chore(skill-evals): commit the audited skill eval set and measurement harness - #457
Draft
hyesungoh wants to merge 3 commits into
Draft
chore(skill-evals): commit the audited skill eval set and measurement harness#457hyesungoh wants to merge 3 commits into
hyesungoh wants to merge 3 commits into
Conversation
|
|
Hi! I'm the It looks like you correctly set up a CI job that uses the autofix.ci GitHub Action, but the autofix.ci GitHub App has not been installed for this repository. This means that autofix.ci unfortunately does not have the permissions to fix this pull request. If you are the repository owner, please install the app and then restart the CI workflow! 😃 |
autofix.ci
Contributor
|
Size Change: 0 B Total Size: 97.9 kB ℹ️ View Unchanged
|
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## main #457 +/- ##
=========================================
Coverage 100.00% 100.00%
=========================================
Files 58 58
Lines 1664 1664
Branches 500 500
=========================================
Hits 1664 1664 🚀 New features to boost your workflow:
|
…ply the merge-review findings - packages/plugin/skills/react-simplikit/evals -> skill-evals/ at the repo root: everything under packages/plugin is fetched by the documented install commands, and the skill directory itself is what future measurement runs snapshot, so the answer key must live outside both - restore the auditor's per-eval audit_note (grading rulings + fixture specs) in evals.json - commit the missing harness assets: arm/grader prompt templates and the three fixture apps (node_modules excluded; the two knowledge evals' versions stay pinned in package.json) - README: disclose the 54/56-vs-53/56 pre-audit counterfactual, the 7/10 prompt reuse, the single-seed Opus grader, and the contamination rules; scope the findings to fd312f5 - check_imports.mjs: stop flagging type-only exports (public-exports.json keeps a 'type ' prefix that the import parser strips); add collect_public_exports.mjs so the list is regenerable - exclude skill-evals/ from VitePress pages and llms.txt
hyesungoh
changed the title
…d layout The template was committed as used in iteration 2, still pointing at that session's scratchpad for the mechanical import check and at an absolute home path for the ground-truth source — so the one step that produces every published number could not be re-run from the repo. Paths are now repository-relative, and the README documents how to produce the eval_metadata.json the grader reads.
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
3 participants
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.
Commits the consumer-skill evaluation set and its measurement harness at the repository root (
skill-evals/), so the artifacts stop living only in a session workspace. Nothing here runs in CI;yarn test:skillis unaffected.Two reviews (OSS-maintainer lens, benchmark-integrity lens) ran against the first revision; both verdicts were "merge only with changes", and this revision applies them:
packages/plugin, and future measurement runs snapshot the skill directory itself — the answer key must live outside both, so it now sits at the root asskill-evals/.audit_noteper eval (binding grading rulings + fixture specs that the first revision dropped).node_modulesexcluded; the knowledge evals' package versions stay pinned).check_imports.mjstype-export false positives and addedcollect_public_exports.mjsso its input is regenerable.skill-evals/from VitePress pages and llms.txt.Contents
evals.json— 10 audited prompts, 56 assertions (5 discovery / 2 knowledge / 3 out-of-scope traps), written by a repo-blind author and audited by a catalog-aware agent with reword/strike-only powers.harness/— arm & grader prompt templates, three fixture apps.scripts/— trigger measurement over the whole event stream, mechanical import checks, per-bucket rollup, public-exports regeneration.README.md— method, both iterations' results with the margin counterfactual, re-run guide.Iteration-2 headline (2026-08-31, no-install fixture)
Trigger accuracy 6/10 in an isolated environment. Sharpest finding (as of
fd312f5): SKILL.md states no package version, and the skill arm declared invented^1.xranges in 7 of 8 manifests it touched — follow-up is to stamp an install line into SKILL.md at generation time.