Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.summand.com/llms.txt

Use this file to discover all available pages before exploring further.

Surprise Finder runs after the semantic layer and looks for places where your data deviates from what the model expects. Each finding — a “surprise” — comes with a confidence score, a feature-level explanation, and the rows that drove it.

​
What counts as a surprise

Summand is looking for interpretable deviations, not raw outliers. A surprise typically falls into one of:
  • Trend shifts. A segment’s behavior changed materially over a time window.
  • Unexpected effects. A feature value pushes the prediction in the opposite direction from what the global model suggests.
  • Concentrated anomalies. A small slice of rows accounts for a disproportionate share of the residual.
Each finding is ranked by confidence and includes the temporal interval (where applicable) and the contributing features.

​
What you can do with a surprise

  • Review and accept. Triage findings; mark the ones worth acting on.
  • Edit the explanation. Refine the auto-generated narrative before sharing.
  • Create your own. Add a manual surprise when you’ve already noticed something the model didn’t flag.
  • Share with collaborators. Surprises live with the dataset and inherit its sharing settings.
  • Drop into a report. A Surprises block can be added to any report so the latest findings render on every export.

​
Versioning

Each run of Surprise Finder is a versioned task — you can re-run it after new data arrives, compare results across versions, and pin a specific run to a report.

​
How it relates to the semantic layer

Surprise Finder consumes the EBM model and feature-effect graphs from the semantic layer. If a dataset’s semantic layer hasn’t been computed (or failed), Surprise Finder won’t run. Re-trigger the semantic layer first and Surprise Finder will follow.