How do I tell whether a reasoning model's scratchpad actually drove its answer?

Frontier models that emit visible chain-of-thought traces often arrive at an answer before or independently of those steps, then generate plausible-looking reasoning as post-hoc rationalization. Existing faithfulness metrics disagree with each other depending on how the classifier is constructed, which means there is no accepted ground truth for what a faithful trace even looks like. No production tooling flags unfaithful reasoning at inference time or attaches any confidence to whether the trace caused the output. Regulated industries and safety reviews that treat visible reasoning as an explanation of model behavior are relying on something that may be a narrative constructed after the fact.