# E114 governance report

Principle: AGENT PROPOSES → CAIN DECIDES → MCPGATE ENFORCES → SYSTEM EXECUTES → EVIDENCE REMEMBERS.
Keys: the authority signs grants, a human approver signs high-risk approvals and promotions, and an independent kill key signs KILL/CLEAR. The director holds none of them.

- Adversarial suite: **44/44** across a2a, allocation, approval, evidence, fail_closed, info_flow, kill, memory, policy, post_model_call, post_tool_call, revocation, spawn, tool.
- KILL race: 16 processes, 6400 pre-issued grants: 859 effects before the order, **0 after**, 0 duplicates.
- Replay race: one grant presented by 12 processes, admitted 1.
- A KILL cannot be cleared by the authority (director-side) key; corrupted kill state fails closed.
- Spawn attenuation: scope, clearance, expiry and depth can only shrink; revocation cascades to descendants.
- Every lab experiment ran under a grant: 29 governed experiments, 29 admitted.

Locality: in-process decision point + multi-process file-locked enforcer on one host; NOT production MCPGate. Microsoft Agent Control Specification: NOT TESTED (Microsoft Agent Control Specification reviewed as a design reference only). E114 enforcement is modelled on it, not compatible with it.
