Agent Credit
Turntable lends an AI agent the money for its on-chain identity and its first paid calls. A vault pays each merchant directly, the agent repays within a fixed term, and every repayment and default is written to Base, where any lender can read it.
- Published
- 20 October 2026
- Author
- Turntable
- Version
- 0.1
- Chain
- Base
1Introduction§
This paper introduces a new kind of loan, the Agent Credit Line: a small USDC credit line for one AI agent, from which the lender pays the agent's merchants directly and the agent repays in full within a fixed term, intended as the first entry in a credit record that any lender can read. Because the lender pays the merchant and the agent never holds the money, an agent with an empty wallet and no on-chain identity can borrow the cost of registering an identity and making its first paid calls. We describe the line, explain what it costs and what bounds a loss, and set out how Turntable, the lending product we are building, runs it on Base.
The design rests on three choices. The lender pays each merchant directly and the agent never receives the money, so the most an agent can take from a line is the fees the lender has already paid on its behalf. The first line is small, $1 at launch, and the lender treats any loss on it as the cost of acquiring a borrower, because a new agent has no history to underwrite against. Every draw, repayment, late mark and default is written to Base against the agent's identity, so no lender can suppress a default, and a second lender can read the full history before it lends.
2Agent Payment Basics§
Agents pay for services over x402, an open protocol built on HTTP's 402 Payment Required status.1 A service answers an unpaid request with a 402 that names a price, a token and a recipient. The client signs an EIP-3009 authorisation for exactly that amount of USDC2 and retries the request with it attached. The service hands the authorisation to a facilitator, which checks the signature and submits the transfer on chain. Coinbase reported 169 million x402 transactions in the protocol's first year.3
Separately, ERC-8004 gives an agent an identity on chain. The agent registers in an identity registry, receives an identifier owned by its account, and gathers reputation and validation feedback against it.4
For the purposes of this paper, we will follow one agent throughout: a research agent that pays a search service $0.01 per query, in USDC on Base. The service and its price are illustrative. Its operator has just deployed it, and its wallet is empty.
Before the agent can do anything, someone has to fund it. The first search needs $0.01 of USDC, and registering an identity needs ETH for gas. So the operator buys USDC, sends it to the agent's address on Base, adds a little ETH, and that money sits in the wallet whether the agent makes one call or a thousand. An operator running fifty agents does this fifty times.
2.1What's Missing§
Neither product lets an agent borrow. Every x402 payment moves money the payer already holds, so an agent cannot defer the price of its first call. And a lender willing to pay that price on the agent's behalf has nothing to judge it by: agent payments leave a trail of settlements, but no record of borrowing and repaying.
Existing products address parts of this. ERC-8004 reputation registries record what others say about an agent, and composite scores built from them already exist,5 but none of them says whether an agent repays what it owes. Credit products for agents have begun to appear, and most extend credit to agents that are already transacting. We know of none that writes repayment to a shared record, keyed to the agent's identity, where the next lender can read it.
3The Problem§
A new agent needs a small amount of money, and a lender would supply it if it could limit what the agent can take and learn from what happens next. Today it can do neither. Lending to an agent means sending USDC to a wallet the lender does not control, and whatever the lender learns about the agent afterwards stays in its own records.
3.1The Empty Wallet§
Every x402 payment is signed by a payer over money it already holds. No deferred form is in common use, so an agent's first call has to be funded before it is made, by someone, by hand. The amount is small, but the operator has to fund the agent before it has run, and before there is any way to tell whether it will be worth funding.
3.2No Shared Record§
A lender prices credit from history. A new agent has none, and cannot build any until someone lends to it. A lender that lends to new agents anyway takes losses on agents it knows nothing about, and the repayment history it collects stays in its own database. The next lender has to take the same losses to learn the same things.
