Aegis Turns AI Policy Into Replayable Evidence

Field notes

The page demonstrates its method through a synthetic Meridian Labs run rather than relying only on consulting language. A recorded sequence exposes policy and baseline probe outcomes, including explicit refusal, source provision, escalation, and an ambiguous case flagged for review. The displayed aggregate uses n=20 and reproduces a 95% Wilson interval of 0.839–1.000 at a stated floor of 0.800. A configuration hash and SHA-256 manifest frame the run as replayable evidence, while the page separately marks live runtime outcomes as roadmap work.