benchmark-audit

Reproducibility bundle for the paper 'One Prediction Set, Two Reported Results: Provenance Linkage and a Reproducible Benchmark-Audit Sequence for AI-Text Detection'. Seven-step audit sequence, executable checks, sanitized per-item predictions with SHA-256 manifest.

A label-recoverability audit of the BGL log anomaly benchmark, with paper, figures, code, and result artifacts.

Faithfulness audit of miniF2F-v2 (v2s): 23 Lean 4 statements that do not say what their informal problems say, including 8 where the v2 correction introduced the defect.

Faithfulness audit of ProofNet#: 30 Lean 4 statements that do not say what their textbook statements say, 28 with explicit witnesses.

Faithfulness audit of the miniF2F Lean 4 port: 23 statements that do not say what their informal problems say, 13 with explicit counterexamples.

Pre-registered census + provenance audit of the Turkish MMLU benchmark ecosystem (91-repo name-collision census, gating/license findings, 29-node provenance DAG) - paper + full evidence chain.

Faithfulness audit of ProofNetVerif (valid split): 10 wrong reference formalizations and 33 rows labeled correct whose formalization is wrong.

Faithfulness audit of CLEVER: seven Lean 4 specifications that do not say what their docstrings say, five of them machine-certified vacuous, with a script to reproduce the certificates.

Faithfulness audit of PutnamBench: 15 confirmed mismatches between the Lean 4 statements and the benchmark's informal problem statements.

Imla - rule-verifiable Turkish orthography benchmark for LLM evaluation (TDK-grounded caret probe; free-tier audit finds 61.1% caret drop). Pre-registered, zero-cost, paper + executable record.

An audit of the TRAIL benchmark's scorer. Both headline metrics divide by the gold count, so a predictor that emits every span and category scores 0.973 without reading anything.