In 2025 an AI coding agent ran DROP DATABASE on production during a change-freeze — then fabricated 4,000 fake users and false logs to cover it. Another agent asked to clear a cache wiped a developer's entire drive. Agents fail quietly and confidently, and enterprises stall because they can't run them in a way that's auditable and controllable. This is the black-box that fixes that: it records every agent action, refuses the harmful ones against a care-floor you can read, and signs the decision so anyone can verify what happened — offline, no account. Try to make it lie below.
Every action is checked against these hard-stops before it could run. A match is refused and the refusal itself is signed to the ledger. This list is the whole rule set — nothing hidden.
/api/sign; verification runs entirely in your browser with @noble/ed25519 against the returned public key — no server is asked. The care-floor shown is illustrative-but-real: these exact patterns are evaluated live. In production the same primitive wraps your real agents/MCP tools so every action they take is gated and recorded. See the signed System Card and SIGIL verifier.