Under what conditions does monitoring a language model's chain-of-thought reasoning reliably reveal the reasoning behind its behavior?

The graph holds assessed claims on chain-of-thought monitorability that pull in different directions (models struggle to evade monitors when reasoning is necessary; reasoning is often unfaithful on natural prompts; learned reward hacks went unverbalized in one RL study; training pressures and adversarial evasion may reduce monitorability), and no question node under which they are read against one another. The distinction between faithfulness and monitorability, and which conditions (task difficulty, necessity of reasoning, training pressure, adversarial intent) account for differing results, is what careful readers of this literature need to ask.

The proposal asks under what conditions monitoring a language model's chain-of-thought reliably reveals the reasoning behind its behavior. The wording is neutral and presupposes little. It does not assume monitoring works or fails, and it names no private person. It is an open question of the kind §7a describes: one that careful readers of the chain-of-thought monitorability literature would need to ask, with candidate answers organised by conditions such as necessity of reasoning, task difficulty, training pressure and adversarial intent. It also separates faithfulness from monitorability, a distinction the literature itself treats as central. The cited sources (arXiv 2507.05246, 2503.08679, 2603.05706) are plausible, relevant sources on this topic; I did not open them, and nothing in this decision depends on them, since a question is judged on its wording alone. Novelty and placement are the Matcher's call, and the Steward will do the work of decomposing it into claims and subquestions. The supporting rationale in the content field does not enter the graph.