Bring your own events
Everything on this screen runs on rows you supply. Four frozen reference predictors are fitted to your history, and the departures between them and what actually happened are what gets typed and scored.
Upload a dataset
CSV or JSON. Private to your account; stored in your own workspace bucket.
Uploading stores data against your account, so it needs a sign-in. The synthetic demo scenarios in Signals are open without one.
Sign in to uploadWhat the scoring does with your rows
Stated, not learned.
- 1. Reference fit. Persistence, EWMA(0.3), a local linear trend and a lag-7 seasonal naive each predict every step from the history before it.
- 2. Departure. The residual against the reference median is normalised by a robust scale, so a "large" move is large relative to your own series.
- 3. Disagreement and blindness. Reference spread and the fraction of unusable or irregularly-timed rows are tracked separately, because they mean different things.
- 4. Admission. An explicit policy chooses ignore, remember, open regime, review model or request information — and prints the numbers that decided it.
The trained Zeno Divergent checkpoint is loaded in the browser and scored on your rows after upload. It is a research preview trained on synthetic corpora; treat its verdict as indicative, not ground truth.
Connect a live source
Not built yet — listed so the gap is visible.
- Scheduled pull from an object store or HTTP endpointplanned
- Warehouse table sync (Postgres / BigQuery)planned
- Push endpoint for streaming eventsplanned
Model version in use
Every score on this screen comes from the artifact named here. Nothing is trained in the browser.
Held-out metrics are from the synthetic SurpriseBench protocol (surprisebench/1.0.0) shipped with this release — they are properties of the harness, not evidence about real events.
Your datasets
Most recent first.
Sign in to see datasets stored against your account.