Auditability is a chain of records, not a citation count

A citation proves that a source exists. It does not prove that the source supports the claim, that the right record entered the calculation, or that the calculation produced the released recommendation.

Follow the chain

ARiDA auditability begins with the decision frame. The frame identifies the question, scope and accountable owner. Acquired records enter a source ledger. Derived outputs carry compute manifests and artifact hashes. Assumption sets identify the values accepted by the model. Canvas objects point to their producing artifacts. Release state records review and approval.

The chain is therefore: question, source, transformation, assumption, result, action and release. A reviewer can enter at any point and move backward.

Make invalid states visible

The platform validates governed work at dispatch, tool call and terminal output. Unsupported final render types show a blocked state rather than a decorative fallback. Missing source-ledger records fail validation. Empty acquisition does not become a “ready” result. Artifacts are promoted only after their bytes and hashes are verified.

This is auditability as system behavior. NIST describes transparency and accountability across the AI lifecycle and calls for systematic documentation, clear roles and evaluation in context (AI RMF). A reference list alone meets little of that standard.

Preserve change over time

Decision Cases and Asset 360 retain the accepted evidence cut, assumptions, alternatives, dissent and reopening triggers. New evidence creates a comparison with the prior state. The organization can see which input changed and which downstream conclusion moved.

An auditable system does not promise certainty. It gives uncertainty an address, an owner and a consequence.