Model training
Artifacts trained per UTC day.
Six domains connect observations, contextual evidence, state layers, forecasts, and products—without hiding uncertainty or pretending every source means the same thing.
Domain is one axis. Knowledge stage is another. Keeping them separate makes it clear whether you are looking at a measurement, an inferred state, a forecast, or an application.
Choose a domain to see the measurements that matter, the state it can describe, and the models or contextual products it can honestly support.
These charts use only timestamped rows retained by the platform. A missing series stays visibly unavailable instead of becoming a smooth placeholder.
Artifacts trained per UTC day.
Scored forecast rows and stored target-unit error.
Every transition is explicit. A healthy provider does not automatically make a model eligible, and an available model is not automatically a defensible claim.
High-signal summaries replace the wall of variables. Counts come from the live registry; unknown evidence remains unknown.
Where each domain currently has observed evidence, State Layers, forecasts or scenarios, and product surfaces.
Declared families are not counted as ready unless persisted evidence and claim gates support them.
Representative lineage shows where evidence is complete, partial, or unavailable—from a Watch experience back to the source record.
The bounded public catalog retains complete channel, product, timing, provenance, recipe, model, and stage metadata.
Open observatory catalog →
Open corpus metadata →
Open Global Context inventory →