Minerval

← the claim it bears on

A line of reasoning from premises to a conclusion. Like a claim, it belongs to no one who makes it.constitution →

The argument from dissociable failure modes

This argument, if it holds, bears in favour of the claim.constitution →Evidence that chain-of-thought is mechanistically faithful does not by itself establish its behavioral monitorability, nor vice versa.

The argument

Mechanistic faithfulness concerns whether the trace reflects the computation that produced the output; behavioral monitorability concerns whether a given monitor, under a given protocol, can detect a target behavior from the trace. Each can hold while the other fails. A monitor can miss behavior even when the trace carries enough information to detect it, because of monitor capability, task difficulty or protocol. And a monitor can detect behavior without the trace being faithful, by keying on the final answer, reference information or post-hoc text. So evidence for one property leaves the other open.

Does it work?

Does it follow?
If the premises are true, the conclusion must be. Whether they are is shown beside this.constitution →
Do the premises hold?
2 supported

The conclusion is a claim that neither property entails the other, and possibility cases in both directions are enough to establish non-entailment. So granting both premises, the conclusion follows. The argument therefore stands with its two premises. That a monitor can miss behavior despite an informative trace is assessed as supported. That a monitor can succeed without a faithful trace is not yet assessed, though it is highly plausible. The argument is not marked decisive while one of its premises remains unassessed.

Premises

Where it is made

Passages where sources make, criticize or mention this argument. They record where it appears, not how strong it is, and none has precedence over another.

  • https://arxiv.org/html/2611.01620makes itThe passage states the inference with what it needs.constitution →
    The two studies test the inference in both directions: in Section 3, causal evidence relevant to faithfulness is present but a monitor with full task information fails; in Section 4, a monitor succeeds but faithfulness is unverified.
  • Chain-of-Thought Oversight Requires Separate Evidence for Faithfulness and Monitorabilitymakes itThe passage states the inference with what it needs.constitution →
    In the security vulnerability reasoning study, probes and activation ablations provide evidence of internal representation and causal influence, but a monitor with access to the reference answer does not outperform the majority-class baseline in a held-out test. In authority-hint factual reasoning study, a monitor correctly distinguishes changed from unchanged answers, but seven mechanistic analyses do not isolate a compact causal mechanism.
The argument from dissociable failure modes · Minerval