Smaller on purpose: the case for a number that can go down

Most dashboards have a number in the top-left corner that only ever gets bigger. Total volume, total transactions, total users — up and to the right, forever. It feels like health. It is usually just arithmetic: if you only ever add rows and never audit them, the sum has one direction it can move.

Ours went down. On purpose. We shipped a correction, watched roughly half a million dollars leave the attributed column, and published the smaller figure with the reasoning attached. This note explains that decision — not the mechanisms behind it, which we’ve covered separately, but the discipline: what it means to publish a settlement number you can defend transaction by transaction to someone who disagrees with you.

A measurement is not a scoreboard. The moment your number can only go up, it stops describing the economy and starts describing your reluctance to subtract.

The correction that cost us half a million dollars

The short version. Some payment wallets in the x402 economy are shared infrastructure: a routing layer receives money on behalf of many downstream services, and each of those services — honestly, with no one lying — advertises that same wallet as its payment address. For a while, our attribution bridge resolved each wallet to a single name. That rule quietly turns shared plumbing into someone’s reputation.

wallet 0x66fa4d79…  ·  $415,797  ·  31,548,003 tx

claimed, truthfully and independently, by:
  aisa.one
  lnpay.ai
  paywall402.com

all three route through it. naming any ONE of them
fabricates a reputation out of the other two's traffic.

the full contested class: 309 wallets
  $920,236 of settlement · 55.4 million transactions

So we changed the rule. A wallet with multiple credible claimants is now contested — attributed to no one automatically, all claimants shown, resolvable only by first-party proof. That is what moved the headline: settlement that used to sit under a single name moved into the contested column, and the attributed total fell from roughly $52.74 million to $52.26 million.

The rule is not nihilism, and it can move money back the other way. BlockRun’s router at 0xe9030014… carries $356,667 across 21,834,453 transactions — and the live payment challenges on BlockRun’s own domain return that exact wallet. That is first-party proof, so a curator named it, and those dollars are attributed. The bar didn’t become “never name anything.” It became “name it when the evidence is yours to stand behind.”

Getting more honest made the number smaller. We want to be very clear that this is the expected direction. Attribution errors are not symmetric: an over-claim silently inflates the total, while an under-claim shows up as an uncomfortable, visible gap that someone eventually investigates. A pipeline that fixes what it finds will therefore mostly find inflation. If years of corrections had only ever pushed our number up, that would be the alarming outcome.

Three numbers, and what each one is for

We publish a split, not a headline. As measured on 2026-08-10:

total settlement            $54.05M   (everything that cleared)
attributed to a named
  service                   $52.26M   (96.69% of total)

of the ~$1.79M not attributed to a name:
  ~$1.17M carries a behaviour label
  the remainder is a genuinely unknown, stated gap

the three parts close to the cent.

The total is what the chains settled. The attributed figure is the subset where we can put a service’s name on the money and show you why. And the difference is not a rounding embarrassment we hope you won’t notice — it is a first-class output with its own taxonomy.

Because here is the thing about a wallet you cannot name: you can still characterise it. We maintain 39 behaviour classes, each defined by a stated SQL predicate anyone can re-run. One example: wallet 8wYeHi8k… is “a $1.00 flat-rate endpoint that took 20,369 payments from 1,243 distinct payers in 2 days and has never moved the $20,369.” We don’t know who that is. We know precisely what it is, and the label says exactly that much and no more.

NULL is not zero. An unknown wallet is not a wallet worth nothing, and a guessed name is worse than an honest gap — because a guess, once printed, is indistinguishable from a fact. The unattributed ~$1.79M shrinks only when evidence arrives. It never shrinks because a plausible answer was available and the quarter was ending.

The peer comparison

There is another public tracker of this economy, x402scan, and its comparable settlement figure is $52.77 million. Put the three numbers on one line and something instructive happens:

Agents Trust  total settlement       $54.05M
x402scan      comparable figure      $52.77M
Agents Trust  attributed to a name   $52.26M

The peer’s single number lands between our two. We measure more settlement than they report, and we attribute less of it to named services than their figure implies. Both halves of that sentence are doing work, and neither is a scoring of who is “right” — the honest reading is that a single undivided number has to sit somewhere inside the region an itemised split makes explicit.

The “we measure more” half is concrete. Facilitators — the relays that actually move x402 payments — define what a tracker can see at all. x402scan publishes a list of 31; we track 30 of those plus 6 they lack: GoPlausible, AEON, Polygon Foundation, Orbis API, Vistara-Labs/b402, and xpay.sh. More rails watched means more settlement observed, which is part of why our gross figure is the larger one.

The “we claim less” half is this whole post. Our attributed number sits about half a percentage point of the economy below the peer’s figure because roughly that much settlement is money we deliberately moved out of the named column — contested wallets, behaviour-labelled unknowns, a stated gap. If we applied a pick-a-winner rule to shared wallets and rounded our unknowns up to their most plausible names, our attributed figure would clear the benchmark comfortably. We know, because it used to.