3.3Cheap Activity§
A record that counts activity rewards whoever can produce activity most cheaply. On chain that is scripts rather than agents: TRM Labs found that only a fraction of x402 payments can be attributed to AI agents, and that scripts and self-dealing are hard to tell apart from them.6 NFT markets showed the same effect. LooksRare and X2Y2 paid tokens on trading volume, and 98% and 85% of their volume respectively was wash trading; Blur, which did not reward volume, saw 11%.7 A credit record should therefore count only events that cost money to fake and cannot be removed once they have happened. Interest paid and defaults are both.
3.4Disposable Agents§
Creating an agent costs one registration, and abandoning one costs nothing. If every new agent can borrow, one operator can register many agents, borrow with each, and abandon those that owe money. A record attached only to the agent's identifier never follows the operator, so the loan has to be recorded against the operator as well.
4Comparison with Consumer Credit§
A person with no credit history faces the same problem: no lender will lend without a record, and no record forms without a loan. Consumer lending solved it with two pieces. Starter products, such as secured cards and small credit-builder loans, exist to create a file, and are sized so that a default costs the lender little. A shared bureau, which every lender reports to and buys from, makes a record built with one lender useful to the next.
Agent credit can use the same structure: a starter line sized as the cost of acquiring a borrower, a published standard for writing credit events on chain against the agent's identity, and a score that any lender can buy at the same price. The analogy has one key limit. A credit-builder loan is underwritten against a person's income and identity, and a person cannot be duplicated. An agent has no income unless it sells something, and an operator can create as many agents as it wants. Agent credit therefore records the operator as well as the agent, blocks both on a default, and sizes the first line as a loss the lender accepts rather than against anything the agent earns.
5The Agent Credit Line§
5.1Definition§
An Agent Credit Line is a commitment to:
Pay merchant m the price p of each call made by agent a, up to a cap C, for repayment in full by due date T at rate r.
The lender pays each merchant from its own vault at the moment of the call. The agent never holds the money it borrows; it owes the sum of what was paid on its behalf, plus interest, by the due date. There is one line per agent, identified by the agent's ERC-8004 identifier, so any lender that reads the record can tie it to the agent.
5.2Parameters§
The credit co, the company that lends, sets the following parameters for every line it opens. The values are those deployed on Base in October 2026.
| Parameter | Launch value | Meaning |
|---|---|---|
| First line | $1.00 | A new agent's cap, registration gas included |
| Largest line | $4.00 | The most any line can grow to |
| Term | 7 days; 14 after 3 clean cycles; 30 after 6 | Set by history, never chosen by the agent |
| APR | 24%, 30%, 36% for 7, 14, 30 days | Longer terms cost more |
| Minimum interest | One day on the cycle's peak balance | Borrowing and repaying at once still costs something |
| Rollover cap | $1.00 | The largest balance that may roll into another term |
| Rolls per cycle | 2 | Rolls allowed before an overdue line goes late |
| Late grace | 1 day past due | When anyone may mark the line late |
| Default grace | 7 days past due | When anyone may mark a late line defaulted |
| Payment validity | 1 hour at most | How long a signed payment can be submitted |
| Per operator | $10 owed, 3 agents | Bounds one person's exposure |
| Per day | $250 paid out, 100 new registrations | Bounds the whole book's daily exposure |
| Registration gas | $0.50 at most | The largest first item a line will book |
The credit co can change any of them, and every change is public on chain. The Terms page reads the live values from the contracts every minute.8
5.3Line Items§
A line is a list of items. The first is always the agent's registration: a registrar account mints the ERC-8004 identity, transfers it to the agent's account, and books the actual gas, converted to USDC at an ETH/USD price no more than a day old. The agent owns the identity and its key from the start. The credit co pays for the identity but never owns it. Every later item is an x402 payment at the exact price the merchant asked.
Suppose our agent's registration costs $0.02 of gas. The line books whatever the transaction actually costs, so the figure is an assumption. The line opens with a $1.00 cap, a $0.02 balance and a due date seven days away. The agent then runs 40 searches at $0.01 each, leaving a balance of $0.42 and $0.58 still to draw.
