Make the lender’s evidence packet inspectable.
Small suppliers can have a legitimate purchase order and still lack the cash needed for materials, labour, packaging, testing, and logistics before delivery. The hard part is not producing another loan dashboard. It is establishing which commercial facts can be trusted, keeping the evidence current, and preventing an automated recommendation from exceeding policy.
LoomCredit tests a narrow protocol: a buyer or marketplace posts an order guarantee on a source chain; a worker waits for finality and Attestcoin attestation, obtains a USC proof, and submits it to Creditcoin; a structured agent proposes a quote from typed evidence; and RiskGuard decides whether that quote can reserve accounting-only sandbox liquidity.
The prototype’s contribution is a visible failure boundary. The local adversarial policy lab shows a safe 30% proposal passing and an unsafe 80% proposal being rejected against the deterministic cap. The recorded live bundle also includes a signed model quote and a backwards-compatible QuoteApproved receipt; the deployed bytecode does not expose the newer QuoteDecisionAudited event. The design does not claim to prove physical delivery, legal enforceability, off-network duplicate financing, or repayment ability.
Evidence infrastructure for a regulated lender.
The first customer is not an anonymous borrower looking for a public lending pool. It is a marketplace, export aggregator, or trade platform that already knows the order context and wants a regulated lender to review supplier-finance cases with stronger, lifecycle-aware evidence.
Commits the order context and guarantee.
Verifies evidence and bounds the recommendation.
Owns agreement, disbursement, and collections.
Receives working capital under the lender’s terms.
LoomCredit is infrastructure and workflow software in this model. Calling the current prototype a lender would be inaccurate.
Separate evidence, intelligence, and control.
Buyer-backed commitment on the source chain.
Attestcoin proof and exact source-field validation.
Typed, attributable recommendation from evidence.
Deterministic policy gate before reservation.
The separation is deliberate. Cryptographic verification establishes what happened. The agent proposes what to do. A contract decides what is allowed. No model output can directly withdraw capital or bypass evidence validation.
From source commitment to evidence registration.
3.1 Source commitment
OrderGuaranteeEscrow records a minimal buyer-backed commitment and emits OrderGuaranteed. The event carries the order identifier, buyer, supplier, settlement token, order and guarantee amounts, delivery deadline, terms commitment, identity commitments, and nonce. The same source contract defines cancellation, dispute, and settlement events.
3.2 Attestcoin and USC
The worker discovers the source event, waits for the required finality and attestation state, requests a Merkle and continuity proof, and submits it to Creditcoin. TradeEvidenceUSC calls Creditcoin’s native query-verifier precompile. The precompile establishes inclusion and finalized-chain continuity; the application contract then checks receipt success, the trusted emitter, the exact event signature, every decoded business field, and replay status.
3.3 Evidence registration
The registry stores only normalized facts and commitments needed by policy. A replay key binds source chain, block, transaction, log, and emitter. A commercial fingerprint binds the order and terms so the same obligation cannot create a second active facility inside participating LoomCredit integrations.
This protection is deliberately scoped. It cannot discover a facility originated outside integrations that share the registry.
AI recommends. Policy decides.
The agent receives typed, attributable evidence—not a raw document to interpret as authority—and returns a schema-bound quote containing decision, advance and fee basis points, expiry, risk tier, reason codes, evidence identifiers, policy version, model version, and a signer-bound nonce.
Agent may propose
- Approve, refer, or reject.
- Advance and fee basis points.
- Risk tier and explainable reason codes.
- Model and policy versions.
RiskGuard must enforce
- Signer, evidence, nonce, and expiry.
- Facility state and replay protection.
- Advance, guarantee, tenor, and liquidity limits.
- Buyer concentration and available liquidity.
These are demonstration controls, not a credit model and not a promise that a lender would use them in production. Model unavailability, timeout, or malformed output routes to REFER; a quote that exceeds policy is rejected. The recorded approval is testnet-only and the deployed bytecode still lacks the extended audit event.
Evidence stays attached to the facility lifecycle.
Cancellation, dispute, and settlement evidence can invalidate or close a facility, including direct settlement from an open evidence or quote state and cancellation after a dispute. The USC boundary compares the complete lifecycle payload. Expired quotes and policy rejections remain off-chain outcomes; UI state is never authoritative.
Fail-closed cases
- Untrusted emitter or failed source receipt.
- Proof replay, duplicate fingerprint, or terms substitution.
- Stale, expired, or unauthorized quote.
- Invalid cancellation, settlement, or state transition.
Keep each trust boundary legible.
| Boundary | Establishes | Does not establish |
|---|---|---|
| Attestcoin / USC | Inclusion and chain continuity | Delivery or legal enforceability |
| Worker | Discovery, proof retrieval, retry | Permission to bypass validation |
| Agent | Structured recommendation | Authority over policy or capital |
| RiskGuard | Encoded action limits | Off-chain fraud absent from evidence |
Full purchase orders, legal agreements, tax records, contact data, and KYB material should remain encrypted off-chain. Public commitments must be salted; predictable identifiers should not be hashed in a way that enables a cheap dictionary attack. The worker signing key remains outside the web application.
Wallet connection plus verified sign-in.
The browser exposes a real Connect wallet entry point using an injected EIP-1193 provider. After connection,Sign in with wallet requests a one-time server nonce and asks the wallet to sign a human-readable EIP-191 message. The server verifies the exact message, chain, expiry, nonce replay state, and recovered address before binding it to an account and server-controlled role.
Successful sign-in creates an opaque, expiring HttpOnly session; the raw token is not stored in the database. Sign-in failures, successful sign-ins, and sign-outs are persisted in the auth audit log. This checkout has no capital-moving privileged web action yet; future privileged handlers must require the session and audit the action before executing.
Claims carry a status, not a costume.
LoomCredit keeps local inspection, testnet evidence, production infrastructure, and future work separate. The current browser console uses deterministic fixtures and explicitly does not invent a transaction hash, proof, receipt, or model response.
LOCAL_FIXTUREDeterministic browser data; not a transaction.TESTNETA real request or receipt with explorer evidence.LIVEProduction infrastructure with reviewed controls.PROPOSEDA design or pilot requirement not implemented here.Earn the next claim with the next piece of evidence.
OCR, multi-model ensembles, public lending pools, tokenomics, arbitrary chain support, and complex dispute arbitration are intentionally below the critical path.
Verify the evidence. Bound the action.
LoomCredit does not claim that a blockchain removes trade risk or that an AI model should control lending. Its narrower thesis is practical: verify the commercial evidence first, bind it to lifecycle state, let software propose an auditable action, and keep the final action inside deterministic policy.
Attestcoin proves the evidence. The agent proposes the action. RiskGuard controls the action. A regulated partner owns the loan.