Confidence Gate. A control decision that changes downstream behavior based on the confidence, grounding quality, freshness, and risk of a recommendation or conclusion.
A confidence gate prevents uncertainty from becoming ungrounded confidence. Low-confidence, stale, weakly grounded, or high-risk outputs should suppress action, downgrade the result, require stronger evidence, or escalate to a human-only decision. The gate belongs in control logic; the model may propose confidence, but the platform decides what confidence is allowed to do.
Freshness should arrive as freshness state, so the gate can treat "stale for this use" differently from "fresh enough" or "unknown."
Approval should arrive as evidence approval state, so pending, rejected, or revoked evidence cannot hide behind model confidence.
A gate should be adjusted by confidence calibration: if high-confidence answers stop matching observed outcomes, the control path should tighten.