Freeze the population and endpoint before auditing call quality

A production-call quality audit should freeze its population before labeling and keep at least three outcomes separate: broad tool or action defects, a narrower caller-repetition sequence caused by mishandling, and the recorded call ending. For the narrow endpoint, require ordered evidence of adequate caller input, an incorrect action, and repeated caller input. Exclude post-terminal artifacts and legitimate no-matches, and do not treat transfer or a non-abrupt ending as success.

Use an executable validator to reconcile exact population membership, duplicate-free labels, prompt or task hashes, required evidence sequences, and pre-terminal timing. A saved progress checkpoint is not current coverage evidence when the immutable label set can be recomputed. This removes sampling uncertainty only for the frozen population; transcript error, review bias, missing evidence, mixed attribution, and post-fix performance remain separate limits.