This is also why we don’t frame the comparison as a race. Two trackers of the same economy should land near each other, and they do — the gap between any pair of these figures is on the order of one percent. What distinguishes measurements at that point isn’t the third significant digit. It’s whether the difference is explainable: whether each dollar of daylight between the numbers has a reason you can read. Ours is itemised on the Transparency page, address by address.

How a number goes down honestly

A downward revision is not one thing. In our history it has come in three distinct shapes, each with its own logic. We’ll keep the mechanisms brief — they have their own write-ups — because what matters here is the epistemics of each cut.

Cut by provenance, not by convenience

Two of our early data loaders swept up all Base USDC movement, not just x402 traffic. The decisive fact: 302,035 of their transactions, worth $6.35 billion, are dated before 2025-05-09 — the first x402 transfer that has ever existed on Base. That money cannot be x402. But we excluded those rails’ Base rows by provenance — which loader wrote them — not by date, because the loader was the thing that was broken. And the same rail names carried 1,792,636 legitimate Polygon transactions worth $54,346, which we kept: evidence justifies exactly the rows it was gathered on, and our evidence indicted the Base sweep, not the Polygon one. The clean-class subset of those capture rails survives at 892 transactions and $447,298. Meanwhile the production rail, cdp_evm_cdp_sql, runs 98.99% attributed; the capture rails ran 99.99% not. The cut followed the failure, no wider and no narrower.

Cut by shape

Some settlement is real on-chain and empty economically: an echo harness whose inflows and outflows both total $203,879.13 with a net of one cent; a circulation loop whose $314,337 of “revenue” came from buyers who spent 99.995% of their money on the wallet that funded them; a 23-node ring passing $85,547 among itself; a single November week holding 69.5 million payment legs — 17 times a normal week — most of it never moving onward. These aren’t under-performing services to be scored lower; they aren’t services, and the honest treatment is removal with the evidence attached. The full taxonomy is in the forgery note.

Cut by humility

The shared-wallet correction above: settlement that is perfectly real and perfectly x402, where the only dishonest thing was our confidence about whose it was. The money stays in the total; it leaves the named column until proof arrives. This is the cut that moved the headline this week, and the one we suspect other trackers will eventually have to make too — routing infrastructure is only becoming more shared.

The ratchet problem

Here is the argument in its general form, because it applies to far more than x402.

Every measurement pipeline makes errors in both directions. It counts things it shouldn’t (contamination, wash, double-claims) and misses things it should count (unwatched rails, unparsed formats). Corrections, therefore, should push the published number both ways over time — down when an inclusion fails audit, up when a blind spot gets covered.

Now look at a metric that has only ever gone up. Either that pipeline has never once made an inclusion error — which no pipeline achieves — or its corrections are being applied asymmetrically: blind spots get fixed, because fixing them feels like growth, while inclusion errors accumulate, because fixing those feels like loss. The number has stopped being a measurement and become a ratchet.

A headline that only ever goes up is measuring the popularity of the metric, not the health of the economy.

And the incentives run one layer deeper. A tracker’s number is also its marketing — a bigger economy makes a more important tracker. Every subtraction we shipped made our own product look smaller. We think that conflict of interest is exactly why the subtraction is the credibility: anyone can publish a number that flatters them. The number that can embarrass you is the one worth checking against.

Questions to ask any settlement number

None of this is proprietary. Here is the checklist we would apply to any on-chain economy figure, including our own — five questions, all answerable from the outside.

1. Has it ever gone down? Not dipped with the market — revised down, with a stated reason. A yes is weak evidence of auditing. A no, over a long enough history, is strong evidence of the ratchet.

2. Is there an “unknown” line? Every attribution system has residue. If the published breakdown sums the named parts to exactly the total, the residue was rounded into someone’s reputation. Ours is printed: ~$1.79M unnamed, of which ~$1.17M is at least characterised.

3. Does the split close? Named plus labelled plus unknown should equal the total — to the cent, not to the vibe. A split that closes is a split you can audit; a split that doesn’t is a chart.

4. What does a zero mean? Measured-and-absent and not-measured print identically. Ask what the pipeline structurally cannot see — in our case, any facilitator outside the allowlist — and whether the publisher says so unprompted.

5. Can they show you one transaction? Pick any dollar in the total and ask for its receipt: the transaction, the rail that captured it, the rule that classified it, the evidence for its name. A defensible number survives this at every row. A scoreboard survives it nowhere.

The number we’re building toward

$52.26 million is not the biggest figure we could print today. We could print $52.74M by un-shipping one correction, or clear the peer benchmark by adopting a winner-take-all rule we know to be wrong. The reason we don’t is not virtue; it’s that the entire value of a census is that someone who distrusts it can check it. A number optimised for size fails that test at exactly the rows where it was optimised.

So the commitment is this: the total will move when the economy moves. The attributed share will move when the evidence moves — in either direction. And every gap between those two will stay itemised in public, on the Transparency page and the Integrity view, where the living versions of every figure in this dated post actually reside.

If our number is smaller than you expected, good. That was the work.