Visual reference · White paper v4.2 · Toolkit package v4.2.2
Six gate decisions for whether a defined AI-supported quality workflow may proceed, be narrowed, or be stopped.
A defensible authorization decision applies to a defined workflow, not AI in the abstract. A model test or completed course does not establish that the combined workflow will produce acceptable work. Testing supplies evidence; the decision also depends on the system’s authority, whether people can perform their assigned roles, and whether the organization can find and correct failures.
This page reflows for phones and tablets. The original HTML preserves the three-page artboard layout and is best viewed on a larger screen. That visual reflects the v4.2.1 package; v4.2.2 adds only the separate print guide.
Overview
Each gate names a decision. Work that changes or fails returns for assessment; passing through the sequence once is not permanent permission.
Controlling question
What work is authorized now, under what limits, on what evidence, and what finding would require the organization to restrict or stop it?
Decision boundary
Inventory is discovery. Pilot authorization and routine-use authorization are separate decisions. Expansion requires a new explicit decision.
Restrict and requalify: return from routine use to workflow testing when performance or conditions demand it.
Change or failure: reassess scope and authority before a changed workflow proceeds.
Decide before the work proceeds
What evidence would justify authorizing this workflow?
What finding would make us restrict or stop it?
Name the task and accountable owner, examine representative results, and agree on stop conditions.
Gate-by-gate requirements
These cards make the handoff visible. The named owner and local quality system determine the recorded outcome for a specific workflow.
Question
What AI is in use, including unapproved use, and who owns it?
Possible decisions
Toolkit record: AI Use Inventory.
Question
What task will AI support, and what could failure affect?
Possible decisions
Toolkit record: Gate Decision Record, light path.
Question
What may the system read, recommend, change, or approve?
Possible decisions
Enforced, not just instructed: the agent’s credentials must technically enforce its authority envelope. Each approval stays tied to the record state reviewed. If that state changes before the agent acts, review again only where the change undermines the approval’s original basis.
Toolkit record: Gate Decision Record, Authority Envelope.
Question
Can the system and people perform representative work reliably, within a reviewer burden they can sustain?
Possible decisions
Test whether reviewers detect seeded and naturally occurring errors under realistic workload. Set the reviewer-burden limit as a preset pass or fail condition alongside the omission criterion.
Toolkit records: Workflow Evaluation Scoring Sheet and Adjudication Protocol.
Question
Does pilot evidence justify routine use, and does performance hold?
Possible decisions
The Authorization Register carries the Stage 3 burden limit and current burden. Its action threshold applies even when the omission rate appears acceptable. Expansion requires its own recorded decision.
Toolkit records: AI Authorization Register and Independent Effectiveness Review Record.
Question
What changed, what was affected, and did the response work?
Possible decisions
Toolkit record: Gate Decision Record and management actions.
Worked example
The deviation chronology figures below illustrate a prespecified acceptance criterion for one bounded pilot. They are not a universal pass mark.
Deviation chronology example
Zero critical omissions in 59 independent holdout cases yields a 95% one-sided upper bound of about 4.95%. One critical omission would require at least 93 total cases with no further failures to support the same below-5% rate claim. A below-1% claim requires 299 zero-failure cases. These are separate statistical bounds, not alternate ways to pass: one omission still fails this example’s preset zero-omission acceptance criterion.
Cases must be independent, representative of the authorized scope, and protected from tuning or rehearsal. Set the reviewer-burden limit in the same acceptance criterion as a pass or fail condition. A result above that limit fails under the preset monitoring rule, even if the omission rate remains acceptable. Carry the limit and actual burden into Stage 4 monitoring and the AI Authorization Register.
Reference toolkit package v4.2.2
They work as a connected control system. No single artifact authorizes a workflow or demonstrates continuing effectiveness.
Identifies actual use, ownership, status, and disposition. Does not authorize.
Records the bounded decision, evidence and approvals, Authority Envelope, accepted risk, and reconsideration triggers.
Records test evidence, critical-failure handling, holdout integrity, and gate recommendation.
Freezes the critical-omission classification method, calibration, and blinded process before the holdout is scored.
Aggregates current state, decision history, and management actions. Indexes; does not replace source Gate Decision Records.
Tests whether authorization, monitoring, and actions actually work in practice.
Explains the connected control sequence, artifact crosswalk, and implementation boundary.
Gives each of the eight role groups an actionable starting point.
Accountability
The visual summarizes starting responsibilities. Local assignments, qualifications, and approvals must be established by the organization.
Own the work and acceptable outcome.
Do: Define the bounded task, baseline, downstream decision, and record path.
Never: Authorize through an inventory or register entry alone.
Decide whether evidence and controls support the regulated use.
Do: Approve thresholds, critical-failure mapping, and the Adjudication Protocol before results are examined.
Never: Approve after the fact.
Define and enforce what the workflow may access and do.
Do: Maintain and technically enforce the approved Authority Envelope.
Never: Rely on a prompt to prohibit an available action.
Run the scoring sheet and produce Stage 3 evidence.
Do: Freeze the rubric and calibrate on seeded cases before scoring the protected holdout.
Never: Confirm your own mapping as the only reviewer.
Judge evidence independently and document the basis.
Do: Complete independent assessment before seeing the AI conclusion when required.
Never: Score tone or polish instead of evidence.
Keep organizational evidence complete, current, and traceable.
Do: Reconcile the Use Inventory, Authorization Register, and decision history.
Never: Replace local process and quality owners.
Test whether authorization and monitoring controls work in practice.
Do: Test actual use against the Authority Envelope and authorized scope.
Never: Review work you designed or operate without disclosed safeguards.
Convert organizational evidence into accountable decisions and resources.
Do: Decide whether to restrict, suspend, requalify, resource, or retire affected workflows.
Never: Accept a portfolio average that hides a failed subgroup.
Foundation and boundaries
The cycle sits within existing regulatory responsibilities and supporting guidance. The references below do not by themselves authorize a specific AI-supported workflow. Read the sourced control-basis discussion.
The visual is an overview. The white paper explains the stages, evidence, assumptions, and sources; the reference toolkit shows the connected records in worked examples.