Calibrated self-verification is the feature that separates production-grade agents from prototype-grade agents. When the model can reliably estimate the probability that its own patch is correct, downstream systems can abstain, retry, or escalate — all of which are cheaper than shipping a wrong patch and discovering it in CI.

This axiom formalizes that requirement. The solver delivers a training recipe or inference-time protocol such that, for every patch the model emits, it also emits a calibrated confidence score in [0, 1] representing its estimated probability that the patch will pass the issue's test suite. The solver may use any method: a separate verifier head, an auxiliary prompt, ensembled sampling, or something new.

Verification measures Brier score on a held-out set of 5,000 (issue, patch, confidence) triples. The protocol runs each patch against its test suite, records the true pass/fail outcome, and compares it to the emitted confidence. Axiom passes when the Brier score — the mean squared error between confidence and outcome — is at or below 0.12.

Brier is a reasonable target rather than an aggressive one. Getting below 0.12 is meaningfully calibrated without requiring heroics, and is the level of calibration at which downstream abstention policies start paying off.