The Week That Broke the Percentile
GIE opened the 2026 season at 39-4 in games where it had a read on the outcome, a 90.7% accuracy rate that flattered neither the model nor its users. The real story lives in the buckets: four games where the engine carried conviction above 70% wound up in the loss column, a distribution that defied the calibration it has spent two seasons building.
Start with the clean part. In the 90-plus confidence tier, where GIE should hit 90% of the time, it went 21-1. That 95.5% closure rate suggests the engine's highest-conviction calls are operating exactly as designed. The 75-90 bucket held at 90%, one miss in ten, which lands squarely in the expected range. Below 75%, the model compressed into tighter contests — the 60-75 band went 4-1, the 50-60 band 5-1 — and did what it should have done.
The problem was not distribution. It was that four losses all clustered in the 58-84 confidence range, a zone where GIE is supposed to be right more often than wrong but retained enough doubt to flag something amiss.
Rutgers to Massachusetts stands alone. The engine had the Scarlet Knights at 99% to win outright and they lost 37-21. This was not a close game redefined by a late sequence or a turnover chain. GIE saw something that did not materialize — either a projection of Rutgers competence that the roster couldn't sustain, or a Massachusetts team it fundamentally underrated. At 99%, there is no daylight for variance. The model failed to incorporate something real.
The Oklahoma State-Tulsa miss (Oklahoma State favored at 84%, lost 10-24) and the Georgia Tech-Colorado result (Georgia Tech at 71%, lost 13-14) fit a different category: close outcomes that could have landed on either side and happened to break wrong. The model saw them as lean-wins for the favorite. The field voted lean-wins for the underdog. That happens. What matters is whether it happened at the expected rate, and one close miss in the 71-84 range is within tolerance.
California-UCLA (GIE at 58%, lost 45-24) was the only one where the model correctly hedged. A coin-flip game with a slight edge to Cal is, by definition, a place where UCLA wins occasionally. This one just did it emphatically.
The 43-game sample is too small to hang a larger diagnosis on. One week can be variance. But the Rutgers collapse — and it was a collapse, not a shootout — suggests the engine either misread a roster's floor or constructed a projection that didn't account for a second-order factor it usually catches. That's worth monitoring in week two as more data arrives.
The rest of the calls held. GIE did not predict tight games correctly in Week 1 at some elevated rate. It simply missed one outlier badly and will need to diagnose why.