01Observed run
Start from the run
Review begins with the events Arkna received in order, rather than recollections and exports gathered after the event.
Evidence ready before the question is asked
Arkna preserves observed AI-agent activity while it happens, so the people examining an action later begin with a record made at the time.
An agent's own account should not be the only account.
When an AI action is questioned, the evidence may be spread across engineering tools, people and outside providers. Arkna preserves a durable, readable record for the people who have to establish what happened.
The tools used to run the system still do their job. The record is built for the different question that comes later: what happened, and what evidence supports that account?
Open the record. See what it holds.
02Provider involvement
See provider involvement
When activity from a connected service reaches Arkna, it appears beside the rest of the run. The record also states what fell outside Arkna's view.
03Shared evidence
Give reviewers the same evidence
Risk, assurance and audit can work from the same record without asking engineering to translate several systems into a new account.
04Limits of the record
Know where the record ends
An exported record can be checked for later changes. It also shows what Arkna observed and where its view ended.
05Incident record
Arrive with the record already made
Investigators can spend the first hours examining the event instead of building its record after disruption.
Testing the record in practice
We are speaking separately with people who build AI systems and people who approve them. The study asks whether both can reconstruct the same account from the records available six weeks after a system goes live.
Talk to us