1 · The problem
Alice must pick a seller
- Fake buyers: the money comes back to Mallory. She pays only a fee, about $0.017.
- Bad service: a sale shows that Alice paid. It does not show that Bob delivered.
Rating providers for AI agents from the payments they already make.
AI agents buy services from each other. They pay for each request, often one cent at a time. The payments settle in stablecoins on a public blockchain. The median payment is \$0.01.
Anyone can list a service. An agent cannot tell a good provider from a fraud.
Every known fix taxes the market: a curator, credentials, a paid referee, or reviews nobody writes. Anduin adds none of them. It reads only the payments that already settled, and it makes a high rating cost real money. Four pictures explain the idea; the sections after them have controls you can move. The pictures use toy numbers, and every measured number comes from the paper.
Circle Research
How it works · four pictures
Alice is an AI agent that buys services. Sellers list them, and Mallory runs one of the sellers. Step through each picture with Next and Back, or the ← → keys. The dollar amounts in the pictures are toy numbers; the measured ones follow.
1 · The problem
2 · Where trust starts
3 · Why a flow
4 · Why amplified
The test: 813 real buyers whose next payment went to a seller they had never paid. Did the rating put that seller first?
amplified flow. The standard rule, co-user count, gets 23.2%: a draw.
to top all 813 buyers at once: 23× the $358 the standard rule asks.
x402 stablecoin settlements, 16 blockchains, 157 days.
The rest of the page has controls you can move, and uses the paper's words: Alice is the buyer, a seller is a provider, the sellers she already pays are her known providers, co-users are customers, and a new seller is a candidate.
Deeper 1 The usual fixes
A buyer like Alice sees endpoints, and the descriptions their owners wrote. There are no brands and no reviews. Marketplaces have proposed other ways to choose, and each one changes the market. Select one to see its shape.
Sources, in order: curated marketplaces (x402 Bazaar); identity credentials; paid evaluator contracts that hold the payment; voluntary onchain ratings; and usage ranking. The last two figures come from prior work on deployed rating systems, cited in the paper's introduction.
Deeper 2 Drawing the market
The pictures above are a small piece of a real record. Blockchain payments are public and permanent. Anyone can read who paid whom, and how much. Buyers pay providers. Providers never pay providers. So the record sorts into two columns, and the paper calls this a bipartite graph.
One edge is one buyer's whole relationship with one provider. Its weight is the sum of their payments in the window, fees included. A thicker edge carries more dollars. This graph is the only input.
Two sets matter, and they are marked below. $K(b)$ holds the providers the buyer already pays. $C(b)$ holds the other buyers that pay those same providers. Deeper 3 uses nothing else.
Hover or tap an edge to read its weight. The weights come from Figure 4 of the paper. The rest of the market is texture and carries no values.
Deeper 3 Scoring a stranger
Buyer $b$ pays two providers. Those two are its known providers, the set $K(b)$. Other buyers pay the same providers. They are co-users, the set $C(b)$; this page calls them customers. Both customers also pay $p'$, which $b$ never paid. $p'$ is the candidate.
A customer that pays your providers needs what you need. Its next choice is evidence. That evidence moves from $b$ to a customer through the providers they share. Each shared provider carries no more than the thinner of its two edges. Add those up and you get the reach to that customer. The customer's own payment to the candidate then caps what arrives.
Move the sliders. These two reaches are what every Anduin rating is built from.
bridge \$4, capped by its \$8 to the candidate
bridge \$5, capped by its \$4 to the candidate
K(b) holds 2, C(b) holds 2
Each edge is a pipe. Its bore is the weight. The water is what gets through. The paper's Figure 4 uses one known provider; a set needs two, so this page carries a second. The second button restores the figure, and the three ratings then read \$4, \$5 and \$8 as printed.
Deeper 4 The attack
Mallory's trick in the first picture has a name. One party posing as many is the Sybil attack, named in 2002. Entry is free, so a provider can create its own buyer identities.
