RESEARCHED PROVIDER · PRIMARY SOURCES

Cocoon

A TON-based compute network where GPU owners execute model requests inside confidential workers.

INPUT / OUTPUT PRIVACY

Stage 1

Partial proof

ASSESSED DEPLOYMENT

Backend client to attested network

RESEARCH REVIEW

2026-10-09 · primary-source review

Provider website

Verification labels describe capabilities established by the cited sources. This review did not perform live cryptographic attestation, reproduce production builds or complete an independent security audit.

Who actually runs the models?

GPU owners run model workers using Cocoon software. Proxies select those workers and settle payments on TON; they route within a model-execution network rather than reselling unrelated inference APIs.

Deployment and TEE coverage

The documented client library runs on the application backend. The guarantee begins there; a Telegram or other application's front-end/backend path and its own data retention are outside this record.

Production proxies and workers use Intel TDX. Workers verify GPU confidential-compute state at boot. Remote parties verify the measured VM transitively; they do not directly verify a separate GPU report.

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.

STEPS FOR STAGE 2

2 of 3 requirements for Stage 2 met. Missing proof: Logs & storage.

  1. Request pathMet

    Private inputs and outputs across the request path, with E2EE or an equivalent attested channel that protects against intermediaries and binds keys to the accepted workload.

    Technically verifiable. Backend → proxy → worker

  2. Inference executionMet

    Protected inference with a verifiable workload identity and protection covering the CPU, GPU and every place content is processed.

    Technically verifiable. Registry + boot-time GPU check

  3. Logs & storagePartial

    No content logging, persistent storage, human review or training. Any temporary content-derived cache must have verifiable isolation and enforced removal.

    Partial technical proof. Engine lifecycle unresolved

Identity and operational metadata are assessed separately below. They do not set the content stage; prompt or response content in telemetry remains part of the content-retention assessment.

Request path

Technically verifiable

RA-TLS chain. Public RA-TLS code binds the certificate public key to the TDX quote; clients check certificate evidence and accepted image measurements before using the protected connection. The application backend verifies the proxy, and the proxy verifies the worker. Production policy must constrain registry/image identities and reject fake-TEE or permissive testing policies. The application's device-to-backend path remains outside this record.

Inference execution

Technically verifiable

Attested worker. Worker-image and model hashes are checked against the TON registry. Public boot code gates GPU use on local verification, and the current worker startup opens the model disk using a verity commitment and launches a digest-pinned SGLang engine. The CPU image therefore supplies transitive GPU evidence. Registry updates and current proxies remain team-controlled; clients must review the accepted image/model policy rather than trust a network label alone.

Logs & storage

Partial technical proof

Transport logs inspectable. The public worker transport logs request IDs, methods, URLs and completion statistics rather than dumping prompt/response payloads in its normal request path. This provides partial inspectable logging evidence. Its current SGLang launch enables cache reporting and reuse, without establishing tenant cache isolation, bounded KV removal or a complete diagnostic/trace policy. Integrating-application storage is excluded, not evidence against this backend scope. No complete in-network content-removal lifecycle was established.

Identity & metadata

Documented only

TON payment records. The network uses wallet-linked payment contracts and records payment state. Encryption does not hide traffic or settlement metadata. The reviewed sources do not establish a complete linkage and deletion policy for off-chain operational data.

Remaining trust assumptions

  • Intel TDX and NVIDIA confidential-computing isolation and verification.
  • Measured proxy/worker code, strict RA-TLS policies and model hashes in the registry.
  • Cocoon's current registry governance and the integrating application's backend.

Limits of this assessment

  • Distributed GPU workers coexist with team-operated proxies and centralized registry updates.
  • The application backend is a trusted client; this is not automatically device-to-model privacy.
  • TON settlement and network metadata are outside content encryption.

Primary-source record

9 sources · reviewed 2026-10-09

Back to providers

Sources reviewed 2026-10-09