Evidence guide for auditors

Use this page to answer an auditor’s “show me”. Each stream below says what it records, where to read it, how it leaves the platform and how long it is kept by default.

Evidence streams

Stream	Records	Read it	Export	Default retention
Run step records	Every model call, tool call, retrieval, approval, delegation, cap, error and output of an agent run, in order, with the agent release, hash-chained per run and anchored in the security audit log	Reconstruct an agent run	Audit export (run_steps stream)	Kept (step_record class); content per the step_content class
Administrative audit trail	State-changing and refused requests on admin and security routes: who, what, when, outcome. Hash-chained since the upgrade that introduced run step records.	Audit trail	Audit export (OpenTelemetry logs or a signed webhook)	Per the admin_audit retention class
Security audit log	Agent identity, credential grant, delegation, redemption and decision events, hash-chained per tenant	The security-audit verify CLI (credential governance)	Audit export	Kept (security_audit class)
LLM gateway request log	One row per model call: model, tokens, cost, latency, outcome and run id. Prompt and response text are not stored.	LLM proxy analytics	Database read	92 days (gateway_request_log class)
Conversations and run events	Messages, tool calls shown in chat, and execution events	The conversation itself	—	Per the conversation class (kept unless set)
Suspensions	Every stop and resume, with scope, reason, actor and source	Runtime safety	Audit export	admin_audit class
Retention and legal hold changes	Policy edits, and holds placed and released	Audit evidence and retention	Audit export	admin_audit class

The export schema is documented in Audit event schema. Every exported event carries its source, its actor and, for chained streams, its chain fields.

Verify that nothing was altered

Three streams are hash chains. Each record’s hash covers its content and the previous record’s hash, so a changed, removed or inserted record breaks the chain at that point.

Chain	Verify with
Security audit log, one chain per tenant	The verification command on the credential governance page
Run step records, one chain per run, anchored in the security audit log when the run ends	GET …/run_steps/prompt_lib/{project_id}/{run_id}/verify (Reconstruct an agent run)
Admin audit trail, one platform-wide chain	GET /api/v2/elitea_core/audit_chain_verify/administration (Admin audit trail chain)

Each answers with the first record that does not match.

In your SIEM or archive, a gap in sequence numbers or a hash that does not chain shows that events were lost or altered on the way.

Warning

A hash chain makes tampering detectable, not impossible. For records that must be immutable, export to storage with object lock (WORM) and keep the copy of record there. See Shared responsibility.

Reconstruct what an agent did

Every agent run has a chained step record: each model call, tool call, retrieval, approval, delegation, cap, error and the final answer, in order, tied to the exact agent release, with the content of each step kept by reference. Follow Reconstruct an agent run to go from a chat answer to its steps, read what went in and out, and verify the record.

For spend and latency per model call, the gateway request log rows carry the same run id.