Clearance Decision

A clearance decision is the named outcome of an AI security review. Drel uses a five-state model rather than a binary pass/fail because real AI systems present situations where neither extreme captures reality: most systems need to ship under conditions, run in restricted pilot, or wait for information that has not yet arrived.

The five states: Proceed means unconditional clearance for the reviewed scope. Conditional means clearance contingent on named conditions being met before production. Restricted pilot only means clearance for limited deployment with explicit user/data scope and a re-review trigger. Hold means more work is required before any deployment decision can be made — typically because evidence gaps are material. Decline means the system should not be deployed as designed.

Each state has implications. A Conditional clearance with no named conditions is not Conditional; it is Proceed. A Restricted Pilot Only with no defined scope is not restricted. A Hold with no defined unblocker is indefinite. Each state's defensibility depends on the specific content that goes with it.

Clearance decisions are scoped. The clearance applies to the system as reviewed — not to changes after the review. Adding a tool, swapping a model, expanding the user base, or routing different data through the system triggers re-clearance.

Clearance decisions are dated and authored. An undated decision cannot be referenced in evidence. An authorless decision has no accountability. The disposition memo records both, alongside the signatories from the AI Committee.