Authorized but Refused: Six Months of Runtime Governance Telemetry from an Autonomous Enterprise

Six months of production security telemetry from a live autonomous enterprise (~50 software agents under a written constitution, no human in the daily loop). Over 2026-01-09 to 2026-07-05 the append-only, hash-chained audit log recorded 451,163 events, reconciling exactly into four buckets: 316,440 request logging, 94,014 perimeter/identity, 30,496 governance interventions, and 10,213 other. The load-bearing finding is not the total but the ~30,496 governance interventions in which the system refused its own authenticated, authorized agents — the overwhelming majority not classified as external attack or adversarial activity, but policy refusals of ordinary agent outputs under granted authority. Identity and authorization answer who an actor is and what it may reach; neither answers whether an authorized action should occur. That third question — runtime governance for delegated non-human actors — was the dominant residual control problem observed. Includes independent re-verification of the figures from durable object-storage archives, four-bucket reconciliation, and threats-to-validity. Companion to the plain-language field note and empirical basis for the Enterprise Agent Architecture Control-Plane and Governance layers.