The Verified Execution Cycle: gates, evidence, and the commit boundary

An agentic action is defensible or it is not, and the difference is decided before it commits.

The gates

A directive passes an ordered sequence. Intent is captured in plain language through the Directive Interface. It is bound to a named professional's role through the Role Identity Fabric, so authority attaches to a person and not to a service account. Executable policy is applied by the Logic Scrubber. The proposed action is verified against the systems of record it claims to be consistent with. Only then does it commit, and the commit is hashed into the Immutable Audit Ledger.

Evidence at each gate

Each gate emits evidence rather than a log line: the directive text, the role and its scope at that moment, the policy version that ran, the records consulted, the routing and cost decisions taken, and the resulting hash. The evidence is contemporaneous, which is the only kind that survives scrutiny.

Why the commit boundary

Post-hoc review scales badly and arrives late. A reviewer reading a completed action cannot unmake it; a gate that refuses to commit an inconsistent action never creates the exposure. Human-in-the-Loop places a person after the fact and calls it oversight. Human-in-the-Role places a named authority structurally before the fact, and the record proves it.

What this does not claim

The cycle does not make a model correct. It makes the conditions of an action explicit, checked, and permanently attributable — which is the property enterprises are actually liable for.