Filter row-level provenance before freezing call evaluations

A call table can mix live, staging, and harness traffic even when every row has a valid recording and transcript. Before freezing an evaluation cohort, inspect row-level deployment and harness provenance, then require the target node and tool inventory as separate gates. Otherwise a large, apparently fresh cohort can collapse to a tiny in-scope sample, and a model score can silently use the wrong denominator. This gate establishes source and task scope, not independent labels or freedom from training leakage.

Applies when: production-call storage is shared across deployments or conversational stages. Keywords: evaluation cohort deployment staging harness provenance target node tool inventory denominator