Loops HouseProvable AI Compliance

NEAR HACKATHON
Links
Team
1 member- VAOwner
Vadim
Overview
Pick a compliance framework, and every AI request your company makes is checked against it automatically: allowed, blocked or escalated to a human, with cryptographic proof that the controls were enforced.
The usual approach, a proxy with a filter and a log, is only a vendor's claim. We replaced the claim with proof, end to end.
Select a framework: EU AI Act Art. 5 and Annex III, GDPR Art. 9, or a company rulebook. Each is a versioned policy pack in a NEAR contract that publishes but never edits. Any other regulation is one more pack.
Programmable controls, inside a TEE: the prompt is encrypted in the browser to an Intel TDX enclave on OutLayer, where every rule is evaluated.
- Allow: a NEAR AI confidential model (Intel TDX + NVIDIA GPUs) answers, and its hardware signs the answer.
- Block: a prohibited practice stops before any model sees it.
- Escalate: a hiring or credit decision becomes a task, opened from the enclave, for the officer the company named. The officer approves with a NEAR wallet signature, and a second attested run answers.
Proof: every run carries an Intel quote over code, input, output and caller, which names the exact binary. Each request ends with a label signed by the company's and the employee's keys, from NEAR MPC confidential key derivation (CKD), bound to this project; the code never sees them. The label lists the framework hashes, the verdict and the approver, with no prompt text. Anyone can verify it from public data.
NEAR end to end:
- frameworks in a publish-only contract;
- people as NEAR accounts;
- company rules encrypted on chain under access conditions (names, DAO roles, account patterns);
- approvals by wallet;
- signing keys from NEAR MPC;
- worker approval checked on NEAR;
- payment keys on the NEAR contract.
Check without keys: cd web && npm ci && npm test, then DEMO_EVIDENCE=$PWD/public/report/evidence npx vite-node ../demo/verify.ts --strict Report with data: https://zavodil.github.io/near-attested-ai/report.html