One asymmetry carries the whole defence. A payment to the attacker's own provider costs only the fee, about \$0.017, because the money returns. A payment to a provider it does not own costs the full amount, and that money is gone. And no attacker can create a payment from the honest buyer.
A marketplace might reach for one of these three. All three read fake edges as real. All three fall for fees alone. Add fake buyers and watch.
What the attack costs, in transaction fees
One axis, one unit: dollars. Each bar carries the score that rule gives the attacker.
Usage ranking counts distinct buyers, so each fake pays the attacker itself: one fee. Co-user count needs each fake to pay a known provider first, to qualify: two fees. Co-user volume sums what customers paid the candidate, and the attacker pays itself: two fees in all.
Deeper 5 Three flows
The rule: trust reaches a candidate only on paths that start at the buyer's own payments, and a path carries no more than its thinnest edge. A fake that paid a known provider one cent carries one cent. More fakes do not help.
Three constructions apply the rule. They differ in how much of the buyer's pipes each customer may use. The sliders in Deeper 3 drive all three, and the page solves the real max-flow for each one.
Amplified flow widens each buyer pipe by $n_p$. That count is per provider: how many customers pay that provider and the candidate. Known provider 1 has two, so its pipe becomes 2 × \$4. Known provider 2 has only one, so its pipe becomes 1 × \$1.
one pipe. Both routes compete for it.
one pipe per customer. Add the routes.
one pool. The fattest route takes it all.
The vessel shows what reaches the candidate, on one scale for all three. An empty pipe is a route the construction could not supply. The page builds the query network of the paper and runs Edmonds-Karp on it, so the numbers are the algorithm's, not an approximation of it.
The paper proves two bounds. The first: if the attacker pays $m$ dollars to providers it does not own, no Anduin rating scores it above $m$. The identity count does not matter, and the bound is tight. So its only lever is money it gives away. Split a budget across more fakes and the score holds still while the fee bill climbs.
capped at the money it gave away
\$0.03 forfeited, plus 6 fees at \$0.017
each thin first edge throttles its route
The second bound is stricter, and covers maximum flow alone. Its score never exceeds the buyer's own spend, whatever the attacker's budget. Deeper 3 shows that ceiling.
Deeper 6 Results
The test. 157 days of real agent payments through x402, which carries most machine payment volume: \$6.1M against roughly \$270K for the next rail. Every rating reads the 90 days up to June 21, 2026. 136,646 buyers were active. 3,656 paid again within 30 days. 813 of those paid a provider they had never paid before. Those 813 are the test set, and each sees 381.4 candidates on average. Did the rating put the right provider first?
Was the first pick right?
Share of the 813 buyers whose next provider the rating ranked first. Higher is better.
Amplified flow is right for 22.8% of buyers. Co-user count, the standard rule, gets 23.2%, so on accuracy the two are a draw. Usage ranking, the marketplace default, gets 3.7%. Maximum flow ties at the top for 78.0% of buyers, so it rarely names one pick.
What first place costs to buy
Cheapest spend to reach the top 1 for all 813 buyers at once
The 813 buyers spent \$41,331 in all. To top every one of their lists costs \$8,259 under amplified flow. That is a fifth of their spend, and 23× the \$358 the standard rule asks. Co-user volume costs two cents. Usage ranking looks dear at \$217,539, but that is 12.8 million fees at \$0.017, and it is right for 3.7% of buyers.
Both measures at once
A rating belongs in the top right: accurate, and dear to fake. Only summation and amplified flow are there.
Figure 6 of the paper places the six ratings by quadrant, not to scale. “To scale” plots their measured coordinates, cost on a logarithmic axis, as an earlier draft of that figure did. Both readings are in the repository.
