OUR EVALUATION FRAMEWORK

A claim is a starting point.
Evidence is the standard.

Every entry should answer one question: why should you believe the provider or infrastructure operator cannot see your inference data?

The provider directory contains primary-source research on real model hosts and compute networks. Source review establishes documented capabilities; it is not a completed live attestation or independent security audit.

What qualifies for inclusion?

To establish technical privacy, a provider must use a mechanism intended to prevent a provider or infrastructure operator from accessing sensitive plaintext during inference: a protected execution environment, an attested encrypted channel, or a cryptographic private-computation construction. We identify exactly which operator is excluded, what data is protected, and which assumptions remain.

Providers must run the models themselves or supply the software and compute network where operators run them. API aggregators such as OpenRouter are excluded. For services with both hosting and aggregation, only the documented hosting tier qualifies, and unresolved per-model hosting stays explicit. The directory also compares hosts with policy-based privacy, such as Venice's own Private tier, at Stage 0. No logging, no retention, no training, dedicated compute, compliance certifications, privacy policies, and contractual promises alone do not establish technical privacy.

Privacy by policy

“We promise not to access your data.”

The operator may retain the technical ability to inspect plaintext. The protection is a behavioral or contractual commitment.

Privacy enforced by technology

“The architecture prevents access under stated hardware or cryptographic assumptions.”

The operator must break those assumptions, compromise a trusted component, or escape the protected boundary to read the data.

The five questions we ask

  1. What is protected? Inputs, outputs, intermediate values, and the explicitly documented scope.
  2. Who is prevented from seeing it? The host, provider administrators, individual MPC participants, or another precisely named party.
  3. What enforces that protection? Hardware isolation, verified workload identity, key binding, ciphertext computation, or secret sharing.
  4. Who and what must still be trusted? Hardware vendors, provider code, a verifier, a key broker, or cryptographic and non-collusion assumptions.
  5. How can the user verify it? Retrievable reports, expected measurements, public tooling, protocols, and reproducible evidence.

Content privacy stages

Every inference provider has a numbered stage from 0 to 3. A stage applies to the stated model tier, client, configuration and deployment path. It assesses inputs and outputs across the request path, computation, and content logs and storage. Identity and operational metadata remain visible as a separate assessment.

Stage 0: Privacy claimed

Privacy is claimed or promised by policy, but independently checkable technical proof of content protection has not been established for the assessed path.

Stage 1: Partial proof

At least one content layer has independently checkable technical evidence, including partial proof within a layer. The full privacy boundary is incomplete. The stage tooltip distinguishes partial findings from fully met requirements for Stage 2.

Stage 2: Full content privacy proof

All three content layers have technical proof: private inputs and outputs in transit, protected inference, and no content logging or persistent storage. Intermediaries are mitigated through E2EE or an equivalent attested channel. Any temporary cache has verifiable isolation and removal under an enforced lifecycle.

Stage 3: Independently auditable

Stage 2 plus independently inspectable implementations for every content layer, a checkable link to deployed artifacts, and verifiable update controls. Identity and operational metadata remain a separate assessment.

What Stage 1 requires

At least one of the three content layers below must have independently checkable technical evidence. This is partial proof: an attested inference worker can earn Stage 1 while its request path or retention remains unresolved. Policy promises and TEE marketing alone stay at Stage 0. If no technical proof is established, the provider still has Stage 0 and its factor assessments explain the evidence gaps.

Partial technical proof within one layer also earns Stage 1. For example, checkable VM isolation with unresolved GPU protection is partial execution evidence. It does not mark the complete execution requirement as met or qualify for Stage 2.

What Stage 2 requires

All three steps must be proven for full content privacy across the assessed path. Stronger evidence in one step cannot compensate for a missing step.

  1. Private inputs and outputs in transit. E2EE or an equivalent channel bound to the expected protected workload must prevent an intermediary from reading content. Verification happens before secrets are sent and fails closed. Ordinary HTTPS to a plaintext gateway is insufficient.
  2. Protected computation. Technical evidence must identify the approved workload and cover every component that processes plaintext, including CPU and GPU workers where applicable. Key custody, model identity and hardware or cryptographic assumptions must be explicit.
  3. Enforced content handling. Inputs and outputs must not be logged, persistently stored, used for training or exposed through traces, files or backups. Protected temporary caches are allowed when isolation and removal are verifiable against the accepted implementation and configuration. Content-derived cache state counts as sensitive content.

Remote attestation identifies running code and configuration. To establish cache removal, the accepted code must enforce its documented lifetime or clear operation. Memory-pressure eviction, a possible future restart, or rotating a cache salt does not by itself establish when old entries are removed. We do not infer deletion from a TEE badge or read-only disks.

Steps within each stage

Each provider's stage tooltip shows its current stage, the next stage's requirements, and the evidence status of each step. Stage 1 needs proof for any one content layer; Stage 2 needs all three. Stage 3 additionally requires independently auditable implementations across all three. Missing evidence is not proof of failure, but it cannot satisfy a step.

