ChainAim Outcome AI
Regulated businesses need intelligence from large volumes of sensitive data, but frontier AI asks them to trust it with their data, their numbers and their budget. Outcome AI replaces that trust with proof. Your files and their rows stay on your PC; only masked text goes to an attested NEAR AI enclave; the agent pays on NEAR before the work, and gets its money back automatically if privacy is not proven. Press Play to hear a real run, step by step (NEAR testnet, 14 Nov 2026): run A in 10 steps (tenant t02, with one counter-offer, job 0ed4fe96…), run B (refund) in 4.
Who sees what
| Party | Sees | Never sees |
|---|---|---|
| Tenant's private runtime (ChainAim data plane) | Your files, the rows, the question and the answer | ChainAim's signing keys (-Planes checks every container) |
| ChainAim's signer (control plane) | Job ids, file sizes, hashes, prices, the mandate, the payment | The question, SQL, rows or answer (the signer refuses unknown fields by name) |
| NEAR AI (TEE) | The masked question and result table, inside an attested enclave | Your files; unmasked personal data |
| NEAR chain (anyone) | Who paid, how much, when; a refund if there was one | Any content: only the payment is on NEAR |
Narration uses your browser’s built-in voices. The smoothest are the “Natural” voices in Microsoft Edge on Windows (Indian English: Neerja, Prabhat; US English: Aria, Guy). Chrome offers Google voices.
Tenant's private runtime: the tenant's machine runs its own AI agent, its files and the ChainAim data plane (console, analyst, gateway with Presidio, ca-attest and the paywall), all in Docker on 127.0.0.1. ChainAim signer and control plane: for the hackathon a separate container (its own Docker project and keys) on the same tenant machine, so the separation is by design, not by hardware. Next: ChainAim hosts the signer and control plane (cloud, then a TEE), and the paywall with its relayer key moves there too.
Demo and proof links
Run it, see it
- Run A, paid outcome after one counter-offer: payment 8E6mhkZP…
- Run B, refund: payment 6tFBUcSS… → refund EyD53Bnv…
- Every payment and refund so far, live: chainaim-payto-seller.testnet
- Code: github.com/chainaimdev/chainaim-private-ai
Rules for judges
- No keys, no account, no install: Node.js 22.22+ and the repository.
- Run
verify-receiptondocs/judge/sample-receipt-2026-11-14T0729Z.json: expect four OK lines (receipt, money trail, attestation, egress) and "counter of". - Run
observe-receipt: the same payment, read from NEAR. - The Docker stack is optional; the video shows it running.
Built to extend
Modular by design: the data plane, ChainAim's control plane and settlement are separate services with signed, allow-listed interfaces. After the hackathon they move into dedicated cloud services, then into TEE workloads (NEAR AI / Shade Agents or Phala Cloud), with full ZK-PRET integration for verifiable trade-finance outcomes.
Check it yourself
node scripts/verify-receipt.ts --file <receipt.json> checks the receipt, the signed quote and limits, the agent's mandate, the money trail, the attestation bundle and the egress line offline, against public keys pinned in the repository. node scripts/observe-receipt.ts --file <receipt.json> reads the same payment from NEAR testnet. Run A's transaction: 8E6mhkZP… on nearblocks.