Tier 2 — gates that inspect what a model may do, not what it says
Tier 2 · #7
Frontier Control Gate
Sits between a frontier model and the world. The model proposes an action; the gate decides ALLOW / BLOCK / REQUIRE_APPROVALbefore it executes. Nothing the model emits runs unchecked.
"A harness executes; it does not license." Sits between an intent and its action; returns exactly LICENCE or REFUSE with a measured rationale.
intent: deploy v2.1 to production (tests green, change advisory approved)
intent: delete backups older than 30 days (no retention policy cited)
Tier 2 · #9
Robotics Safety Gate
Wraps the causal engine into a C API (es_safety_gate.h). One question: "Given what is observed now, would this action cause a bad outcome?"DEFAULT-DENY — returns SAFE only on positive evidence.
Earlier gates (Moral / Behavior / Harm Marker) inspect text. A frontier model can produce perfectly polite text that proposes a catastrophic action. Action-level gates close that gap: they sit on the actuation boundary, where text becomes a consequence. Every gate here is deterministic and reason-coded — the model does not get to grade its own action.