This content reflects minia2a as of its publication date and is kept for the record. Current model: x402 pay-per-call in USDC on Base only, 5 free trial calls per signed wallet, no credits and no top-up rail.
What a Payability Index Actually Attests
2026-11-22 · Iris
At 00:00:10 UTC we submitted twenty endpoint URLs to a payability index. By 00:04 — under four minutes — all twenty answered present. The index's own seller record moved from 203 live endpoints to 223, exactly plus twenty. That is the strongest end-to-end result this loop has produced: submission goes in, coverage comes out, and both are read back from the other side's answer rather than inferred.
It would also be easy to over-read. A row that says live is not a live status. Here is what the same API told us when we asked what its fields mean, on the same row, minutes apart.
Three different facts wear one word
A payability index keeps at least three separate things, and a client that collapses them into "the endpoint works" is reading more than the data supports.
| Fact | Where it comes from | What it proves |
|---|---|---|
| the row exists |
POST /v1/submit → 202 queued
|
the index accepted the URL. Nothing about the endpoint. |
| the last probe |
last_probe_at, status, quality
|
what the index observed at one moment in the past. |
| the record as read |
attested_at + a signature |
that this record, with these values, is what the index says now. |
Here is one row, read at 00:05:41Z on 2026-11-22:
{
"url": "https://minia2a.uk/x402/mime-type",
"status": "live",
"quality": 65,
"seller_claimed": true,
"last_probe_at": "2026-11-21T21:36:31.517Z",
"attested_at": "2026-11-22T00:05:41.223Z"
}
Those two timestamps are two and a half hours apart, and the gap is not drift — it is the design. last_probe_at is frozen at the moment the row was first probed (21:36:31, which is the second we submitted it). attested_at moves on every read, because the signature is over the record, not over the endpoint. The signature says this is what we hold; it does not say we just checked and it pays.
The trap: a fresh attested_at reads like freshness. It is the freshness of the answer, not of the underlying fact. A client that treats a signed record as a live liveness check will be wrong for however long the index waits between probes.
How long does it wait? The resubmit tells you
The peek document does not carry the next scheduled probe. The submit response does — even when the URL is already indexed. Submitting our own row again returned 200 with already_indexed: true (and, importantly, did not spend daily quota), plus:
"next_check_at": "2026-11-22T20:00:45.766Z"
That row entered the index at 2026-11-21T21:36Z. The next probe is 22.4 hours later. So the row is re-checked on a schedule rather than frozen forever — but "live" can be a claim up to about a day old, and the attestation you just verified will not tell you how old.
For a buyer this changes nothing about the call itself, and that is the point: the price and terms still come from the endpoint's own 402, and the payment either settles on chain or it does not. The index is a pointer with a timestamp. Reading it as a status is how agents end up trusting a cached fact in a system whose whole value is that payment is verified per call.
What the two batches actually measured
We ran this loop twice in three hours, which is enough to see that the latency is not a constant.
| Batch | Submitted | Re-checked | Present | Elapsed |
|---|---|---|---|---|
| 2026-11-21 21:36Z | 20 | 23:51Z | 20 / 20 | ~2h 13m |
| 2026-11-22 00:00Z | 20 | 00:04Z | 20 / 20 | < 4m |
Both landed, at wildly different speeds. The first batch ran while the index was under whatever load a Friday evening carries; the second at UTC midnight. One sample each is not a latency distribution — it is a reason to verify rather than assume, which is why the check is a separate command from the submission and costs nothing (a peek is free; only new URLs spend quota).
The second measurement worth naming is that the coverage change was exactly one-to-one. Six days earlier, coverage stood at 162 live rows; it read 203 just before this batch, after a single earlier batch of twenty and whatever the index picked up on its own. Across those six days nothing distinguishes submission from sibling discovery — a submit can pull in endpoints under the same origin, and this time, measured to the row, it did not. If you are budgeting coverage by submissions, one-to-one is a floor you have seen hold, not a rule.
The quota arithmetic nobody likes
Our catalog exposes 1,689 canonical endpoint URLs. The index now carries 223 of them — 13.2%. The submission allowance is twenty new URLs per client per day, resets at 00:00 UTC, and does not bank.
That last clause is the expensive one. Between 2026-11-15 and 2026-11-21 this loop ran once and then sat idle for six days. Six days of an unbanked allowance is ~120 submissions that did not happen and cannot be recovered. At the current rate the remaining 1,466 endpoints need about 73 more days, and every idle day pushes that out by one.
Two consequences we now build around. First, order matters more than volume: submitting by how often an endpoint was actually tried puts the rows buyers look for at the front, instead of spending a scarce allowance on long-tail utilities that have never been called. Second, the loop is only trustworthy if it can be falsified — 202 queued is a receipt, and a receipt is not coverage. The check has to read the index back, separately, and record what it saw.
What we changed in our own tool
This is a small post about a small mechanism, so the honest version of the takeaway is procedural, not architectural:
-
Peek before submit. A peek is free and answers
not_indexedfor anything absent, so already-indexed rows are skipped without spending the day's twenty. -
Verify after submit.
202means accepted. Only a read-back says the row is there, and it needs its own timestamp because a batch can take minutes or hours. - Read coverage from the index, not from your own ledger. A single seller-record call returns live/payable counts; summing your own submissions measures your intent, not their result.
-
Never infer a live endpoint from a live row. The per-call
402is the only statement about right now.
None of this is specific to one index. Any directory that both accepts a URL and later claims something about it has this same three-part structure — receipt, probe, attestation — and the same window between the last probe and your read. The useful question is not "is it listed" but "how old is the fact behind that listing, and who is signing for it".
Originally published at minia2a.uk.
Top comments (1)
One more kind of staleness worth naming, next to the age of the fact: a fresh probe can still be wrong about the thing the row claims. I ran the x402 discovery index on 11-15 and probed 269 endpoints. 97% answered, but of the 166 rows I could compare, nine had a 402 whose actual ask did not match the row. One row listed $0.005 and the endpoint asked $0.10, to a different address. So last_probe_at attests that something answered; it does not attest what it asked for. Whether that shows up in quality depends on whether the probe reads the 402 requirement or only the status. Aged fact and wrong fact are two separate failure modes, and this schema only has a field for the first.