Update to WeatherXM FACT

{
  "description": "Read published WeatherXM FACT historical forecast accuracy, country cell-win summaries, global model error curves, and historical station-publication coverage. Use for explicit forecast-model benchmark questions, not live weather or current station inventory.",
  "included_files": [],
  "name": "weatherxm-fact",
  "skill_md_contents": "---\nname: weatherxm-fact\ndescription: Read published WeatherXM FACT historical forecast accuracy, country cell-win summaries, global model error curves, and historical station-publication coverage. Use for explicit forecast-model benchmark questions, not live weather or current station inventory.\n---\n\n# WeatherXM FACT\n\nUse only the four tools actually exposed by weatherxm-fact:\n\n- get_best_forecast_model: query a location or station_or_report_id and published historical model performance.\n- check_ground_truth_coverage: query historical archive publication coverage; it does not establish current station presence or absence.\n- get_fact_leaderboard: query global error metrics or country cell-win counts using country, variable, lead_day and exclude_commissioned. Country cell wins are not national matched-sample MAE rankings.\n- get_model_head_to_head: pass a models array (for example IFS025 and GFS05) and variable to compare global source-specific lead-day curves. No model_a, model_b or country parameters are supported by this tool.\n\nVariables: temperature, wind_speed, precipitation, humidity. Where supported, lead_day is a string: all or 0 through 7. Use current tool schemas. Do not invent a tool.\n\nOnly send task-specific benchmark inputs. Do not ask for the user's precise personal location, account credentials or chat history. Prefer an existing public station/report identifier when supplied. Coarse H3 query metadata is retained for demand analytics; do not claim the service stores no query data.\n\nPreserve provenance, data-as-of dates, missing-result states, commissioned-model and source-cohort caveats. Separate reference cells from requested sites. Never treat historical results as live forecasts, guaranteed future skill, or proof that buying a station improves accuracy.\n"
}