feat(eval): eval-run shortcut + runner label injection, calibrated baselines, full eval docs (#15)#69
Merged
Conversation
…cript (#15) (#58) Scoped pilot, claude runner, --runs 2: all three measured precision 1.00 / recall 1.00 (±0.00), 0 clean-FP. Sets precision_baseline/recall_baseline to measured; floors left conservative (erosion gate now effectively precision >= 0.90). Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
#59) Overnight sweep, claude runner, --runs 2: csharp, dart, java, kotlin, php, ruby all measured precision 1.00 / recall 1.00 (±0.00), 0 clean-FP. Sets *_baseline to measured; floors unchanged. Deliberately EXCLUDED (measurement artifact, not committed): rust/shell/yaml seen and java held-out scored 0.67/0.50 because the runner doesn't enumerate each language's valid category labels — the model emitted correct-line, correct-concept findings with synonym labels (panic->unwrap-panic, quoting->SC2086, coerce-> norway-problem, cat#2->swallowed-interrupt). Root-cause fix + re-measure tracked separately. swift came back PARTIAL (transient) and was not written. Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
…ow + safety) (#15) (#61) Single mental-model overview at the top of the canonical eval doc: the four parts (validator/harness/critic/runner), the corpus→runner→sentinel→grader→gate data flow, and the load-bearing safety properties. No new file — avoids a third source of truth drifting against the README and the design spec. Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
… (#63) Convenience entry point for the maintainer harness: locates detect.sh, auto-selects a runner (claude preferred, codex fallback) if DEFECT_SCAN_EVAL_RUNNER is unset, and forwards all flags to 'detect.sh eval-run'. README leads with the shortcut; CI sh -n covers it; bats verifies forwarding (stub) + the no-runner exit-3 path. Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
…/yaml/swift (#15) (#67) * fix(eval): runners inject the language's valid label set (eval-categories) (#15) The eval-mode prompt named 'a language-specific label' without enumerating the valid ones, so for custom-label languages the model invented synonyms (panic-> unwrap-panic, quoting->SC2086, coerce->norway-problem) that the exact-match grader scored as FP+FN — artificially depressing rust/shell/yaml to 0.67 and java held-out to 0.50 (a FALSE overfit flag). Runners now run 'detect.sh eval-categories <lang>' and inject the exact valid set. Verified: rust unwrap now emits '3:panic' (was 'unwrap-panic'), an exact match. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * chore(eval): record measured seen baselines for rust/shell/yaml/swift (#15) Re-measured with the label-fixed runner: all four now 1.00/1.00 (±0.00). rust/shell/ yaml were the 0.67 synonym-label artifacts (now resolved by injecting eval-categories); swift's earlier PARTIAL transient cleared. Completes measured SEEN baselines for all 13 languages. java held-out (0.25, persistent overfit FLAG) intentionally NOT committed — fixture-quality issue tracked separately; it keeps its conservative placeholder. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
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
1 participant
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
Completes the #15 self-improving eval: it's now real-model-validated, fully calibrated, and documented.
Genuinely-new since v1.7.0:
scripts/eval-runwrapper — auto-selects a runner, forwards all flags (feat(eval): scripts/eval-run shortcut for the harness (#15) #63).eval-categories) so the model emits grader-matchable labels (fix(eval): runner label injection + measured baselines for rust/shell/yaml/swift (#15) #67) — fixed a measurement bug that scored correct finds as FP+FN for custom-label languages.seenbaselines for all 13 languages at 1.00/1.00 (chore(eval): record measured baselines — pilot (python/go/react-typescript) (#15) #58, chore(eval): measured baselines for 6 more languages (csharp/dart/java/kotlin/php/ruby) (#15) #59, fix(eval): runner label injection + measured baselines for rust/shell/yaml/swift (#15) #67) — replacing conservative placeholders; the gate now enforces a measured erosion floor (~0.90 precision) instead of a guess.Validation
End-to-end against real models: every language's
seencorpus measured 1.00/1.00. The exercise surfaced and fixed a real harness bug (label injection) and confirmed the safety machinery (overfit FLAG + PARTIAL-refuses-baseline both fired correctly).Known follow-ups (filed, non-blocking)
Test plan
🤖 Generated with Claude Code