AI availability monitor

Route evidence

Exact decision inputs

A status badge needs evidence

The four fictional routes produce deterministic synthetic checks. Each observation has a start time, completion time, duration and outcome. A check contributes only after its result is reported. Unknown means no result has arrived; insufficient means the rolling window contains fewer checks than the chosen minimum. Stale means the latest observation is too old. Unhealthy means fresh, sufficiently numerous observations exceed the error or latency budget. Healthy evidence still cannot route an embedding request to a chat-only endpoint.

Control the evidence

Incident scenario chooses healthy traffic, regional degradation or a primary outage from model second fifteen up to but excluding forty. Required capability applies a separate compatibility gate. Check interval schedules probes; Probe timeout turns an overlong response into a timeout at the configured deadline. Rolling window includes completions after its lower boundary and through the current instant. Maximum observation age rejects stale telemetry; Minimum completed checks prevents one recent result from implying a well-supported status. Maximum error fraction includes errors and timeouts. Maximum p95 check duration uses nearest-rank p95 across all recent completed checks, including timeout durations. It is not the p95 of successful customer requests.

Try an outage and a blind monitor

Choose Primary outage and seek to second thirty with the default policy. The primary fails its budget, the regional route is too slow, and the compatible backup is selected. Choose Embeddings instead: the dedicated embedding route can qualify, but the chat routes cannot. Restore defaults and enable Lose primary telemetry after second twenty. Seek to forty: the primary becomes stale rather than automatically unhealthy, because missing observations do not prove failed service. Increase the freshness budget and compare how long old evidence remains acceptable. Set the check interval to ten, rolling window to five and minimum checks to two to make sufficient evidence impossible in this fixture.

Inspect and interpret the policy

Select a route card or the inspector menu to read exact timestamps and gate inputs. Play runs sixty modeled seconds at normal speed; Pause freezes the experiment, Step advances one second, Restart rewinds, Speed changes inspection pace and the time slider scrubs the same reproducible observations. Restore defaults resets all policy inputs. The preferred primary is retained when eligible; otherwise the lowest-p95 compatible healthy route wins, with a stable identifier tie-break. At second fifty-nine the recovered primary should qualify again under the default window. No real provider requests, customer traffic, credentials, rate-limit reservations, data residency rules or production failover actions occur. Small-sample p95, fixed capabilities and this immediate switching rule are teaching simplifications, not an operational SLO or a recommended production routing algorithm.