EVE for Audit and Assurance

For internal audit and external assurance teams that need to test AI-agent controls with evidence they can re-verify themselves, rather than take on trust.

The problem

Auditing an AI agent means testing a moving target: outputs vary run to run, and the model's narrative of its own behavior is not admissible as a control record. Auditors need tamper-evident artifacts and a way to reproduce a decision on demand.

How EVE helps

  • Deterministic, replayable decisions. EVE CoreGuard computes a deterministic ALLOW / BLOCK / MODIFY decision before a governed action executes, with no LLM in the decision path. The same inputs produce the same decision, so a control test is reproducible.
  • Signed evidence bound to each decision. Every governed decision can produce a signed evidence record bound to the decision (Ed25519 in the production configuration, HMAC-SHA256 fallback) — records useful for review and re-verification.
  • Independent, offline verification. An auditor verifies signed evidence independently and offline in Python, Node, or the browser without trusting the EVE server, checking integrity, signature, and schema.
  • Substitution detection. For file inputs, scan evidence binds the artifact digest to the decision, so an artifact substituted after approval is detected at verification.
  • Customer-runnable adversarial suite. Teams can run a packaged adversarial regression suite (single-step and multi-step chains) against their own configuration; blocked attacks provably cause zero unauthorized side effects, and results are signed.

What re-verification proves

Verification confirms that an evidence record is authentic and unaltered — it does not attest to the correctness of the underlying decision, only that the decision, reason codes, and bound inputs are exactly as recorded.

Readiness

  • EVE SDK core is a RELEASE CANDIDATE WITH EXCLUSIONS.
  • Signed decision evidence and offline verification (jcs-1 canonicalization) are SUPPORTED.
  • The core governance architecture is pilot-validated.
  • Artifact and red-team subsystems are customer-pilot ready at supported boundaries.

Limitations

  • EVE produces records useful for review and supports evidence collection; it does not render an audit opinion or certify a system.
  • Independent verification requires the Ed25519 public key; the HMAC fallback is symmetric and not independently verifiable.
  • The red-team suite is representative, not exhaustive; customers add their own fixtures. Standards mappings are descriptive, not certification.
  • Deterministic scanners are pattern/structure-based; OCR/vision/QR-barcode extraction is probabilistic, disabled by default, and reported as unavailable — incomplete extraction is reported as partial and cannot masquerade as complete.
  • The pip client governs via hosted mode (requires a reachable endpoint); embedded evaluation is the EVE service.
Part of the EVE AI Core control plane Deterministic AI Governance Control Plane → Policy decisions that return the same result for the same input every time, before execution.