Reckoning: Typed, Calibrated, Auditable Decision Governance Under Model Churn

AI systems are usually evaluated on whether they can perform a task. Operational systems must answer a different question: whether they should, under uncertainty, accountability, and changing model behavior. Fast one-pass decision models offer attractive latency and cost, but deploying them directly into operational loops introduces silent failure modes. Aggregate accuracy can conceal confidently wrong behavior, and a model upgrade can invalidate the evidence that justified automation. Reckoning is an operational decision governance layer designed to prevent silent automation failures in fast operational decision loops during model churn. It maps typed operational questions to raw label distributions from a serving backend, applies calibration and policy to derive a gated action (act, escalate, or abstain), and writes an append-only receipt with application-level tamper evidence before the decision is returned to the caller. Confidence scoring, calibration, abstention, and audit logs are each established techniques. The contribution is treating operational judgment as a contract that outlives any model: the same typed questions and thresholds govern both live decisions and model adoption, and trust earned by one checkpoint is not inherited by the next. We describe three authority planes governed by this contract: authority to act, authority to serve, and authority to receive data. We report an illustrative baseline on a public benchmark (N = 231) in which the current default backend reaches 0.584 aggregate accuracy yet scores 0 of 8 on an adversarial trap family at 0.64 mean confidence, and a synthetic calibration run (an uninformative stub backend, error measured on held-out rows) in which temperature scaling reduces expected calibration error from 0.240 to 0.088 on its largest bucket. We state which mechanisms are enforced today by code and which are configurable or planned.