EVE for Model Risk Management
For model risk teams that must show a validator or examiner what was decided, by which policy, and with what evidence — not just that a model produced an output.
The problem
Model risk frameworks put agentic and generative systems under scrutiny: what actions can a model trigger, under what controls, and how is each decision recorded? An LLM's own reasoning is not a control artifact — it is non-deterministic and cannot be independently replayed.
How EVE helps
- Deterministic decision boundary. EVE CoreGuard makes a deterministic policy decision (ALLOW / BLOCK / MODIFY) before a governed action executes. The verdict is computed by deterministic policy code — no LLM is in the decision path — so the same inputs yield the same decision.
- Signed decision evidence. Every governed decision can produce a signed evidence record bound to the decision (Ed25519 in the production configuration, HMAC-SHA256 fallback). This supports evidence collection for a model risk file.
- Independent verification. Signed evidence can be verified independently and offline — in Python, Node, or the browser — without trusting the EVE server, so a validator checks integrity, signature, and schema on their own.
- Non-authoritative candidate evaluation. Authenticated customers can evaluate a candidate policy in shadow — live or by historical replay — and see divergences without changing the authoritative decision. Shadow reports are signed and carry modeled-effect estimates.
EVE produces records useful for review and can map controls to frameworks descriptively; it does not issue regulatory approval.
What a decision record contains
A GovernanceResult carries the action taken (ALLOW / BLOCK / MODIFY), reason codes, a
decision id, and a signed evidence envelope. Verification confirms the record is authentic and
unaltered — it attests to the evidence, not to the correctness of the underlying decision.
Readiness
- EVE SDK core is a RELEASE CANDIDATE WITH EXCLUSIONS.
- The core governance architecture (CoreGuard decision, signed evidence) is pilot-validated.
- Signed decision evidence and offline verification (jcs-1 canonicalization) are SUPPORTED.
- The Shadow subsystem is authenticated pilot-ready.
- The Python client (
eve-coreguard0.2.2) and TypeScript client (@eve/coreguard0.1.0-rc1) are release-candidate core clients, distributed privately for pilots and not published to public registries at this stage.
Limitations
- EVE supports evidence collection and can map controls to frameworks for review; it does not determine regulatory outcomes or issue approval.
- Independent (asymmetric) verification requires the Ed25519 public key; the HMAC fallback is symmetric and not independently verifiable.
- The pip client governs via hosted mode (requires a reachable endpoint); embedded in-process evaluation is the EVE service, not the client wheel.
- Shadow evaluation is structurally non-authoritative and cannot alter the active verdict; divergence figures are modeled effects from replay/sample, not observed production outcomes.
- Statistical anomaly detectors are advisory unless wired to a deterministic enforcement action.