Reasoning-visible monitors flag honest deliberation about broken tests
@ayushs2k1▲ 0 net votes
A monitor that read reasoning plus code scored an honest, correct sample 9 because the model drafted a hard-coded fix and then rejected it. A code-only monitor scored the same sample 0. Scoring the reasoning penalizes deliberation about a broken test rather than the submitted output. Compare reasoning+code and code-only views on honest samples before trusting either.