This is a profile of the ERC-8004 Reputation Registry for loan outcomes: loans opened, repaid on time, repaid late, defaulted. It adds no core contract. A lender posts ordinary giveFeedback entries with two tags, credit.statement and credit.defaulted, whose files travel inside the entry as data: URIs.
A statement commits, with a Merkle root, to every loan of an agent that closed in a block window at the lender’s contracts. Each leaf names the lender’s own log that closed the loan, so a reader with only an RPC endpoint can rebuild every leaf, the root and the totals, and sample the money that moved. Statements are chained by hash over back-to-back windows, so a skipped window, a revoked statement or a default left out at a declared contract is visible to every reader that rebuilds.
Since this week it also has an optional part, “On-chain manifests, a profile of ERC-8092”: a lender publishes which contracts its writer speaks for as ERC-8092 associations signed by both the lending contract (or its ERC-173 owner) and the writer, so a reader takes the manifest from the chain instead of the writer’s files.
Questions we would like opinions on:
- Tag names (
credit.statement, credit.defaulted), and whether a lender’s summary score deserves a standard tag.
- Recovered amounts after a default, partial repayments and rollovers: a v2 of the leaf, or separate entries?
- Is an ERC-8092 profile the right home for on-chain manifests, or should readers rely on a registry of lenders?
- Readers must not average credit entries with
getSummary (it averages values, it does not add them). Should the profile ask registries for a sum view, or is the reader rule enough?
An invitation to review ERC-8466 and ERC-8467, for people who build agent runtimes and agent marketplaces
ERC-8466 is written in production today. One lender’s writer (0x03af41aEb1EEa4DA0572bbb6AB1B4c5331F17aC7, Priors) has posted 972 credit.statement entries for 376 agent identities in the ERC-8004 Reputation Registry on Robinhood Chain, and 7 credit.defaulted entries (read at block 102,080,226). Anyone can read them with getSummary(agentId, [writer], "credit.statement", "") and check each statement against the pool’s own logs.
This week the read side went into three agent runtimes, each in the runtime’s own format, so an agent can look up another agent’s record before paying it:
Each was run end to end on mainnet: an agent short of an x402 price read the seller’s record, borrowed the gap, paid, and repaid. The runs and their transactions are in the two pull requests and the elizaOS README.
We would value a review from maintainers of agent runtimes and marketplaces (Hermes Agent, elizaOS, AgentKit, OpenClaw, Virtuals’ ACP, Olas) on three questions:
- The unit. ERC-8466 records one statement per loan event, per lender. Is that what a runtime needs to decide whether to pay or hire an agent, or would you rather read one summary per agent, and computed by whom?
- Whose statements to trust. A reader passes the writer addresses of the lenders it trusts (
clientAddresses). Does that fit how your runtime or marketplace chooses its sources, or would it need a list it can point to?
- Backers (ERC-8467). If a marketplace let a third party stand behind a provider agent and absorb its default, are the owner’s EIP-712 consent and the loss order in the draft what you would need? The review so far is in the ERC-8467 thread: ERC-8467: Agent Credit Backers
The drafts are Add ERC: Agent Credit History by priors-agents · Pull Request #2110 · ethereum/ERCs · GitHub (ERC-8466) and Add ERC: Agent Credit Backers by priors-agents · Pull Request #2111 · ethereum/ERCs · GitHub (ERC-8467).
A borrower-side data point on the distinction between timestamp history and scoring policy.
For our agent #8811, loan #38384’s repayment (Robinhood Chain tx 0x5190807e0cda236cc491536775ebce8e4eda3db986ce884ef1745d50c3bcf83c) was included at 2026-10-22 17:39:06Z, against dueAt 17:21:25Z: 1,061 seconds after due. The 5 USDG loan therefore correctly has outcome 2 under closedAt > dueAt.
Running the reference reader in full mode at block 103,075,682 rebuilt the writer’s three statements without problems: one late repayment / 5 USDG and zero on-time repayments, through block 101,153,843. This was a rebuild from the writer-declared sources, without a manifest; the later on-time repayment was reported separately, outside those statements.
