mirror of
https://github.com/garrytan/gbrain.git
synced 2026-07-29 19:01:39 +00:00
Garry TanandClaude Opus 4.8
8464691496
Surfaced by the SkillOpt real-LLM eval (Track B). The reflect step was shown
only a pass/fail score and the agent transcript — never WHAT the benchmark
judge rewards. On a skill judged by structure (e.g. "must include a
Confidence: line") the optimizer proposed plausible-but-off edits ("close with
a synthesis") that never satisfied the literal check; every candidate scored 0
on D_sel, the validation gate rejected them all, and the skill text never
changed (optimized === baseline === 0).
Fix: render each benchmark Judge (rule checks / llm rubric / qrels) into
plain-English criteria via new exported describeJudge / describeJudges, and
thread them into the reflect prompt (a SUCCESS CRITERIA block) for both the
loop reflect calls and the one-shot-rewrite path. The orchestrator computes the
distinct criteria across train+sel+test once. The optimizer system prompt now
instructs it to satisfy the criteria through genuine content, never empty
keywords — reward-hacking stays defended by the independent held-out gate
(cat32 confirms the gate catches a keyword-stuffing hack).
End-to-end this took a deficient skill from 0.00 to 1.00 on a held-out set it
never trained on. Pinned by test/skillopt/reflect.test.ts (describeJudge per
kind, describeJudges dedup, criteria present/absent in the prompt). Folds into
the open v0.42.9.0 PR (#1759).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>