Start with System Intake.
Define the purpose, boundary, people, data, roles, initial risk, and whether the work should proceed.
Records that make practice reviewable
Choose records by the decision stage. The website explains when each record is needed; the repository retains the copyable source templates.
How to use the toolkit
Need the process before the paperwork? Follow the Implementation guide.
Define the purpose, boundary, people, data, roles, initial risk, and whether the work should proceed.
Use the directory below to select evidence, authorization, pilot, repair, change, and transparency records as needed.
Store project records in a versioned, access-controlled evidence repository. Never put personal, confidential, classified, or exploit-enabling detail in a public repository.
Frame
Purpose, system map, data, initial risk, accountable roles, readiness, and the decision to proceed, narrow, use a non-AI alternative, or stop.
Open template ↗Affected groups, benefits, burdens, meaningful choice, correction, appeal, displacement, and unresolved concerns.
Open template ↗Verify, compare, authorize
Claims, sources, integrity, unknowns, contradictions, and the evidence threshold required by consequence.
Open template ↗Purpose, authority, evidence, alternatives, selected action, safeguards, approvals, outcome, and learning.
Open template ↗Hazards, preventive and detective controls, owners, evidence, metrics, thresholds, responses, and residual risk.
Open template ↗For structured records, the repository also provides a machine-readable Decision Record schema ↗.
Pilot and observe
Hypothesis, boundaries, stages, gates, measures, safety, remedy, authorization, and the final adopt/revise/stop decision.
Open template ↗Definition, interpretation, baseline, thresholds, required response, counter-metrics, privacy cost, and validation.
Open template ↗Correct and repair
Correction is not complete when a model output is merely edited. Reachable downstream effects and affected-person remedy remain part of the record.
Change, provenance, transparency
Problem, affected people, alternatives, safeguard impact, evidence, provenance, validation, and review.
Open template ↗Identity, rights, permission, information boundary, verification, attribution, and publication decision.
Open template ↗Purpose, AI role, data, oversight, performance, limitations, incidents, repair, open risks, and next review.
Open template ↗