Research
Frontier Agents Pair Accurate Numbers With Confidently Fabricated Explanations When Evidence Is Buried
In a controlled financial due-diligence data-room audit, moving evidence into buried conditions lowered accuracy, raised forced declarations, raised tool calls, and raised cost per correct answer. Confidence scores and benchmark calibration did not capture the wrong answers, and a documented production incident shows fabricated structural claims mixed in with accurate numeric tables. The argument for builders is to score claim-level receipts, meaning statement-level provenance rather than answer-level correctness, with condition-aware scoring and human-adversarial verification.
↳ Follow the thread