Invigilator

• validate-cerberus-pass

  Validate a strong agent (Cerberus, sanctions disposition). Expect a clean pass with a high passRate; fairness not applicable.

• validate-spark-fairness

  Validate a lending agent (Spark, credit decisioning) where fairness IS in scope. Expect pass/conditional with a fairness section and a real passRate.

• adv-unknown-target

  Adversarial: a target the validator cannot run (not in the supported set). It must NOT rubber-stamp an unverifiable agent as 'pass'; it must issue a non-pass opinion with a cited condition.

Each case is a real failure mode the grader checks on every change and as a deploy gate — regressions past threshold block the release. Adversarial cases probe the failure modes the golden set doesn't.

Briefing

• Independently validates another Cadre agent the way an SR 11-7 validation team would: skeptical by default, with evidence it generates itself.
• Never takes the target's self-report on trust — it re-runs a fresh sample of the target's own golden cases and reads every per-case grade.
• Challenges how the target worked, not just whether it was right: a tool call the manifest does not allow, a commit that never happened, or a run that decided suspiciously fast are all findings.
• Checks fairness handling only where the target declares a prohibited-basis guardrail; where that does not apply, it says so plainly instead of inventing a finding.
• Issues pass, conditional, or fail — and anything short of a pass must cite a concrete failing case or anomaly, or the opinion is not supported.