Post-hoc protocol v1 — frozen before computation

Source: S01 cuda-q1-first-v1, producer 21baeeba9af979f3825de4deca78ee5e4a0e5ab2, retained evidence-access map and hashes. Primary final checkpoint; secondary validation-selected checkpoint. No reselection. Native history is observational chronology. No CUDA/Mac causal comparison.

Authenticate all13 pinned metadata files and both checkpoint identities, frozen source, dataset ordering and terminal receipts. Native domain admission precedes numerical reporting. Reject changed roles/support/identities.
Admit selected/final original-label validation rows and recompute their native reductions. New fixed-weight eval uses authenticated weights and full frozen TRAIN/dev stores, eval mode, float32 model/geometry, CUDA when available. CPU parity checks use a bounded same-input sample if feasible. Freeze evaluator source identity separately from original producer. Failed population admission stops that numerical branch and records the exact contract.
Stable grain: attempt/checkpoint/split/physical scene/native chain/root/factual acquisition index/candidate identity. Compare identities before values, report complete planned/admitted/excluded support. Do not pair independent generated candidate tables by index. Compare same native rows using one joint mask.
Candidates average within state, states within scene, scenes equally. Preserve native training weighting in a separate column. Root/later and acquisition position/family/scene/label-spread views are descriptive, not independent trials. Full-TRAIN versus dev is an unpaired distribution contrast. The constant is fitted from TRAIN only using the existing scene-equal fitter; uniform choice and deterministic constant ties are separate controls.
Metrics: Huber(delta1),MAE,MSE,signed residual,state-offset squared,centered residual variance,prediction/label spread,Huber-region support,pairwise ordering, selected regret,best-action margin and ties; root/later error concentration. Verify MSE decomposition algebraically before aggregate reduction. Retain zeros, missing, nonfinite and unsupported states separately.
Paired selected/final development contrasts use physical-scene bootstrap 95% intervals,10000 draws,seed20261015. No candidate-row bootstrap or training-seed uncertainty. Conditional sparse/degenerate support remains descriptive.
Inspect actual retained output/gradient telemetry before claiming it absent. Successful updates differ from attempts; retain source row ordinals and multiplicities. Fixed checkpoints cannot recreate historical optimizer paths.
Fixed-weight feature probe, conditional on successful full-input admission: use a fixed TRAIN scene roster selected by native sorted identity before loss inspection (first three eligible scenes, root and last recorded state). Compare ordinary forward with exact repeated control; bounded activation traces; source-qualified removal of selected-depth evidence only if the native actor interface can express consistent availability/causal prefix. Preserve poses, acquisitions, labels and hard support; label this an out-of-distribution reliance probe, not an alternate acquired state or retrained architecture. If no safe interface exists, report that limit rather than fake an ablation. Optional no-update gradients use this same roster and restore RNG,modes,buffers.
Historical Mac/Q1/QH results retain original roles/populations. QH extension only when it changes a named next-run decision; moving fitted targets do not become policy metrics. No new fit, selection sweep, collection or optimizer step.

Outputs: typed domain-admitted tables and immutable reporting snapshot; concise answer-first report with causal limits, thesis/code reading map, source appendix, telemetry availability matrix and at most three next interventions. A confirmed implementation defect gets its smallest owner patch, regression and rerun proof. No architecture/default/hyperparameter change is inferred solely from correlation.

Telemetry coordinate clarification (before final telemetry reduction)

The decoded transport carries repeated diagnostics across later logging rows. For train-diagnostics/* and same-cadence train-output/*, group by the explicit train-diagnostics/actual_updates, require identical observed values within each coordinate and retain the first source row plus transport multiplicity. Conflicts stop that metric rather than being averaged. This defines unique captured updates, not independent training replicates. Optimizer-callback fields lack that same coordinate: retain transport ordinals and observed extrema, but do not infer a fresh observation rate or align them using a carried task-diagnostic counter. Preserve native trainer/global_step, epoch and _step alongside source rows.

Feature-probe realization (before probe execution)

Use final weights and the first chain of each of the first three lexically sorted TRAIN physical scenes. Keep the complete admitted16-state chain for each forward; summarize root and last state. Withhold selected depth returns by replacing only depth_m and valid_mask with zeros. Keep calibrated cameras, factual prefix identities, selected poses, history, support, horizon and labels unchanged. Native prefix validation still applies; no valid return contributes a surface/free vote. This is a missing-return input-reliance control, not an omitted acquisition or physical counterfactual. Root has no selected observation and is the negative control. Use repeated ordinary forward and bounded native evidence capture; require repeat and root-control agreement within rtol1e-5/atol1e-6. Inspect CPU/GPU agreement on the first complete chain at the same tolerance. Record actual deviations/choice changes even when tolerated. Verify untouched parameter and buffer bytes, training flags and RNG. Do not compute a feature-importance or population benefit from three selected scenes. No optimizer or backward step.

Follow-up on the same fixed roster: distinguish no measured support from no in-map query support by inspecting the native volume states and sampler. Retain surface/free voxel counts, valid input-return counts, integrated observation identities, in-map target/segment fractions and feature availability, separately for materialized and action-valid rows. No scene expansion or outcome-selected replacement. The first parity probe failed its frozen tolerance (maximum3.2e-6); retain that failure, measure full development discrepancies and choice changes, and keep fresh comparisons on their declared GPU evaluator. Do not widen the tolerance or claim backend equivalence.

Full frozen-population query-coverage census (before census execution)

The same-roster follow-up found a state with fused selected evidence but no in-map action-valid query. Census all145TRAIN and20development chains at root/last state, using the original fixed chart, target lattice and segment fractions. Reuse native initialization, query construction and voxel sampler, without encoder/prediction calls, fitting or label changes. The chart geometry is constant through a chain, so initial-volume geometry applies at later states. Retain each physical scene/store/rollout/state identity; count materialized and action-valid candidates separately. Report27target-query in-map fraction,16segment-query fraction and the number/fraction of valid candidates with no in-map query at all. No actions means missing candidate fractions, not zero coverage. Report chain, state,candidate and physical-scene support and distinguish a finite-population census from extrapolation. Verify the same six scene/state geometry observations against the earlier probe. Exact zero is a diagnosed inaccessible input path; no arbitrary positive coverage threshold or causal learning benefit is selected.

Frozen evaluator repeat (before repeated inference)

Independent final review identified four package digests across the first six fresh receipts: the package tree changed while unrelated analysis/report modules were being authored. No pre-inference file inventory establishes which edits were irrelevant, so that batch is superseded rather than treated as one frozen source. Repeat all six declared evaluations under one package digest, and stop if the package or producer script changes during any evaluation. Bind that digest in all receipts and the manifest; the publisher rejects mixed or unguarded batches. Do not change checkpoint roles, labels, populations, metrics, intervals or select new outcomes. Keep the original batch privately for audit. This is a provenance repair, not an independent training replicate or a positive parity result.

Descriptive geometry strata (before the stratum reduction)

Using already produced scalar state metrics and the declared geometry census, compare root states by exact zero/nonzero target-query access and last states by all versus not-all action-valid candidates lacking every query. Retain original scene/store/rollout/state identities and selected/final/control roles. Report state/chain/scene/candidate denominators and scene-equal descriptive MSE,MAE, regret,uniform regret,prediction/label spread within each stratum. Do not fit a threshold, select a checkpoint or compute a causal effect from these different scene/difficulty groups. Sparse groups stay descriptive; this reduction can weaken the inference that inaccessible queries dominate aggregate error.