Climate Tuner

Use the project's native Godot runner; do not attempt to call the Kotlin tuner directly with java or Gradle.

Workflow

1. Locate planet/skills/tune_climate.ps1 if the runner, paths, or options may have changed.
2. Confirm the user intent:
   • Use a new batch unless the user asks to continue, then add --resume.
   • Add --apply only when the user explicitly authorizes changing climate_config.json.
3. Read the prior report.json before selecting a new hypothesis. Prioritize, in order:
   • the largest reference-to-simulated confusion;
   • the weakest latitude or elevation band;
   • the averaged Hersfeldt classifier inputs for that confusion; and
   • the baseline deltas from previous candidates (corrected versus regressed tiles and classes).
4. Select a focused parameter set. With a budget of 9 and no interactions, use at most four parameters because the baseline plus both directions needs 1 + 2N evaluations.
   • Temperature: baseTemperature,baseTemperatureInsolationScalar,oceanBaseTemp,oceanInsolationScale
   • Moisture: moistureToMm,startingMoistureMultiplier,moisturePropagationMultiplier,landPrecipitationScalar
   • Circulation/precipitation: choose up to four relevant entries from climate_tuning.json.
5. Add --interactions A+B only for a concrete paired hypothesis. Each pair reserves up to four evaluations, so raise the budget or reduce the number of coordinate parameters accordingly.
6. From the planet directory, run ./skills/tune_climate.ps1 with --parameters, --max-evaluations, and any authorized --resume/--apply flags. Use an explicit output/report directory for experiments that should not overwrite the default run.
7. Read the generated report.json and inspect the baseline/best files under its artifacts directory. Report baseline and best loss, exact-match percentage, changed values, corrected/regressed tiles, leading remaining confusion, evaluation count, and whether climate_config.json changed.

Safety and interpretation

• Each evaluation reloads save/earth.json and simulates all months. Do not claim a run is quick; use the requested budget and wait for completion.
• Lower loss is better. Prefer the report's bestLoss; use match percentage as a companion metric.
• White or transparent pixels in the reference are an unscored ocean/background mask. Never interpret referenceMisses as climate errors or optimize them through climate parameters; use referenceCoveragePercent only as an import/reference alignment diagnostic.
• A tile's reference target is the majority painted class in its fixed geometric footprint. Use confusion, geographic bands, classifier inputs, and candidate deltas to form the next hypothesis.
• A run without --apply must leave climate_config.json unchanged. When --apply is used, the runner writes a backup before its first replacement.
• If no improvement is found, retain the existing configuration and state that outcome plainly.
• For follow-up tuning, inspect the prior report and use --resume only when continuing from its best candidate is appropriate.