Current health and official status
Official status targets 60-second HTTP polling when supported; independent current health targets five-minute slots in Seoul, N. Virginia and Oregon. A check expires at six minutes and becomes unavailable immediately after the next expected slot misses its 60-second grace. Republishing a snapshot never renews measurement age. Official API and consumer components remain separate. Google official machine-readable status is unverified, so it stays unknown with an official page link.
Calendar reports and continuity
Continuous monitoring uses separately approved seven-day leases prepared before expiry, with previous end equal to next start. Raw timestamps stay UTC; Daily, Monday–Sunday Weekly and Monthly reports use America/New_York IANA time, including 23/25-hour days. Incomplete periods cannot become completed indexed reports.
Current state
This is a pre-launch preview with bounded regional observations. Freshness, source and limited history remain explicit; no fixtures or synthetic examples are published as observations.
Timing
TTFT starts at dispatch and ends at the first nonempty visible text delta on a monotonic clock. First event, last text and completion are separate. No-text, partial, timeout and parser outcomes remain explicit.
Tokens and reasoning
Billed and reasoning tokens are not interchangeable with visible tokens. Throughput is withheld because reliable visible counts are not established. An event is not treated as a token.
Baselines
Provider, model revision, region, profile, prompt version, reasoning, token limits and cache policy define a baseline. Health, speed and coding distributions remain distinct.
Reliability
Every attempt preserves its initial outcome. Authentication, quota, credit and monitor failures are separate from provider service errors. Retries never erase first failures. Raw records are immutable.
Coverage and scores
The first day is collecting baseline; days two through seven remain limited history. Freshness and sample counts are checked per comparable group. p95 is withheld below 30 valid timing samples. The composite score is unvalidated and disabled.
History
Best Time requires seven days, distributed hourly evidence and held-out validation. Elapsed time alone is insufficient. Timezone conversion must handle daylight saving and preserve absolute timestamps.
Cost and independence
No visitor action starts a paid call. Benchmarking requires an atomic budget reservation and a readable enabled control gate. Secrets stay outside browser assets. Advertising is disabled.