Each run is evidence, not noise.

A model run should preserve inputs, outputs, warnings, audit events, and approval status.