CoreGuard

AI Policy Enforcement

EVE enforces policy at the point of action. A proposed tool call, model action, or MCP request is evaluated against policy packs, and EVE returns a deterministic ALLOW, BLOCK, or MODIFY verdict before the side effect occurs.

How enforcement works

  • Policy packs — CoreGuard evaluates requests against policy packs (29 packs ship in-tree covering domains such as lending, banking/AML, and PII handling), using deterministic threshold and rule functions.
  • Three dispositions — ALLOW proceeds, BLOCK stops the action, MODIFY forwards only the modified request.
  • Reason codes — every non-trivial disposition carries deterministic reason codes and a decision identifier.
  • Fail-closed — if the governance service is unreachable, enforcement resolves to BLOCK or raises; there is no silent permissive fallback.

SDK example

from eve_coreguard import CoreGuardClient

client = CoreGuardClient(api_key="eve_sk_...")

result = client.evaluate(
    request_id="req-001",
    tenant_id="org_abc",
    proposed_action={"type": "export_customer_pii", "format": "csv", "rows": 20000},
    model_output={"decision": "export", "confidence": 0.88},
    context={"principal_id": "svc_export", "session_id": "sess_44"},
    policy_set="pii_handling",
)

match result.decision.status:
    case "ALLOWED":
        do_export()
    case "MODIFIED":
        do_export()  # only the modified (forwarded) request proceeds
    case "BLOCKED":
        log_blocked(result.decision.action)

Readiness

Policy enforcement via CoreGuard is PILOT_READY (128 CoreGuard tests passed). Fail-closed deployment across embedded, hosted, sidecar, MCP-gateway, and sovereign modes is PILOT_READY. The SDK core is a RELEASE CANDIDATE WITH EXCLUSIONS.

Limitations

  • Enforcement covers the governance verdict; the governed application is responsible for actually gating on the returned action.
  • The pip client enforces via hosted mode and requires a reachable endpoint; embedded enforcement is the EVE service.
  • Standards mappings that accompany policy packs are descriptive; they are not a certification and do not by themselves make a system compliant with any framework.

Next step

Bring a policy set into a pilot and watch ALLOW / BLOCK / MODIFY decisions land on real requests with signed evidence.

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.