Monitorability decomposes into faithfulness and verbosity, and they trade against each other
A trace can be honest and too terse to catch anything, or exhaustive and unfaithful. Treating chain-of-thought monitoring as a single property hides the trade-off that determines whether it works.
Separating the two is the contribution. Faithfulness asks whether the trace reflects the actual computation. Verbosity asks whether enough of it is written down to act on. A monitoring system needs both and most evaluations measure neither cleanly.
The measurement itself turns out to be unstable. Work on classifier sensitivity finds that faithfulness scores depend materially on how faithfulness is operationalised, which means cross-paper comparisons of these numbers are weaker than they look.
This lands on something concrete. OpenAI's Astra safeguards read the chain of thought and interrupt high-risk activity mid-run. That control is exactly as good as trace monitorability is, and the literature is currently arguing about how to measure it.
arXiv — Measuring chain-of-thought monitorability through faithfulness and verbosity → · arXiv — Measuring faithfulness depends on how you measure: classifier sensitivity in LLM chain-of-thought evaluation → · arXiv — Beyond the commitment boundary: probing epiphenomenal chain-of-thought in large reasoning models →