Skip to main content
Batoi AIF Docs

Audit and Evidence

Capture accountable records for AI execution and review.

Version Foundation release Early foundation work Reviewed Jul 21, 2026 Next review Sep 21, 2026

Owner: Batoi Engineering Reviewed by: Batoi AIF Maintainers Source revision: cd680e8

AIF is the public engineering framework. For the managed intelligence capability used across supported Platform and Flex workflows, see Batoi Intelligence.
Browse Batoi AIF Documentation Audit and Evidence

Audit and Evidence

Every governed terminal state attempts one audit record, including success, policy denial, review required, provider failure, evaluation failure, and audit failure signaling. AIF uses canonical hashes and correlation fields so reviewers can connect the decision to the request without exposing raw sensitive content by default.

Request User, purpose, input Policy Allow, deny, review Execution Provider metadata Evidence Audit and review

Audit Record Contents

Audit records should include:

  • audit UID and terminal status
  • canonical request and response hashes
  • user, workspace, and trace correlation
  • operation
  • provider and model
  • prompt key and version
  • policy decision and version
  • provider request UID
  • latency, usage, token, and cost metadata where available
  • evaluation and review evidence
  • classified error code and safe message

Privacy Boundary

KeyNameSensitiveDataClassifier identifies common sensitive fields before evidence is retained. Do not log credentials or full sensitive content unless policy explicitly requires it and the storage path is permission checked. Prefer canonical hashes, references, labels, and redacted excerpts.

JsonLinesAuditExporter supports portable evidence export. RadPdoAuditLog provides MySQL persistence with immutable call-log protections. Retention remains a host-owned policy implemented through the audit retention contract.