 alchain
|
5bfc6b48dd
|
v2: rubric expanded to 9 dimensions (SkillLens meta-skill validated)
Major changes (validated by 6 independent judge agents, 86.05 → 92.7, +6.65):
* Rubric: 8 → 9 dimensions, based on SkillLens paper (arXiv 2605.23899)
- dim3 失败模式编码 (10→12 weight): require explicit "if X fails → Y" branches
- dim5 可执行具体性 (15→17 weight): ban softening words (建议/可选/根据情况)
- dim9 反例与黑名单 (NEW, 6 weight): require "what NOT to do" lists
* Phase 1/2/2.5 checkpoints now have visible 🔴 CHECKPOINT · 🛑 STOP markers
(HL-1 lesson: LLM parsers scan visual markers, not 必须 prose)
* New auto-break rule: stop hill climbing after 2 consecutive rounds Δ<2
(HL-4 lesson: marginal gains signal over-engineering)
* New "darwin 操作反例黑名单" section: 8 anti-patterns from real踩坑
- same-context self-eval, git reset --hard, padding for scores,
skipping test-prompts, multi-dim per round, dry_run > 30%,
silent error skipping, ignoring dimension correlation cluster
* New HL speedrun: 4 high-leverage operations from results.tsv
(huashu-gpt-image +10.85 / huashu-weread-advisor +14.9 / claude-design +16.5)
* References split: empirical evidence + runtime neutrality moved to
references/ to keep SKILL.md under skill-creator's 500-line guideline
* SkillLens controlled study: 5 independent judges blind-tested V1 vs
degraded V2 huashu-research → 5/5 V1>V2 high confidence, Δ +46.5
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
2026-05-27 22:41:16 +08:00 |
|