Are the production line's week-to-week revisions over-sized, are FantasyPros' calibrated, and is any over-sizing an UPDATE defect (last week's line carries information this week's discards) or plain over-dispersion that a recalibration already removes?

Owner lane: q-language

Question

Are the production line's week-to-week revisions over-sized, are FantasyPros' calibrated, and is any over-sizing an UPDATE defect (last week's line carries information this week's discards) or plain over-dispersion that a recalibration already removes?

Prediction

Frozen in protocol.json (sha 230c8322) at 0718Z before any outcome: model b_R in [0.70,0.90] with Bonf-6 upper < 1; FP b_R >= model and >= 1 (consensus underreaction, BGMS AER 2020); kappa model in [0.03,0.25], kappa FP <= 0.03; A1-A0 in [-0.15,-0.05] MAE; A2-A1 in [-0.02,0.00] with Bonf-6 interval SPANNING zero (over-sized updates are mostly over-dispersion, not information in last week's line).

Finding and verdict

PARTIAL (frozen rule: model b_R Bonf-6 upper < 1, A2-A1 spans zero). The line's week-to-week revisions are 15% too large (b_R 0.853 [Bonf-6 0.824, 0.887], 47,223 player-weeks 2019-25), not the filed 22%, and the largest 5% are not worse (0.846, filed 0.65). FantasyPros underreacts (1.051 [1.003, 1.101]). On the board's metric a revision shrink is worth nothing: walk-forward A2-A1 MAE -0.0003 [-0.0008, +0.0003]; the shipped blend's revisions are already calibrated. The over-sizing is real in squared error (A2o-A1o -0.068 MSE [-0.130, -0.010], 5/5 seasons) and nowhere a median line can use.

Reasoning

H-193 asked whether a shrink on updates could be shippable. Two things close that: the MAE gain of a walk-forward partial adjustment over a plain recalibration is -0.0003 pts/row with a Bonf-6 interval of +/-0.0006, and the board already serves a 0.40 blend with an underreacting consensus that removes the overreaction (blend b_R 1.003). What remains is a mean-only effect at WR/TE that matters to anything reading the line as a MEAN (the sim, a prop probability off a distribution centred on the line), not to the board.

Next action

(1) Anything that reads the line as a MEAN -- sim.py, prop probabilities, the kernel -- inherits a 15% over-sized revision at WR/TE that the board's blend hides; a mean-consumer walk-forward (prop Brier or sim totals) is the test that could make this pay, and it belongs to those layers. (2) The blend's calibrated revisions are an argument FOR the H-190 finding (consensus weight 0.95): raising the weight moves the blend toward FantasyPros' underreaction, and at w=0.95 the blend's b_R should rise above 1; a follow-up can read b_R along the weight grid. (3) Opponent decomposition of the revision is not reached.