Dividing by possession made the model worse, and only the second score showed it

The arithmetic that breaks the obvious fix fits on one line. A per-90 count behaves like a + b * possession, and the intercept is real, because a player receives the ball and plays it whether or not his team dominates ^[11]. Divide by possession and you get a/possession + b, where the first term is a hyperbola that falls steeply as possession rises, so the operation does not remove the possession signal, it substitutes one of the opposite sign ^[12]. Division is correct only when the relationship passes through the origin, and according to the author it almost never does ^[13].

Only a scored defect exposes that. Uncorrected, the space read 0.695 against a chance baseline of 0.000 ^[7][10]; dividing brought it to 0.572, and the residual method brought it to 0.450 ^[8][9]. That is 17.7 per cent of the contamination removed by division and 35.3 per cent by residualising ^[1], which puts the ten-minute version at almost exactly half of what was available to remove ^[2]. On its own, 0.572 would have passed as a fix.

Recomputing pays elsewhere too. The multiplier quoted for how much more often a Barcelona player touches the ball is 1.74, while dividing the two possession shares as printed gives 1.76 ^[5][6]. Rounding, not error, but you only see it by redoing the division yourself.

What survives the better operation deserves more attention than it gets. At 0.450 the space still carries 65 per cent of its original contamination ^[3], and the writeup does not account for the remainder. Role coherence moved from 76.3 per cent to 76.8 per cent ^[15], so the correction cost nothing in the property the space was built for. Free and harmless is not the same as sufficient.

The earlier attempt is the same failure with a different denominator. The theory was that a ratio cannot scale with exposure the way a count does, so ratios should be immune by construction ^[19]. Passes per 90 has 24 per cent of its variance explained by which club a player is at; pass completion, the ratio brought in to replace it, has 21 per cent ^[16]. The substitution removed about an eighth of the problem ^[4]. Completion ratios inherit team style because a dominant side offers easier passes, while shape metrics stay clean, with final-third touches at 3.2 per cent and shots per 90 at 3.6 per cent ^[16][17].

The part that carries over to requests per user or errors per deploy ^[20] is the choice of diagnostic. Contamination was first tracked as the share of a player's eight nearest neighbours who were his own club teammates: 3.0 per cent against a 1.3 per cent chance baseline ^[18], about 2.3 times random ^[7], which reads as a blemish not worth an evening. The same vectors, scored on possession instead, read 0.695 against 0.000. One defect, two verdicts, and the normaliser written next depends on which measure happened to exist first.