_alpha_for solves the sharpening exponent on the WR room and applies it to the whole squad, so the exponent moves targets out of the TE and RB rooms. Does keeping each room's mass fixed through the sharpening stage improve the split and fantasy MAE?

Owner lane: construction

Question

_alpha_for solves the sharpening exponent on the WR room and applies it to the whole squad, so the exponent moves targets out of the TE and RB rooms. Does keeping each room's mass fixed through the sharpening stage improve the split and fantasy MAE?

Prediction

Registered in the runner docstring, committed 914106d5 (2026-09-23T0719Z) before any run: M WR room gains 0.015-0.040 squad target mass at the sharpening stage, TE+RB lose it; RB gains carry mass 0.005-0.030. P1 rp_same SQ_T better by 0.001-0.004, TE targets line/actual 0.92 -> >0.94. P2 rp_same SQ_C nil |d|<0.001. P3 fantasy MAE nil |d|<0.005 both arms. P4 rp_resolved within 0.001 of rp_same on SQ_T. P5 rp_shuffled worse than rp_same on SQ_T. REFUTED (mechanism) if rp_same SQ_T point >= 0; REFUTED (as a fix) if either arm WORSE on fantasy MAE or SQ_T with Bonf-6 excluding 0. P1 held at ~50%.

Finding and verdict

TESTED -- REFUTED AS A FIX, PREMISE CONFIRMED AT THE STAGE, INERT DOWNSTREAM. Room-preserving sharpening with the shipped exponent is nil on fantasy MAE (-0.002) and SQ_T (-0.0008), both Bonf-6 spanning 0. Its one clearing gain (SQ_C -0.0037) is reproduced by shuffled room labels, so it is not attributable to rooms. Re-solving the exponent inside the room is worse by 0.051 pts (clears). Do not ship either.

Reasoning

The transfer between rooms is created at sharpening and absorbed by calibration; a defect that the next stage cancels is not where the TE/RB split error comes from.

Next action

(1) The SQ_C gain that survives label shuffling says ANY grouped rescale after the power damps carry sharpening in a way that helps -- a carry-side exponent question (the carry exponent 1.20 may be too high), not a domain one. Not filed as a new row: it is one clearing endpoint whose control failed, so it needs its own registered test before it is a hypothesis. (2) The TE under-projection at 0.92 line/actual survives this arm; its source is elsewhere in the chain (calibration or volume), not the sharpening domain. (3) Decision endpoint not measured.