Project TD Tuning — Version 7

# V7 Context V6 rebuilt how the model measures team strength. V7 actually calibrated it. The difference sounds technical, but the results make it concrete. V5 and V6 both opened at 53% in Week 1. V7 opened at 87%. That gap comes from the weights. Earlier versions used weights that were set before several of the model’s inputs were working properly. Better inputs arrived over time — but the model was still interpreting them through an outdated lens. The weights and the inputs weren’t matched. For V7, we fit all of the model’s factors together against the full historical record — over 7,000 games going back to 1999. The result is a set of weights that reflects what actually predicts NFL outcomes, not what we assumed would. Some factors turned out to matter far more than we’d given them credit for. Others that seemed important faded into the background once everything was calibrated together. That’s the point — the data gets to decide. V7 ran twelve weeks. The model has never gone that far. It stopped on a rolling average after back-to-back quieter weeks, not on a single bad week. That’s the version of failure we want — one where the model runs long enough to show you what it actually knows.