A drafting tool

Decision-record builder

Draft the record one decision would leave behind.

Pick one consequential automated decision — the kind that lands hard on a specific person — and answer the prompts below as the practitioner who knows how the system actually works. The tool assembles your answers into a clean, copyable decision record, organized around the four properties: provenance, replay, standing, and contestability. This is a thinking tool and a template, not a product, and not a legal or compliance artifact. It does not certify anything and nothing you enter leaves this page. The point is to make you write down, in one place, what you would need if the decision were ever challenged — and to show you, by the blanks you cannot fill, where the record does not yet exist.

By Brian C. Long · Founder, Summit Cognitive

Decision identity

What the decision was, when it happened, and who is accountable for it.

Decision name or ID

A short label or reference number for this one decision.

Date and time of decision

When the decision was actually made.

The decision made

The action taken — what was actually done to or for the affected party. Keep this distinct from any prediction or score; a risk score is not a decision, declining the application is.

Decision owner (accountable human)

The person or named role answerable for this decision — not the system, and not the vendor.

The question actually asked

The precise question the system was answering at the moment of decision, stated plainly.

Provenance

The actual inputs and rules the decision rested on, captured as they stood at decision time — not summarized later, not reassembled from data that has since changed.

Inputs and evidence consulted

List the specific records, scores, and signals the decision rested on — ideally in the order they were consulted. One per line.

Rules or policy version in force

The governing rules, thresholds, or policy as they stood at decision time, with the version or effective date — not whatever is current now.

Model or system, with version

The model or service that produced the score or decision, and the specific version, if any. Note "none / rules only" if no model was involved.

Replay

Whether the decision could be re-run on the same inputs, through the same rules, and land in the same place — checked rather than merely explained.

Confidence or uncertainty noted

Any score, confidence band, or uncertainty the decision carried — and how close it sat to the threshold.

How the same inputs could be re-run (replay note)

What it would take to reproduce this decision — pinned versions, fixed parameters, a recorded seed, the stored inputs. If full reproducibility is out of reach, say honestly what is fixed and what is not.

Standing

Whether the affected party can learn of the decision, has the right to the record, and a usable path to reach it.

Who is the affected party

The person the decision lands on. Describe by role, not by name — this is a template, not a case file.

How they are notified

How and when the affected party is told an automated decision was made, clearly enough to know there is something to question.

Where the record is stored / who holds custody

Where this record lives, who holds it, and how long it is retained — ideally for at least as long as the affected party can still contest.

Contestability

Whether there is a real remedy — a forum and a route that can actually change the outcome — and a person accountable for acting on it.

How they can contest (route and deadline)

The defined route to challenge the decision, the forum that hears it, and the deadline by which the affected party must act.

The remedy available

What a successful challenge can actually do — reverse the outcome, re-decide, or something narrower. Be honest if it only logs a complaint.

— Dispatches · Summit Cognitive