Skip to content

Add research harness validator and auditor skill - #7

Open
a-green-hand-jack wants to merge 4 commits into
deusyu:mainfrom
a-green-hand-jack:codex/research-harness-validator-upstream
Open

Add research harness validator and auditor skill#7
a-green-hand-jack wants to merge 4 commits into
deusyu:mainfrom
a-green-hand-jack:codex/research-harness-validator-upstream

Conversation

@a-green-hand-jack

Copy link
Copy Markdown

Summary

  • add a deterministic ML research harness drift validator
  • add a research-repo-auditor skill for semantic claim/evidence audits
  • add the ML research repo structure note with the two-layer anti-drift mechanism and update thinking counts/navigation

Validation

  • python3 -m py_compile scripts/check-research-harness.py
  • python3 /Users/jieke/.codex/skills/.system/skill-creator/scripts/quick_validate.py skills/research-repo-auditor
  • bash scripts/check-consistency.sh
  • fixture smoke test on the original branch: valid minimal ML repo passes; missing linked-claims.yaml fails as expected

QLing-yes pushed a commit to QLing-yes/harness-engineering that referenced this pull request Jun 27, 2026
…n and critical analysis

Add the seventh article to the harness engineering reference index — Anthropic's
"Scaling Managed Agents: Decoupling the brain from the hands", which introduces
the meta-harness concept. Includes full Chinese translation and a thinking piece
examining five tensions between meta-harness architecture and existing frameworks.
ilove928612-lang added a commit to ilove928612-lang/harness-engineering that referenced this pull request Aug 13, 2026
- 还原实验结构:Terminal-Bench 2.0 6 档资源配额,1x→3x 修可靠性(p=0.40),3x+ 性质突变(成功率跳 4pp)
- 五点架构启示:INFRA.md 一等公民、3x 折中启发式、资源配额=价值观锁定、运行时污染资源侧补全、时间维度噪声
- 三开放问题:任务集=噪声放大器、折中点基准无关性、紧配额策略可迁移性
- 交叉索引:与 #38/#34/#61/#47/#35/deusyu#23/deusyu#2/deusyu#7 八篇关联文章互链
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant