Once #754 lands, I think we should try and resolve this issue by documenting and standardizing across datasets which of these coordinates are pure "model space" and which are actually lat/lon and then use the new data saved by #754 to map between them rather than the current renaming that's happening to get to spherical_area_weights.
@fomo-bot can you please keep an eye on #754 and put up a PR for this once it and #670 have both landed?
Preserve real LLC coordinates instead of index coordinates
For LLC sources, this renames the integer i/j indices to spatial axes that validate_data() later turns into lon/lat; downstream validation and inference call spherical_area_weights(self.data), so LLC metrics end up using cos(j-index) weights rather than geographic cell locations/areas. Any LLC evaluation or validation run will report area-weighted metrics with the wrong spatial weighting unless the canonicalized dataset carries real LLC center coordinates/areas.
Useful? React with 👍 / 👎.
Originally posted by @chatgpt-codex-connector[bot] in #670
Once #754 lands, I think we should try and resolve this issue by documenting and standardizing across datasets which of these coordinates are pure "model space" and which are actually lat/lon and then use the new data saved by #754 to map between them rather than the current renaming that's happening to get to spherical_area_weights.
@fomo-bot can you please keep an eye on #754 and put up a PR for this once it and #670 have both landed?
Originally posted by @chatgpt-codex-connector[bot] in #670