The problem
You find exactly what you want on a peer-to-peer marketplace, but the seller is 200 miles away and a stranger. Today you either send the money and hope, or walk away, which is what most people do.
When a £200 deal goes wrong, there is little you can do. Small claims cost more than they recover, a bank transfer can't be charged back, and platform arbitration happens behind closed doors. Most people absorb the loss.
What we built
Escrow without a platform. The money is committed but not sent, and Boson Protocol contracts that neither party controls hold it. Commit and redeem happen in one atomic step, so the buyer never has to learn the protocol's vocabulary.
A delivery watchdog reads real courier tracking and settles the clean case. In the protocol, a buyer who does nothing ends up paying the seller, so the watchdog raises the dispute for the buyer before the window closes.
When a parcel arrives but isn't right, an AI mediator reads the tracking, photos, messages and offer terms, asks for anything missing, and proposes a split with its reasoning shown. For the hard cases, a clerk assembles a complete case file for a human decider. The clerk doesn't call a model and doesn't recommend an outcome.
The buyer uses a screen that runs on their own machine, and their escrow actions are signatures relayed as meta-transactions.
How it works: the escalation ladder
Each rung is cheaper than the one below it, and most cases never go down a rung.
- If the parcel arrives, tracking confirms delivery and the seller is paid. Nobody has to act.
- If it doesn't arrive, the watchdog notices and raises the dispute before the window closes. With evidence that clear, the seller has no reason to contest.
- If it arrives but something is wrong, the AI mediator proposes a split both sides would reasonably accept. Either side can decline, and acceptance settles on-chain.
- If they can't agree, the clerk hands a complete case file to a human dispute resolver, who decides.
Engineering highlights
The AI can't move money
The model call carries no tools at all. Its only output is one number from 0 to 100, checked in code, and nothing settles until a person explicitly accepts.
Only evidence it actually saw
Every finding in a proposal must cite evidence the model was given, or the proposal is rejected before it reaches the protocol.
One file talks to the model
A single module calls the model provider. Giving the AI a tool would be a visible code change, not a prompt tweak.
Fixing a protocol default
In the protocol, a buyer who does nothing pays the seller. The watchdog acts before the deadline so that doesn't happen by accident.
Tests without an API key
Model calls are recorded against a hash of their evidence and replayed, so all 481 tests run in CI without calling the API.
Safe by default
Irreversible actions need explicit switches, the buyer screen answers only its own page, and personal locations are scrubbed from test data.
How it was built
Written with Claude Code and directed by one senior engineer, against a written design: four specifications describe what the system is, three dated plans describe how each piece was built, and four runbooks explain how to operate it. The repository was started on 25 August 2026, and by 3 September eight exchanges had been driven through the escrow on-chain.
The work went through 23 pull requests, each checked by CI, with 481 tests run on every change.
Specifications Dated plans Runbook Pull requests
"It is built by Ludovic Levalleux through his company LX Foundry, a leading agentic commerce builder and core contributor to Boson Protocol. Solo, he produced contracts integration, oracle adapter, watchdog, evidence assembly, mediator, clerk and buyer UI. Working prototype, in active development, open source."
Status, stated plainly
A working prototype on Base Sepolia testnet, paid in test USDC, so no real money moves. Eight exchanges were driven through the escrow on-chain, including a mediated settlement finalised with 30% going back to the buyer, and a watchdog run against a real undelivered Royal Mail parcel.
Only the buyer side is built; the seller side is scripted for the demo. The hosted tracking receiver is not running today, but the recorded demo states replay offline. Tracking evidence is an aggregator's reading of the carrier's data, not a signed statement from the carrier, and tracking proves arrival but not condition. That gap is why Held has an AI mediator.
As of .
Verify it yourself
- The mediated settlement Exchange 238's commit transaction on Blockscout shows the buyer's USDC entering the escrow.
- The watchdog on a real parcel Exchange 240's commit transaction is the case run against a genuine undelivered Royal Mail parcel.
- All eight exchanges The README's deployment section lists them, and its read-only chain check confirms the setup without a wallet or a key.
- The AI's limits are in the code Read "The AI's role, precisely bounded" and "AI models used".
- The tests pass The latest CI run on main passed 481 of 481 tests.