Polish: audit follow-ups (CLAIMS wording, evidence hygiene, cwd-robust scorers)#1
Merged
Conversation
…t scorers) From the artifact stress-test audit. No reproduced number changed; all data was already correct — these fix wording, evidence-explorer hygiene, and scorer portability. - CLAIMS.md C11: reword escalation economics as three distinct options over different doomed sets (Haiku-doomed ~$2.20; Sonnet-doomed ~$3.07-$7.51), not a single sequential cascade (the paper's Table 11 was already correct). - CLAIMS.md C12: baseline is the signaled-budget cell, not "Floor"; state the precise +6.7pp lifts (Sonnet 37.3->44.0%, codex 20.7->27.3%). - CLAIMS.md menu row: name the source JSONL for every cell, including the leak-free "quality 77" file (exp4_lever_floor100_siginject.jsonl) that was unnamed. - build_evidence_index.py: add "ratelimited" to SKIP so the quarantined RATELIMITED-MIXED held-out file no longer surfaces in the evidence explorer; regenerate evidence_index.json. - test_runtime_e2e_replay.py: fix stale comment that named a nonexistent *.CONTAMINATED-MIXED.jsonl (actual quarantined file is *.RATELIMITED-MIXED.jsonl). - agreement_signal.py, value_of_resolve.py: anchor evals/ globs on the repo root so these standalone-invokable scorers reproduce from any working directory. Verified: 12/12 claims reproduce; both scorers produce correct numbers from a foreign CWD; offline suite 483 passed, 1 skipped, 3 deselected. Co-Authored-By: Claude Opus 4.8 (1M context) <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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Follow-ups from the artifact stress-test audit. No reproduced number changes — every paper claim was already data-correct (independently recomputed from raw
evals/and cross-checked against the PDF). These fix wording, evidence-explorer hygiene, and scorer portability.Changes
exp4_lever_floor100_siginject.jsonl) that was previously unnamed in the evidence column.ratelimitedtoSKIPso the quarantined*.RATELIMITED-MIXED.jsonlheld-out file no longer surfaces in the public evidence explorer; regenerateevidence_index.json(that one entry removed). Note:mixed_workload_*is intentionally kept (it's a labeled experiment, not a corrupted file).*.CONTAMINATED-MIXED.jsonl(the real quarantined file is*.RATELIMITED-MIXED.jsonl). Test logic was already correct (exact paths, not globs).evals/globs on the repo root so these standalone-invokable scorers reproduce from any working directory (previously CWD-relative → empty results if run from elsewhere;verify_criteria.pywas unaffected since it runs withcwd=ROOT).Verification
benchmarks/verify_criteria.py→ 12/12 claims reproduce./tmp(foreign CWD).🤖 Generated with Claude Code