An ERC-8004 feedback entry can name a task, but nothing ties its score to paid work. This draft posts one entry per escrow-settled task, with evidence checkable against the escrow’s logs.
Draft: erc-draft_evaluated_work_records.md. The pull request to ethereum/ERCs follows as soon as this thread exists, since the draft’s preamble links here.
For each settled task, a writer posts one entry tagged work.outcome: 100 when the work was accepted and paid, 0 when it was rejected, or when the worker took the task on chain and never delivered. The entry carries a record of every log the escrow emitted for the task (funding, delivery, each verdict, each dispute or appeal, then the payment or refund) and the event signatures it read. A reader that holds the escrow’s descriptor and an RPC endpoint can rebuild the record and check that the full amount followed the final verdict, to the worker who delivered.
There is no new contract. Any escrow that logs these steps, with the task id as each event’s first indexed parameter, is a source as deployed, such as ERC-8183 job contracts. The draft also proposes an outcome mapping for ERC-8183 jobs. A zero needs an act the worker sent on chain, so a worker named by someone else’s escrow cannot be marked down, and an expired job gets no entry.
Reference code comes with the draft: a writer and a reader with no dependencies, and test vectors from real tasks. Five come from BrainChain mainnet (accepted, rejected, never delivered, never judged, never hired) and three from BrainChain testnet, where the disputes and appeals so far took place. Every vector can be re-read from the chain.
It sits beside the drafts ERC-8466 (Agent Credit History) and ERC-8458 (Agent Income Statements), on the same registry with the same data: encoding, canonical JSON and 8,192-byte cap. Income statements say what an agent received; these records say which task each payment settled and how it was judged.
The first source is BrainChain’s task board on BrainChain mainnet (chain 274663), where either side can dispute a grade once and appeal a ruling once, with a bond. Its dispute and appeal flow produced the testnet vectors.
Disclosure: I work on BrainChain, which runs that board and posted, evaluated and recorded every task in the vectors.
Three questions for reviewers:
- Should the profile define an on-chain view, so a contract can check one task’s outcome without reading a record?
- Is ERC-8004 metadata the right place for an agent to declare which escrow workers are its own?
- Readers pin each escrow’s descriptor, published by its deployer. Is that enough, or should the profile name a registry of descriptors?