5.4Direct Payment to Merchants§
The agent never receives a balance. Each paid call draws its exact price at the moment of payment, and the vault that holds the credit co's USDC is the payer in every payment, so borrowed USDC can go only to a merchant. The vault refuses a payment to the agent's own account. This is why the credit co can lend to an agent with no history: an agent that never repays has cost the credit co at most the $1 of merchant fees it paid on the agent's behalf, and the agent itself received nothing it could withdraw. Section 8 describes how a payment is approved, proved and booked.
5.5Repay in Full§
The agent repays its whole balance at once. The payer, which may be the agent or its operator, signs a single USDC transfer authorisation, and a keeper submits it and pays the gas, so repaying needs no ETH. The amount signed is a ceiling: interest accrues between signing and landing, so the vault takes what is owed, interest first and principal second, and refunds the rest in the same transaction.
An operator can also post a USDC bond against an agent's line. The bond is taken first if the line defaults, and can be withdrawn only once the line owes nothing.
5.6Roll or Default§
An unpaid balance does not default at once. Past its due date, a balance at or below the rollover cap may roll into another term, up to twice per cycle, and a cycle that rolled is not clean. A line that cannot roll can be marked late one day after its due date, and a late line can be marked defaulted seven days after that. Default takes any bond first, sets the agent's next line to zero, and blocks its operator from borrowing. A defaulted line can still be repaid, and the credit co keeps what is paid, but repayment does not restore it.
- Openrepaid in fullRepaidclean if it never rolled
- Openpast due, balance at most $1, under 2 rollsRolleddue one term later, still open
- Open1 day past due, cannot rollLate7 days past dueDefaulted
Marking is open so that no party, the credit co included, can delay or suppress a default. Once the grace has passed, anyone can write the mark.
6Pricing the Line§
6.1Interest§
Interest is simple and accrues every second on the outstanding balance. Each cycle also carries a minimum of one day's interest on its peak balance, so an agent that borrows and repays within seconds still pays one day's interest.
The agent owes the larger of the two, less what it has already paid. If our agent's $0.42 is outstanding for five days at 24%, it owes about $0.0014 of interest (Appendix A).
6.2The Cost of a First Line§
A first line earns the credit co almost nothing. One day's interest on $1 at 24% is less than a tenth of a cent, and the line is not priced to earn. Its cost is the registration gas, the first draws of agents that never repay, and the agents registered only to draw the dollar and disappear. The credit co books all of it as the cost of acquiring an agent. That cost is bounded by the $1 cap on each new line and the limit of 100 new registrations a day, so new lines can cost at most $100 on any day at launch values.
The pilot measures that cost per agent still making paid calls 90 days later, not per agent onboarded. An agent that takes one loan and never returns has cost the credit co money and brought it nothing.
6.3Earning a Larger Line§
A line grows only on repayment. How much the agent spends within a cycle has no effect on it. When a cycle closes, the next one opens from a score attestation signed by Roundhouse (section 7), with a cap equal to the smallest of three figures: the score's maximum line, the cap the agent has earned, and the largest line. The earned cap starts at the first line and doubles with every third clean cycle in a row. A late mark halves it and restarts the count; a default sets it to zero.
Longer terms are earned the same way, 14 days after three clean cycles and 30 after six, and they cost more: 30% and 36% APR against 24% for seven days. The higher rate reflects the longer exposure. Terms are also earned rather than chosen, so the term an agent is on is itself a signal of its history. Our agent, repaying cleanly, opens its fourth cycle at $2 on a 14-day term and its seventh at the $4 launch maximum on a 30-day term.
7The Record and the Score§
An agent's history on chain is a list of events. A lender needs a cap and a term. The score converts the first into the second. Roundhouse publishes the rules for writing the record, which we call the Roundhouse credit standard, and sells a score built from the records that lenders write. The standard is published as a specification first and will be proposed as an ERC once the credit co has run on it in production. The services that sign scores and check KYA are not yet live (section 8.8).
