Earlier today we started publishing a payability index: a JSON file generated straight from our own delivery ledger, with the window, the exclusions and a list of things it does not claim. The first version said this, over 30 days:
paid_calls.delivered 62
paid_calls.distinct_paying_identities 15
A reader would reasonably take "62 delivered paid calls from 15 identities" as demand. It is not. When we broke those same rows down by the User-Agent the caller declared, 38 of the 62 delivered paid calls say they are a probe or a scout in their own UA string. An indexer paying one cent to prove that an endpoint settles is a real payment โ it settles on chain, it is revenue โ but it is not a buyer choosing to spend. If a marketplace counts settlements as buyers without saying so, its "paid calls" number is mostly measuring its own verification traffic.
So the file now publishes the split. We did not subtract the probes from the headline.
Rule, as published in the file: the caller's declared User-Agent contains probe, scout, crawl or monitor, case-insensitive. Everything else is other. 30-day window, platform-internal traffic and two known test wallets excluded โ the same filters as the headline numbers.
| paid calls, 30 days | declared probe/scout | other | all |
|---|---|---|---|
| attempts | 39 | 44 | 83 |
| delivered | 38 | 24 | 62 |
| identities (per class) | 5 | 10 | 15 |
| trial calls, 30 days | declared probe/scout | other | all |
|---|---|---|---|
| attempts | 124 | 1098 | 1222 |
| delivered | 60 | 402 | 462 |
| identities (per class) | 104 | 815 | โ |
Trial traffic is mostly other: 124 of 1,222 attempts declare themselves probes. Paid traffic is the opposite. That makes sense if you think about what a probe is for. A free trial proves the handler works; a paid call proves the whole rail works โ challenge, signature, verify, settle. The cheapest way to test a rail is to pay one cent for it once, per endpoint. Thirty delivered paid calls from a single declared-probe identity is exactly that shape.
Removing probes would make the headline nicer and the file less honest. A settled payment is a fact about the rail; a buyer is a fact about demand. Collapsing them hides the second fact. Reporting both โ 62 delivered paid calls, of which 38 declare themselves probes โ lets a reader decide which number answers their question. If you are asking "does the payment rail work", 62 is the number. If you are asking "is anyone buying", the probe-split column is the one to read, and even then see the caveat in the next section.
other.other is not evidence of a buyer. It contains real agents, our own tooling that is not excluded by the wallet list, and anything that did not announce itself.self flag is not a buyer test. It separates platform-internal traffic by source IP. It cannot tell a third-party probe from a buyer, which is why the caller-declared UA is the only split the ledger can offer.as_of, not the number in an old quote.The generator used to parse the temp file and rename it over the served one. Parsing proves the JSON is valid, not that the numbers agree with each other. Since this file is meant to be cited, a document whose own breakdown does not add up is worse than a stale one, so the writer now re-reads what it just built and refuses to publish when:
by_outcome does not sum to attempts
delivered != by_outcome.ok
success_rate != delivered / attempts
success_rate_excluding_caller_refusals != delivered / (attempts - refused)
caller_ua_class does not cover every row
On a contradiction it leaves the previous as_of in place and exits non-zero, so the hourly cron fails loudly rather than serving a number that cannot be recomputed from its own fields. That is the whole test โ it is local and cheap, and it deliberately cannot catch a wrong SQL predicate. For that, the check is the independent recompute readers can do with the method block.
A probe that pays one cent and a buyer that pays one cent are the same row in a settlement log. The difference is only recoverable if someone keeps the caller's metadata. So, to anyone who runs an x402 directory and reports calls: publish your as_of, say what you exclude, and if you settle test payments yourself, say how many of your "paid calls" are your own probes. That one field is what would make two directories' numbers comparable โ and we would rather ask for it than guess how yours are counted.
Current snapshot: as_of 2026-11-22T19:02:52Z, 30-day window, 1,940 active services of which 251 are third-party listings. Refunds over the same window: 100 payouts / 9.64 USDC across 9 days, with 90 of them on 2026-11-17 โ published separately because a refund is a payout, not a failed delivery. Fetch the live file (or read the human version); the numbers here are a snapshot and the file moves hourly.