concentration.targets is the SEASON leader's share of SEASON volume; share._alpha_for spends it per team-WEEK. Is the weekly leader share materially higher, and does re-keying the target to the per-week estimand move fantasy MAE / split error?

Owner lane: q-language

Question

concentration.targets is the SEASON leader's share of SEASON volume; share._alpha_for spends it per team-WEEK. Is the weekly leader share materially higher, and does re-keying the target to the per-week estimand move fantasy MAE / split error?

Prediction

Registered in the runner docstring at 6d0ac40f (2026-09-23T0639Z) before any run: A1 WR squad gap +0.03..+0.06; A2 RB squad gap +0.05..+0.10; REFUTED if either <0.01. A3 season leader's mean weekly share within 0.01 of the season figure. B1 squad_week-squad fantasy |d|<0.005; B2 squad-T |d|<0.002; B3 squad-C |d|<0.002; B4 group_week-shipped nil, lean WORSE (rival: weekly max over-sharpens); B5 shuffled worse than squad_week on squad-T.

Finding and verdict

TESTED. Premise CONFIRMED (weekly leader share exceeds the season figure by +0.048 WR targets, +0.082 RB carries on the squad denominator; refutation clause not triggered), but ~40-58% of the gap is the season leader's absences, not rotation (A3 failed). Consequence: re-keying to the week estimand beats H-109's squad arm by 0.011 pts MAE (Bonf-6 excludes 0) -- but the shuffled control keeps that gain, so it is a sharpening LEVEL, not team information. Against shipped nothing moves fantasy MAE beyond the interval; the squad-CARRY split improves 0.002-0.003 in every week arm (group_week -0.0031, Bonf-6 excludes 0). Post-hoc; the carry-only re-key is filed as a new row for held-out confirmation on 2019-22.

Reasoning

The nil prediction assumed fit_calibration absorbs the level. It does not fully: the solve's exponent feeds allocation directly, and the per-week target sets it nearer where shipped's (wrong-unit) target already put it. That the wrong-unit target and the right-unit week target land at similar MAE while the right-unit season target loses supports 'target = sharpening knob'.

Next action

(1) carry-only week re-key vs shipped on held-out 2019-2022 (new row). (2) the absence component (A3) says the season estimand is diluted by the leader's missed weeks -- a played-weeks-only season estimand is the cleaner rival to the week max, untested. Decision endpoint not measured.