Pick the role closest to your work to see the standards, tools, and first steps that apply to it.
Or start from a symptom you see · an industry term · the one-sprint plan
Does this apply to your system?
The framework earns its cost where four things are true.
Which systems this is for: six kinds of automated decision the standards are built for, where to start in each, and six where the framework does less.
If a system decided about you
Three checks: can the decision be challenged, can a challenge change the outcome, and who can reverse it. Check a decision about you
If an AI agent acts for you
Three checks: when its permission ends, what it does without asking, and what you can check it against that it does not write. Check an agent that acts for you
Start by role
Build decision systems whose errors reach someone who can change the rule: a log entry for every decision, permissions that expire, and challenges the system has to answer, written as code, not policy.
Turn the standards into obligations a regulator or a contract can enforce: mappings to the law, numeric targets for challenges, and clauses a vendor has to sign.
Check a system from its records, decisions, and halt logs, not from what its operator says.
Keep the queues where errors land answerable: count the work they make, give each stalled case a deadline, and treat a backlog as a reason to change the system.
Turn standards into flows that ask for consent, show people how to escalate, and use words they can act on.
Use the same defined terms the standards use, so your findings can be cited in them.
See what failures a deployment will cost the organization before you commit to it.
Start by time available
Answer six questions about one workflow. See which permissions nobody can justify, and whether its decisions can be undone.
Three drills: can the people who run the system halt it without penalty, can its decisions be reversed and the people harmed repaid, and who absorbs the cost when it strains.
An appeal path that reaches the rule, a decision log that records dissent, and a named person who can halt the system, each with an owner and an effort estimate.
Vocabulary crosswalk
Compare common AI terms with the controls and records the standards ask for. Each card names a check you can make.
Common use: A human reviews or approves an automated output.
Check whether the reviewer can change or stop the action.
The delegation standard asks who holds authority, which actions are permitted, and when that permission expires.
Common use: A control halts an automated process.
A kill switch is one mechanism for a halt. Check whether operations can continue after it is used.
The standard on people's time includes a right to stop. The operator needs a tested fallback so halting the process does not leave people without a service.
Common use: A project states principles or checks model behavior against them.
Check whether a person can challenge a consequential decision.
The contestability standard specifies a challenge path, a response deadline, and a route to remedy.
Common use: An audit compares outcomes across groups.
Group-level results do not explain an individual decision.
The decision-record standard specifies a record for each decision. Its higher levels add the reasons, who may object, and where an objection goes.
Common use: A model produces an unsupported answer or its performance changes.
Check whether the evidence still supports the permitted actions.
The delegation standard ties permission to act to stated evidence and to events that force a review. A model’s output does not itself grant permission to act.
Common use: Staff report exhaustion or difficulty keeping up with changes.
Count repair work when assessing the system’s workload.
The human impact standard asks how a system affects people. Record the time staff spend fixing outputs rather than treating those repairs as free capacity.
Common use: A team reports that it has completed required checks.
Ask for records another reviewer can inspect.
The decision-record standard sets the record format. The record checker grades an exported log against it; that does not establish that the decisions were justified.
Three safeguards with owners and effort estimates, a record format for each decision, and four things to ship.
Support / Trust, with the rule's owner
Gives people a way to challenge a decision inside the product, and counts upheld challenges by rule. A set number sends the rule back to its owner for review.
Open mechanism →Product / Policy
Records why a decision was made and who disagreed, so changes are visible.
Open mechanism →Ops / Reliability
Gives a named person a tested way to halt the automation while a correction is made.
Open mechanism →Example receipt
{
"schema_version": "1.0.0",
"receipt_id": "RCP-2026-0842",
"issued_at": "2026-04-11T18:12:09Z",
"owner": { "name": "Risk Operations", "role": "System owner" },
"action": { "class": "WRITE", "description": "Eligibility update", "reversible": true },
"clocks": { "ack": { "hours": 2 }, "review": { "hours": 24 }, "remedy": { "hours": 72 } },
"appeal_path": { "url": "/appeals/eligibility", "channel": "in-product form" }
}Check receipts against the published receipt schema so tools can read them.
Sprint deliverables