EVE AI Core
The Infrastructure of No.
EVE AI Core
The Infrastructure of No.
Guardrails AI validates model output after it is generated. For actions that carry regulatory or safety consequences, the decision has to be made before the action runs — and it has to leave an auditable record. This guide maps Guardrails concepts to EVE CoreGuard, then walks a parallel-run cutover that keeps your existing validators where they still fit.
You control which decision fields are sent; in VPC and on-prem deployments, governed content and evidence stay inside your boundary. As EU AI Act obligations phase in and, although SR 26-2 (the 2026 interagency model-risk guidance that superseded SR 11-7) excludes generative and agentic AI from its formal scope, banks must still govern technologies outside it, the record has to exist at the moment of the decision. Trust Center · Security
Framework mappings describe technical control support and evidence capabilities. They do not constitute legal advice, regulatory certification, or a guarantee that a customer’s deployment is compliant. “SOC 2 readiness” is a preparation status, not a completed audit.
Guardrails AI and EVE CoreGuard solve different problems and sit at different points in the pipeline. Guardrails is an open-source validation library that checks an LLM output against a schema, regex, or validator function after generation. EVE CoreGuard is a deterministic, fail-closed governance boundary that returns an ALLOW/BLOCK/MODIFY decision on a proposed action before it executes, with a signed, offline-verifiable certificate for every decision. The goal of a migration is not to delete your validators — it is to move the consequential decisions behind a governed boundary and keep the rest where they are.
A field-by-field map from what you have today to the equivalent construct in EVE CoreGuard.
| Guardrails AI concept | EVE CoreGuard equivalent |
|---|---|
| Validator (Python function or Hub validator) | A rule inside a versioned policy pack — or keep the validator and feed its verdict to CoreGuard as an input signal |
| Post-generation validation of an output | Pre-execution decision on a proposed action, before it runs |
| Pass / Fail on a prompt or response | ALLOW / BLOCK / MODIFY on the action |
| Deterministic for schema/regex validators; probabilistic for ML-based validators | Deterministic rule evaluation — same input under the same policy version always yields the same decision |
| Self-hosted library, or Hub validators you run and keep available | Managed API with an enterprise SLA, or self-hosted in your VPC / on-prem — on the managed tier, enforcement availability is EVE’s responsibility |
| Validation log line | Ed25519-signed decision certificate, verifiable offline without calling EVE |
| Behavior on validator error is left to your integration | Fail-closed by default |
Nothing here requires a big-bang switch. You add EVE alongside Guardrails, compare decisions, then move the enforcement boundary.
Steps 3–4 are where a migration earns its keep: the shadow run tells you exactly where post-generation validation and pre-execution enforcement would have diverged, before anything is enforced in production.
The same intent, moved from “validate the output after it exists” to “decide on the action before it runs, and sign the decision.”
The model has already produced the output; the validator decides whether to keep it. If the action already fired, the validator runs too late to prevent it.
# output already generated, then validated from guardrails import Guard guard = Guard().use(SomeValidator, on_fail="exception") result = guard.validate(model_output) # pass / fail on the text
The proposed action is evaluated against a versioned policy pack before it executes. The response is a deterministic decision plus a signed certificate you can verify offline.
# decide on the action BEFORE it runs curl -X POST https://api.eveaicore.com/v1/decisions/evaluate \ -H "Authorization: Bearer eve_<key>" -H "Content-Type: application/json" \ -d '{"proposed_action": {...}, "context": {...}, "policy_set": "lending_v1"}' # -> {"decision":{"status":"ALLOW|BLOCK|MODIFY"}, "audit":{... signed ...}}
You do not have to choose one tool. Use this split to decide, guard by guard.
Reuse-not-rewrite: EVE’s assurance layer already ingests external finders (for example Promptfoo and Garak) as signals. The same adapter pattern lets a Guardrails validator’s verdict feed CoreGuard as an input signal while EVE remains the authoritative pre-execution decision — so existing validation logic is not thrown away. A packaged Guardrails signal adapter is not shipped today; it is a supported integration pattern.
Pick a single consequential guard, shadow it against a CoreGuard policy pack, and inspect the signed decision certificate. No credit card required to evaluate.