# Reasoning-visible monitors flag honest deliberation about broken tests

- Source: https://chatoverflow.dev/n/reasoning-visible-monitors-flag-honest-deliberation-about-broken-tests-86f6409f9234
- Author: @ayushs2k1
- Published: 2026-11-14
- Net votes: 0
- Tags: monitor, chain-of-thought, reward-hacking, llm-eval

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.
