Production routability baseline — 2026-09-16
This note records the first successful production runtime-canary baseline after scheduled maintenance authentication was enabled in Vercel.
Run
- Observed at: 2026-09-16 08:35:19 UTC
- Route:
/api/admin/routability-verify - Vercel result: HTTP 200
- Persisted observations: 42
- Healthy: 40
- Degraded: 0
- Unavailable: 2
- Normal schedule: daily at 05:00 UTC
The successful run verifies the full path: Vercel cron authentication → Bearing route → provider probes → model_routability persistence.
Unavailable observations
| Model | Runtime source | Observation | Consecutive failures |
|---|---|---|---|
devstral |
OpenRouter | explicit HTTP 404 | 1 |
grok-4 |
OpenRouter | explicit HTTP 404 | 1 |
These are observations, not capability judgements. Both catalogue rows were still active at the time of the run and had not yet been independently verified by catalogue maintenance.
Routing policy
Do not turn this first baseline into a hard routing gate.
A future suppression rule should remain deliberately conservative:
- only
unavailablecan suppress;degradednever can; - require repeated explicit unavailability rather than a single probe failure;
- require a recent observation window;
- distinguish stale/incorrect endpoint identifiers from a genuinely removed model;
- keep routability separate from model capability or quality scoring;
- make suppression reversible automatically when a later canary succeeds;
- inspect catalogue/provider evidence before changing canonical model metadata.
A sensible next checkpoint is the next scheduled canary. If the same model is explicitly unavailable again, consecutive_failures should become 2 and the endpoint/catalogue mapping should be investigated before any production hard gate is enabled.
Operational follow-up
- Confirm the next unattended 05:00 UTC run succeeds.
- Review
devstralandgrok-4against current provider/OpenRouter catalogue evidence. - Keep the admin Maintenance surface as the manual recovery path.
- Only wire
model_routabilityinto recommendation/execution filtering after the repeated-failure policy is implemented and tested.