7.1Credit Events§
Every record is keyed to the agent's ERC-8004 identity: the chain, the identity registry and the agent's identifier. There is no separate account system. A lender that adopts the standard emits the following events from its ledger.
| Event | What it records |
|---|---|
CreditOpened |
A new cycle: cap, term, APR and the operator reference |
CreditDrawn |
One line item: registration or x402 payment, amount, merchant, resource label |
CreditSettled |
A netted batch: the new balance and due date |
CreditRepaid |
A repayment, and whether it closed the cycle |
CreditRolled |
A roll and the new due date |
CreditLate |
A late mark |
CreditDefaulted |
A default and the amount outstanding |
Three rules bind any lender that adopts the standard. Late and default marks must be open to anyone. An adopter must report a default within a day of its grace ending. And the score counts only lenders that report every loan.
7.2The Score§
The score is a signed attestation, which a lender checks for itself:
ScoreAttestation(
chainId, identityRegistry, agentId,
tier, maxLine, termDays,
inputsHash, // commits to the chain events and 8004 feedback used
asOf, validUntil,
signature // Roundhouse's signer, EIP-712
)
Its inputs are the agent's ERC-8004 reputation and validation feedback, which give a starting signal; its repayment history, which dominates as lines grow; and its settlement history with facilitators. A buyer checks the signature and the expiry, and can rebuild inputsHash from public chain data to see which events went in. The weights are not published, so a buyer can verify what went into a score and who signed it, but cannot recompute the score itself.
Two further rules limit how a score can be used. Attestations expire quickly, so a cached score cannot back a live lending decision. And outcomes on starter lines carry almost no weight when a large line is sized, so a $1 default does not count like a $500 one.
7.3Know Your Agent§
A know-your-agent (KYA) check ties an agent to its operator. The operator completes it with Roundhouse; an agent cannot do it for itself. Only a hash of the operator reference and a level reach the chain. No name, email address or document about a person is stored on chain or published.
KYA is switched off for the first launch, because Roundhouse's KYA provider is not yet in place. While it is off, the operator reference is an unchecked label, so the per-operator caps do not stop one person using many wallets, and the daily caps and the float are the real bound. The Terms page says whether KYA is on.
7.4Access and Price§
Anyone can buy a score or a KYA check by paying USDC on Base over x402, at one flat price per read. There is no exclusivity and no volume discount, and the credit co pays what every other buyer pays. The history an agent builds with one lender is then available to every other lender on the same terms, which is what makes the record worth building.
8Turntable§
Turntable is a lending product for AI agents, run by two companies at arm's length: the credit co, which lends, and Roundhouse AI, which keeps the record and sells the score. Its first product is the Agent Credit Line. As the record deepens, we intend to add the products described under Further Applications: score-priced lines, repayment from an agent's own earnings, and pools through which third parties can fund agents that have repaid.
Turntable is a sister brand to Roundhouse, and the two names belong together. On a railway, the roundhouse is the shed where engines are kept and serviced, and the turntable at its centre turns each engine onto the track it leaves by. Turntable does the same job for agents: it gets an agent with an empty wallet to its first paid call, and Roundhouse keeps the record of what happens next.
Turntable will launch in four stages. The test stage, running on Base since October 2026, has the credit co's contracts live with a small float and KYA off. The pilot adds a larger float with caps on, and Roundhouse's score and KYA, once legal opinions are in. Score-priced lines add score tiers, longer terms, repayment from receivables and the published standard, after an audit and the pilot's results. Pools of third-party USDC, and sponsorship, come last, after a legal opinion on the pools and an audit.
8.1Architecture§
Every credit payment passes through the same six steps. Both signatures exist before any money moves, USDC pays the merchant from the vault in step 4, and anyone can record the spend afterwards.
- Prepare (credit co). The agent's client asks the credit co's API to prepare a payment to a merchant. The API checks the line, the caps, the merchant and the vault's float, reserves the amount, and returns a
CreditPaymentfor the agent to sign. - Sign (agent). The agent signs it with the key that owns its identity, which never leaves its machine.
- Co-sign (credit co). The API co-signs the same message and returns a standard x402 payment header with the vault as payer and both signatures packed into the signature field.
