The system audits itself in the open
Calibration, coverage, and drift are recomputed on every build from real one-step forecasts against what the feeds then reported, and shown publicly — not as a footnote, but as a feature. A model that hides its own performance is asking for blind trust. TerraGuard asks for the opposite.
Calibration diagram
Nominal coverage vs. what actually happened on held-out data. Points on the diagonal mean the intervals are honest, neither over- nor under-confident.
Coverage
The share of held-out values that fell inside the 90% interval. A coverage far below 90% is the model over-promising; the Trust Core widens or abstains when it sees this.
Drift (PSI)
Population stability of the day-to-day change distribution vs. training. Below 0.1 is stable; rising drift is an early warning that a model may need recalibration.