it and report the coefficient' the panelist asks for is unmet. Until a measured program-budget control is entered between-rover, the G1/G2/G3 contrast remains, on this record, indistinguishable from a funded-mission-class index, exactly the F2 endogeneity charge. (raised by goswami_garretson)
- **[economics]** REFUSED: no retrieved source this turn supplies either marginal cost curve in the required units or the dollars-per-meter-per-sol crossing. The Space Economy corpus returned no marginal-cost-per-kg-vs-software-development crossover figure (queries returned empty), ACTA returned no rover autonomy cost/mass-tradeoff or mission cost-estimating-relationship hits, and no TechPort dollar cost record was retrievable. The candidate's own design concedes it estimates a productivity coefficient, not a cost-per-unit-productivity on either channel, and computes no break-even (R1 economics finding, dissertation lines 184,1138,1227). Because the decision a program manager faces is a budget-constrained marginal-dollar allocation and the within-rover coefficient holds the entire budget fixed and therefore prices nothing, the 'buy productivity with software not mass' conclusion cannot be elevated to an allocation rule on the retrieved record; it remains a within-budget variance decomposition with an economic label. The F1/F5 break-even objection is sustained for a second round. (raised by goswami_garretson)
- **[rival]** The within-platform autonomy increments named (MER directed-vs-autonomous tradeoff; Perseverance Enhanced Navigation post-landing) establish that the discriminating variation EXISTS, but retrieval this turn did not independently corroborate a MEASURED per-sol productivity STEP for either under a documented flat-or-declining program budget with mass, compute, and downlink certified constant: the NTRS 'Tradeoffs' abstract was truncated and gives no per-sol delta, and the Verma et al. 2025 Enhanced-Navigation per-sol figure rests on the candidate's own citation (ACTA/AMOS returned zero on Enhanced Navigation, per R1). So the discriminator is advance-specifiable and the falsifier is named, but the quantitative within-budget step that would convert it from a clean identification strategy into a delivered result is not yet in the retrieved record. (raised by goswami_garretson)
- **[identification]** Whether the PDS drive-sol panel actually has the cross-terrain-class variance and sample size to estimate the autonomy-by-terrain interaction with enough power to reject the additive specification is an empirical-design fact about the PDS archive. No retrieved corpus (AMOS, ACTA, Space Economy, meadows brain) characterizes PDS drive-sol statistical power or terrain-class coverage, so the adjudication question is left unanswered rather than asserted. (raised by meadows)
- **[measurement]** Whether TechPort or NTRS records actually contain fields from which a coarse planning-cycle-effort-per-meter proxy (commanding burden, uplink-cycle counts, ground-team effort) could be constructed is an archive-content fact. No retrieved corpus indexes TechPort/NTRS commanding-effort metadata, so the existence and construction of that alternative stock-proxy measure is left unanswered. (raised by meadows)
- **[governance]** Which specific PDS or NTRS observable would let the design adjudicate whether the high-leverage gain is the commanding-paradigm shift (goal-level reorganization of mission commanding) rather than the drive-distance scalar is an empirical-design fact about those archives. No retrieved corpus names such an indicator, so the adjudication observable is left unanswered. (raised by meadows)
- **[empirics]** No retrieved source this turn supplies the specific NET quantity the question demands: recovery-sols-lost and autonomy-attributable abort/safe-mode frequency PER UNIT of autonomous-fraction gained, drawn from PDS/NTRS fault-anomaly-recovery records. The NTRS citation API returned zero hits for rover autonomy fault/recovery queries this turn, and AMOS/ACTA/Space-Economy returned nothing on rover-autonomy operational faults. The candidate must therefore COMMIT to building that net ledger from PDS/NTRS; it cannot be asserted as an existing result, and the magnitude (let alone the sign) of recovery-sols-lost per unit autonomous-fraction is unsupported by retrieval. (raised by meadows)
- **[measurement]** No PDS/NTRS/TechPort variable in the candidate's panel measures the propositional base directly (onboard slip/sinkage model accuracy, stereo-hazard geometry coverage, or planning completeness as a per-generation series). The independent observables exist in the literature (Angelova RSS 2006; slip/sinkage estimation 2010) but are NOT in the design, which lets the autonomy-generation coefficient stand in for Omega. Until that series is added, the Mokyr propositional/prescriptive frame is untested by the design and risks being decorative. (raised by mokyr)
- **[economics]** The realized access cost of retrofitting an autonomy increment onto a landed rover (per-asset flight-software version-history lag distribution; fraction of autonomy improvements actually uplinked-to-surface vs frozen-at-next-launch) is NOT estimated in the design, it is asserted 'in principle.' The generational delivery pattern (MSL then Mars 2020) and the STI-diffusion record (NTRS 19960052732) suggest most step changes shipped at launch, which would collapse the software/hardware retrofit asymmetry. The candidate has not produced the version-history measurement that would settle whether the autonomy lever is genuinely retrofittable or effectively frozen-at-launch. (raised by mokyr)
