Research
An Audit-Repair Episode Already in Context Makes an LLM Verifier Measurably More Lenient, in 15 of 15 Configurations
With the present task held byte-identical, a completed audit→repair episode sitting earlier in the model's context lowers false alarms on human-verified-correct ProcessBench traces in all 15 model-by-wording combinations, by 2.8 to 11.5 percentage points against a length-matched non-audit control (a 9-25% relative reduction). Signal-detection analysis locates the shift in the decision threshold rather than discrimination — the criterion moves in 15 of 15 and survives correction in 13, while d′ survives in none — and the effect persists with reasoning enabled. For anyone wiring a checker model and a fixer model into one pipeline, prior turns are silently retuning the checker's threshold.
↳ Follow the thread