Metadata is assessed separately

Accounts, IP addresses, API keys, token totals and billing records do not automatically lower the content stage. Their collection, linkage and retention remain explicit in the metadata column. Telemetry containing prompts, responses or content-derived state is still assessed as content retention; calling it metadata does not exempt it.

Stage 3 requires independent inspection of all three content factors, their deployed-artifact linkage and update controls. Public source alone is insufficient. Verifiable and auditable describe available capabilities under stated assumptions; this index has not completed a live attestation or security audit.

How to read verification labels

Each factor has its own evidence label. The directory separately assesses request protection, inference execution, logs and storage, and identity and metadata. A source-review date records research, not successful verification of every live deployment.

Independently auditable

The content-critical implementation can be independently inspected and linked to the deployed artifact, with explicit update controls. This describes an available capability; it does not mean this index has completed an audit.

Technically verifiable

Checkable technical evidence establishes the assessed factor under its stated hardware or cryptographic assumptions. Other factors may still have unresolved requirements.

Partial technical proof

Checkable technical evidence establishes part of a factor, while significant parts remain unresolved. This can earn Stage 1 but does not satisfy a full Stage 2 requirement.

Documented only

Sources describe a policy or design, but independently checkable enforcement has not been established for the assessed factor.

Operator access

The assessed architecture permits an operator or intermediary to access plaintext. Its behavior remains a trust dependency.

Unknown

The reviewed sources do not establish this factor. Missing evidence does not demonstrate a failure, but cannot satisfy a stage requirement.

Assurance describes the approach

Cryptographically private

The stated inference construction uses ciphertext or secret shares without exposing complete inputs to an individual computation provider, subject to its scope and cryptographic or non-collusion assumptions.

Attested confidential inference

Plaintext exists inside a TEE. The client can verify the expected workload, and secrets are bound to successful attestation. Hardware, approved application behavior, and verification code remain trusted.

Hardware-enforced confidentiality

Hardware isolates inference from the host. Workload identity or key release is not fully independently verifiable. This label does not establish that the provider cannot decrypt inputs.

Hybrid confidential computation

The pipeline combines mechanisms. Every stage has its own plaintext boundary; partial encrypted computation does not establish ciphertext-only processing throughout.

Insufficient evidence

The claimed mechanism may be relevant, but available sources do not establish the end-to-end guarantee. The factor assessments and numbered stage make the unresolved requirements explicit.

These approaches are not an absolute security ranking. FHE, MPC, and TEEs have different protected scopes, threat models, implementation risks, and trust assumptions. We do not collapse those differences into a score.

Evidence standards

We prioritize inspectable enforcement over an unsupported assertion. The list below describes progressively weaker direct support for an implementation claim; a cryptographic protocol and a hardware report answer different questions and are not substitutes for one another.

  1. User-retrievable attestation evidenceSigned reports a customer can obtain, with freshness checks and a clear hardware endorsement chain.
  2. Public attestation verification procedureInstructions identifying exactly what is checked, by whom, and before which secrets are sent.
  3. Attestation-bound key release designEvidence that a session key or decryption secret is released only to the intended, verified workload.
  4. Open-source verifierInspectable verification code, expected-measurement handling, failure behavior, and session-key binding.
  5. Open workload or reproducible measurementA connection between reviewed source, the deployed workload, and the measurement a customer checks.
  6. Published cryptographic constructionA protocol, threat model, protected computation scope, and explicit security assumptions for FHE or MPC.
  7. Independent security analysisA scoped analysis tied to the relevant implementation and version; audits do not replace checking deployment identity.
  8. Detailed architecture documentationA traceable plaintext boundary, key lifecycle, and account of all trusted components.
  9. Vendor technical claimUseful as a lead for investigation. A claim alone does not establish customer-verifiable enforcement.

Each source states which fields it supports. Vendor claims are labeled separately. A link to documentation is evidence of a documented design, not proof that every deployed instance implements it.

Unknown means unknown

We do not infer attestation from TEE support, workload identity from a boot measurement, key isolation from encryption, or end-to-end private inference from ordinary HTTPS. If a source does not establish an answer, we display “Unknown.”

“Not applicable” is reserved for a property absent from the architecture, such as hardware attestation for a pure FHE circuit. False and unknown values never satisfy affirmative verification filters.

Trace the entire plaintext boundary

Ordinary TLS protects data while it travels to a TLS endpoint. If that endpoint is a provider gateway, plaintext may exist there before inference. An attested encrypted inference channel instead binds the destination key to a verified protected workload and decrypts only inside that boundary.

For FHE, we ask which operations use encrypted values and whether results return encrypted. For MPC, we ask what each participant sees, how many may be corrupted, and whether collusion breaks privacy. Hybrid designs must document the boundary at every transition.