#import "../generation_evidence.typ": load-generation-evidence #let generation = load-generation-evidence() #import "../draft_markers.typ": development_only, scientific_core_todo, prune_todo, review_note, review_strike = Discussion #prune_todo( [Complete this interpretation when the confirmatory comparisons are available. The present discussion interprets bounded renderer, runtime, collection and one-step results; the retained development hypotheses are not results.], source: [@sec:thesis-results and the frozen analysis plan], gate: [Each interpretation identifies a reported observation, a justified scope, and any competing explanation; no missing outcome is treated as zero.], ) == What the available evidence establishes *Measurement is conditional on the acquisition protocol.* The completed renderer contrast supports candidate cropping only where full-scene selected observations, successor histories and endpoint scoring remain equivalent. Its earlier failure shows why faster rendering alone cannot qualify that change. The separate runtime diagnostic found too little preparation work to justify a larger mesh cache. Because the clocks cover different devices and nested operations, it identifies further measurements rather than a universal bottleneck or throughput guarantee. *Generation yield and independent experience differ.* The historical collection demonstrates retained continuations under its original evaluator. The newer collections add targets and states within the same development scenes. Neither a longer suffix nor another target draw adds an independent scene, creates dense future labels, or admits a new observation protocol. Successful input joins are necessary for fitting the declared objective, but cannot establish predictive usefulness. Summed costs from collections with different recipes also cannot replace the matched renderer comparison. The post-hoc census makes this distinction concrete: its incremental TRAIN chains broaden target and state exposure on already represented scenes. @tab:generation-posthoc-coverage retains the independent-scene denominator. Candidate support varies by family, and supported choices and finite immediate-gain variation contract at later recorded acquisitions. These are measured properties of the incumbent recipe. Scene geometry, target exposure, unequal quotas and conditional continuation coverage can all contribute; none is isolated as a causal explanation. Repeating supported choices within familiar scenes may help fitting, but unseen-scene usefulness remains a separate evaluation. *Ranking signal need not improve the selected action enough.* The frozen one-step screen missed its practical-effect threshold, while exploratory analysis found a ranking signal. Ranking many pairs correctly can coexist with an error at the maximum that determines the chosen view. The constant reference's tie rule further affects that choice. Post-hoc tie analysis helps design another screen; it does not overturn the frozen negative result. Neither absent useful signal nor representation loss has been established as its cause. The larger fixed-seed analysis further separates action choice from numerical prediction error. The final checkpoint has less regret than uniform choice on the exposed scenes, while its mean absolute error exceeds that of the constant predictor. Its regret advantage over the selected checkpoint remains uncertain. A model can preserve a useful ordering while assigning inaccurate gain magnitudes, so neither prediction error nor action regret substitutes for the other. These conditional observations do not identify state aliasing, an architecture limitation or poor collection as the cause, and leave the frozen checkpoint selector unchanged. *Support and completion condition the conclusions.* The earlier proposal audit remains a negative result for its own profile, even though later collection succeeded. Radius limiting retained inspected family support without proving wearer-motion realism or policy utility. Likewise, the B0/S endpoint contrast is conditional on shared-complete chains; its small population and undefined admission preservation leave the challenger unqualified. Retaining the baseline is therefore a bounded development decision, not proof of optimality. The interrupted continuing-value fit supplies a numerical failure, rather than a comparison of successful multi-step estimators. A rising loss against a changing bootstrap target can reflect target movement as well as prediction error; it is not an independent measure of policy value. Repeating a qualified fit with fixed-label checks and independent continuation outcomes is needed before interpreting the learning method, representation or collector as its cause. The principal held-out multi-step prediction and learned-policy endpoint results remain unavailable. An implemented observed-target route does not supply an evaluated actor-visible population. == What the unresolved comparisons must distinguish Measured reconstruction and actor information can evolve differently. After an acquisition, the oracle evaluation cloud includes the selected surface, while an actor restricted to compact root features and pose history may receive no corresponding geometric evidence. Two histories that acquired different surfaces can then appear identical to the scorer. This is a possible source of state aliasing, not an observed explanation of a learned-policy result that has yet to be measured. An oracle may also choose a continuation using information unavailable to the actor, so a gap to its choice need not be approximation error alone. On a fully labelled successor shell, comparing an actor-selected continuation with the oracle maximum isolates the consequence of the two selection rules under that actor. It does not identify either the best possible actor or an irreducible information