DISPATCHES · Summit Cognitive

← All dispatches

A self-assessment

Is this decision admissible?

A thinking tool for one decision you already make.

Pick one consequential automated decision your organization makes — the kind that lands hard on a specific person, the kind someone might reasonably demand to contest. Then answer the questions below honestly, as the practitioner who knows how the system actually works rather than how the diagram says it should. This is a thinking tool, not a compliance audit and not legal advice. It will not certify anything. What it does is hold one decision up against the four properties of an admissible record — provenance, replay, standing, and contestability — and show you, in plain language, where the gaps are. The point is to surface the gaps, not to grade you. Nothing you enter leaves this page.

By Brian C. Long · Founder, Summit Cognitive

Provenance

Were the actual inputs and rules preserved as they stood at the moment of the decision — not summarized later, not reassembled from data that has since changed?

Are the actual inputs the decision rested on captured and stored — the specific records, scores, and signals that were in front of the decision, not a later summary of them?

Are those inputs frozen as they stood at decision time, so that later changes to the underlying data do not silently rewrite what the decision faced?

Are the governing rules, thresholds, or policy in force at the time captured inside the record — rather than reconstructed afterward from whatever version is current?

If the model or service can change beneath you, is the specific version that made this decision recorded, so you could tell which one was actually used?

Replay

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

Could you take the captured inputs and rules and re-run this decision end to end today, without depending on the current live system state?

When re-run on the same inputs, does the decision reliably reach the same outcome — through pinned versions, fixed parameters, or a recorded seed where randomness is involved?

Does the record let someone check whether the outcome actually follows from the inputs and the rules — rather than relying on the system's own after-the-fact explanation of itself?

Where full reproducibility cannot be reached, does the record state honestly what was fixed and what was not — rather than implying a determinism it does not have?

Standing

Does the affected party have the right to demand the record, and the practical means to reach it — or is the record answerable only to the operator?

Is the affected party told that an automated decision was made about them, clearly enough that they could know there is something to question?

Does that party have a recognized entitlement to demand the record of the decision — not merely the operator's discretion to share it if asked nicely?

Are the means to obtain the record real and usable — a path that an ordinary affected person could actually follow, not one buried behind cost, delay, or expertise?

Is the record scoped and retained for at least as long as the affected party can still contest the decision — so it exists when they reach for it?

Contestability

Is there a real remedy — a forum and a route that can actually change the outcome — or only a complaint box that changes nothing?

Is there a defined forum or process where the affected party can challenge this decision and have the challenge genuinely considered?

Can a successful challenge actually reverse or change the outcome — not merely log a complaint that leaves the original decision standing?

Does the record give the challenger something concrete to argue over — the inputs, the rules, the replayable path — rather than a polished account with nothing to grip?

When a challenge succeeds, is there a person or role accountable for acting on it — so the remedy is owned rather than left to no one?

— Dispatches · Summit Cognitive