MAS AI Agent Compliance Checklist
12 verification checks for AI agents operating under MAS oversight. Structured around the four MAS FEAT principles: Fairness, Ethics, Accountability, and Transparency.
Use this checklist before submitting evidence to a MAS examination or onboarding a new AI vendor. Each item maps to a verifiable Armalo Trust Oracle dimension.
Fairness
Verify the agent operates within declared scope boundaries and produces equitable outcomes.
Does the agent operate within its declared scope? Is the scope-honesty dimension score ≥ 70?
Has the agent been evaluated for disparate outcomes across protected characteristics?
Does the agent's pact explicitly define prohibited output types?
Ethics
Confirm the agent has been stress-tested against adversarial inputs and factual grounding requirements.
Has the agent passed a red-team eval with ≥ 80% adversarial resilience score?
Is the accuracy dimension score ≥ 75? Is there a documented test for factual grounding?
Is the model version locked and compliant with the agent's declared capability tier?
Accountability
Establish a tamper-evident record of commitments, audit logs, and incident response protocols.
Is there a signed, versioned behavioral pact with a verifiable hash?
Does every agent session produce a tamper-evident interaction log?
Is there a documented protocol for agent failures? Is it tested?
Transparency
Ensure external parties — including MAS examiners — can independently verify agent behavior.
Can an external MAS examiner verify the agent's score via a public API endpoint?
Is there a 90-day behavioral score history available for examiner review?
Has the agent passed at least one multi-LLM jury evaluation with a documented verdict?
Check your agents against all 12 dimensions automatically
Armalo runs adversarial evaluations, scores your agents across every FEAT dimension, and produces a Trust Oracle record that MAS examiners can independently verify — no manual checklist required.
Check your score free14-day Pro trial · no credit card required