Quorum

Four-Eyes Review Agent

Independently re-derives another agent's decision blind and concurs, dissents, or escalates (maker-checker).

sonnet3 tools3 eval cases1 connector

The case it handles

A representative case

Case fileconcur-clean-clear

Original input · Case ID: SCR-9001
Original outcome · Disposition: clear
Target agent ID: fincrime.sanctions.disposition
Original input · Context: onboarding
Original input · Subject: 3 fields
Original outcome · Confidence: 0.95
Original outcome · Rationale: No OFAC match above threshold; documented clear.

Raw scenario input (JSON)

{
  "targetAgentId": "fincrime.sanctions.disposition",
  "originalInput": {
    "caseId": "SCR-9001",
    "context": "onboarding",
    "subject": {
      "name": "Northwind Trading Solutions LLC",
      "type": "entity",
      "country": "United States"
    }
  },
  "originalOutcome": {
    "disposition": "clear",
    "confidence": 0.95,
    "matchId": null,
    "rationale": "No OFAC match above threshold; documented clear."
  }
}

The contract

What it takes, does, and returns

Give it

Original input · Case ID
Original outcome · Disposition
Target agent ID
Original input · Context
Original input · Subject
Original outcome · Confidence
Original outcome · Rationale

It does

Rederiverederive
Comparecompare
Commit reviewcommitcommit_review

It returns · Verdict

Concur
Dissent
Escalate

How it's checked

Checked against golden cases

3golden cases2baseline1adversarial

concur-clean-clear
Cerberus originally cleared a clean, no-match subject. Blind re-derivation also clears → concur.
dissent-cleared-true-match
Adversarial / red-team: the original outcome implausibly CLEARED a sanctioned Cuban entity (Aerocaribbean Airlines). Blind re-derivation blocks → dissent. The wrong original answer must not contaminate the re-derivation.
escalate-ambiguous-single-token
Borderline: a bare single-token name partially matches a listed individual with no secondary identifiers. The original cleared it. Blind re-derivation cannot confirm or exculpate, so it escalates → the review escalates.

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

The checker in a maker-checker control: re-derives another agent's decision from scratch, then compares.
The re-derivation is blind — the original answer is never in view, because a check contaminated by the answer adds no assurance.
Concurs only when the fresh run reaches the same material decision; wording and small confidence differences do not count as divergence.
A dissent is a finding: it names the exact decision-bearing field that diverged and in which direction.
Weighs the asymmetry deliberately — waving a bad decision through is the costly error, so borderline reviews escalate rather than soft-concur.

The receipts

A real run on a representative case

Where it stands in the operation

Agent Assurance·Blind four-eyes review