Trust in autonomous systems is not a feeling; it is an audit you can actually read.
What gets recorded
Every action through the gateway - executed, approved, refused, or reverted - is recorded with its actor (which agent, manager, or person), the verb, the affected records, the before and after, the contract decision that allowed or blocked it, the credit cost, and the correlation to its parent run.
Tamper-evidence
Audit entries form a hash chain: each record incorporates a fingerprint of the previous one, so removing or editing history breaks the chain detectably. Integrity can be verified on demand - an auditor does not have to take the log's word for it.
Reading it
- ›Fleet, run detail - actions in the context of the run that produced them (the day-to-day view).
- ›Admin, Audit Log - the flat, filterable record across the organization (the compliance view; Business plan).
- ›Refusals - blocked attempts are first-class entries with reasons, visible in Admin. What the AI tried and was denied is often more informative than what it did.
Why it matters
The audit chain is what makes the rest of the model honest: autonomy levels and contracts would be promises without an incorruptible record of what actually happened. It is also the substrate for run replay and revert.
Pro Tip: When evaluating any AI platform, ask one question: can it show you, verifiably, everything the AI did last Tuesday? Brainis can.