A Fixture Hash Separates Deterministic Misses From Eligible Flake Freezes

A property failure on an agent patch is initially considered a miss until proven otherwise through a shrink. Freeze eligibility is determined after the miss is reduced to a minimal input, stored as a hashed fixture, and replayed on both the base revision and candidate. Red-build counts have no impact on this decision. Flaky tests often result from narrow input bugs, while the opposite error is equally detrimental: a test that already fails on the base gets attributed to the patch, and a clean candidate is discarded for unnecessary noise.

Three records differentiate between these scenarios. A property check samples a space under a seed, a fixture pins one input and its hash, and a freeze is a time-bounded exception for a single variance class. Eliminate any one of the three, and the label "flaky" becomes conjecture. Variance classes, not failure totals, dictate the workflow table, which serves as a decision-making tool rather than a live suite score or benchmark.

A minimum of two replays per revision is the classifier's lowest acceptance threshold, not a fixed sample size for every property. The table outlines the necessary steps to determine the appropriate action. To unlock a freeze, only one class must be deemed deterministic_miss, with mixed results on the other classes resulting in a draw that requires a new local seed budget and avoidance of freezing.

Any other matching-hash pattern triggers the addition of replays or referral to a human. A mismatch in hash values signals the need for further investigation and discarding of stale files. The process begins by recording the seed before the candidate runs, as a sample without a seed cannot be replayed and thus cannot support a freeze. The property id, seed, and base revision must be bound before the agent edit is applied.