concentration.targets measures the top man's share of his POSITION GROUP and alpha_for_subset spends it as his share of the WHOLE SQUAD. Does rebasing the target to the squad denominator improve allocation and fantasy MAE?

Owner lane: decision

Question

concentration.targets measures the top man's share of his POSITION GROUP and alpha_for_subset spends it as his share of the WHOLE SQUAD. Does rebasing the target to the squad denominator improve allocation and fantasy MAE?

Prediction

Registered in the runner docstring at 42af5c13 (2026-09-15T04:42Z), before any run: mechanism squad WR 0.21-0.24, RB 0.47-0.51, target fallback ~0.27 -> <0.05, carry ~0.09 -> <0.03, median solved exponent <1.2; P1 squad-T split delta in [-0.004,-0.001]; P2 squad-C nil |d|<0.001; P3 fantasy MAE nil |d|<0.005; P4 squad_shuffled worse than squad on squad-T. REFUTED if squad WORSE on squad-T split or fantasy MAE with Bonf-3 excluding zero, or better on MAE by >0.005.

Finding and verdict

REFUTED (as a proposed fix). Premise confirmed, consequence reversed: rebasing to the squad denominator is worse on fantasy MAE by 0.016-0.019 pts (Bonf-3 excludes zero in both code states) and on the squad-T split by 0.002-0.004. Do not ship the relabel. What remains open is WHY a wrong-unit target beats the right-unit one -- that is H-150's renormalisation-domain question.

Reasoning

A unit mismatch that the rest of the chain has been fitted around is not an error to delete on its own; the oracle arm shows the target is not being used as a concentration measurement at all.

Next action

H-150 (renormalisation domain) now carries the question: test whether solving alpha over the WR room only (not the whole squad) with the squad-denominated target recovers or beats shipped. H-202 (season vs week estimand) is the time-axis sibling and predicts a level the calibration absorbs. Decision endpoint (flex start/sit vs consensus) not measured; with MAE worse by 0.016 pts it is unlikely to reverse, but it was not tested.