# Reconstruct an AI decision

Preserve what was available when a consequential choice was made.

id: https://miskir.com/guides/ai-decision-records/
version: 2026-10-15.1
published: 2026-10-15
status: published_guidance
record_type: editorial_guidance
attestation_status: not_a_signed_receipt

## Begin with the decision boundary

Identify the actor, delegated authority, question, permitted actions, and relevant policy version. Record the evidence available at the decision time, including meaningful absences. Later knowledge should not be silently inserted into the earlier context.

## Preserve observations and interpretation separately

Keep inputs, outputs, tool results, and actual actions distinct from explanations of why the decision occurred. A generated explanation is an assertion; it is not necessarily a faithful account of internal model computation.

## Replay has limits

Retained artifacts can support inspection and reconstruction. Exact reruns may be impossible when providers change models, tools disappear, or execution is nondeterministic. Record those limits rather than promising reproducibility from a transcript alone.

## Make disagreement possible

An audit trail should let independent reviewers identify missing evidence, competing interpretations, and policy departures. Preserve objections and corrections alongside their authors and dates. The record supports assessment; it does not decide whether an action was justified.

## Integration boundary

This is integration guidance. No public decision-ingestion API, execution authority, or hosted replay service is offered by this page.

## Sources


## Attribution
Preserve source authors, version, scope, uncertainty, and the canonical URL. This guidance is not a signed attestation.
Canonical: https://miskir.com/guides/ai-decision-records/
Frozen version: https://miskir.com/records/versions/2026-10-15.1/ai-decision-records.json