A concrete walkthrough

This is deliberately mundane. No hypotheticals, no edge cases.

Pilot Scenario: Automated Benefits Eligibility System

(U.S. State Agency)

Current state (baseline)

  • Vendor-provided eligibility scoring system
  • Automated triage flags high-risk claims
  • Caseworkers review flagged cases
  • Oversight via audits and post-hoc appeals

Authority is effectively embedded in:

  • model thresholds
  • routing logic
  • continuity of deployment

ACP intervention points

1. Claim identification

System output: “Claim flagged as high-risk.”
This is recognized as a claim, not a fact.

2. Ratification requirement

Before denial or delay:

  • a designated authority must ratify reliance on the claim
  • ratification is logged, time-bounded, and revocable

No ratification → no adverse action.

3. Continuity check

Every fixed interval:

  • continued use of the scoring system requires re-ratification
  • justification must address observed outcomes, not just design intent

Failure to re-ratify pauses automated routing.

4. Refusal pathway

If ratification is refused:

  • system does not escalate automatically
  • alternative process (manual review) is triggered
  • refusal is recorded, not penalized

5. Accountability shift

When harm occurs:

  • responsibility attaches to ratifiers
  • not vendors
  • not governance teams
  • not “the system”

Authority and responsibility converge.

What changes (and what doesn’t)

Does not change

  • model architecture
  • vendor relationship
  • statutory eligibility criteria

Does change

  • where authority sits
  • how continuity is justified
  • who must own decisions

Failure mode (acceptable)

If no one is willing to ratify:

  • the system cannot operate
  • that failure is visible
  • governance is functioning

Closing alignment across artifacts

  • Artifact 8 establishes constraint
  • Artifact 9 makes it legally legible
  • Artifact 10 proves it is operationally real

None of these promise safety; all of them remove evasion.