Are the SERVED room shares over-dispersed relative to their own skill -- and is the registered HHI endpoint able to see that?

Owner lane: allocation

Question

Are the SERVED room shares over-dispersed relative to their own skill -- and is the registered HHI endpoint able to see that?

Prediction

Frozen in research/scientist/run/h110_served_overdispersion.prereg.txt at 2026-09-27T01:17Z before any number: E1 WR 2025 served excess +0.002 to +0.008 (H-110's >+0.010 bar FAILS on magnitude); E3 S-T positive for WR and TE, WR interval excluding 0, RB still flattened; E4 beta_S(WR) 0.85-0.97, beta_T(WR) 0.95-1.05, beta(RB) > 1 on both rows. Addendum (post-hoc, prediction written in the runner docstring before it ran): WR/TE served shrink by prior-season beta reduces room half-L1 by 0.002-0.006; RB expansion also reduces it; trained moves less than served.

Finding and verdict

TESTED. H-110 as registered is REFUTED: on the served row the WR room is as concentrated as a draw from its own p, 2025 +0.0019 [-0.0066, +0.0101], pooled 2023-25 -0.0009 [-0.0056, +0.0038]. But that endpoint cannot test the premise in H-110's title: HHI is invariant to WHICH man holds each share, and scoring it on projections shuffled within the room returns the identical +0.0019. The endpoint that does see skill -- the within-room calibration slope, the standard 'predictions too extreme' diagnostic -- says the premise HOLDS for receivers on the served row and is REVERSED for backs: served beta WR 0.861 [0.838, 0.883], TE 0.780 [0.750, 0.810]; trained 0.956 / 0.969; RB 1.098 served, 1.169 trained (flattened, H-029's direction). So a served receiver share moves the actual share only 0.78-0.86 of the way; the dispersion MATCHES reality's but is only partly placed on the right man. The sharpening is the served INPUT (one-game-stale history + CP-H49 last-3-played shares), not the chain: apply_shares-stage beta equals the final-line beta for TE (0.969 / 0.782 vs 0.969 / 0.780) and the lines_for reconciliation slightly FLATTENS WR (0.927 -> 0.956 trained, 0.841 -> 0.861 served). 'Three sharpening stages' is not what the chain does. The slope is stable across seasons (served WR 0.864, 0.843; TE 0.809, 0.764; RB 1.107, 1.074) and walk-forward shrinkage by last season's beta cuts share MSE per man on 2024-25 (TE -0.0056 [Bonf-3 -0.0078, -0.0036], WR -0.0005 [-0.0007, -0.0003], RB expansion -0.0009) but WORSENS H-029's room half-L1 (TE +0.0124 [+0.0071, +0.0176], WR +0.0045 [+0.0025, +0.0064]); RB improves on both (-0.0058). Whether a shrinkage stage is worth building therefore depends on the consumer's loss, and neither endpoint here is the fantasy line or the start/sit call. Separate: at this release RB rooms are flattened -0.096 to -0.101 in HHI excess, 2.7x H-029's -0.0375 at the 09-13 release (TE +0.029 and WR +0.000 replicate H-029 exactly).

Reasoning

H-110 was rewritten to an HHI-vs-own-draws endpoint after H-029. That endpoint measures dispersion against REALITY. 'Over-dispersed relative to skill' is a claim about dispersion against the model's correlation with reality, and only a calibration slope (or equivalent) measures it; a model whose spread matches reality's but whose correlation is below 1 is over-dispersed by construction under squared loss.

Next action

(1) The deciding test is through the line: shrink served WR/TE shares by prior-season beta INSIDE the production chain (before lines_for reconciliation) and score fantasy MAE and the flex start/sit call on SEL_S (H-089's runner) -- the room statistics disagree in sign, so only the consumer's loss settles it. (2) The served beta gap (0.86 vs 0.96 WR, 0.78 vs 0.97 TE) is H-222's repair target measured on a new endpoint: a served row that includes the last game should close it. (3) The RB HHI flattening deepened from -0.037 (09-13) to -0.10 at b6b4ce0d; find which commit (model HHI 0.549 -> 0.489).