Operation Entitlement: External Governance and the Making of a 12B Clerk

Operation Entitlement extends the answer-entitlement framework from assertion-like model output to model-mediated action. It asks not only whether an AI agent can detect useful work, but whether it is warranted in acting on the operational surface it has detected. This manuscript presents a concept-building operational case study from the MOBIUS/MMV project. A local Gemma 12B Medium secretary was run as an overnight documentation-patrol clerk over a 130-file repository surface. In its initial patrol, the clerk detected a real 26B/12B discrepancy but formed an unsafe repair path: it treated feature-branch runtime evidence as if it authorized rewriting release-facing documentation. The failure was therefore not merely factual, but an action-boundary failure. The revised governed configuration supplied external role, canon, WIP/release, escalation, and review-path boundaries through a Box B governance-document layer, ME5/FAISS indexing, startup retrieval, and hard patrol guards. On rerun, the same 12B model still detected the discrepancy but routed it to release-state verification rather than proposing release-surface rewrite. The evidence is presented as a concept-building case study, not as a mechanism-identifying ablation. Its contribution is to define operation entitlement as an authority-axis evaluation layer for AI clerks and to position it as complementary to least privilege, role-based access control, separation of duties, capability security, and agent evaluation.