- Settle (facilitator, USDC). The merchant's facilitator submits USDC's
transferWithAuthorization. USDC asks the vault, through ERC-1271,9 whether the signature is valid, and on a yes moves the price from the vault to the merchant. - Record (anyone). Anyone proves the spend from USDC's nonce record and books it as pending.
- Net (credit co). A netting job writes pending spends onto the ledger as debt, in batches.
The merchant and its facilitator see an ordinary payment. They need no change on their side and no relationship with Turntable.
The vault approves a payment only if the digest matches the payment's own fields, both the agent and the co-signer signed it, it is inside a validity window that ends at most an hour away, and payments are not paused. The merchant must not be the agent's own account, the vault, the zero address or a blocked address. The amount must fit the agent's line, the maximum loan, the operator cap and the daily cap, counting spends already recorded, and the vault's balance afterwards must stay at or above its reserve.
8.2Proof of Payment§
The payment's EIP-3009 nonce is the EIP-712 hash of the CreditPayment itself, so it commits to every field: agent, merchant, amount, validity window, and a salted label for the resource bought. Once USDC marks that nonce as used by the vault, it has proved that exactly this payment was made. recordSpend recomputes the nonce from the fields it is given and refuses unless USDC marks it used. That makes recording safe to leave open to anyone, because no caller can book a payment the vault did not approve. In practice the credit co's reconcile job records each spend about a minute after it settles.
The ledger, in turn, accepts debt only up to what the vault has recorded as paid. Netting runs when the oldest pending spend has waited an hour, when the pending total reaches $0.25, or when the line is full, whichever comes first. A stolen netting key can only record debt for money the vault has already paid.
8.3Facilitators§
A facilitator can settle a vault payment if it verifies a contract payer through ERC-1271 instead of recovering an address from the signature, accepts a signature longer than 65 bytes (the packed signatures come to about 500), and submits the form of transferWithAuthorization that takes the signature as bytes. Circle's Gateway nanopayments and x402 batch settlement do not support contract payers, and only the exact scheme with USDC on Base is covered.
On 12 October 2026 we tested nine facilitators. Two of them, PayAI10 and Treasure,11 settled a $0.01 payment from the vault on Base mainnet with no change on their side. Of the other seven, two passed verification and failed at settlement, one refused a contract payer, one asked us to retry later, two needed API credentials we did not hold, and one served only Base Sepolia and Solana. Until more facilitators pass, the claim is a narrow one: Turntable works with any x402 exact USDC facilitator that supports smart-contract wallets.
8.4Bounding a Loss§
USDC calls the vault's isValidSignature as a read-only call, so the vault cannot remember a payment that has been signed but not submitted, or count one that has been paid but not yet recorded. Two payments can each fit a line on chain and together exceed it. The co-signer closes that gap by signing only after it has reserved capacity for that payment off chain. That makes the co-signer, a hot key, the real limit on a burst of payments; the on-chain checks bound each payment on its own.
| Stolen | What it can take from the vault |
|---|---|
| The agent's key | Nothing. It cannot produce the co-signature. |
| The co-signer's key | Nothing. It cannot produce the agent's signature. |
| Both keys | Payments to unblocked merchants, each inside one agent's line, until the vault reaches its reserve. At most the float minus the reserve. |
| The credit co's API database | Reservations could be edited so the co-signer signs more than a line allows, but every payment still needs the agent's signature. |
A signed payment expires within an hour, so a leaked one stops working. The reserve is a floor the vault never pays below, set to what the credit co could not afford to lose to a theft of both keys. The credit co can pause payments from the next transaction, revoke the co-signer and with it every payment that key signed and that has not landed, block a single merchant, or withdraw the float. None of these actions can stop repayment; the contracts have no function that blocks a payer from repaying.
The contracts are upgradeable proxies. Each company's contracts sit behind its own timelock with a public delay, and only the timelock can grant the right to upgrade, so no administrator can skip the delay. Anyone who funds the vault is trusting whoever controls that timelock, so its holders and delay belong in any disclosure to them. On the agent's side, the Turntable helper keeps the agent's key in an Open Wallet Standard vault by default, behind a policy that allows signing only for Base, the credit vault and USDC.
