An independent check on the Visa/Artemis seller rule

AQUIX Check #01 · measured 2026-09-12 · Base · 21,082,341 settlement legs

← index · Methodology · Audit Report #1 · Corrections

Why this exists

In July 2026 Visa and Artemis published a wash-adjusted x402 figure — $135.7M raw reduced to $15.0M adjusted, about 89% of volume removed — and, unusually, they also published the rule that produced it: a seller counts when it has at least 3 adjusted transactions from at least 2 unique buyers.

A published rule is an invitation to be checked, and in a year nobody took it. Most participants in this market publish a number without a method, or a method without the per-address evidence to test it. We hold a complete settlement store for Base, so we can run their rule directly and say whether it reproduces.

This is not a criticism of Artemis. It is the thing their transparency was for.

The method

SOURCE   x402_evm_settlements — every EIP-3009 settlement AQUIX has reconstructed on
         Base, each carrying recipient, payer, amount, tx_hash, log_index
READ     21,082,341 settlement legs · 203,529 distinct recipients
         walked by keyset on (tx_hash, log_index), no sampling, no estimation
RULE     a recipient PASSES when  transactions >= 3  AND  distinct payers >= 2
         implemented exactly as published, with no additions of our own
NOT      we did NOT apply our scope test, our classifier, our sybil gate or our
APPLIED  exclusions to this calculation. The point is to run THEIR rule, not ours.

A correction to our own first attempt. This check was first run on 2026-08-13 and keyed its walk on tx_hash alone. The settlement primary key is (tx_hash, log_index), and a batched settlement puts two or more legs on one hash — so every page boundary that fell inside such a group silently skipped the rest of it. That walk missed 757 legs. The figures below come from a corrected walk that cannot skip. We publish this because a check that will not check itself is not a check.

Result 1 — the rule reproduces, in direction

recipients that PASS      2,598 of 203,529          (1.3%)
dollars that PASS         $30,520,884 of $110,083,015  (27.7%)

their rule REMOVES  72.3% OF DOLLARS  and  98.7% OF SELLERS

Visa and Artemis published about 89% of volume removed. We measure 72.3% on a different dataset, a different chain scope and a different period. The direction and the order of magnitude agree; the seventeen-point gap does not, and we have not isolated what causes it. Our first run of this check reported 80.3% and called a nine-point gap a corroboration — on a larger, corrected store the gap is wider, and the honest statement is narrower.

★ The finding is not that they are wrong. It is that almost nothing on this rail survives a three-transaction, two-buyer test. 98.7% of the addresses paid on this rail fail it. That is their story, not ours, and our numbers support it.

Result 2 — where the two methodologies disagree

AQUIX publishes                        420
the Artemis rule admits              2,598
in BOTH                                199
AQUIX only (their rule excludes)       221
ARTEMIS only (WE exclude)            2,399     worth $25,481,960

Only 199 addresses satisfy both rules. Two serious methodologies answering the same question — which sellers on this rail are real? — agree on under 8% of the admitted set. Anyone quoting an "adjusted" x402 figure is quoting one methodology's answer to a question on which two of them disagree by a factor of six.

Result 3 — why we exclude those 2,399, stated whether or not it favours us

Our first run left this question open and marked it [PENDING]. It stayed open for twenty-nine days. Here is the answer.

    10    $18,080,662   excluded on purpose: INFRASTRUCTURE
  2,147     $6,193,129   never graded — no provenance record at all
    239     $1,207,536   our own provenance grades IT x402_direct
      3           $634   graded none

$18.08M of the $25.5M — 71% — is ten addresses we exclude deliberately as infrastructure: routers, facilitators and payment plumbing that move money rather than earn it. That is a documented decision, not an oversight, and it is most of the disagreement.
239 addresses, $1,207,536, are a contradiction between two of our own instruments. Our provenance grader says x402_direct; our scope test excludes them. Our own constitution records why: the scope test requires the authorising party to be the credited party, which can only hold at the address the buyer signs to — it measures "is this the front door of a rail", not "is this x402". Every merchant one hop down fails it by construction. Their rule makes no assumption about payment topology. Ours does, and the assumption is wrong.
2,147 addresses, $6.19M, we never graded at all. The grader works to a compute budget and stopped before reaching them. Until 2026-09-12 nothing on our pages distinguished "we ran out of budget" from "we checked and found nothing" — that is now fixed, and the budget raised.

Stated limits — first-class output, not caveats

Different datasets. Visa and Artemis measure across their coverage; we measure what AQUIX has reconstructed on Base. Part of the gap is this and we do not claim to have isolated how much.
Our store begins 2026-06-28 and holds nothing before it. Any address whose x402 life is mostly older than that is understated here.
Within the period it does cover, our store is about 99% complete, not 100%. Independent sampling from chain located x402 payments the store does not hold, at roughly one in a hundred.
Coverage is USDC on Base. Our own measurement shows a large share of published addresses move real non-USDC value we do not count.
Neither rule can see intent. Three payments from two addresses is a behavioural proxy, and one entity with free wallets can satisfy any such rule. Neither their rule nor ours is proof; both are evidence.
Our published error rate applies to all of the above and is on the front page.

How to reproduce

Walk any complete x402 settlement store on Base by keyset on (tx_hash, log_index) — not on tx_hash alone, which is not unique. Group by recipient; count transactions and distinct payers; admit where tx >= 3 AND payers >= 2. No sampling. If your numbers differ from ours, we want to know, and we will publish the difference.

AQUIX is an independent observatory. No protocol pays to change a verdict. Every correction is permanent and history is never rewritten. This check was run because Artemis published their rule — which is the behaviour this field needs more of, not less.