Glossary

Vocabulary used by this deployment. Policy pol.sepsis-micu 1.2.0, contract 2.0.0.

Verdict

What the policy decided about one output. There is no PASS.

no violation in checked scope: Internally `pass`. None of the configured checks fired. Not a statement that the output is correct or that the patient is well.
flag: A check disagreed, or a review condition was met. The output is not stopped.
hold: Policy requires the output to be withheld, rerouted or corrected before it proceeds.
unable to verify: Required evidence, authority or capability was unavailable. The verifier declines rather than guessing.

The verifier's own view, separate from what the policy then does about it.

clinically corroborated: Independent screening criteria computed from the same observations agree with the model's alert state. No clinician reviewed the case.
clinically disputed: The criteria and the model disagree, in either direction: criteria met with no alert, or an alert with no criterion met.
review required: Material ambiguity, or a condition the policy says a person must resolve.
unable to verify: Too few criteria components were available to form a view.

Whether the verdict turned out to be right, once the outcome matured.

confirmed: The configured truth source supports the verdict.
overturned: The truth source contradicts it.
inconclusive: Evidence exists and cannot settle it. Used when the reference label and the independent criteria disagree.
unresolved: The outcome window has not matured, or the signal was never collected.
not sampled: No truth contract selected this verdict for adjudication. Nothing is owed on it.
truth debt: Verdicts whose window has matured with no truth signal. Carried as portfolio risk.
truth contract: The rule that says in advance how a class of verdict will be shown right or wrong: source, window, sampling, owner.
maturity window: How long the resolver waits before an outcome is considered settled.

Whether a verdict reached the workflow and closed. A correct verdict nobody reviews changes nothing.

delivered: The verdict reached its configured destination.
reviewed: An authorised person opened it.
acted: It was corrected, escalated, held or otherwise handled.
overridden: An authorised person rejected the disposition, with a reason.
closed: The required action and available outcome evidence are recorded.
failed to land: Delivery, review, action or closure missed the service level.

Reported separately, with denominators. Never combined when a factor is missing.

Coverage: Verdicts completed within the service level, divided by consequential outputs the verifier received. An output never sent is invisible to it.
Confirmed Validity: Confirmed stays divided by adjudicated stays. A stay is one fact, confirmed only if nothing about it was overturned. Contracts resolved against the verifier's own arithmetic are excluded, and the interval resamples patients.
Landing: Actionable verdicts that reached the closure state the policy requires, inside its service level, divided by actionable verdicts issued.
EVC: Coverage × Confirmed Validity × Landing. Withheld when any factor has no denominator: a product with a missing term is not a smaller number, it is not a number.

Everything the hospital owns. Each versioned and printed on the record.

claim contract: A vendor claim rewritten so it can be tested: endpoint, decision timepoint, population, inputs, output meaning, metric with a target, named reference standard.
authority bundle: The versioned sources a check may cite. Model output can extract or infer; promoting it to a rule takes clinical approval and a new bundle version.
failure mode profile: What is known to go wrong for a claim, with a detector, a severity and a default action.
verification policy: The executable configuration: checks, decision table, dispositions, truth contracts, service levels.
decision mode: How a check decided. Direct deterministic is a plain function over structured evidence; statistical calibrated carries a measured error rate; a model's self-reported confidence is never used.
evidence coverage: How many components a check needed against how many it had.
decision proof: The method, inputs, authorities and rules used, so the result can be replayed.