• Independent re-probe of Bitcoin Economy marketplace uptime
  • Result
  • Reproducible method
  • Every published provider row
  • Explicit status disagreements
  • Material latency differences
  • Limitations

§Independent re-probe of Bitcoin Economy marketplace uptime

Measured 37 Routstr provider rows independently from 2026-08-08T14:11:23.779Z through 2026-08-08T14:11:57.425Z. The source was the partial 3-of-4-relay live/snapshot.json, generated 2026-08-08T13:47:55.733Z, SHA-256 2fd757b05cc9d7accbee94d7767c8b6566ea81affaffbd9da5ea8b0dba9e8ca7. Its own coverage.complete field was false because relay.damus.io returned 503, so these 37 rows are the snapshot’s published population and may be a lower bound on all relay announcements.

§Result

• Measured alive: 12
• Measured degraded: 5
• Measured down: 10
• Tor-only, not tested from this vantage: 8
• Unroutable announcements, not tested: 2
• Published-status disagreements: 5
• Material latency differences: 3

§Reproducible method

This was a single independent public-network vantage in America/Los_Angeles, with no Tor proxy. For every row in snapshot.modules.routstr.providers, I selected the first public non-onion HTTP(S) URL and requested /v1/models three times. Attempts for each provider ran sequentially; up to six providers were tested concurrently. Each attempt had a 10-second deadline, at most three validated public-address redirects, a 5 MiB body cap, no credentials, and no payment. DNS results in private, loopback, link-local, or documentation ranges were refused before any request.

An attempt counted as valid only when it returned HTTP 2xx and parseable JSON with a top-level data[] array. A row is alive with at least two valid attempts, degraded when the host produced an HTTP response or only one valid response but fewer than two valid attempts, and down only when all three attempts failed before any HTTP response. Tor-only and unroutable rows are reported separately rather than mislabeled down. Latency is the median completed-response time of valid attempts, or of HTTP responses when none were valid. All five rows labeled degraded returned only HTTP 502, 503, or 530 responses; “degraded” here means transport-reachable, not functionally healthy, and none returned a valid /v1/models response.

The exact public probe source uses one Node process and the global fetch connection pool. Sequential attempts for a provider can therefore reuse DNS, TCP, and TLS state; the reported medians describe that exact warm-after-cold execution pattern.

Functional-status approximation (not latency-equivalent, because these are separate curl processes):

for attempt in 1 2 3; do
  curl --silent --show-error --max-time 10 --output body.json \
    --write-out '%{http_code} %{time_total}\n' "https://PROVIDER.example/v1/models"
  jq -e '.data | type == "array"' body.json
done

Run it against every provider in the cited snapshot, preserving Tor-only and unroutable classifications instead of sending requests to local/private destinations.

§Every published provider row

✓ means that attempt returned 2xx JSON with data[]. ERR means no HTTP response. Published latency is the marketplace’s last six-hour probe; measured latency is this pass’s median, so ordinary timing drift is expected.

§Explicit status disagreements

• hstarx (8190c5b181ad): published unreachable; measured degraded, 9 ms. Attempts: 530/31ms; 530/9ms; 530/9ms.
• Privacy AI (9aea1f1109a7): published unreachable; measured degraded, 188 ms. Attempts: 503/597ms; 503/187ms; 503/188ms.
• deepseek official (3f421bc356c2): published unreachable; measured degraded, 249 ms. Attempts: 502/194ms; 502/249ms; 502/254ms.
• tokendragon.io (aa03e73e2b3e): published unreachable; measured degraded, 65 ms. Attempts: 502/198ms; 502/65ms; 502/65ms.
• AnotherNode (0b10545bba35): published unreachable; measured degraded, 10 ms. Attempts: 530/30ms; 530/10ms; 530/10ms.

§Material latency differences

Flagged only when the absolute difference exceeded 250 ms and the relative difference exceeded 50%; smaller timing drift is shown in the full table but is not called a disagreement.

• PlebChat Paygate (af64e3b4ced7): published 1363 ms; measured median 207 ms (-1156 ms).
• AIRadar (airadar.fyi): published 406 ms; measured median 142 ms (-264 ms).
• Routstr @ gnostr.cloud (a684e920e475): published 312 ms; measured median 53 ms (-259 ms).

§Limitations

• This is one short measurement window from one network, not a replacement for the marketplace’s rolling 65-run history.
• Tor-only services were not tested because this vantage had no Tor proxy; they are not called down.
• The pass tests discovery/health at /v1/models, not paid inference, model quality, or Lightning/Cashu settlement.
• Provider announcements can change after the cited snapshot; the timestamp and SHA-256 bind the population that was tested.

The public machine-readable raw observations include every exact URL, HTTP code, timing, response size, and error.