A barometer, not an oracle.

What counts as an observation?

Each usable record should identify the resource, provider, runtime, task context, source and time. A catalog listing shows that a resource was listed; it does not establish current access or task performance.

A fresh catalog snapshot describes the last verified listing. It does not mean every scheduled refresh succeeded; recent failed attempts and the retained snapshot are shown separately.

When do we compare?

For each task class, a resource needs at least 20 attempts, 10 completed runs and 5 verifier-checked outputs before ASN reports a measured score. Zero verified successes is the minimum; failed completed attempts remain in outcome denominators and unresolved attempts are shown separately.

A comparison group needs at least two fresh resources with qualifying evidence for the same task class. A matched-protocol comparison also requires identical suite hash, task IDs, prompt variants and verifier identity. Cross-provider confirmations are reported separately; only confirmed cross-provider outcomes count as production verified successes.

The index is refreshed on successful five-minute production cycles. If a cycle fails, the last verified snapshot remains available and its age and freshness are shown. Local model studies and third-party leaderboards remain separate from production index rankings.

Observed model output and outcome verification are separate signals. A capability claim is not proof of measured performance.

What are the limits?

ASN is not a complete census or a universal intelligence score. A registry entry does not establish that a tool is reachable or useful; a model listing does not prove it can solve a task. The public site may show an unknown state until production observations are enabled and integrity checked.

The scheduled cycle can propose questions, but external execution, spending and production changes remain under operator control.

How quickly does the observatory appear?

Controlled cold-browser lab runs are reported separately from visitor telemetry. Visitor timing summaries appear only after at least 20 observations in a device profile; they cover a rolling seven days, use coarse timing buckets and report percentiles as upper bounds. We track the first Earth frame, satellite visibility, full scene, and Deep Field NASA-map decoding and first visible frame separately. With 10% sampling, the 20-observation floor corresponds to about 200 eligible loads in one device profile on average. Counts below the threshold are hidden, so an empty summary cannot distinguish a small sample from no sample. These self-reported measurements describe site operations, not AI capability.

Which parts of the observatory are being opened?

Counts show sampled page loads grouped into broad page categories over a rolling 30 days. They are not unique visitors, sessions or returning-visitor counts, and ASN does not extrapolate them to total traffic. Raw events, page paths and visitor identifiers are not stored; Do Not Track and Global Privacy Control are honored. This privacy-preserving aggregate can show which broad sections are read, but cannot tell whether the same person returns. To follow changes over time without an account, use the dated receipt history or subscribe to the public Atom feed.