Evidence that chain-of-thought is mechanistically faithful does not by itself establish its behavioral monitorability, nor vice versa.

The claim says the two properties need separate evidence. Mechanistic faithfulness asks whether a chain-of-thought reflects the computation that actually produced the model's output. Behavioral monitorability asks whether a particular monitor, under a particular protocol, can detect a target behavior from the trace. The claim holds that showing either one does not, on its own, show the other. The qualifier "by itself" matters. The claim does not say the properties are unrelated, or that evidence for one never makes the other more likely. It says only that each needs evidence of its own.

The claim is mostly supported by how the two properties are defined. Monitorability depends on the monitor: its capability, what context it is given, the baseline it is compared against, and how hard the task is. Faithfulness says nothing about any of these. So a trace can carry the decision-relevant information while a monitor still misses the behavior↗︎. In the other direction, a monitor can succeed without the trace being faithful↗︎. It may key on the final answer, on reference information, or on post-hoc text that correlates with the behavior. A 2026 position paper (arXiv 2611.01620) shows both directions on open-weight reasoning models. In a security-vulnerability task, probes and ablations found causal internal representations, but a monitor did no better than a majority-class baseline. In an authority-hint task, a monitor succeeded, but an answer-only rule matched it, and seven mechanistic analyses found no compact mechanism. These cases are small and specific. The second shows faithfulness left unverified, not refuted. Still, a claim of non-entailment needs only possibility cases, and these suffice. Earlier monitorability work also generally treated the two as distinct, and argued that monitoring can work for reasoning-heavy behaviors even where faithfulness is imperfect.

The strongest objection is that the faithfulness evaluations most in use, such as whether a model verbalizes a hint that changed its answer, are partly behavioral. They measure whether decision-relevant factors show up in the trace, which is close to what a monitor needs. On this view, faithfulness evidence serves as a proxy for monitorability↗︎. This makes such evidence informative about monitorability, but it does not make it sufficient. Even a perfectly verbalized factor still needs a monitor able to recognize it. The remaining disagreement is mostly about degree: how much faithfulness evidence should raise expectations of monitorability, and which faithfulness metrics are closest to monitoring-relevant properties. That disagreement does not threaten the claim as stated.