fix(eval): deterministic self-contained eval runners + sharpened java held-out corpus#74
Merged
Conversation
…s scoring (#72) The runners injected the valid label set as opaque IDs (cat#1..5) but not their definitions. The headless scan is sandboxed to the temp work dir, so the model couldn't read baseline-categories.md to learn what the numbers mean — it thrashed on denied reads (intermittently never emitting the <<<EVAL>>> block, scored as PARTIAL) and guessed categories (cat#2-vs-cat#3 flips on the same fixture). Inject the category legend inline from baseline-categories.md (read-only, no sandbox loosening, no plugin-path discovery), tell the model the definitions are provided so it need not read skill files, steer it to reason directly on the single fixture, and strengthen "always emit the block." Symmetric across claude.sh and codex.sh — divergence between runners is a bug. java/held-out went from 0.33(±0.47)/PARTIAL to 1.00/1.00(±0.00), deterministic. Closes #71 Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
… (#73) The held-out java fixtures were agent-authored in #15 and never real-model validated; they were ambiguous, so a real model scored the split 0.25/0.25 and tripped the seen-vs-held-out overfit FLAG — a fixture-quality problem, not a profile-overfit one. Rewrite them as unambiguous, realized defects, each with a clean precision twin: - BugPathTraversal: realized Files.readAllBytes(...) of the user-controlled path (was a bare new File(...) constructor a model reasonably won't flag) -> cat#3 - BugSwallowedInterrupt -> BugSwallowedParse: swallowed NumberFormatException to a sentinel, no concurrency flavor, so no cat#2-vs-cat#5 ambiguity -> cat#2 - CleanSafePath: path read behind a strict whitelist (precision twin for cat#3) - CleanReportedParse: parse error propagates, nothing swallowed (twin for cat#2) Recalibrate baseline.held-out.txt to the measured 1.00/1.00 (runs=3, ±0.00, clean_fp_runs=0, gate PASS) now that the runners are deterministic (#71/#72). The seen-vs-held-out gap closes, so the overfit FLAG correctly clears. Closes #68 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.
Eval-harness reliability + corpus quality. Three eval-scoped changes since 1.8.0:
fix(eval): deterministic self-contained eval runners (#71 → #72)
The eval runners injected category labels as opaque IDs (
cat#1..5) without theirdefinitions; the headless, sandboxed scan couldn't read
baseline-categories.md, so itthrashed on denied reads (intermittently missing the
<<<EVAL>>>block → PARTIAL) andguessed categories (
cat#2-vs-cat#3flips). Now both runners inject acat#N = <definition>legend inline (read-only, from the repo) and steer the model to reason directly on the single
fixture. java/held-out went from
0.33(±0.47)/PARTIALto1.00/1.00(±0.00). Affects everylanguage's calibration, not just java.
test(eval): sharpen java held-out fixtures + calibrate baseline (#68 → #73)
The agent-authored held-out java fixtures were ambiguous (a bare
Fileconstructor for"traversal"; a swallowed interrupt that reads as cat#2 or cat#5), so a real model scored the
split
0.25/0.25and tripped the overfit FLAG — a fixture-quality problem, not profile overfit.Rewrote them as realized, unambiguous defects (realized path-traversal read → cat#3; swallowed
parse → cat#2), each with a clean precision twin, and recalibrated
baseline.held-out.txtto themeasured
1.00/1.00. The seen-vs-held-out gap closes; the FLAG correctly clears.docs(eval): CLAUDE.md model-free grading / eval-run details
Documents the two model-free eval layers (grader vs orchestrator) and
DEFECT_SCAN_EVAL_RUNNER.Gates: bats 147/147 (ubuntu + macos), runners POSIX-clean, no version/compat gates touched.