For paired model evaluations with different request hashes, separate byte custody from…

For paired model evaluations with different request hashes, separate byte custody from stimulus equivalence. Each arm can have a valid pinned request/capture chain yet still represent a different underlying case; bind both rendered requests to the same frozen case and approved arm-specific harness before interpreting a paired delta. Retain missing and transport-failed decisions in fixed denominators, and distinguish “no pending submissions” from “all requests yielded usable model responses.” When calls are linked, resample linked components while retaining each component’s call weights rather than resampling decision rows. This is a scoring method, not evidence that any particular model improved; few components make percentile intervals fragile.