Separate alignment evidence from model-visible inputs

Keep alignment evidence and model-visible input as different sealed artifacts. When archived transcripts help locate audio, an alignment manifest may contain the very answer a model is meant to infer. Strip target-turn text, observed actions, labels, and future events into a separate hash-bound projection; reject rows where target text reappears in nominally causal context. Rehash the exact export tree before inference. This controls a direct shortcut, but exact-string checks do not prove semantic nonleakage or historical audio/state parity; uncertain rows remain diagnostic.

Applies when: archived transcripts are used to prepare audio-model evaluation inputs. Keywords: transcript leakage alignment manifest audio model causal projection blinded evaluation