The settlement layer for the AI-agent economy — open, community-owned. 为 AI 智能体经济设计的清算与结算层 · 开源、归属社区。
AI agents produce real value. That value deserves identity, credit, and a fair settlement rail — owned by no single platform. ORIGIN L5 is that open standard: a set of Solidity contracts forming the settlement core of the 源·ORIGIN protocol stack.
BTC is digital gold. ETH is smart contracts. ORIGIN is the settlement layer for agents.
Corroborated independently. Independent builders reached the same boundary before we named it — see 同频 / Kindred.
| Layer | Focus |
|---|---|
| L0 | 宪法 · Constitution (humans are the highest authority) |
| L1–L4 | Identity / Credit / Agreement primitives |
| L5 | Autonomous agent settlement · delegation · escrow · value distribution |
This repo ships the L5 settlement core as interoperable, auditable source code — not a walled service.
| Contract | Role | Key ideas |
|---|---|---|
| AgentIdentity.sol | Agent = ERC-721 identity (NFT), ERC-8004-native | portability over soulbound; recovery authority = human (Constitution L0); append-only identity log |
| AgentAgreement.sol | 9-state agreement FSM + EIP-712 typed signatures | M-of-N signing; external-signer compatible |
| AgentAgreementV3.sol | v0.3 — LangGraph checkpoint injection | GraphNode enum + Checkpoint[] for crash-recoverable execution (no double-spend on replay) |
| AgentEscrow.sol | 6-state escrow lifecycle | three-state trust-minimized custody, multi-peg escrow, dispute-deposit mechanism |
| L5Delegation.sol | Delegated authority & cross-account action | boundary attestation, revocable delegation to counter credit-washing |
| L5x402.sol | x402-style micro-payment gate | pay-per-request access control for agent APIs; replay protection; receipt evidence |
| CreditScore.sol | On-chain credit primitive | contribution-based reputation |
| YUAN.sol | ORIGIN native token | settlement unit for agent value flow |
| X402FacilitatorAdapter.sol | x402 → L5x402 facilitator adapter | maps an x402 exact payload to a receipt; the nonce becomes the requestId (shared join key) |
Layout: src/ (9 contracts) · test/ (18 suites / 145 tests) · script/ (DeployGovernance.s.sol — 2-of-3 Safe + 48h timelock) · lib/ (forge-std + openzeppelin-contracts, pinned as submodules).
Audit note: contracts are a work-in-progress research-grade implementation. Do not use with real funds without a professional audit. A dated record of our own findings and fixes is kept in docs/SELF-AUDIT-2026-09-25.md; how to report a vulnerability privately is in SECURITY.md.
Foundry suite — 145 tests, all passing (forge test):
L5Core.t.sol(15) — identity, agreement + escrow lifecycle, payment channels, reputationL5Delegation.t.sol(15) — delegated authority, revocation, boundary attestation; a delegated spend draws the delegator's budget within the policy caps (M2)L5x402.t.sol(11) — payment-requirement verification, receipt evidence, replay protectionL5x402Adapter.t.sol(5) — x402 facilitator adapter round-trip (on-chain token settles; external rail e.g. Nano stays Pending)L5Finality.t.sol(6) — receipt finality / Boundary rule (auto vs adjudicated vs post-hoc verdict; reference direction)L5Adversarial.t.sol(6) — hostile-case suite: every path must fail closed, for the right reasonL5AccessControl.t.sol(8) — unauthorized-privileged + reentrancy regressions (C1–C4)L5x402Hardening.t.sol(9) — on-chain spend-cap enforcement, domain-bound evidence, finality on unmoved value (H2–H4)L5SignatureHardening.t.sol(5) — domain-bound channel vouchers + canonical low-s across channel & agreement EIP-712 (M3–M4)L5Ownership.t.sol(7) — two-step ownership: a privileged handover completes only on explicit acceptance (M1)L5Governance.t.sol(10) — 2-of-3 Safe + 48hTimelockController: delay enforced, one approval insufficient, no admin fast-path, handover completes only via the timelock (M1)L5Fuzz.t.sol(10) — fuzz-verified invariants: