What changed
The last rolling 30-day validation window showed the totals model over-projecting the 2–3 goal band by roughly 3.4 percentage points relative to settled outcomes. The under-projection concentrated in low-scoring domestic cup fixtures.
Response
Per-league weights in prediction_weights were re-derived from the settled sample. Mid-band confidence has been reduced on cup fixtures where the historical sample skews lower.
Where you can verify
Every calibration change lands in the public archive. The transparency board on the homepage shows the current 30-day accuracy for each model layer alongside the sample size behind it.