Files
Trevin Chow 8985e68861 fix(ce-optimize): confirm the selection's own gain on a disjoint holdout
Address PR review feedback (#1708):

- experiment-prompt-template.md, loop.md, experiment-log-schema.yaml: a
  remote worker commits its mutable-scope change as the candidate
  (head_sha), then commits result.yaml alone on top. Collection requires
  the tip to add only result.yaml with head_sha as its parent; the keep
  merges head_sha. The old protocol asked result.yaml to name the commit
  that contained it, which no worker could satisfy.
- decide.mjs: held-out confirmation agrees only when the holdout improves
  an objective the selection comparison improved; a holdout gain in some
  other objective no longer confirms the keep. Regression test added.
- optimize-spec-schema.yaml, loop.md, spec.md, guide: the
  confirmation_seed judge sample is drawn only from items the selection
  sample did not draw, so a different seed alone no longer counts as held
  out; an output with nothing left cannot be confirmed that way.
- wrap-up.md: the final confirmation measures the original baseline
  revision against the final state, and repeats that pairing on the
  holdout, instead of leaning on per-keep incremental holdout pairs.
2026-09-27 20:33:48 -07:00
..