What the collection experiments teach together

Reliable learning experience needs valid measurement, useful alternatives and supported continuation. Passing one of those tests does not establish the other two. This synthesis connects experiments; it does not pool their cohorts or reclassify historical pilots as confirmatory results.

Measurement first

The initial renderer comparison failed even with matching selected actions. The camera-stable follow-up checked observations, histories and endpoints and passed on its declared recipe. This establishes why a rendering speedup needs transition-level checks.

More labels are not more independent experience

The generation census adds target/state exposure on already represented scenes. Its bearing-family support gap motivated two separate interventions. The R1 pilot then retained more labels on three matched roots, with heterogeneous endpoint effects and no tested learning benefit. An older geometry-only support audit and later finite-label census answer different questions.

Late factual states contain fewer supported alternatives, but some historical H12→H16 completions still add gain. Recording length, requested prediction horizon and genuine termination must remain distinct. The post-hoc protocol specifies the missing common-risk-set cost/endpoint evidence rather than inferring an optimal cutoff.

Score accuracy and chosen action can disagree

The earlier Q1 analysis has conditional action signal despite calibration drift. The matched Mac study does not clear its frozen retention gates. The failed QH attempt's teacher discrimination and pair ordering do not establish useful continuing-value policy behavior. Backend, objective, exposure and checkpoint-role differences prevent one shared leaderboard.

What to investigate next

1. Matched collector coverage and independent endpoints: highest immediate value for RQ1 because census nominations now have a small executed pilot. Freeze separate quota/radius interventions and include all attempted cost.
2. Fixed-checkpoint calibration and feature-use controls: high value for RQ2/RQ3 because they distinguish observed drift from untested representation/optimizer explanations, before spending on another fit.
3. Factual-record thinning: feasible as a data-coverage experiment, with no established learning benefit. Preserve each retained record's actual observation history and horizon. Omitting acquisitions is a different intervention requiring reconstructed state and recomputed labels/returns.

The September evidence map, October generation map and scorer owner expose failed and unpublished diagnostic routes as well as primary reports. Archived orbit/motion work is discussed in the collection overview; its control-parity failure does not support adaptive-controller success.

Stable history and reuse

Current evidence routes: score-floor reanalysis, F24/G60 throughput screen, and historical QH matrix producer. The producer is historical and does not reproduce its figures on current main.

The full conceptual source preserves the original detailed narrative and historical alternatives. Existing reporting domain builders, cold readers and gallery entries remain the numerical/plot owners. Archived products whose current producer or raw inputs are unavailable are historical evidence, not reproducible current pipelines.

Finding → research question → thesis

Scientific snapshot and attempt identities remain those of the cited experiment. Report synthesis: 15 October 2026; restored to the domain/date owner: 16 October 2026.

Exact archived HTML source retains its original bytes; publication generates the historical view above at its original resource base.