gap. This distinction limits what a future finite-horizon result can establish without prescribing a larger architecture in advance. A policy may choose useful actions despite biased value estimates, so endpoint success requires a separate prediction and calibration analysis before it supports a claim about value accuracy. == Data collection before action-space replacement #review_note(kind: "scope", id: "X04", position: right, note: [Keep collection conditional on measured continuation coverage. Distinguish a matched collector robustness test from claims of reduced value bias or policy improvement.], )[Additional interaction is relevant when the available proposals are useful but their consequences are poorly represented in replay.] DAgger motivates collecting supervision at learner-visited states @DAgger-ross2011, and Ball et al. demonstrate retaining offline data during online off-policy learning @RLPD-ball2023. Oracle-labelled transition aggregation would use that distributional motivation, not implement the same imitation-learning algorithm or inherit its guarantees. The proposed comparison therefore includes an equal-cost exploratory-collection control on training scenes. // evidence: // - @DAgger-ross2011 -> https://proceedings.mlr.press/v15/ross11a/ross11a.pdf, Sec. 3, Algorithm 3.1 (learner-mixture rollouts and aggregation). // - @RLPD-ball2023 -> https://proceedings.mlr.press/v202/ball23a/ball23a.pdf, Sec. 4, especially 4.1 (offline and online experience within off-policy learning). The TRAIN-only analysis froze two collector nominations before validation was exposed. The first is a matched quota test, retaining all four families and the same proposal budget. Its frozen prescription is: #quote(generation.nominations.at(0).proposed_change.replace("G60 quotas", "proposal counts")) The second is a separate matched bearing-radius test: #quote(generation.nominations.at(1).proposed_change .replace("radius", "radius ") .replace("incumbent", "incumbent ") .replace("unchanged", "unchanged ") .replace("m with", " m with") .replace("m step", " m step") .replace("m height", " m height") .replace("→", " → ")) Both nominations subsequently received the bounded three-scene Mac pilot in @ssec:r1-collector-pilot. More labels were retained, but heterogeneous endpoints and the absence of a learning comparison leave a collector winner unestablished. Further independent scene/seed tests remain prospective. Preserve root, target and assignment matching, complete attempted cost and independent endpoints; testing quota and radius separately distinguishes allocation from geometry. The validation readback selected neither intervention. The matched Mac learning comparison in @ssec:mac-q1-comparison likewise shows why more trajectories on the same scenes cannot be equated with more independent experience. Its negative retention result and localized initial-state error motivate fixed-checkpoint calibration and feature-use controls before another architecture or optimizer claim. No active auxiliary-loss competition was needed for drift to occur in that comparison. #review_note(kind: "scope", id: "URS01", position: right, note: [Choose a recording-length or augmentation comparison only after aligning factual risk sets, acquisition costs and independently measured outcomes. Neither late support contraction nor record thinning establishes a learning benefit.], source: [@ssec:generation-supervision; source-bound research consolidation, 15 October 2026.], gate: [Freeze recording length separately from requested horizon, retain missing continuations, and qualify altered-history targets before a new fit.], )[Late support contraction does not identify an optimal recording length.] Selecting fewer training records can preserve the actual observations that produced each retained state. Skipping acquisitions changes those observations and may change subsequent actions. Such an altered history requires rebuilding the observed or fused state and recomputing affected rewards, targets, validity and returns. Joining factual transitions must retain elapsed-acquisition and horizon semantics together with intervening returns. The learning benefits of record subsampling and altered acquisition histories remain separate hypotheses. Changing the action space addresses a different limitation. GenNBV's continuous camera action and sequential coverage objective provide a precedent @GenNBV-chen2024, not evidence that continuous control will improve this thesis's target-specific metric. Such a comparison needs its own feasibility, observation, simulator, and cost assumptions. New experience, updated model weights, causal scene updating, and continuous actions are distinct interventions. The optional fusion comparison associated with @ssec:rq2 keeps model weights fixed. Changing the simulator or observation model would test a broader claim than generalization to new ASE scenes. It can change candidate feasibility, calibration, depth noise and reconstruction behaviour together. Those changes need separate specification and comparison before transferring conclusions about policy utility or numerical value accuracy. // evidence: // - @GenNBV-chen2024 -> docs/literature/tex-src/arXiv-GenNBV/3-Method.tex:13-25,76-95 (camera action, sequential observations and coverage reward). The next evidential step is therefore narrower than an architecture change: qualify the intended learning objective on the retained labels, evaluate predictions against independent references, and complete matched policy comparisons. The renderer endpoint validation already closes its bounded acquisition contrast; the collection audits verify stored endpoint bindings rather than independently recomputing their endpoint statistics. A supported negative result can answer the bounded planning question. Missing support, unreliable measurement, and unfinished runs cannot.