- **[mechanism]** The strongest identifying variation (within-rover autonomous-drive-fraction) measures utilization/operational trust, not knowledge-base widening, so the Mokyrian extensibility mechanism is unidentified and remains observationally equivalent to prescriptive habituation (same fixed technique switched on more often). No model-accuracy or planning-completeness series is supplied to break that equivalence, leaving the trial-and-error stagnation rival alive. (raised by mokyr)
- **[rival]** No retrievable source this turn supplies the realized post-landing autonomy-retrofit count: the number of MER/Curiosity/Perseverance flight-software builds that materially changed AutoNav/ENav driving policy after landing, the sol of each uplink, or the attributable change in autonomous-drive fraction or meters-per-sol. AMOS, ACTA, and Space Economy corpora returned zero on rover-specific autonomy version histories; OpenAlex/NTRS/Crossref returned only adjacent Mars 2020 instrument/science-campaign papers, no flight-software-version record. The retrofit premise therefore remains asserted, not estimated; REFUSED rather than confabulated. (raised by mokyr)
- **[governance]** No retrievable source this turn discriminates technical infeasibility (Explanation A) from incentive-driven suppression (Explanation B). The mission-assurance review records, driving-policy-uplink sign-off authority, abort/safing-risk gate thresholds, and the technical-feasibility-vs-risk-acceptance basis of rejected or never-proposed autonomy upgrades were not present in any corpus or vault source queried. Without that record the principal-agent discriminator cannot be supplied; REFUSED. (raised by mokyr)
- **[economics]** The retrofit-path expected-loss term Q3 ultimately demands (residual probability that an authorized driving-policy uplink degrades or loses a landed asset, with magnitude drawn from actual review records and documented uplink anomalies) was not retrievable this turn. The certification-cost term is shown to exist (claim mokyr_r2_c3), but the rover-specific magnitude and the anomaly-based loss probability are unmeasured here; that portion is REFUSED rather than invented. (raised by mokyr)
- **[identification]** No retrieved source establishes that the candidate's PDS command logs or NTRS Sol-by-Sol AutoNav reports contain an exogenous shifter of the autonomous fraction (downlink window, staffing, Earth-Mars geometry, or a documented operations-policy autonomy-posture directive) usable as an instrument, nor any result showing the H1 coefficient survives instrumenting on it. Identification of the machine-competence channel separate from the reliance-calibration channel is therefore unresolved on available evidence. (raised by parasuraman)
- **[mechanism]** No retrieved source shows that, in the candidate's PDS/NTRS data, the autonomous-distance share is separable from the commanded drive-distance ceiling and the conservatism of the commanded route. Without that separation the meters-per-sol gain cannot be attributed to the software channel rather than to the ground team commanding bolder drives as trust grows. (raised by parasuraman)
- **[measurement]** No retrieved source demonstrates a competence-net outcome (meters per sol after charging autonomy-attributable faults, anomalies, safing, and aborts) constructed from PDS fault/event records, nor any robustness check of H1 against it. Whether the headline coefficient survives netting the commission/fault tail is unresolved on available evidence. (raised by parasuraman)
- **[measurement]** No retrieved source (AMOS=0, ACTA=0, Space-Economy=0, NTRS=1 unrelated, OpenAlex=general autonomy surveys only) supplies the candidate's dated PDS/NTRS within-rover time path of the autonomous fraction against sols and against cumulative autonomy-attributable aborts/faults/ground-overrides. The claim that the autonomy effect does or does not survive conditioning on a trust-accrual trajectory cannot be asserted; the candidate owes this empirical construction. Refused per no-confabulation contract 3.3. (raised by parasuraman)
- **[mechanism]** No retrieved source supplies the NTRS/PDS engineering decomposition of AutoNav versus Enhanced Navigation across the four processing stages. Whether the claimed productivity gain attaches to a stage-4 (action-implementation timing / think-while-driving) change or to a stage-2/3 (analysis/decision) capability change cannot be settled from retrieval; the candidate owes the stage-by-stage type-and-level table. Refused per contract 3.3. (raised by parasuraman)
- **[identification]** No retrieved source supplies the PDS command-log / NTRS operations-report data needed to construct a commanded-drive conservatism measure (command sparsity, route-margin, post-anomaly blind-segment insertion) or to test whether the autonomous-fraction coefficient is separable from a time-varying ground-team conservatism response. The separability finding cannot be asserted from retrieval; the candidate owes this dynamic-confounder construction and test. Refused per contract 3.3. (raised by parasuraman)