One knob matters in deployment: how far the rating looks back. A longer window reaches more of the market. It also lengthens every candidate list, so first place is harder to win.
the true provider is somewhere on the list
25.1% of the buyers it could reach at all
mean length of the candidate list
Amplified flow, fee-inclusive dollar weight. The slider snaps to the six measured windows; nothing between them is interpolated. Reach climbs all the way to 90 days. Accuracy peaks at 30. Use 30 to 90 days.
Deeper 7 Shipping it
Four things to know before you ship it. Two are proved. Two are awkward.
A score costs money, not identities. Pay $m$ dollars to providers you do not control, and no Anduin rating scores you above $m$. Any number of fakes gives the same result. The bound is tight, so that is the real price.
Maximum flow is capped by the buyer. Its score never exceeds what that buyer has spent with its own known providers. The attacker's budget does not change this.
Nothing has to change. The input is payments that already settled. No party must agree a convention before it transacts. Every other proposal needs that agreement.
The cost to compute is local. Summation flow precomputes one value per customer, then answers each candidate in constant time. Maximum and amplified flow solve one small max-flow per candidate, over the buyer's own neighbourhood and never the whole market.
Cold start is real. A new buyer has no known providers, no customers and no candidates, so every rating is undefined. Collaborative filtering has the same hole. One settled payment starts it. A new provider stays unrated until one customer pays it.
Shared wallets blur. One address can be the treasury for several agents. Their histories pool into one node, and the ratings get noisier. Per-agent subaccount keys — ERC-4337 session keys, BIP-32 derivation — would split them again.
Ship amplified flow. It has the best accuracy of the three, and the highest attack cost of all six at rank 1. Avoid maximum flow in a product: its guarantee is the strongest, but it ties at the top for 78.0% of buyers, so it cannot name a first pick.
Next Future work
Cold start is the open problem, and it bites the buyer first. A buyer that has never paid anyone has no known providers. With $K(b)$ empty there are no customers and no candidates, so every rating is undefined. An undefined rating cannot be ranked.
One way out: publish a small curated set of trusted providers. A new buyer credits each one a fixed amount — say \$1 — and those artificial edges become its $K(b)$. Nothing else is invented. The real customers of those providers become $C(b)$, the providers they pay become the candidates, and Anduin ranks them exactly as it would for any other buyer.
The seed size is the whole trade. It is also the ceiling: maximum flow can never score any candidate above what the buyer has spent, and at the start that is only the seed. So a small seed gives a weak but honest ranking that real payments quickly outweigh. A large seed hands the curated set the ranking, which is the gatekeeper Anduin set out to remove.
amplified flow, through the curated set's own customers
same buyer, same seed, a different neighbourhood
ranked from real payments the curated set never made
maximum flow can never pass 3 × \$1 until the buyer pays for real
None of this is in the paper. Nothing here is measured. The curated providers, their customers and the two candidates are a sketch; only the arithmetic is the paper's, run as a query network per candidate. Set the seed to zero to see the hole it fills. The questions we would test next: how large a seed can be before the attack cost falls, and how fast real payments outweigh it.
Reference
All six ratings on the 813 buyers, fee-inclusive dollar weight. Anduin's three are shaded.
To capture every buyer at once is one attack. To buy only the biggest spenders is another. The top 79 buyers hold 91% of the spend.
66,073,201 x402 stablecoin settlements across 16 blockchains, March 23 to August 26, 2026, worth \$6.1M at a median payment of \$0.01. Self-payments are removed. Ratings read the 90 days before June 21, 2026, 00:00 UTC — 24,429,942 payments worth \$3.83M — and are scored on the 30 days after. A provider is mature, and so rateable, once two distinct buyers have paid it: 3,649 of them. The base fee is \$0.017, the mean settlement fee across the chains that carry x402. Every rating on this page uses the fee-inclusive dollar weight, so accuracy and cost always come from one graph. Every figure here is read from the paper.
The paper behind this page is under review. This walkthrough presents the idea and the measured results. It is not the paper, and the paper governs.