The larger earlier Q1 fit has conditional action-ranking signal despite late calibration drift. The joint QH attempt failed and supplies no completed objective comparison. These are different outcomes in the 7 October analysis.
On 20 exposed validation scenes with paired R0/cropped-R1 support and one fitted seed, final Q1 regret minus uniform is −0.004718 [−0.008551, −0.001345]. Regret measures the labelled gain lost by choosing the model's maximum. Final MAE minus the TRAIN-fitted constant is +0.004087 [0.001618, 0.006680]; numerical calibration worsens despite the action signal. Final minus selected regret remains uncertain. Fixed-weight scene intervals omit seed and selection uncertainty.
The constant predictor has a deterministic tie choice: its regret differs from uniform choice. Beating that constant alone would not establish useful ranking. The original figure and exact contrast table retain scene dots, paired intervals and selected/final roles.
Joint QH seed 41 failed after 2,467 of 15,400 intended updates. Three remaining cells were unlaunched. Fixed-label H1 Huber first decreases and then increases across three validations; mixed-horizon fitted loss rises against changing targets. That moving-target loss cannot serve as independent evidence of policy value or identify the failure mechanism.
The optimization diagnostics, partial-checkpoint evaluation and utility analysis are supporting views of this failed attempt, not three successful independent training experiments. The later Mac comparison has different training exposure and cannot be pooled with this earlier result.
Next: establish numerical stability with frozen controls before a score-only versus joint-feasibility comparison. Preserve failed-attempt charges. Precision recovery is a new attempt with its own receipt; it cannot convert this failed run into a completed comparison.
The scorer archive owner binds the
canonical snapshots and native source admission.
ScientificReportConfig with scorer_study, its
domain builder and read_report_bundle preserve tables,
figures and checkpoint roles. Public summaries support this historical
account; unavailable failure-time diagnostics remain explicit. No
retained living QH mission is supplied in this publication.
| Finding / evidence | RQ | Thesis section | Coverage / gap |
|---|---|---|---|
| Q1 calibration versus decisions | RQ3 | sec:thesis-results | Covered as bounded development evidence |
| Failed continuing-value fit | RQ3 | sec:thesis-results | Partial: failure described; completed objective/policy comparison missing |
Previous primary report retains its complete narrative, exact tables and original figures. This 15 October revision is an authored synthesis, not a new experiment or a change to the numerical snapshot.
Experiment index and evidence routes. Missing/private parents remain explicit in the native campaign owner; historical versions retain their original dates and scientific identities.
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.