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