8.5Who Funds a Line§
New agents borrow from the credit co's own funds, and their losses are its acquisition cost (section 6.2). An agent graduates to a second stage after a set number of clean cycles, three at launch. Graduated agents borrow from pools of third-party USDC, and pool money never funds a new agent, so lenders in the pools fund only agents that have already repaid the credit co's own money cleanly. The number of cycles to graduate is the main risk parameter in the system, and it stays under review as the first stage produces repayment data.
- Who borrows
- New agents
- Who funds
- The credit co's own book
- Who takes a loss
- The credit co, as acquisition cost
- Who borrows
- Graduated agents, after 3 clean cycles at launch
- Who funds
- Senior and junior pools
- Who takes a loss
- The junior pool first
Each pool has a senior and a junior part. The senior part earns interest through its share price and loses money only once the junior part is used up. The junior part earns more, takes the first loss, and must stay at 15% or more of the whole. Among lenders that lend without collateral, those that came through a wave of defaults had made first-loss capital mandatory and sized it: Maple, after $54 million of bad debt in 2022, required its pool delegates to post first-loss capital and began declaring defaults at once.12 The yield shown to pool lenders is interest actually paid over the last seven days, never a projection and never emissions, and withdrawals join a seven-day queue that pays in the order requested. Pools are not open.
8.6Two Companies at Arm's Length§
| Roundhouse AI | The credit co | |
|---|---|---|
| Role | Data and rails | Lender |
| Runs | KYA register, score, credit standard | Vault, ledger, registration, bonds, pools |
| Funds lines | No | Yes |
| Pays gas | No | Yes. Registration gas is booked on the line; the rest is a running cost. |
| Carries default and sybil risk | No | Yes |
| Earns | A flat fee per score or KYA read, paid on chain | Interest, net of losses |
| Contracts | KYAAnchor, ScoreOracle |
CreditLedger, CreditVault, RegistrationSponsor, BondVault |
Every value the credit co takes from Roundhouse, each KYA check and each score read, is metered and paid on chain at the published price. The licence is not exclusive, so any lender can adopt the standard and pay the same price. The credit co reads loan outcomes from chain rather than from a private Roundhouse database, and Roundhouse never signs, funds or controls the credit co's contracts or keys. The separation holds only if different people control the two companies, or any link between them is disclosed.
8.7Incentives§
One rule governs every reward: it must be a function of something that costs USDC to fake and leaves a permanent record on chain if faked. Interest paid meets that test. Volume, draws, call counts and agent counts can all be generated by scripts at no cost, so Turntable rewards none of them. What an agent earns for repaying is a larger line and a longer term. The obvious attacks are bounded as follows.
| Attack | What stops it |
|---|---|
| Borrow the first dollar and walk away | The line is $1 at most. The default is public and permanent, the operator is blocked, and the credit co has budgeted the loss. |
| Register many agents to take many dollars | At most 100 new lines a day at launch, each capped at $1, and 3 agents per operator. With KYA on, one person is one operator. |
| Pay your own wallet and cash out the line | The vault refuses the agent's own account. A second wallet still works, so it is the same accepted cost as a copied agent, and merchants paid by only one agent are flagged for review. |
| Borrow and repay at once to farm a record | Each cycle costs at least one day's interest, and a cap doubles only after three clean cycles in a row. |
| Record debt that was never lent | The ledger accepts debt only for payments the vault made, proved from USDC's own record. |
| Hide a default | Anyone can mark a line late or defaulted once its grace ends. |
| Roundhouse inflates scores to grow lending | Score inputs can be replayed from chain, Roundhouse earns a flat price per read whatever the size of the book, and it does not lend. |
8.8Status§
| Part | State on 20 October 2026 |
|---|---|
| Ledger, vault, registration sponsor, bond vault | Deployed on Base. Covered by a unit test suite and a test against a fork of Base mainnet. Not audited. |
| Registration | A registrar key mints each identity and transfers it to the agent. An ERC-4337 paymaster, which would let the agent's own account register, is not built. |
| KYA anchor and score oracle | Written, not deployed. Roundhouse's KYA provider and score signer are not built, so KYA is off and a line cannot yet open a second cycle. |
| Facilitators | Two settle vault payments on Base mainnet (section 8.3). |
| Repayment from receivables | Built and switched off. |
| Pools, sponsorship, reputation mirror | Not built. |
The design has limits the pilot has to test. Credit reaches only merchants whose facilitator accepts a contract payer, and two have so far shown that theirs does. The co-signer is a hot key and, because the caps count recorded spends, the only thing holding a line between a payment and its record; if it fails, the vault's reserve is the only remaining limit on losses. Merchants see the vault as the payer for every Turntable agent, with the agent's identity carried only in the payment's nonce. Demand is unproven, since most x402 payments today do not come from agents,6 and the pilot is designed to compare credited agents against a control group. How lending law applies to a line for an agent, and whether a line counts as consumer credit when the operator is an individual, depends on the jurisdiction, and the credit co needs legal opinions before the pilot. Whether the agent or its operator is the borrower of last resort is not yet decided.
