Libero reads roles as a Gaussian mixture per position: 27 features (league-adjusted per-90 rates and pitch coordinates), z-scored, projected onto principal components, then a full-covariance mixture with k from 1 to 6 chosen by BIC. Seasons of 900+ minutes, five seasons. Wingers give three roles that survive resampling (adjusted Rand 0.89 against refits on 80% subsamples), central midfielders three (0.74). Attacking midfielders were expected to split at least as finely: the ten, the second striker and the inside creator are distinct ideas.

Sample size is the first suspect: the pooled fits have nearly four times the rows, and a full-covariance component is expensive. A 35-point loss at k=2 does not rule out roles too weakly separated to pay for themselves at this n. The more telling check is what the pooled roles are made of.

Against central midfielders, being an attacking midfielder is itself the axis: the mixture finds one component for them and spreads central midfielders over the other three. Against wingers the label carries no separating information; the three mixed roles are winger styles, and attacking midfielders are assigned to whichever one their numbers resemble. Neither pool finds structure inside the position; the roles it appears to have are its neighbours'.

Both pooled fits are less stable than either neighbour alone (0.64 with central midfielders, 0.70 with wingers), which is what a mixture straddling a continuum looks like. Libero ships the winger pooling, so its four attacking-midfield roles are winger roles that attacking midfielders are measured against. Three of them are the winger model's own roles again, led by the same players (Haji Wright and Kevin Schade, André Geraldes and Robert Brady, Bukayo Saka and Lamine Yamal), and they carry the same names.