Are x402 health labels accurate?
The short answer
Not reliably. A large share of the endpoints one directory flagged as failing returned a valid payment quote on every independent check we made, and some it flagged as working never produced one.
As published in the issue
Of 955 endpoints 402index labelled
degradedordown, 754 — 79% — returned a valid payment challenge on every independent probe we made in late August. 52 labelledhealthynever returned one.
(2) The OBSERVED population: every endpoint on probe402's seed that produced at least one graded observation inside the sealed window of 21–31 August 2026, joined to a third party's published label for the row whose address is the exact route we probe; an endpoint we did not grade, or whose label read failed, is outside every joined figure and is counted rather than dropped.
On 1 September the
last_checkedstamps on the disagreeing labels were hours old, so staleness does not explain what follows.
The sentence above is quoted from a published issue and is not re-written here. Its date, its population and the record behind it are below.
Why it matters
Agents, and the people who buy agent infrastructure, route on these labels. A label that disagrees with observed behaviour sends traffic away from endpoints that answer and towards endpoints that do not.
What we measured
We compared the directory's published labels against our own observations of the same routes across the observed population inside the sealed late-August window, and counted the rows where the label and the observed payment-wall behaviour disagreed.
What the evidence shows
- Sealed days behind the archive figures
- 11
- Observations in that window
- 29,022
- Issue
- Machine Commerce Reality Index: #001 — August 2026
- Published
11 days
2026-08-21 · 2026-08-22 · 2026-08-23 · 2026-08-24 · 2026-08-25 · 2026-08-26 · 2026-08-27 · 2026-08-28 · 2026-08-29 · 2026-08-30 · 2026-08-31
All 11 days behind these figures (21–31 Aug 2026) are sealed, each with its published hash on the Integrity page; each date opens that day's manifest.
What this does not say
- Third-party figures are those providers' own published numbers, quoted with their own denominators and the date we read them. They are not our measurements of those providers.
Method and reproduction
The runs behind the figures, the record each one produced, and the terms this question uses.
What ran, when, and the record it produced.
| What ran | When | Record hash |
|---|---|---|
| The join behind the issue's opening finding: every independently graded endpoint of the sealed window set beside one fresh dated read of the health-label surface, covering all 1,588 hosts we cover, with that surface's own last-checked stamp carried per row. | 158940a9769dd5b44614f9f9201235f708388e7b0fdae0a589e3fd063798be1a |
The terms
The terms this question uses are defined in the glossary.
Corrections and disputes
This answer names a party we measure. The route to dispute a reading is on the corrections page.
Corrections to this issue are on the corrections register, and this answer reads from the record rather than from a copy.