A periodic review in which an AI system's observed behaviour is checked for continued correspondence with its intended goals and constraints; distinguished by a documented review occurrence that compares behaviour against a stated specification.
EVENT. One occurrence would be coded when a documented review compares sampled system behaviour against its stated goals/constraints and records a conformant-or-deviation verdict. Routine use without such a documented behaviour-versus-specification comparison is not counted.
Proposed measurement protocol (not yet empirically validated): Audit-record coding: count of completed congruence reviews per period, each yielding a conform/deviation verdict, with deviation severity rated 1-5. Two coders rate a verdict sample; agreement on the conform/deviation call could be reported as Cohen's kappa.
https://andreasehstandlicenseofclarityloc.github.io/augmanitai-periodic/#topic-safety-alignment