Trust Center ยท Auditability
Every answer is traceable to its evidence.
An agent that cannot evidence the origin of a conclusion is not deployable in a regulated environment. Traceability is written into the execution path rather than reconstructed from logs after the fact.
Forensic tracing
OpenTelemetry spans across every hop: ingress, retrieval, model call, tool execution and write-back. A single trace id reconstructs an entire agent decision end to end.
Cryptographic audit record
Each action is written as a signed record. Tampering is detectable because the chain of records is verifiable independently of the application that wrote it.
Source traceability
Every answer carries the documents, records and prior decisions it drew on, so a reviewer can follow a conclusion back to its evidence.
Algorithmic kill-switch
Automated halt conditions stop an agent class immediately when behaviour crosses a defined threshold, without waiting for a human to notice.
Red teaming
Adversarial evaluation of prompts, tool access and data exfiltration paths, run in development and repeated against production behaviour.
Continuity and recovery
Defined recovery objectives with tested restore paths for the graph, the memory store and the audit chain.
Data Lineage
What an audit record contains
One record is written per agent action. Together they form the cryptographic chain an auditor reads.
Correlates every span in one agent decision
The human or agent identity that initiated the action
The scope resolved for this request, and the policy that granted it
Documents and records read, with their own identifiers
Model, version, and whether it ran privately or via a provider
Hash of the produced output, for later verification
Cryptographic signature over the record