How the collection is run

Mean uptime

99.95%

30 days · across 52 sites with a full 30 days of probe history (149 of 151 sites have a probe)

Code changes shipped

3,277

8 weeks · 4,417 deploy runs, 123 failed (2.8%), 5 cancelled (not counted as failed)

Error-tracker records closed within 24 hours

83.2%

30 days · 731 of 879 error-tracker records closed (records, not distinct errors, from 67 sites; one noisy site can dominate)

Servers with a completed nightly backup

12 of 12

last night (8 Oct 2026) · of 12 servers in the nightly backup run (the job’s completion status, not a restore test)

Last restore test

Passed

6 Oct 2026 · one sampled database on one server per month, rotating

Certificates expiring within 14 days

0

next 14 days · of 151 certificates tracked; soonest expiry in 31 days

SEO fixes verified working

23.2%

90 days · 47 of 203 judged fixes (527 applied, 324 still pending)

• Only 47 of 203 judged SEO fixes (23.2%) are verified working over the last 90 days: 119 made no measurable difference and 37 made things worse. A further 324 are applied but not yet judged.

• Uptime probe coverage is 149 of 151 sites (98.7%). The other 2 get an HTTP check only, so their uptime is not measured the same way. This stays listed until coverage reaches 100%.

• Only 52 of the 149 sites with an uptime probe have a full 30 days of probe history, and only those are in the fleet uptime figure. Newer probes show their own shorter window, or nothing until they have 7 days.

• 123 of 4,417 deploy runs failed over the last 8 weeks (2.8%); a further 5 were cancelled, which is not counted as failed.

• 6 Oct 2026 — On 6 October the daily SEO report email silently failed to send; we could not prove the cause because the logs had been overwritten. A catch-up job now re-sends and alerts if it happens again.(written by hand)

• Nightly backups

  12 of 12 servers completed

  last night (8 Oct 2026)

• Monthly restore test

  Passed

  on 6 Oct 2026; one sampled database on one server per month, rotating

• Workstation watchdog

  30 min since last heartbeat

  measured at the time of the snapshot

• Certificate monitoring

  151 tracked

  0 expiring within 14 days; soonest in 31 days

• Uptime probes

  149 of 151 sites

  2 sites have an HTTP check only; 52 with a full 30 days of history

• Servers whose last nightly backup job completed, out of all servers backed up. This is the job’s completion status at server level; on its own it does not prove a backup can be restored.
• Once a month one sampled database on one server is restored, rotating through the servers. This is the most recent test. It does not mean every server was restore-tested.
• A watchdog on the workstation reports in regularly; this is how long ago it last did.
• Certificates on the monitored sites are tracked for expiry.
• Sites with a real uptime probe, out of all monitored sites. The rest get an HTTP check only.

Plus 44 members-only sites, counted in the totals above. And 4 other monitored sites that are not part of this directory. 151 sites are in this snapshot in all.

• October 2026in progress1 snapshots received

1. Caught 1 of 5

   A health reporter that passed its test but could never send

   A health reporter passed its own test mode, but it could never actually send anything: the test skipped the send step, so a green result proved nothing about delivery. The checker ran the real send step against a mock server and showed that it failed.

2. Caught 2 of 5

   A noise filter that would have hidden a real bug

   A noise filter meant to hide scanner traffic would also have hidden a genuine “API key not set” bug. The checker caught it before it shipped.

3. Caught 3 of 5

   A status page that would have named private sites

   A public status page would have listed sites that are deliberately kept private. It was caught before publishing, and names are now opt-in: only sites this directory lists as public can be named.

4. Caught 4 of 5

   A migration that would have silently skipped a table

   The ordering of a database migration would have silently skipped creating a table.

5. Caught 5 of 5

   A results endpoint that rejected every passing result

   A results endpoint rejected every passing result it was sent, so a failure alert could never clear.

The point is not that the first attempt is wrong. It is that nothing is believed until a different party has failed to break it.