x402 Uptime Observatory

Method (pre-registered): top ~200 resources by lastUpdated on the public CDP x402 discovery catalog, refreshed daily; GBLIN's own endpoints are always included and judged by the same rules. Probe: GET, accept: application/json, 8s timeout, follow redirects; if the GET returns 400/402 without a parseable challenge/403/404/405/501 (a POST-only route) one POST retry with an empty JSON body; 17 endpoints are probed every 3 hours, so each endpoint is probed about every 36 hours. Alive = HTTP 402 whose challenge exposes a non-empty accepts[] array — read from the PAYMENT-REQUIRED header (base64 JSON, the x402 v2 form) or from the response body — or any 2xx, within the timeout. probes never pay anyone, never judge quality — liveness only.

Correction log. Rule v2.3 (2026-09-24 → null): as v2.2, plus a GET answered with 403, or with 402 but no parseable challenge, also triggers the single POST retry. found by a hand check the same day of the 14 not-alive results in a random sample of 400 of the 17,071 catalog listings: 3 of them (two answering 403 to GET, one answering 402 without a parseable challenge to GET) carried a valid challenge on POST.
Rule v2.2 (2026-09-24 → 2026-09-24): as v2.1, plus: (a) a listing whose path keeps a route placeholder (/:name or {name}) counts as alive if it answers with a challenge and is otherwise left out, since a refused literal placeholder is not evidence that the service is down; (b) a probe that gets no HTTP answer at all (network error or timeout) is shown as unconfirmed and left out of the percentage until a second consecutive probe also gets none. hand check on 2026-09-24 of the 4 endpoints reported not alive out of 187: 3 had failed a single probe with no HTTP answer and, re-checked by hand from a separate network, answered with a valid 402 challenge; the 4th was a placeholder path (':symbol'). Also corrected: the probe cadence was published as 'roughly every 2 hours', while the cadence in force is 17 probes every 3 hours, about 36 hours per endpoint.
Rule v2.1 (2026-08-18 → 2026-09-24): as v2, plus HTTP 501 added to the statuses that trigger the single POST retry (aligns with the independent cross-check method published by M. Oliva, x402 Slack, 2026-08-18). no effect on the 201 endpoints tracked on 2026-08-18 (none returned 501); recorded so that the two methods are reproducibly identical on the verb-fallback rule.
Rule v1 (2026-08-16 → 2026-08-18): plain GET only; alive = HTTP 402 with a parseable accepts[] in the BODY, or any 2xx. v1 measured dialect, not liveness: it never read the PAYMENT-REQUIRED header and never retried POST-only routes. Flagged by a third-party re-run in the x402 Slack (#wg-domain-discovery, 2026-08-17). Our own cross-check on the same 276 targets, 2026-08-18: rule v1 33.7% alive, rule v2 98.9%; 139 targets carried the challenge in the header only, 41 were POST-only, 0 were unreachable. The published 36.7% (2026-08-16) was therefore wrong as a liveness figure. Cross-check data: https://github.com/gblinproject/x402-catalog-probe. v1 consecutive-fail counters were reset at migration; last_ok timestamps observed under v1 remain valid (v1 was strictly stricter).