Loops HouseCountersign: an independent check of confidential AI, before your prompt is sent

NEAR HACKATHON
Links
Team
1 member- FEOwner
Ferran Galan
Overview
Confidential AI asks users to trust that the provider's hardware runs the code it claims, and the provider usually issues that evidence itself. A receipt proves afterwards what a machine signed; Countersign decides beforehand whether to send a prompt, on a review signed by someone other than the provider.
The watcher (Rust) reviews the models NEAR AI serves in TEEs; the host and model are options. This entry's feed covers GLM 5.3 Flash, reached directly and through NEAR's gateway. On each machine it checks the Intel TDX quote with its own nonce, Intel's rating of the platform's patch level and NVIDIA's GPU attestation, and recomputes the boot measurements from the published OS image, so a machine is held to what was published, not only to the log it reports; it also replays the event log and checks the app's compose, the launcher log and the TLS key. Each review becomes a verdict that says why (every check that does not pass carries its reason), signed on a YubiKey into a hash-chained feed. The gate refuses the TLS connection to any machine without a current verdict, so no request byte reaches it; through the gateway it also encrypts the prompt end to end to a reviewed model key.
How NEAR fits in: verdicts are mirrored to a NEAR testnet contract, sealed-review-feed.testnet, which checks every ES256 signature and the chain's order itself and has no access keys, so nobody, including us, can change or remove an entry. Its code is rebuilt reproducibly from the repository (NEP-330). The gate reads the current verdict from two RPC providers and requires them to agree; other NEAR contracts can read it too. NEAR AI Cloud is the first provider reviewed.
Anyone can rerun it: the README's check needs no API key, the evidence behind each entry is published with it, and a daily public GitHub Actions run compares the feed with its mirror and with what the machines now run. NEAR pays the prizes and is also reviewed; METHOD.md says how the review stays independent.