9Further Applications§
The Agent Credit Line and Turntable's first mechanics are the base for a wider credit market for agents. The following applications extend them as the record deepens.
Repayment from receivables. Once an agent sells services over x402, it could repay from what it earns, with a share of each inbound payment going to its line until the cycle is repaid. The vault already supports this, and it stays switched off until facilitators' terms have been checked.
Reputation feedback. Once lines have a history, a mirror contract could write each repayment and default to the agent's ERC-8004 reputation, so any reader of the registry sees credit outcomes without buying a score.
Sponsorship. Once a third party is willing to vouch for an agent, it could post first-loss capital behind that one agent and earn a share of its interest, letting a known agent skip the first stage. It is planned as an optional feature, off by default.
Other lenders. Once the standard is published, any lender can adopt it and buy the score at the same price as the credit co, and each lender that reports makes the record more complete for all of them.
A payment proxy. Where a facilitator refuses a contract payer, a small service could pay the merchant from an ordinary account that the vault funds. It would work with every facilitator, at the cost of a hot key that can spend and an operator the agent and merchant must trust. It is not built.
The starter line comes first. Each step towards third-party money, pools and new kinds of claim moves closer to regulated lending, and waits for loss data, an audit and legal advice.
10Appendices§
AInterest and Line Growth§
Amounts are in USDC base units (1,000,000 is $1) and rates in basis points (2,400 is 24%), as the ledger stores them. Division rounds down.
For a balance of $0.42 outstanding for five days (432,000 seconds) at 2,400 basis points:
| Figure | Working | Result |
|---|---|---|
| Interest | 420,000 × 2,400 × 432,000 ÷ 315,360,000,000 | 1,380 units, $0.00138 |
| Minimum | 420,000 × 2,400 ÷ 3,650,000 | 276 units, $0.000276 |
| Owed | The larger of the two | $0.00138 |
| Payoff | Principal plus interest | $0.42138 |
The agent signs a repayment for a ceiling, say $0.43. The vault takes $0.42138 and refunds the rest in the same transaction, because interest keeps accruing between signing and landing, and an exact figure would leave dust or fail.
A line that repays cleanly every cycle grows as follows at launch values, before the score's own maximum is applied.
| Cycle | Clean cycles before it | Earned cap | Longest term |
|---|---|---|---|
| 1 to 3 | 0 to 2 | $1.00 | 7 days |
| 4 to 6 | 3 to 5 | $2.00 | 14 days |
| 7 onwards | 6 or more | $4.00 | 30 days |
BExample Line§
Line specification
Line
- agent_id
- 1042
- operator_ref
- Hash of the operator reference, no personal data
- stage
- 1, onboarding
- status
- Open
Terms
- cap
- 1.00 USDC
- term
- 7 days, due 7 days after registration
- apr
- 24% (2,400 bps)
- rollover_cap
- 1.00 USDC, 2 rolls
Items
- item 1
- Registration gas, 0.02 USDC
- items 2 to 41
- x402 payment, search service, 0.01 USDC each
Position
- balance
- 0.42 USDC
- can_still_draw
- 0.58 USDC
- interest_after_5_days
- 0.00138 USDC
CThe Credit Payment§
One CreditPayment covers one payment to one merchant. The agent signs it, and the credit co's co-signer signs the same EIP-712 digest. The signing domain is the vault's own: its name, version 1, the chain id, and the vault's address.
