# GeoXplain

> GeoXplain is an interactive Python-based visualization toolkit for exploring geospatial attribution maps across climate variables, atmospheric pressure layers, and forecast time. It features a model-agnostic viewer for geospatial XAI attributions. It loads raw arrays and compatible result bundles without importing a model. `geoxplain_aurora_adapter` is the first packaged backend and has a separate documentation section for Aurora computation and deployment.

## GeoXplain core

- [Quickstart](getting-started/quickstart.md): load a `.xia.npz` result into Jupyter or a standalone viewer.
- [Visualize results](guides/visualize-results.md): accepted source forms, levels, targets, options, export, and screenshots.
- [Weather overlays](guides/overlays.md): NPZ, NetCDF, array, and compatible backend inputs.
- [Architecture](concepts/architecture.md): the boundary between model-specific computation and model-agnostic presentation.
- [Data model](concepts/data-model.md): result protocols and persisted interchange formats.

## Aurora backend

- [Backend overview](backends/aurora.md): responsibilities and the end-to-end flow.
- [Compute attributions](guides/compute-attributions.md): Aurora targets, methods, batches, and rollouts.
- [Attribution methods](concepts/methods.md): Aurora implementations and cost controls.
- [Remote execution](guides/remote-execution.md): Aurora listener modes and Python dispatch.
- [Adapt to another model](backends/adapt-another-model.md): build a backend for a different model by cloning the adapter.

## Exact interfaces

- [GeoXplain Python API](reference/geoxplain.md)
- [Aurora adapter Python API](reference/aurora-adapter.md)
- [Aurora HTTP API](reference/http-api.md)
- [Generated API inventory](api-reference.json)
- [Consolidated generated reference](llms-full.txt)

## High-level selection rules

- Use `GeoXplainWidget` in Jupyter and `GeoXplain` for scripts or standalone browser viewing.
- Pass `XiaResult` directly to `add_attribution()`; do not repeat its method, target, timestamp, level, normalization, or labels.
- Use exactly one of `level` and `pressure_level` for a single raw attribution grid.
- GeoXplain accepts any method label; its viewer is not limited to the methods implemented by the Aurora backend.
- For the Aurora backend, omit `remote` only when the caller has a visible GPU and the local profile.
- Aurora's `run_rollout()` currently implements only `saliency` and `ig`.