0118Z (decision) found the season-to-date target-share term (A1 = walk-forward OLS fantasy ~ 1 + A0 + f + RB + WR per week band) improves the same-position start/sit ORDER of the RAW score_arm line (flex_sym 2023-25 +0.46 pp, Bonf-2 lower bound -0.000). The board does not serve the raw line: forecast.build multiplies each current-week line by consensus.factor_for(ours, fpapi, pos, level_for(rows being blended)) at WEIGHT 0.40, MAX_MOVE 0.60. Does the gain survive in the order a manager actually reads?

Owner lane: serving

Question

0118Z (decision) found the season-to-date target-share term (A1 = walk-forward OLS fantasy ~ 1 + A0 + f + RB + WR per week band) improves the same-position start/sit ORDER of the RAW score_arm line (flex_sym 2023-25 +0.46 pp, Bonf-2 lower bound -0.000). The board does not serve the raw line: forecast.build multiplies each current-week line by consensus.factor_for(ours, fpapi, pos, level_for(rows being blended)) at WEIGHT 0.40, MAX_MOVE 0.60. Does the gain survive in the order a manager actually reads?

Prediction

B1 - B0 flex_sym accuracy positive with point in [+0.0005, +0.0040] and ratio to raw in [0.3, 0.9] (blend puts 0.6 weight on ours for covered men); the Bonf-2 interval CONTAINS 0. B0r - B0 is NOT exactly 0 (the blend is not scale-invariant because lvl and the clamp read our level) but |acc| < 0.002. B1 - B0r within 0.001 of B1 - B0. Shuffled nil, planted detected. Uncovered-pair gain larger than both-covered. Refutation: B1 - B0 interval excluding 0 on either side, or ratio outside [0.3, 0.9].

Finding and verdict

REFINED (serving). The 0118Z order gain reaches the served product, but a quarter of the served B1-B0 is the blend re-weighting itself, not the share term. f-specific served gain: +0.29 pp flex_sym (edge of zero, as raw), +0.21 pp close 2023-25 (excludes 0). New: the served order is recalibration-sensitive because 10.2% of served lines are clamped at MAX_MOVE and the realised consensus weight is 0.36 not 0.40. H-021 end to end on MAE still MEASURED NOTHING.

Reasoning

The lane owns every number a person reads. 0118Z measured the raw order; a manager reads the blended one. The two differ in a way that changes how any line-change gate on the order must be scored.

Next action

(1) Replicate on the full rostered board per week (level_for on every rostered row, incl. non-players) to fix the denominator, and with conservation, via forecast.build on historical weeks if it can be pointed at 2023-25. (2) Prospective: 2026 served board weeks 5+ (after 10-13) scored flex_sym with and without the term. (3) A frozen test of whether the clamp at 0.60 costs order on genuine rows (clamped rows only), which is a served-transform question this lane owns.