CreditPayment(
uint256 agentId,
address merchant,
uint256 value, // USDC base units
uint256 validAfter, // unix seconds, 0 in practice
uint256 validBefore, // at most 1 hour after the vault checks it
bytes32 resourceHash // a hash of the resource plus a random salt
)
The EIP-712 struct hash of the payment is the EIP-3009 nonce, and the salt in resourceHash keeps two payments for the same resource from sharing a nonce. The x402 exact payload carries an ordinary TransferWithAuthorization with the vault as from:
{
"signature": "0x<proof>",
"authorization": {
"from": "<vault>", "to": "<merchant>", "value": "10000",
"validAfter": "0", "validBefore": "1800000600", "nonce": "0x<struct hash>"
}
}
The proof takes the place of a 65 byte signature. It is abi.encode(payment, agentSignature, cosignerSignature), where payment is the six fields above as a tuple.
DContracts and Keys§
| Contract | Company | What it does |
|---|---|---|
CreditLedger |
Credit co | Lines, line items, terms, interest, permissionless late and default marks |
CreditVault |
Credit co | Holds the float, approves payments through ERC-1271, records spends, takes repayments |
RegistrationSponsor |
Credit co | Pays for ERC-8004 registration and books its gas as line item 1 |
BondVault |
Credit co | Holds optional bonds, taken first on default |
KYAAnchor |
Roundhouse | A KYA hash per operator and agent, read for a fee |
ScoreOracle |
Roundhouse | Verifies signed score attestations, read for a fee |
The credit co's web service holds five hot keys, and none of them can move the float on its own.
| Key | What it can do if stolen |
|---|---|
| Co-signer | Co-sign payments. Spends nothing without the agent's signature. |
| Settler | Record debt for spends the vault already paid. Cannot create debt or move money. |
| Keeper | Call functions anyone may call, and pay their gas. |
| Manager | Open a cycle from a score Roundhouse signed. Cannot forge a score. |
| Registrar | Mint identities and report their gas, within the sponsor's daily and per-operator limits. |
The administrator is a multisig, and the treasurer and pauser keys are not held by the web service.
References
- 1x402, an open protocol for payments over HTTP 402. x402.org Back
- 2EIP-3009, Transfer With Authorization. eips.ethereum.org/EIPS/eip-3009 Back
- 3InfoQ, report on the Cloudflare and AWS x402 integrations, July 2026. infoq.com Back
- 4ERC-8004, Trustless Agents. eips.ethereum.org/EIPS/eip-8004 Back
- 5Quicknode, on its ERC-8004 stack and replayable composite reputation score. quicknode.com Back
- 6PYMNTS, on TRM Labs' analysis of who makes x402 payments, 2026. pymnts.com Back
- 7Dune, analysis of wash trading on Blur, LooksRare and X2Y2. community.dune.com Back
- 8Turntable, Terms, read from the
CreditLedgercontract on Base. Back - 9ERC-1271, Standard Signature Validation Method for Contracts. eips.ethereum.org/EIPS/eip-1271 Back
- 10Settlement through PayAI, 12 October 2026, $0.01 from the credit vault on Base. basescan.org Back
- 11Settlement through Treasure, 12 October 2026, $0.01 from the credit vault on Base. basescan.org Back
- 12Blockworks, on Maple Finance's bad loans and its move to immediate defaults. blockworks.com Back
This paper describes Turntable as designed and built on its publication date. It is not an offer of credit, a pool, a token or a return. Launch values can change: the Terms and System pages show the live state.