EngineSep 24, 202685 replayed slates

What each engine channel earns

We switched off, or halved, every hand-set piece of the engine and asked whether the numbers got better or worse on games the engine was never tuned on.

Why we ran it

The engine has several hand-set channels layered on top of its core rates: the pitch-mix matchup, the park, the bullpen, platoon and the home-plate umpire. Each one sounds reasonable. The question is whether each one earns its place on games it has never seen.

What we found

  • The pitch-mix channel double counts the hitter. It scores a hitter’s results against each of tonight’s starter’s pitches. That mostly re-measures how good the hitter is overall (it correlates .61 with his talent) and the platoon edge, which the engine already prices. Removing the part talent and platoon explain was the only change that improved 1+ hit and home runs in both test windows.
  • Strikeouts: no add-on moves the number measurably. The strikeout model’s strength lives in its core rates, not in these channels.
  • Park, bullpen quality, platoon and umpire: nothing clears zero. The umpire table touched only 83 of 1,145 replayed games.

Two inputs that did nothing

The audit also found two inputs the engine was reading that nothing ever wrote: weather runs and an arsenal-confidence scale. No served number had ever used either. They are now named constants, so the code says plainly what the engine does, and switching either on will need its own held-out test.

What it means

The engine is not broken, but most of its hand-set pieces are too small to measure. That points the work toward what does move numbers: the core rates, the starting point for each player, and how long each starter lasts.