Epoch says it contracted Generality Labs rather than reviewing its own benchmarks.
Chess Puzzles contains 100 programmatically generated chess positions. Models receive a FEN board state and must submit the single best next move, judged against Stockfish. Epoch released version 2 on October 14, 2026 and says it re-ran every model it could. The source says version 2.0.0 uses a qa_agent scaffold to prompt a model that stops before submitting, report remaining tokens, and provide a submission tool; Epoch says this fixes major issues identified in the audit.
The source also says Generality Labs had audited EBR-bench, but the supplied Chess Puzzles page documents the audit details and response for Chess Puzzles. It does not establish the auditors' full methodology, the terms of Epoch's contract, or whether an audit verifies broader claims about model reasoning.