Every forecast Hermes publishes is checked against history before you are asked to trust it. This page runs a walk-forward backtest, reports calibration, shows the reporting-lag correction the corpus implies, and measures how many spatial clusters DBSCAN would find on shuffled noise.
If the model is honest, ~50% of actual months fall inside the dark band and ~95% inside the light band. Large deviations mean the intervals are either over-confident (too narrow) or sand-bagged (too wide).
Forecasts describe report-volume behaviour (a function of observers, media, and seasonality), not UAP activity. Calibration metrics are walk-forward, mimicking real-time conditions. Spatial shuffling within the observed bounding box is a conservative null; population density inflates cluster counts in noise, so signal-to-noise above ~2.0 is where we treat a cluster as robust.