A system can be observable and still be hard to answer for.
Operational logs are built to run software. Consequential AI also needs an account prepared for the person who did not build or operate it.
When an AI system affects a customer, employee or citizen, the difficult question is rarely whether some logs exist. It is whether another person can use the available record to understand what happened and challenge the explanation.
Engineering logs are necessary. They help teams operate, debug and secure complex systems. But their structure, retention and vocabulary follow the needs of the operating team. A risk committee, auditor or affected person arrives with a different question.
The reader changes the record.
A technical trace may show a request, a tool call and a response. A reviewer also needs to know which policy applied, whether an exception was raised, who approved it, what source produced each fact and which parts of the workflow were not observed.
This is why Arkna separates the underlying activity from the account prepared for review. The aim is not to disparage the source logs. It is to preserve their references while making the relevant sequence legible to a different reader.
Independence is not a word on the cover. It is a question about control, sources and who can check the result.
Separate does not mean omniscient.
A record built from a gateway can describe interactions crossing that gateway. It cannot reveal a supplier agent's internal reasoning if the supplier never exposes it. A record built from an SDK can include the events the integration emits. It cannot establish that an unreported event never occurred.
A credible account therefore states its boundary. It distinguishes integrity from completeness: whether included data still matches the record, and whether the record included everything the reviewer needed, are different questions.
Accountability remains human.
A record does not decide whether an outcome was fair, lawful or appropriate. It gives responsible people a better basis for making that judgment and a clearer way to explain it afterwards.
That is the purpose of Arkna: make consequential AI activity easier to examine without asking the reviewer to trust the system, the operator or Arkna more than the evidence supports.