The member portal, the employer confirmation site, the caseworker dashboard, and the service behind them, checked every 5 minutes.
Last checked 11/21/2026, 3:35 AM CT. Uptime over the last 24 hours: 100.00%.
The last 24 hours
The last 90 days
Uptime by calendar month
| November 2026 (so far) | no checks | No checks recorded | |
| October 2026 | 99.168% | 50 min down | Below the 99.9% target |
Every 5 minutes the checks above run; a month's figure is the share of those checks where every part of the service answered. The target is 99.9% per calendar month.
Services
| verifyhours.com | Normal | 170 ms | |
| confirm.verifyhours.com | Normal | 171 ms | |
| dashboard.verifyhours.com | Normal | 172 ms | |
| API and database | Normal | 230 ms | |
| Outreach job runner | Normal | Last run null minutes ago (never). |
Incidents
10/02/2026 to 10/20/2026: false "down" pages from the status page itself (no service interruption)
Verify Hours went live on 09/23/2026. No service interruption has been recorded since. What happened in October was the
status page misreading two things and paging the operator for them.
- The hourly job runner was judged by the 5-minute rule. The idle switch (10/02/2026) runs the outreach job once an
hour while no production tenant exists. At the top of every hour the API reported the last run as 60 minutes old
without saying it was idle, so the status page applied its 15-minute daytime limit and paged "Outreach job runner" every
daytime hour from 10/02 to 10/20, and counted every hourly sample as downtime. The uptime figure for October read
50.429% (1445 minutes "down") on 10/20/2026; the hourly samples from 10/02 to 10/20 were corrected to 0 failures the
same day (old counts, day: checks/failures: 10/02 22/22, 10/05 22/22, 10/06 20/20, 10/07 2/2, 10/08 23/23, 10/09 30/20,
10/12 12/12, 10/13 13/13, 10/14 23/23, 10/15 24/24, 10/16 24/24, 10/17 24/24, 10/18 24/24, 10/19 24/23, 10/20 13/13).
- One failed retention run stayed on the board for 4 days. On 10/16/2026 at 4:00 AM CT the nightly proof-file
retention job hit a transient database error on the wi-test tenant and recorded a failed signal. The job only rewrote
its signal on nights with files to delete, so the failure was never replaced, and from 10/16 5:00 AM CT the status
page paged "Proof-file retention job" at the top of every hour, day and night.
- Each page was followed by a recovery ten minutes later. Between hourly ticks the API reports no operations
signals at all, which the pager read as all green: a DOWN text and email at :00, a recovered text and email at :10,
every hour (48 of each a day from 10/16).
Fixed 10/20/2026: the hourly heartbeat carries the idle flag and is judged by a 65-minute limit; the retention job writes
its signal every night and the signal is classed as daily (repeats every 2 hours); a signal the pager could not read is
carried forward, not read as recovered. The 10/16 retention row was cleared by hand the same day.
2026-10-22 19:00Z and 2026-10-23 06:00Z: "API and database" pages from the hourly wake outrunning the 2-second health budget
Two page cycles: 10/22 19:00:42Z (recovered 19:10:40Z) and 10/23 06:00:44Z. Each time GET /health answered 503
"database unreachable" in ~2.2 s while every site check and operations signal was green (tenant databases 2 of 2
reachable at their hourly check).
Cause, read off the code and the probe records: the idle switch runs the health route's database leg only at the top of
the hour, so that leg almost always lands on the control database waking from scale-to-zero, and the leg gave one
heavyweight query (listTenants, which loads every tenant's policy) a single 2-second budget with no retry. A wake
running past 2 s paged the operator although the database was healthy; the next probes read the idle answer and sent the
recovery ten minutes later.
Fixed 2026-10-23: the health leg is one cheap registry query with a 3-second budget and one retry (controlHealth,
apps/api/src/tenants.ts, unit-tested in apps/api/test/health.test.ts); worst case stays under the status Worker's
8-second probe timeout, and a genuine outage still fails both attempts and answers 503.
2026-10-23 16:00Z: one more wake page, and the architectural resolution — the pause and the activity switch
The 10/22-10/23 fix above (budget + retry) reduced the false pages but did not eliminate them: at 16:00:53Z one more
"API and database" page fired (recovery 16:05:56Z) — the hourly wake from scale-to-zero occasionally outran even the
widened budgets, and the confirm-before-alerting retry landed on the same cold wake.
Resolution, same day (owner decisions, docs/decisions.md 10/23/2026): the premise was wrong, not the budgets. A paused
product's databases should never be woken by their own monitoring. While the system is asleep (the activity switch:
no real API request and no claimed work for 48 hours), no cron runs, no health route opens a database, and the status
page's uptime roll-up writes nothing; the three public-site checks stay at 5 minutes. The top-of-the-hour wake — and
with it this entire class of page — no longer exists while paused. While awake (real use), full monitoring runs with
the 10/22-10/23 hardening: a cheap health query with budget and retry, confirm-before-alerting on every probe, and a
failed database reading that latches until a good reading replaces it.
2026-10-24 16:50Z to 2026-10-26 15:55Z: a two-day repage cycle about jobs that were off on purpose
One page cycle opened 10/24 16:50Z and repaged every 2 hours until 10/26 15:55Z — about the "Daily end-to-end test"
and "Nightly backups" signals, whose GitHub workflows were deliberately disabled with the 10/23 pause. Two compounding
defects, both in the 10/23 activity-switch design: (1) the stamp counted any successful API request, so one forgotten
portal tab's background polling held the system awake from 10/24 onward; (2) once awake, the staleness rules read the
deliberately-disabled jobs' old rows as outages and repaged on the daily cadence.
Fixed 2026-10-26: the stamp counts only deliberate acts (non-GET API requests, excluding provider callbacks; the job
runner never stamps), and disabling a scheduled job is now a pair of acts — the workflow disable plus clearing its
stored signals (apps/api/scripts/clear-disabled-job-signals.ts) — so off-on-purpose reads as unmonitored while a
silently-broken job at production still goes stale and pages. The stale rows were cleared, the pager observed green
and sent its recovery at 15:55Z, and the leftover activity stamp was removed so the system sleeps again.
This page runs separately from the product, so it stays reachable when the product is not.