Record the cap64 reversal once the larger benchmark finished
Constraint: Strategy guidance must now reflect that cap48 and cap64 produce different winners under verified runs Rejected: Keep high_energy as the generic default | The completed cap64 run shows hybrid winning clearly at a larger subset size, so the docs must acknowledge scale sensitivity Confidence: high Scope-risk: moderate Directive: Do not present a single global default strategy again until bucketed and style-aware benchmarks explain the cap48/cap64 divergence Tested: Verified cap64 report.json, progress.json, high_energy eval.json, and hybrid eval.json; confirmed cap64 winner=hybrid with top1 0.875 vs high_energy 0.625 Not-tested: Multi-seed cap64 aggregates, bucket/style-aware benchmarks, and any revised hybrid training design
Showing
4 changed files
with
34 additions
and
10 deletions
-
Please register or sign in to post a comment