Players split into ten groups from lowest projection to highest. If the ranges only worked for one kind of player, it would show here.

Where it does not hold. Coverage is not flat. Across the ten groups it runs from 73% to 98%, and the headline 82.6% is an average of that spread rather than a description of any one group. Too narrow for groups 5 to 7 and 9, around 23 to 56 projected points, where only 73% to 78% of players landed inside. Those are middle-of-the-board players, which is where most picks actually go. Too wide for group 1, the lowest projections, at 98%. A range that is almost never wrong is not being careful, it is declining to say anything. Both directions are failures on the standard above, and both are ours to fix.

What we said we could not do. Games played is deliberately not a per-player prediction here. It lost to the league average on a season the model never saw, so it is published as a population expectation with a band. That band covered 85.7% of actuals, and on average we were off by 12.5 games. Most of the big misses above are availability rather than talent, which is the honest shape of this problem.