Validation Against Trusted Sources and Accepted Methods

Official public point forecasts are scored against, never ingested as truth. Measured station history, the record back to 1950 and macro ocean state are ingested as input. Seismic, solar and lunar series are treated as correlation witnesses only.

Public feeds return past and future days in one array. Read naively, the model days are taken as measurements and the engine anchors to a prediction instead of to reality — a fault that was present here and ran the projection roughly 21°F warm.

Only days at or before the last completed observation day are now accepted as measurement. After the fix the anchored projection tracks the observed record to within about 5°F on the same test point. It is recorded because it is the most likely way this class of engine fools itself.

The dashboard reports the anchor correction in degrees and the number of measured days behind it, so the fit is visible rather than asserted.

Confidence falls as the projection stretches from the last measurement, and the held window is published with the answer so a peak cannot be cherry-picked afterwards.

Ensembled numerical weather prediction is better at short-range general forecasting, at precipitation and at anything severe. This is not a warning service and does not pretend to be one.

Smoothing destroys exactly the structure an operator needs: the late-season warm window, the false peak before a hard frost, the rebound after a thermal reset. Carrying the signals as interfering waves and collapsing them to a single date keeps that structure instead of averaging it away.

Every published seasonal call is back-run against what happened and the result kept, win or lose. Reference cases are locked in the test suite, so a change that moves them fails the build. Back-runs ship with the broadcast data product.