Runtime & Integrations · Evaluation Core
Authority Resolution
May this metric gate at all?
resolve_authority(AuthorityInputs) → ResolvedAuthority climbs a three-stage ladder:
not authorized ⇒ informational, authorized but evidence unsound ⇒
blocked, fully authorized ⇒
can_gate. The reasons are typed codes, not prose.
This is the code behind Authority. It reads only typed, host-owned inputs — dataset status, metric status, threshold approval, judge calibration, data policy, baseline state — and returns a structured result the Verdict Engine consumes.
The ladder
| Stage | Triggers | Result |
|---|---|---|
| 1 · not authorized | metric not gating; threshold not approved (threshold_proposed); dataset proposed; judge uncalibrated/calibrating. | informational |
| 2 · evidence unsound | policy not in {permitted, redacted}; judge drifted/retired; dataset retired; required baseline not comparable. | blocked |
| 3 · fully authorized | none of the above. | can_gate |
Each non-pass stage attaches typed reason codes — threshold_proposed,
judge_drifted, baseline_not_comparable, policy_forbidden —
that travel onto the Scorecard’s authority block.
Data policy is a hard gate
The accepted egress set is exactly
{permitted, redacted}
(_POLICY_OK). A forbidden, missing, or unknown
policy blocks an otherwise-authorized gate — evidence you may not use cannot back a claim.