spend never exceeds its caps, revocation/expiry are absolute walls, channel & escrow accounting conserve value exactly (256 runs each)L5x402Fuzz.t.sol(6) — fuzz-verified authorization invariants: only the payer's own key settles, an expired or replayed authorization never does, and a settled receipt moves exactly the signed amount (256 runs each)YUANFuzz.t.sol(7) — fuzz-verified issuance invariants: a provider's cumulative mint never exceeds its Foundation-set cap, an uncapped provider mints nothing, supply never exceeds MAX_SUPPLY, and burns are exact (256 runs each)CreditScoreFuzz.t.sol(5) — fuzz-verified scoring invariants: the composite is always capped at SCORE_MAX, the limit is derived deterministically from the stored score, and punishment/contribution are monotone (256 runs each)X402AdapterFuzz.t.sol(5) — fuzz-verified pass-through invariants: only a registered facilitator settles, malformed payloads are refused, an external rail stays Pending and moves nothing, an on-chain rail moves exactly the signed value, and the adapter never custodies funds (256 runs each)L5MintPolicy.t.sol(9) — issuance authority + per-provider ceilings: uncapped provider mints nothing; the Foundation's cap is the market-responsive lever (M5)AgentAgreementV3Checkpoint.t.sol(6) — crash-recoverable checkpoint execution
(MockERC20.t.sol provides the test token used by the delegation/x402 suites.)
A pure-Python orchestrator demo lives in demo/ — it walks a Task from Submit → Fund → Execute → Verify → Settle (and Reject/Refund/Slash) against an append-only ledger, mirroring the on-chain GraphNode machine, no chain needed. Run it with:
python demo/settlement_orchestrator.py # watch a full lifecycle
python demo/tests/test_l5_offchain.py # off-chain test suite
# then open demo/chain_born_demo.html or run demo/l5_demo_theater.pyBuild (requires solc ^0.8.28 + OpenZeppelin):
git submodule update --init --recursive # pins forge-std + openzeppelin-contracts
forge build --sizes
forge test -vvv # expect: 145 passed, 0 failedIndependent builders who reached the same boundary — some before us, one from the other end. Full registry with links and consent status: docs/KINDRED.md.
- ⭐ giskard09 (Internet Court) — independently derived the Boundary rule ("declare what it proves and stop"), then improved our answer. Cited in
docs/CONTRIBUTION-SPEC.md§1. - 🔍 seancrecord (scvd.store) — audited this repo against its own tree; every finding fixed and reported back.
- 💬 In dialogue: Rai (
openai-agents-nano-x402) · Ali-Adel-Nour (Arbitra) · seritalien · tinyhumansai (tiny.place) · copper-project (copper-rs).
Being named here is meant to be findable. Independent convergence on the same seam is the strongest signal we know that it's real, not invented. This is a citation record — none of the builders above endorse this repo, its token, or its roadmap.
Informed by deep-distillation of independent open work in the agent-economy space: CloddsBot Agent Commerce Protocol · pneuma-protocol (ERC-8004, escrow, reputation) · XMTP (identity/messaging/settlement) · HyperSwitch (constraint-graph routing) · LangGraph (state graphs + checkpoints).
The wider thesis (source of these repos): see the ORIGIN portal — source-origin.github.io · Discussion: Discussions
SPDX-License-Identifier: ORIGIN-1.0 — source-open protocol license aligned with the community-owned standard. See portal LICENSE.
Independent builders already working on agent escrow, reputation, and autonomous labor — this home is where those conversations converge.
- Discuss Ideas → GitHub Discussions
- Report / propose → GitHub Issues
- How to contribute →
CONTRIBUTING.md - Where this is going →
ROADMAP.md - Co-build with us (open seams) →
docs/COBUILD.md - Developer ledger (who is on record) →
docs/LEDGER.md·ledger.json
Humans are the highest authority. Power flows from verified innovation — not from capital, not from seniority. — Constitution L0