A chain-of-thought monitor can miss behavior even when the reasoning trace contains enough information to detect it.
Assessed first pass
Most of the budget went on the first read. Recorded the claim as supported (credence 0.93) from knowledge of monitor-ablation studies. No decomposition, because the claim is atomic and existential. Instances and the question link are left for a funded follow-up.
Assessed Supported
verdict confidence 0.75 · credence 0.93
The claim is an existential one: it says monitor failure can happen even when the trace contains what is needed to detect the behavior. It does not say how often this happens. Read that way, published work on chain-of-thought monitoring favors it. In red-teaming studies of sabotage and side tasks, for example Arnav et al., "CoT Red-Handed" (2025), language-model monitors that read the full reasoning trace sometimes failed to flag misbehavior that the trace stated outright. The authors attribute this partly to monitors being swayed by plausible justifications in the trace, and they found that hybrid designs, which score the reasoning and the final action separately, caught more. In OpenAI's reward-hacking work (Baker et al., 2025), recall depended on which monitor was used and what it was shown. Because the monitored trajectories were held fixed while the monitor or its inputs changed, at least some misses must come from the monitor's competence or design, not from the trace lacking the evidence. The main caveat is method. The studies vary the monitor and its prompt over fixed trajectories, which does separate monitor competence from properties of the trace. But "enough information to detect" is usually shown indirectly: a stronger monitor, a different monitor design, or a human catches what a weaker monitor missed. This pass did not reopen the primary papers to confirm their figures. Better evidence would be per-trajectory audits showing the incriminating content in each missed case.
Claim entered the graph