dafang 0d046b5098 feat(experiments): P2 — onboarded seed deconfounds artifact metric (all 1.000) ...

- seed stage 06-onboard: .codestable/ skeleton (attention.md + execution
  conventions from the cs-onboard template) as a real evolution commit
- e2e_outcome scorer keeps a .codestable/ tree snapshot in evidence so
  artifact=0 can be triaged without manual reproduction
- rerun of 3 baseline segments (27 runs, zero errors): artifact 0.33/0.78/
  0.89 -> 1.000/1.000/1.000 — sonnet's low score was the un-onboarded seed
  (it was strictly honoring the skill's preflight), not process disobedience.
  Final cs-issue outcome verdict recorded: repair ability at ceiling; the
  process contract is the reliable measurable gain (27/27 vs bare 0/27) at
  +20-30% token cost; skill itself needs no change after two eval layers.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

2026-07-08 08:48:48 +08:00

..

stages

feat(experiments): P2 — onboarded seed deconfounds artifact metric (all 1.000)

2026-07-08 08:48:48 +08:00

build-seed.py

feat(experiments): P2 — onboarded seed deconfounds artifact metric (all 1.000)

2026-07-08 08:48:48 +08:00