Skip to content
Process AtlasModel Governance

Agent Assurance

The roster is a supervised population. Invigilator independently validates fellow agents; Quorum re-derives their decisions blind. Agents governing agents — with the committee holding the pen.

4 agents staffed4 A2A links1 human checkpoint1 open slot

Inventory

Agent registry
Bank system

The governed roster itself — every agent versioned, with evals, guardrails, and a decision ledger.

Assurance fleet

Invigilator
Independent validation

Re-runs a fellow agent's eval sample, challenges its trajectories, and issues a validation opinion.

  • hands to Cerberus via run_eval_sample
  • hands to Spark via run_eval_sample
Quorum
Blind four-eyes review

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

  • hands to Cerberus via rederive
  • hands to Spark via rederive

Subjects under review

Cerberus
Subject: sanctions decisions

Cerberus under review — its eval sample re-run and its dispositions re-derived by the assurance fleet.

Spark
Subject: credit decisions

Spark under review — decision quality and fair-lending integrity checked from outside.

Resolve

Model-risk committee
Human checkpoint

Validation opinions and dissents land with the committee.

Remediation tracking
Open slot

Tracking findings to closure. Open slot — no agent staffs this step yet.

The walkthrough

  1. 1

    The roster itself is a supervised population: every agent ships with evals, guardrails, and a decision ledger.

  2. 2

    Invigilator independently validates a fellow agent — re-running its eval sample and challenging its trajectories — and issues a validation opinion.

  3. 3

    Quorum re-derives sampled decisions blind — concur, dissent, or escalate. Agents governing agents, on the record.

    Replay the real runsc-quorum
  4. 4

    Opinions land with the model-risk committee; remediation tracking is the next slot in the cadre.