Praesidia.AI

Companies promise to limit data retention, restrict exports to authorized users, and stop agent work when permission is revoked. As coding agents change software and business agents execute it, those promises can drift from code and behavior. Praesidia Investigator and Jurator help engineering teams make selected promises inspectable and testable. Jurator uses Meta models to extract commitments with exact policy quotes, connect them to code, and propose reviewable rules. The program tests those rules against planted breaks before acceptance. An optional attacker/judge loop challenges weak rules and attempts bounded repairs. Accepted rules can then check code changes without model calls, and configured Git hooks can block detected regressions. Model-only assessments remain advisory. Investigator complements these code checks with bounded runtime experiments and independently observed effects. Its revocation example detects queued work that still reaches a destination after cancellation is acknowledged, then tests a supported repair while preserving authorized work. Both tools are Apache-2.0 open source and run without a Praesidia account. They help teams shipping agents and AI-written software investigate the gap between written promises and what their systems do.