Corrected result and evidence index
Raw file names and content are unchanged. RevisionEvaluation_Local_20260910/Results/full is the corrected 57-file, 2,850-episode full comparison. RevisionAdditional_20260911/Results separately contains the new-seed comparison, RTG sensitivity, BC diagnostic, and Play-mode timing.
Key original reports under RevisionAdditional_20260911/Reports:
| File | Content |
|---|---|
existing-statistics.json |
Full50 descriptive statistics, effect sizes, intervals and 126 Holm-adjusted comparisons |
heldout-statistics.json |
New200 descriptive statistics, effect sizes, intervals and 54 Holm-adjusted comparisons |
rtg-statistics.json |
72 RTG/BC-reference contrasts |
bc-diagnostic-effect.json |
Six normal-minus-legacy BC diagnostic contrasts |
heldout-reward-components.json |
Return residual analysis and its limitations |
additional-verification-retry1.json |
Completed expanded Unity parity verification, 96 actual input cases |
pytorch-additional-parity.json |
PyTorch verification for four checkpoints, 24 cases each |
latency-benchmark.json |
Standalone inference measurements, including raw call times |
control-timing-summary.json |
Play-mode timing summary |
models-before-additional.json, reference-weight-manifest.json |
Evaluated model and reference-weight provenance |
final-audit.json |
Historical experiment/code/config audit; not an audit of a newly generated manuscript |
RevisionEvaluation_Local_20260910/Reports includes the completed full-phase result, initialization/model-contract checks, model inventory and preservation evidence. Reports of smoke/repeat checks are historical verification metadata; the smoke/repeat raw episodes are not part of this release's full-performance cohort.
Manuscript_Figures contains the already-generated figures copied from the manuscript revision. They are presentation artifacts of the recorded evaluations, not new runs. The source generator also edits manuscript files and must not be invoked merely to recreate figures.
Older work notes and reports preserve what was known at their recorded time. In particular, previously written data-availability notes may predate recovery of the C/model archives during package preparation. Use the top-level README and per-folder availability statements for the current package inventory.