CoreGuard
AI Governance Runtime
EVE is a runtime enforcement layer: it sits in the execution path and decides — deterministically — whether a governed action may proceed, at the moment it is about to run. This is different from a static policy checker or a post-hoc log analyzer. The runtime returns a verdict before the side effect, and it fails closed on error.
Runtime properties
- In-path decisioning — the ALLOW / BLOCK / MODIFY verdict is produced synchronously before the tool call or MCP request executes.
- Deterministic core — verdicts come from policy code, not a model; the decision path contains no LLM.
- Fail-closed — a service error (embedded, hosted, sidecar, MCP-gateway, sovereign) resolves to BLOCK or raises. No silent permissive fallback.
- Signed evidence per decision — the runtime can emit an evidence envelope bound to the decision for independent verification.
- Warm latency — Python embedded govern measures p50 10.8 ms / p95 15.7 ms / p99 17.0 ms (CoreGuard-dominated; hosted round-trip is network-bound).
SDK example
from eve_coreguard import CoreGuardClient
client = CoreGuardClient(api_key="eve_sk_...", base_url="https://api.eveaicore.com") # runtime enforcement via the CoreGuard API
result = client.evaluate(
request_id="sess_2",
tenant_id="org_abc",
proposed_action={"type": "send_email", "to": "[email protected]", "attachments": ["report.pdf"]},
model_output={"decision": "send"},
context={"principal_id": "agent_mail", "session_id": "sess_2"},
policy_set="agent_tools_v1",
)
if result.decision.status != "ALLOWED":
handle_denied(result.decision.status, result.decision.action)
Links
- Determinism explained: /seo/landing/deterministic-ai-governance.md
- Deployment modes and SDKs: /seo/landing/ai-governance-sdk.md
- MCP request binding: /seo/landing/mcp-governance.md
Readiness
Runtime enforcement (CoreGuard decision, fail-closed deployment modes) is PILOT_READY; the core governance architecture is pilot-validated. Signed evidence and offline verification are SUPPORTED. The SDK core is a RELEASE CANDIDATE WITH EXCLUSIONS.
Limitations
- Latency figures are warm-path measurements; hosted round-trip latency is network-bound and depends on your endpoint placement.
- The pip client runs in hosted mode and needs a reachable endpoint; embedded in-process enforcement is the EVE service.
- Wrapper-based framework integrations can be bypassed by calling the underlying tool directly; hard runtime enforcement requires a gateway or sidecar.
Next step
Deploy EVE as an in-path sidecar in a pilot environment and measure real verdict latency on your workloads.