EVE AI Core

Machine learning is absent from the governance path, which is the headline claim, and only from there. In that path 126 enforcement rules evaluate deterministically, so the same input always produces the same verdict. That is a design choice about how governance should work, not an absence of AI from the business. The company exists only because models are in production, and its broader platform runs the AI it governs. A runtime plane handles model routing, agent orchestration, memory retrieval and response generation, alongside persistent identity, long term memory and agentic behavior.

Oversight is built into the architecture, with the control placed before the action rather than after it. Every proposed action is intercepted and risk scored, returning allow, block or modify. The gate is fail closed, so nothing reaches the world unless it clears. Agent governance is properly distinguished from prompt guardrails. The control operates on tool invocations, interface calls and data writes rather than text around a model call.

Unregistered tools are refused outright, and high stakes steps can require human approval before executing. Kill switches and a stakes classifier sit in the pre inference plane. The company names the failure mode it rejects precisely: letting the action through and logging it is exactly the incident one is trying to prevent.

The evidence properties are strong, and each is specified cryptographically rather than described. Every verdict emits a signed certificate that a third party can verify offline with no vendor service in the loop. Decisions append to hash chained trails aggregated into signed tree structures. Each decision is bound to the exact policy version that governed it, so it can be rerun and confirmed.

The evidence survives independently of the vendor, which is what distinguishes it from ordinary immutable logging. The company also names a real and little discussed risk: because behavior is governed by policy rather than by whichever model sits behind the decision, silent provider updates cannot silently change outcomes.

No customer, institution or deployment is named, and no adoption figure, action volume or outcome measure is published. Eighty-five patent filings and a public code repository demonstrate investment and allow technical inspection, neither of which evidences that a regulated institution has put this in its decision path. For an enforcement layer that sits between AI intent and execution, that absence is the central gap.

Two structural features help: the control plane can be self-hosted, and evidence is verifiable offline without the vendor in the loop, which together limit dependence on the company. No boundary statement accompanies them. A system that intercepts every AI action across customers observes what regulated institutions attempt and what gets blocked, and nothing states whether that informs policy packs, is isolated per customer, or is retained.

No data protection agreement, retention schedule, subprocessor list or redaction mechanism is published. The platform sits in the request path for regulated decisions and retains signed records of each one. What those records contain about the applicant, and how long they persist, is a question the architecture makes unavoidable and the published material does not answer.

A trust center is published, and the security engineering visible in the code is disciplined. The veto engine imports only standard library modules, with no third party packages at all. A deliberate load time module freeze guard adds to that, removing external dependency risk from the security critical path entirely. Signing and attestation use named, standard cryptographic primitives. No independent audit attestation or certification is named, so the control environment around the code is unevidenced even though the code itself is inspectable.

The regulatory reasoning is deep, if narrow in coverage. It engages the current interagency model risk guidance directly, along with its predecessor's history, and makes a sophisticated argument about controls. Where validation identifies that a model may produce noncompliant adverse action narratives under certain input conditions, periodic review of a sampled output log is not an adequate control.

The failure is conditional and sampling will miss it, so the right control is a mechanism that keeps noncompliant output from reaching a decision maker at all. Two regimes are engaged deeply rather than a broad set of obligations being mapped. The company is an infrastructure provider rather than a licensed entity.

Fair lending is addressed at the mechanism level rather than as an aspiration. The platform governs automated credit and underwriting decisions against those obligations and produces a signed record of every approval and denial. The adverse action problem under the Equal Credit Opportunity Act (ECOA) serves as the worked example of why prevention beats sampling. Insurance pricing, claims and eligibility models are covered on the same basis. The platform enforces the policies a customer writes rather than testing for disparity itself, so no bias metric, fairness evaluation or disparity analysis is offered.

There is no published guarantee, indemnity or correction process. The affected consumer benefits indirectly and substantially, since a signed record of every approval and denial, bound to the governing policy, is the raw material for explaining a decision. Nothing states whether an applicant can obtain that record or what happens if the gate wrongly blocks a legitimate action. Nor does anything say what the vendor stands behind if evidence proves insufficient under examination.

Model agnosticism is architectural rather than aspirational. A provider registry and the explicit property that providers can be changed or self hosted without altering the control plane directly address the third party model dependency supervisors expect institutions to manage. The dependency disclosure for its own critical path is precise, naming the exact standard library imports of the veto engine and confirming no third party packages. No model provider is identified individually.

Integration is aimed at the AI stack, with a provider registry, agent orchestration, tool registration and a unified router, and the deliberate design point is that the control plane is independent of the model behind it so providers can change without touching governance. No banking, lending, claims or case management system is named, so how enforcement verdicts and signed evidence reach an institution's own control environment is not described.

Self-hosting is explicitly supported and stated as changing nothing about the control plane, which matters for an institution unwilling to route regulated decisions through a third party, and provider changes are similarly decoupled. Sub-millisecond evaluation implies the gate runs adjacent to the workload rather than as a remote call. No hosted region, residency commitment or infrastructure detail is published for customers who do not self-host.

A pricing page exists in the site structure alongside a trust center and whitepaper, but no pricing, packaging or basis of charge is public. The charging model is material for a control that evaluates every action, since cost could follow action volume, protected applications or seats. None of it is indicated.

Applicability is described across regulated decision types rather than institution types, covering automated credit and underwriting against fair lending obligations, insurance pricing, claims and eligibility models, and trading, with healthcare named alongside. That framing suits an enforcement layer, which cares about the decision rather than the buyer. No institution class, size range or geography is evidenced, and coverage remains a statement of where the control could apply.