Are the arm gates decided below the startable floor? H-189 asserts 111% of SB's pooled-MAE win over S sits on rows no lineup consumes, the arms are indistinguishable above the floor, and the product (start/sit) prefers S.

Owner lane: allocation

Question

Are the arm gates decided below the startable floor? H-189 asserts 111% of SB's pooled-MAE win over S sits on rows no lineup consumes, the arms are indistinguishable above the floor, and the product (start/sit) prefers S.

Prediction

Frozen in protocol.json before any error was computed: E0 SB beats S pooled, CI excludes 0, point in [+0.01,+0.08]; E1 share_below(F1, SB vs S) in [0.30,0.90] and at or below the C1 row-fraction comparator; E2 all six above-floor Bonferroni intervals span zero; E3 startable-pair hit rates within 1pp of S, intervals spanning zero.

Finding and verdict

MEASURED NOTHING, and the premise is void. H-189 locates a win that does not exist: on the stored held-out ledger SB beats S by +0.007 pts [-0.008,+0.024], so its 111% is a ratio with a nil denominator (F1 share 0.35, 95% [-2.30,+3.57]; a random floor gives [-0.32,+1.48]); the figure reproduces only when the floor is ranked by SB's own projection (1.15). Above the floor all six Bonferroni cells span zero (largest SQ +0.021 [-0.024,+0.074]), and so do the below-floor cells. On the decision the product does NOT prefer S: startable-pair order SB-S +0.30pp [-0.40,+0.92], close pairs +0.41pp [-0.86,+1.61]. The four arms are indistinguishable at every level measured -- pooled, above and below the floor, room total (19.41-19.44 pts a team-game) and start/sit. The shipped default SBQ is lowest on close startable pairs (-0.67pp [-1.80,+0.46] vs S), inside noise.

Reasoning

The evidence-arm gates are read as a ranking of S/B/Q information classes. On the stored ledger the classes do not separate on any endpoint a lineup consumes, so no gate verdict on that table -- for or against B or Q -- is a measurement until an arm contrast clears its own pooled interval.

Next action

A referee could rebuild arm_forecast at HEAD (arms.rebuild_forecast) and rerun this unchanged runner on it; if the pooled contrasts stay nil, the arm gates should be marked as measuring nothing on this population rather than read as S-vs-B verdicts.