Assurance & operations

Audit Evidence Integrity

Can an investigator trust the record of what an agent did?

Reference notes by Min · Reviewed October 15, 2026

Working governance definitionAlso: Tamper-evident agent logs · Evidence completeness · Agent audit evidence2 primary references

Definition

Audit evidence integrity is the protection of operational records against unauthorized alteration or deletion, together with enough context to assess their origin and completeness. For agent workflows, it includes the reliability of the event producer and its link to the actual service outcome.

An investigation needs a defensible connection between the requested task, permission decision, attempted action and observed result. Each link may come from a different system with different clocks, retention and access controls. A long transcript does not necessarily supply that chain.

What it covers

Origin
Identify the component producing each record.
Protection
Restrict modification, deletion and logging configuration changes.
Completeness
Detect dropped events, missing stages and collection failures.
Reconciliation
Compare intended actions with downstream receipts and state.

Integrity and completeness are separate questions. A protected record may omit failed calls or events lost during an outage. A signed statement provides evidence about its origin and integrity under the signing assumptions, but not necessarily about the accuracy of its contents.

In practice

Practical design patterns by Min.

  • Emit consequential action records from trusted service boundaries.
  • Separate execution privileges from evidence-management privileges.
  • Monitor collection failures and retain meaningful event identifiers.
  • Reconcile a completion claim with the service that committed the action.

Define a minimal event contract: event identity, task reference, actor, target reference, decision and service result. Avoid copying secrets or unnecessary document bodies. Store collection-health signals separately enough that an agent failure does not also hide the failure of its evidence pipeline.

Research notes

Source-grounded details and editorial applications. Follow each link for the original document.

Protect records and logging tools

NIST SP 800-53 AU-9 addresses unauthorized access, modification and deletion of audit information and protection of logging tools. It also calls for alerts when unauthorized activity is detected.

Source: NIST SP 800-53 Rev. 5: Security and Privacy Controls ↗

Tampering must be detectable

OWASP recommends tamper detection and restricted, monitored log access. The completeness and truthfulness of the original events still require separate consideration.

Source: OWASP: Logging Cheat Sheet ↗

Working example

An agent reports a customer refund

The conversation says the refund succeeded, but the customer reports no credit.

  1. Locate the proposal, approval and tool call using a shared transaction reference.
  2. Check the payment service’s accepted, failed or pending outcome.
  3. Resolve the transaction without assuming the conversational summary is authoritative.
Illustrative scenario, not a reported deployment.

Why the controls matter

The refund may have been accepted while settlement remains pending. Preserve that distinction. Retrying an uncertain operation can create duplicate effects, so reconcile with the service’s transaction state before authorizing another attempt. Record any gaps that prevent a confident reconstruction.

Common pitfalls

  • Calling a log immutable without defining who can delete its storage.
  • Treating a model’s self-report as a service receipt.
  • Assuming signatures prove that a logged action was authorized.

Protected storage is only one part of the chain. Review who can disable collection, change the event schema or impersonate an event producer. Document limitations in incident findings instead of presenting an incomplete reconstruction as a fully observed sequence.

Questions for operational review

Use these questions to decide what evidence your team needs. They are editorial prompts, not a compliance checklist.

  • Which records come from the executing service?
  • How are missing events detected?
  • Who can change logging behavior or delete evidence?

References

Primary documents reviewed for this entry. Examples and implementation patterns are editorial synthesis.

  1. NIST SP 800-53 Rev. 5: Security and Privacy Controls ↗Accessed October 15, 2026
  2. OWASP: Logging Cheat Sheet ↗Accessed October 15, 2026