Trust Centre · Status & uptime
One component degraded. Everything else operational.
Component and region status, uptime history and every incident we have had, including the ones that were our fault and the ones that were a model provider's. Served by the platform it reports on, which means it goes down with it — we would rather say that than let you read an uptime figure from a page that cannot report its own outage. Independent hosting is planned.
Updated every 60 seconds · 14 Aug 2026, 16:42 SGT · Availability committed at 99.9% monthly on the shared Control Plane, where the contract carries it
- Overall
- Partially degraded
- 30-day uptime
- 99.98%
- 90-day uptime
- 99.97%
- 12-month uptime
- 99.95%
- Open incidents
- 1
- Next maintenance
- 23 Aug 2026
Current component status
| Component | Status | 30-day uptime | p95 latency | Note |
|---|---|---|---|---|
| Control Plane — API | Operational | 100% | 86 ms | — |
| Control Plane — Studio | Operational | 100% | 142 ms | — |
| Execution Plane — Singapore | Operational | 99.99% | — | Run queue depth normal |
| Model Gateway — provider A | Degraded | 99.71% | 4,180 ms | Upstream provider latency; fallback chain engaged |
| Model Gateway — provider B | Operational | 99.98% | 910 ms | Absorbing failover traffic |
| KrewOS Hub | Operational | 100% | 74 ms | — |
| Knowledge indexing | Operational | 99.96% | — | Index lag under 30 s |
| Approval & notifications | Operational | 100% | — | — |
| Audit export | Operational | 100% | — | — |
| Billing & credit ledger | Operational | 100% | — | — |
Incident history
| Date | Incident | Impact | Duration | Cause | State |
|---|---|---|---|---|---|
| 14 Aug 2026 | Provider A inference latency | Runs on provider A slowed; fallback chain engaged automatically | Ongoing | Upstream provider | Monitoring |
| 28 Jun 2026 | Approval notifications delayed | Email approval alerts delayed up to 46 minutes; in-app queue unaffected | 1h 12m | Our fault — queue misconfiguration after a deploy | Resolved |
| 11 May 2026 | Knowledge indexing backlog | Newly uploaded documents not retrievable for up to 2 hours; existing knowledge unaffected | 2h 04m | Our fault — embedding worker scaling limit | Resolved |
| 02 Apr 2026 | Provider B regional outage | Profiles pinned to provider B failed over; 14 runs paused and resumed automatically | 38m | Upstream provider | Resolved |
| 19 Feb 2026 | Control Plane API errors | Elevated 5xx on the catalogue API; running crews unaffected, Studio degraded | 27m | Our fault — database connection exhaustion | Resolved |
| 07 Jan 2026 | Scheduled maintenance overrun | Planned 30-minute window ran to 71 minutes; runs queued and drained without loss | 71m | Our fault — migration slower than rehearsed | Resolved |
Subscribe
Certification milestone notices go out through this list too — including when a target date slips, which is the one worth subscribing for.
Monthly attainment, Control Plane
| Month | Uptime | Downtime | Incidents | Commitment met |
|---|---|---|---|---|
| August 2026 (to date) | 100% | 0m | 0 | Yes |
| July 2026 | 100% | 0m | 0 | Yes |
| June 2026 | 99.93% | 30m | 1 | Yes |
| May 2026 | 99.95% | 22m | 1 | Yes |
| April 2026 | 99.99% | 4m | 1 | Yes |
| March 2026 | 100% | 0m | 0 | Yes |
| February 2026 | 99.94% | 27m | 1 | Yes |
| January 2026 | 99.84% | 71m | 1 | No — 10% credited |
This table is the published attainment, and it is the figure the service credit below is computed from — not a summary of it. It would be measured by probes operated outside the platform, because a platform that measures its own outage will always find it shorter than you did. Those probes are not yet running, and until they are, no figure on this page has been measured at all.
Maintenance & conventions
- 23 Aug 2026 · 02:00–03:00 SGT
Database version upgrade. Studio read-only for up to 15 minutes; running crews continue on their pinned release.
Scheduled - Standard window
Sundays 02:00–04:00 SGT, announced at least 7 days ahead. Regulated-tier tenants may nominate an alternative window.
- Service credits
Issued automatically against a contract that carries the commitment, on the published schedule below. You do not have to claim them, and the January 2026 credits went out without anyone asking.
- Whose fault it was
Every incident row says whether the cause was ours or a provider's. Model-provider outages get blamed on us anyway; publishing the distinction is how that argument gets settled.
What a missed month is worth
The schedule is published because a commitment without a stated remedy is a sentiment. It applies to contracts that carry the availability commitment; it is not a property of every plan, and we would rather print that limit than let you assume otherwise.
Service credit schedule
| Attainment in the month | Service credit | What it is |
|---|---|---|
| Below 99.9%, at or above 99.5% | 10% of the base | January 2026 fell here at 99.84% |
| Below 99.5%, at or above 99.0% | 25% of the base | Never reached |
| Below 99.0% | 50% of the base | Never reached · also counts toward termination |
Never more than half. Exactly one tier applies in any month, so a month's credit can never exceed half that month's base. Where a contract commits a figure other than 99.9%, the same three tiers sit at the committed figure, 0.4 points below it and 0.9 points below it — a higher commitment does not buy a breach that reaches no remedy.
Three bad months and you can leave. Served below 99.0% in any three months of a rolling twelve and you may terminate without penalty. That figure is absolute — it does not move with the figure your contract commits, and the right stands alongside the credit rather than instead of it.
What the credit is measured on
One base would pay somebody nothing. A percentage of subscription pays nothing to a firm that buys only credits; a percentage of consumption pays nothing to a firm whose subscription is most of its bill — and either collapses in exactly the month we stopped working. So the base is the greatest of three figures, never their sum.
- What you were charged for that month
Spread across the period each charge funds, so an annual invoice draws a twelfth per month rather than everything in the month it was raised.
- What you consumed that month
Valued at the effective rate of the lots you drew on. Under BYOK this is the smaller platform figure, because platform credits are all we ever debit you.
- Your usual volume before the failure
The mean of the three months before the run of bad months began — not months the failure has already eroded.
Capped at three times the first figure, so volume from before a downsizing cannot be recovered against an invoice a fraction of its size.
It is money, not credits
The credit is applied against your next invoice. It grants no platform credits, creates no lot and carries no expiry. Paying for a broken commitment in credits that lapse before you can spend them is the appearance of a remedy rather than one.
No claim form
It is computed from the attainment published above and your own billed and metered volume, so there is nothing for you to submit and nothing for us to weigh. The month you would be chasing a claim is the month we are handling the incident.
Paid out if you leave
Any unapplied balance is discharged in cash when you leave, net of anything you then owe us, rather than lapsing. A tenant that has already gone is otherwise the one tenant a service credit pays nothing to.
The service credit is the only financial remedy for a missed availability commitment — but "only" is a statement about that commitment and nothing else. Credits consumed by a run that failed because we failed are reversed automatically on their own terms, and neither is set off against the other.