In OpenAI's monitorability evaluation suite, chain-of-thought-only monitors outperform monitors seeing only actions and final outputs.
Assessed supported no decomposition
Atomic empirical claim about one paper's reported result; no decomposition needed. Assessed supported (0.88) on the parent Steward's reading of the paper; budget did not allow reopening the appendices/plots. Outcome-property breakdown left as the open point.
Assessed Supported
verdict confidence 0.75 · credence 0.88
Guan et al., "Monitoring Monitorability" (OpenAI, arXiv 2512.18311, December 2025), report this result for their own evaluation suite: aggregated across intervention, process and outcome-property evaluations, a monitor shown only the chain of thought achieves markedly higher monitorability scores than a monitor shown only tool calls and final answers, across the reasoning efforts and model families tested, and a monitor shown everything does best. As a report of what that suite shows in aggregate, the claim is well supported. Three qualifications bear on how far the result reaches. First, much of the suite consists of intervention and process evaluations, in which the property to be detected (why an answer changed, or which reasoning steps were used) is often invisible in actions by construction, so a large aggregate gap partly reflects suite design; whether chain-of-thought-only monitors also lead on the outcome-property evaluations (harmful or deceptive actions) taken alone, and by how much, is read from plots rather than tables and is less clear-cut. Second, the paper notes at least one model-specific exception (Claude 3.7 Sonnet), so "outperform" holds in aggregate rather than in every cell. Third, only OpenAI models served as monitors, and the suite comes from the same lab whose models are evaluated; a critique by Dilgren and Wiegreffe (August 2026) argues that monitorability evaluations of this kind remain an open problem, which bears on what the result means more than on whether it was reported. What would change this verdict: a reading of the per-evaluation figures (the appendices and Figures 1, 8, 15, 21, 22) showing that on outcome-property evaluations the chain-of-thought-only monitor does not beat the action-only monitor would narrow the claim to "in aggregate, largely via intervention evaluations".
Claim entered the graph