FWD-3 Explaining an AI decision

Fijishi Working Draft 3 – 14 October 2026

Stage 1 of 4: working draft. Not yet open for public comment, not a numbered version, and not the criteria Fijishi certifies against. This draft replaces, as a new version, the earlier paper "Cognitive Integrity Protocols (CIP)" (2023-2026); the earlier text is available on request. Standards in development

The question

If a customer, a regulator or a court asks why an AI decision came out as it did, can the firm give a reason that is specific to that decision and true to how it was made?

Why it matters to firms

Rules in several jurisdictions require firms to tell people about, or explain, decisions made by automated means, and to offer a route to have them reviewed. A general description of how a model works does not answer why one claim was declined. An explanation that does not match how the decision was actually made can do more harm than none. Jurisdiction by jurisdiction: the Index.

Principles

  1. Specific – the reason is about this decision, not the model in general.
  2. Faithful – it reflects the factors that actually drove the outcome.
  3. Understandable – it is written for the person receiving it.
  4. Consistent with the record – it can be traced to the decision's record (FWD-2).
  5. Contestable – it tells the person how to ask for review by a person.

What a firm should be able to show (draft criteria)

  • FWD-3.1 For each decision type, a defined explanation: what is given, to whom, and when.
  • FWD-3.2 An explanation for a sampled decision, produced from that decision's record.
  • FWD-3.3 The method used to identify the main factors in a decision, and its documented limits.
  • FWD-3.4 A review of customer-facing reasons for plain language.
  • FWD-3.5 A route, stated in the explanation, to ask for review by a person.
  • FWD-3.6 A process for correcting wrong input data and re-examining the decision.

How it relates to existing rules and standards

  • It sets out how an explanation or notice requirement can be shown in evidence. Where such a requirement binds a firm, the requirement governs.
  • It is meant to operate inside an AI management system such as one certified to ISO/IEC 42001, and alongside the firm's model risk management.

Where it is used now

It informs the Constitutional Baseline Review and Juris.

What happens next

Public comment follows FWD-1. Comments are welcome now: protocol@fijishi.com