Why a separate record matters

A record prepared for review is a different document from an operational log. Operational logs exist to run and debug software. A reviewer who did not build or operate the system needs an account prepared for that purpose.

Arkna Last updated August 2026

Where an AI system affects a customer, employee or citizen, a reviewer must be able to establish from the record what happened and on what basis. The existence of logs does not settle that.

Engineering logs are necessary. Their structure, retention and vocabulary follow the requirements of the operating team. A risk committee, an auditor or an affected person has different requirements.

What a reviewer needs

A technical trace records a request, a tool call and a response. A reviewer additionally requires:

  • which policy applied to the decision
  • whether an exception was raised, and who approved it
  • what source produced each fact relied on
  • which parts of the workflow were not observed

Arkna therefore separates the underlying activity from the account prepared for review. The separation preserves the references in the source logs while making the relevant sequence legible to a different reader.

Whether an account can be relied on turns on control of the record, the sources it draws from, and whether a third party can check the result.

Limits of a separate record

Record built fromCan establishCannot establish
A gatewayInteractions crossing that gatewayA supplier agent’s internal reasoning, where the supplier does not expose it
An SDK integrationThe events the integration emitsThat an unreported event did not occur

A record therefore states its boundary. Integrity and completeness are separate questions: whether included data still matches the record, and whether the record held everything the reviewer required.

What a record does not decide

A record does not determine whether an outcome was fair, lawful or appropriate. It provides a basis for that judgement and a means of explaining it afterwards.

The activity of an AI system should be examinable without requiring a reviewer to trust the system, the operator or Arkna further than the evidence supports.