Coarse Evaluations Reward Strategic Omissions

This proposed mechanism asks why human-AI editing chains may lose the exceptions and sources that keep a report’s meaning anchored. A fluent central claim can survive each retelling while small details that would limit later reinterpretation quietly drop away. The hypothesis predicts this pattern when the same participant expects continued evaluation and receives only a broad whole-story check. A proposition-level audit, which checks prespecified claims one by one, should retain those details. The comparison must cross both audit styles with a one-time handoff and hold reward, source access, time, length, wording, and policy constant. Support requires better predictions of new, held-out editing chains than calibrated models of ordinary task-dependent editing can provide.