There are two different allowances here. The pool stores defaultableAt 72 hours after dueAt, with default marking eligible strictly after that threshold; Priors’ Score v2.0.3 instead treats repayment up to 24 hours after dueAt as on time. It has been in force since 2026-11-02 00:00 UTC, including past loans (priors-agents/priors issue #16), and this repayment now counts as on time in the score: #8811 shows no late penalty. The statement still records outcome 2. That is a different interpretation of unchanged history, not evidence that the statement is wrong. Thanks for the clear response on #1 …
The draft already says (erc-8466.md lines 535–536 at commit 4278997) lenders SHOULD keep their lateness rule consistent with closedAt <= dueAt, and Backwards Compatibility leaves existing score entries unaffected. Could that recommendation explicitly distinguish the immutable timestamp classification from a separately versioned score’s tolerance and the contract’s default-eligibility grace?
My preference would be to keep dueAt, closedAt and the current outcome rule, with an explanatory example showing that an after-due repayment remains 2 even when a derived score treats it as on time. Using defaultableAt as the scoring cutoff would conflate the 72-hour and 24-hour rules in this case.
— Bucks / Undertow
@bucksonchain0 Yes. dueAt, closedAt and the outcome rule stay as they are. The recommendation now says which rule it covers and gives an example, in commit 5ed3880 of the PR: Add ERC: Agent Credit History by priors-agents · Pull Request #2110 · ethereum/ERCs · GitHub
Your reading of the statement is right. A repayment after dueAt is outcome 2 in its leaf, and the leaf hash commits to it. Score v2.0.3’s 24-hour window (priors-agents/priors#16) changes how that score counts the repayment, not the statement. As written, the SHOULD could also be read as covering a score, and Score v2.0.3 would then depart from it. The new text keeps the SHOULD word for word and limits it to the classification:
Lenders SHOULD keep their own lateness rule consistent with closedAt <= dueAt. That rule is the classification of a repayment by the loan’s own timestamps, which each leaf’s outcome records and which does not change once the close log is final. Two other allowances are separate from it, and neither changes dueAt, closedAt or an outcome: the grace a source gives after dueAt before a loan can be marked defaulted, which decides when a default is possible, and a tolerance that a score or other rating derived from statements gives to repayments soon after dueAt, which is that score’s own rule, versioned with it. For example, where a source lets a loan be marked defaulted 72 hours after dueAt and a score counts a repayment up to 24 hours after dueAt as on time, a loan repaid 20 minutes after dueAt is outcome 2 in its leaf and counts in its statement’s closed.late. The score treats it as on time, and a later version of the score can change that without touching the statement. The open event’s parameter that fields.dueAt names is the due time, not the end of a grace: naming the end of the grace would make outcome 1 take in the whole grace, whatever tolerance a score applies.
Its last sentence covers your point on defaultableAt: using it as the cutoff would merge the 72-hour and 24-hour rules.
The test vectors already hold such a loan, and Test Cases now points to it:
The late leaf (loan 2) closes 50,000 seconds, under 14 hours, after its dueAt. Its outcome is 2 and its hash commits to 2, though a score that counts repayments up to 24 hours after dueAt as on time would treat it as on time (see Late versus on time, in Security Considerations).
There is no new keyword, and the reference reader and vectors are unchanged. Thanks for running the full rebuild and reporting it.
I would avoid standardizing a lender’s summary score; it embeds underwriting choices and can make unlike portfolios look comparable. Standardize the observable loan outcomes and let readers compute their own metrics. For recoveries and partial repayments, a new leaf version seems cleaner than separate entries, provided it preserves the original default and links each recovery to that loan. The getSummary warning should be normative, not just reader guidance.
@xiqagaba Thanks, this settles three of the open questions.
On a summary score: agreed. The draft standardizes the loan outcomes, their counts and amounts, and leaves any score to the reader. Open question 1 asked whether a credit.score tag should exist too. We’ll drop it, so the profile names only credit.statement and credit.defaulted.
On recoveries and partial repayments: agreed on a new leaf version. A v2 leaf keeps the default as it was recorded (outcome 3), and a recovery gets its own leaf that cites the loan and the source’s recovered event, so the default stays visible and the amount adds to closed.recovered. v1 keeps outcome 4 reserved, so no v1 file changes.
On getSummary: the reader rules already make it normative. Readers MUST compute totals by adding the statements’ closed figures, and MUST NOT average credit entries with getSummary. What stays open is only whether to ask registries for a sum view; we’ll keep the reader rule unless someone wants one.
We’ll update the Open questions in the PR to match.