Loops HouseBlindspot

NEAR HACKATHON
Links
Team
1 member- SIOwner
Singupalli Kartik
Overview
Private AI on NEAR has a blind spot. Intel's attestation of a NEAR AI Cloud enclave covers one compose file with four images; the model engine and about fifty other services start later and exist only in a change log that NEAR AI's own verifier hashes but never reads. After the November incident, NEAR's founder wrote that users should verify exactly what happens with their data and prompts. Blindspot is that verification as a product, in five layers. Gate: 20 checks before every request (three Intel TDX quotes, a fresh nonce, the boot log, the change log, NEAR AI's public repo, Sigstore build records, your policy); if one fails, the prompt is not sent. Witness: outside accounts chain every change-log entry in a NEAR contract, one of them signing only with a post-quantum ML-DSA-65 key, so history cannot be rewritten quietly and reviewers can revoke a release for every client at once. Canary: bait NEAR keys that exist only inside prompts sent to one enclave; if one is ever used, its access-key nonce on NEAR moves and the gate shuts. Veil: keys, emails, accounts and names become placeholders before sending and are restored in the answer. Pass: pay with a Groth16 proof verified inside a NEAR contract, so the relay knows someone paid but not who. Measured on NEAR AI's three live enclaves: under NEAR's documented 48 hour rule, 27 of 27 running deployments fail; most went live within minutes of their public commit. Ships as a web app that shows every network call it makes (0 to Blindspot), an OpenAI-compatible proxy, an SDK and MCP tool for agents published on npm as useblindspot, a hosted API with a status badge, a monitor and a GitHub Action. NEAR is the memory and the alarm: witness chains, the release registry, canary commitments and nonces, and pass verification all live on NEAR testnet; the inference target is NEAR AI Cloud.