RAI-Lab
/

Wildfire-FM

+# Integration Handoff Status
+This note answers the implementation questions raised during web/inference
+integration and points to the public files that should be used by downstream
+clients. It is intentionally scoped to the released FireWx-FM repository.
+## What is fully specified now
+| Question | Status | Where to look |
+|---|---|---|
+| Which input channel maps to which dataset? | Specified in human-readable and machine-readable form. Channel order matters. | `data_sources/DATA_SOURCES.md`, `models/wildfire_fm/input_channels.json` |
+| Is value normalization applied? | Released California checkpoints use native source units with no mean/std or min/max normalization. The training pipeline now supports optional train-split z-score normalization for retraining. | `models/wildfire_fm/input_channels.json`, `training/train_cold_tiled_mainline.py` |
+| Which CAPE variable is used? | HRRR surface instantaneous CAPE, selected with `typeOfLevel=surface` and `stepType=instant`; not most-unstable or layer CAPE. | `models/wildfire_fm/input_channels.json`, `data_sources/DATA_SOURCES.md` |
+| What do the validity masks mean? | `firewx_valid` is a cache-level dynamic input-presence mask and is 1.0 everywhere in the released California cache. `static_valid` is the fraction of four static layers valid after reprojection. | `models/wildfire_fm/README.md`, `models/wildfire_fm/input_channels.json` |
+| How are static 30 m rasters resampled to the 5 km grid? | LANDFIRE fuel and canopy use nearest-neighbor resampling; WRC housing density and LandScan population use bilinear resampling. | `training/configs/stage1_cache_regional_hrrr_ca_5km_l12_template.json`, `training/build_phase1_cache_regional_hrrr.py` |
+| Can outputs be shown at county or sub-county granularity? | Yes, by aggregating native 5 km grid probabilities after inference. This does not require changing the checkpoint. | `spatial_serving/README.md`, `spatial_serving/grid_to_polygons.py` |
+| Can HRRR be downloaded automatically? | Yes. A direct NOAA HRRR public-archive downloader is included and has been smoke-tested with an `.idx` sidecar download. | `data_downloader/hrrr_downloader.py`, `data_downloader/README.md` |
+| Is the released checkpoint nationwide-trained? | No. The public checkpoint is California regional. CONUS cache and training templates are included, but no CONUS checkpoint is released yet. | `models/wildfire_fm/training_scope.json`, `training/NATIONWIDE_RETRAINING.md` |
+## Current model-serving boundary
+The released checkpoints are the paper-aligned California regional checkpoints.
+They are valid for reproducing the reported regional model behavior and for
+inspecting the tensor contract. They should not be described as nationwide
+weights.
+For gridded inference, use full-map inference when possible. If tiled inference
+is required, use overlapping tiles with halo cropping or blending. Independent
+non-overlapping 32-by-32 tiles are not the intended serving mode because they can
+introduce tile-position and edge artifacts. The helper
+`models/wildfire_fm/tiled_inference.py` implements the overlap/halo pattern.
+For CONUS or client-facing deployment, the current path is:
+1. Build a local CONUS cache from provider-hosted HRRR, FIRMS, LANDFIRE, WRC,
+   and LandScan inputs.
+2. Retrain using the CONUS template with random-containing positive tile
+   placement and optional continuous-channel z-score normalization.
+3. Serve native-grid probabilities with full-map or overlap/halo inference.
+4. Aggregate native-grid probabilities to county or sub-county polygons using
+   the spatial serving adapter.
+## What is not being uploaded as solved
+A preliminary CONUS normalized/jittered retraining run completed, but seed-level
+behavior was not stable enough to release as a corrected nationwide checkpoint.
+For that reason, the repository exposes the corrected training and serving
+hooks, but it does not publish a new CONUS checkpoint or claim that the CONUS
+serving behavior is fully resolved.