EVE AI Core
The Infrastructure of No.
EVE AI Core
The Infrastructure of No.
One pre-execution layer where every AI decision is checked against policy before it runs — and returns the same verdict for the same input, every time. Deterministic, not probabilistic. Enforced, not advised. Every outcome is recorded in a signed Decision Certificate you can verify offline with the published Ed25519 public key.
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 SR 11-7 examiners ask for model-decision evidence directly, 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.
“Deterministic AI governance means policy decisions return the same result for the same input every time, before execution.”
Two words carry the whole idea. Deterministic — the same input under the same policy version yields the same verdict, so a decision is reproducible and auditable instead of a one-off model output. Before execution — the check happens at the gate, so a disallowed action never runs in the first place. That is the difference between a policy that advises and a control plane that enforces.
In one sentence: EVE AI Core is the deterministic runtime enforcement and cryptographic evidence layer for enterprise AI governance. Most AI governance platforms provide visibility, dashboards, and compliance workflows; EVE enforces the decision before execution and produces signed, replayable proof after it.
An LLM-based filter scores a request and usually blocks the bad ones. “Usually” is not a control. When a regulator asks “was this specific decision authorized, and would the same inputs always produce the same outcome,” a probabilistic system cannot answer. A deterministic control plane can.
A deterministic gate is a pure function of the inputs. Replay the exact request a year later and you get the byte-identical verdict — the property an examiner, an auditor, and a courtroom all require.
The verdict is computed before the action executes. A blocked decision never reaches production — the decision and the consequence are not the same event, so there is a real place to say “no.”
Every outcome — allowed, modified, or blocked — emits a signed Decision Certificate, hash-chained per tenant and verifiable offline. The audit answer is “here is the signed record,” not “trust the logs.”
The two approaches differ on every axis a regulated buyer cares about — reproducibility, timing, and evidence.
| Property | Probabilistic guardrail | Deterministic control plane |
|---|---|---|
| Same input → same verdict | Not guaranteed (sampling, temperature) | Guaranteed — a pure function of inputs + policy version |
| When it acts | After the model generates output | Before the action executes |
| On a violation | Flags or scores; the output may still ship | BLOCKED — the action never runs |
| Evidence | Logs you are asked to trust | Signed, hash-chained Decision Certificate, verifiable offline |
| Replay a past decision | May differ on re-run | Byte-identical from inputs + policy version |
| Auditor answer | “It usually catches it” | “Here is the signed record” |
Each property of the control plane answers a specific examination requirement.
| Regime | What examiners ask | Control-plane property |
|---|---|---|
| SR 11-7 (model risk) | Show the decision and that controls were applied | A signed Decision Certificate per decision, replayable from its inputs |
| EU AI Act (Art. 12 & 13) | Automatic record-keeping + transparency for high-risk AI | Hash-chained, per-tenant evidence retained and verifiable offline |
| ECOA / Reg B | Reproduce and justify an adverse action | A deterministic verdict reproduces the exact outcome on replay |
| NIST AI RMF (Govern / Measure) | Enforced, documented governance — not advisory | Pre-execution enforcement (HARD_BLOCK), not post-hoc scoring |
An application (or an autonomous agent) does not act directly. It proposes an action. The proposal transits a deterministic governance pipeline — a charter compliance check whose HARD_BLOCK vetoes cannot be overridden, cognitive locks, and policy-pack evaluation — which returns ALLOWED, MODIFIED, or BLOCKED and signs a certificate for the outcome. Same engine, same verdict, every time.
check_charter() · check_cognitive_locks() · check_drift_budget() · policy packs · signed Decision Certificate per outcome
veto_core governs a single decision, a tool call, and a multi-agent mission — there is no privileged path around it. HARD_BLOCK charter rules run on every request regardless of caller, tier, or trust score.
The verdict is a pure function of the inputs and the active policy version. No sampling, no temperature, no drift — the same request always resolves to the same ALLOWED / MODIFIED / BLOCKED.
Decisions are gated before the action runs. Charter HARD_BLOCKs cannot be overridden by the model, the prompt, or the operator — the disallowed path simply does not execute.
Every outcome emits a signed certificate (Ed25519 in production), appended to a per-tenant hash chain and verifiable offline with the published public key via EVE Proof.
Because the gate is deterministic, any decision can be replayed from its inputs and policy version to reproduce the exact verdict — the foundation for SR 11-7 model-decision evidence and EU AI Act record-keeping.
Each tenant gets isolated policy, certificate chains, and evidence. No data leaves your tenant — SaaS, VPC, or on-prem — and policy logic is not embedded in your stack.
The deterministic core runs in the request path with a tight latency budget, so enforcement is something you keep on in production, not a batch report you read after the fact.
This is the hub for deterministic governance at EVE AI Core. The products enforce it, the documentation specifies it, and the field notes work through the hard cases — each one links back to this page.
We’ll run one of your real decisions through the control plane, show you the same verdict twice, and hand you the signed certificate your auditors will verify offline.