EVE vs Model-Provider Guardrails

Many model providers ship built-in safety filters and moderation controls at the model boundary. This page compares that category to EVE, sourcing only EVE's capabilities from the EVE claims registry. Category rows we cannot source are marked "not independently verified here." Because "model-provider guardrails" spans several providers with different features, we do not make per-vendor claims.

Positioning

Provider-side guardrails generally operate on model inputs and outputs — moderation, safety categories, and content filtering delivered as part of the model API. EVE operates on the action: it makes a deterministic decision (ALLOW / BLOCK / MODIFY) before a governed action executes, with no LLM in the decision path, and produces signed evidence bound to each decision that is verifiable offline without trusting the EVE server — and without depending on any single model provider.

Capability matrix

Capability EVE Model-provider guardrails (category)
Deterministic pre-execution decision on a tool call / side effect Yes — PILOT_READY not independently verified here
No LLM in the governance decision path Yes — PILOT_READY model-boundary filtering is model-driven; not independently verified here
Signed decision evidence bound to the decision Yes — SUPPORTED not independently verified here
Offline independent verification without trusting the vendor Yes — SUPPORTED not independently verified here
MCP approved-vs-executed binding (at-most-once) Yes — PILOT_READY not independently verified here
Server-derived MCP identity (forged headers rejected) Yes — PILOT_READY not independently verified here
Action-sequence governance Yes — PILOT_READY not independently verified here
Deterministic budget / resource limits Yes — PILOT_READY not independently verified here
Provider-neutral (not tied to one model vendor) Yes provider-specific by nature
Content moderation / safety categories on model text not the focus of EVE's sourced claims not independently verified here

Where EVE fits

Choose EVE when you need governance that is independent of the model vendor and that produces verifiable proof of each decision. Provider guardrails can screen model text; EVE governs the resulting action — a tool call, an MCP request, a file ingestion — and records a signed decision an auditor can verify offline. EVE fails closed on service error in every deployment mode (embedded, hosted, sidecar, MCP-gateway, sovereign): no SDK silently falls back to permissive behavior.

The two layers are complementary: keep the provider's model-boundary filtering and add EVE for deterministic, provider-neutral, action-level governance with independent evidence.

Readiness

  • EVE SDK core is a RELEASE CANDIDATE WITH EXCLUSIONS.
  • Python (eve-coreguard 0.2.2) and TypeScript (@eve/coreguard 0.1.0-rc1) are release-candidate core clients, distributed privately for pilots — not on public registries at this stage.
  • Core governance architecture is pilot-validated; signed evidence and offline verification are SUPPORTED.
  • Framework adapters other than the generic adapter are experimental.

Limitations

  • Only EVE's rows are sourced. Category rows are "not independently verified here" and are not accusations of missing features; the category spans multiple providers with differing controls.
  • The pip client governs via hosted mode (needs an endpoint); embedded evaluation is the EVE service.
  • Independent verification requires the Ed25519 public key; the HMAC fallback is not independently verifiable.
  • MCP binding is at-most-once, not remote exactly-once; unknown remote outcomes are recorded as OUTCOME_UNKNOWN.
  • Framework adapters other than the generic adapter are experimental; wrapper-enforced adapters are bypassable by a direct underlying-tool call, and hard enforcement requires a gateway or sidecar.
  • EVE governs the action and evidence layer; it does not replace a provider's model-boundary content moderation.
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.