ComplianceOS — Regulation-to-Enforcement Compiler

Enterprises deploy AI agents against a wall of regulation — GDPR, HIPAA, the EU AI Act. Today the gap between "the regulation says X" and "the agent is actually prevented from doing not-X" is bridged by hand: lawyers read the text, engineers hard-code rules, and nobody can prove to an auditor that the running system matches the law. ComplianceOS closes that gap. It compiles a regulation PDF into typed clauses, generates a Lobster Trap enforcement policy gated through the real binary, and runs that policy in front of every agent prompt. A two-layer protected plane — Lobster Trap regex plus a Gemini semantic classifier — inspects each prompt and takes the most-restrictive action. On a 932-prompt adversarial corpus, regex alone caught 4–10%; the two layers together reached 99.8%. Every decision is one hash-chained row in SQLite, so the audit log is tamper-evident. An auditor asks a plain-English question and gets back a signed PDF evidence pack where every fact traces to the log, with an embedded integrity proof. A clause-grounded adversarial validator produces a per-clause coverage matrix — proof, not promises. When a regulation phases in, ComplianceOS re-compiles, diffs the policy, re-validates the new clauses, and hot-reloads. Verified end to end with Gemini: a 29-event evidence PDF and an EU AI Act amendment adding 734 clauses at 100% delta coverage.