Repository navigation
feat: OOD corpus (73 real skills) + false-BLOCK gate - #3
Merged
Merged
Conversation
Out-of-distribution evaluation using partial vendored packages from superpowers, ponytail, vercel agent-skills, anthropic skills, and awesome-claude-skills (safe-labeled). - dataset/ood/safe/** + ood_catalog.jsonl + SOURCES attribution - eval --suite ood with safe FPR ≤ 0.05 and n ≥ 40 - docs/BENCHMARKS.md with honest protocol - Precision fixes from real-world FPs: chmod 600, child_process sh, .env co-occurrence, bare GITHUB_TOKEN severity - CI runs OOD gate; 40 tests green; 0/73 false BLOCKs Next priorities remain PyPI + policy packs.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Adds out-of-distribution evaluation: real Agent Skills from public repos, labeled safe, gated on false BLOCK rate.
Also includes precision fixes discovered while scanning those skills (so OOD is usable, not a wall of FPs).
Corpus
dataset/ood/safe/dataset/ood/SOURCES.mddocs/BENCHMARKS.mdMetrics (this branch)
FP precision fixes (still adversarial-green)
chmod 600no longer setuidchild_processneeds strong sink.env+ network needs exfil shapeGITHUB_TOKEN→ MEDIUM;${{ secrets.* }}stays CRITICALExcluded
anthropic/claude-api(multi-lang docs pack, not a typical skill)Test plan
uv run pytest -quv run python eval/run_eval.py --suite all --check --details