← All three profiles

The gate

Safety Gate

Fail-closed gates between model output and action, with escalation to a human. The model proposes; the runtime decides what may proceed.

The Safety Gate sits between a model’s suggestion and any real-world action. Sensitive or high-stakes decisions are flagged and routed to a responsible human; low-stakes paths proceed deterministically. Default is closed: when in doubt, nothing executes.

Every gate decision — allow, deny, or defer-to-human — is recorded with its escalation, so the boundary is not just enforced but accountable.

The artifact it produces

A fail-closed allow / deny / defer-to-human decision, with the escalation and its reason recorded.

What it does not claim

It is not a certified physical-safety controller. It governs the language-and-action decision; certified hardware still owns life-safety. It reduces the blast radius of a bad model output; it does not make the model safe.

Where it fits

  • Agentic tool-calling where every call is a real-world action
  • Family AI — the system proposes, the responsible adult decides
  • Any workflow where wrong action has real-world cost

In practice

Every profile emits its record in AIREP, the neutral interchange format. The profiles compose — see all three.