v0.2.0: ML Pipeline

Corrected 2026-09-23. The uncertainty band was never calibrated and its coverage is far below 95%. The original wording is replaced above; nothing else in this note has changed.

Phase 2: ML Forecasting Pipeline

• LSTM model trained on 20+ years of NOAA Kp index data
• Monte Carlo Dropout spread, not a calibrated interval. ci_lower and ci_upper are
  the mean plus or minus 1.96 standard deviations across the dropout passes, which is
  the model disagreeing with itself and carries no observation noise term. Nothing was
  ever calibrated, so "95% CI" was wrong when this was published, not only in
  hindsight. Coverage measured on 2026-09-22 against the checkpoint then running:
  9.4 / 14.3 / 16.2 / 13.1 percent at 3 / 6 / 12 / 24 hours.
• FastAPI inference server on port 8000
• Historical data download and preprocessing scripts
• 7-feature input: Kp norm, hour sin/cos, month sin/cos, solar cycle sin/cos
• Mean RMSE: 1.04