Caleb Williams

RAISED 22 Aug. The 16 Aug override that stood here CUT him, and both halves of its premise are now false. It said 'Board QB5 on the 12th-best rate and 15.3 games ... he gains +7 ranks from a games spread' -- on the current board his NATIVE line is QB13 on 17.59/g, which is the 13th-best rate, so his rate rank and his VOR rank are identical and there is no games-driven inflation left to correct. The board moved under the override; the override did not notice. Its own caveat also flagged that it argued GAMES and, he being healthy, spent itself on RATE -- the exact defect the `games` column shipped on 17 Aug to fix, and Caleb and Bo Nix were the two rows that cleanup missed. 283.5 lands at 18.06/g over 15.7 games. That sits between the model's native 17.59 and his OWN 2025 rate of 18.53 over a full 17 games, so it concedes part of the regression rather than all of it. NOTE THE REALISED SPLIT, which is not the one intended when the total was chosen: 15.7 games came back rather than the 15.3 the native board carried, so the rate landed under 18.53. The total is the claim and 18.06/g is what it buys. No games value is written because the games are not in dispute: 15.3 against a 14.2 median for the top 24 QBs, with Allen at 15.4 and Maye 15.0, is unremarkable. So this correctly buys RATE. STATED AS A VIEW, with the measurement against it. The model reaches 17.59 by weighting his rookie 15.56/g at 80% of his 2025 18.53/g, and USAGE_DECAY_BY_POS['QB']=0.80 was re-tested 22 Aug on 380 consecutive QB season-pairs 2012-2025: 0.8 is optimal pooled (MAE 2.776 against 2.778 at no decay) and for third-year QBs the best alternative saves 0.015 points a game. So the decay is not a defect and this is a football claim, not a model fix -- that a 24-year-old in year three under a new offensive-minded head coach repeats his own last season rather than regressing toward his rookie year. The model cannot see that: playcaller.apply moves team VOLUME at PC_VOLUME_STRENGTH 0.22 and has no per-play efficiency term. Lands QB9. Market ADP has him 7th off the board (71.0), which would need 19.0/g -- a further leap, deliberately not taken.

This is a view, not a correction — the model said something different and was overruled. See how it works.

Pace extrapolates the selected weeks over 17 games. It is what that stretch was worth, not a projection — no availability, matchup or regression adjustment is applied.