Event data rarely tags a player's exact role — you get a generic G/D/M/F and have to derive the rest from where the player actually stands on the pitch. Our first version of that logic split centre-backs from full-backs using lateral position: if a defender's average y-coordinate fell inside a symmetric central band, he was a centre-back; outside it, a full-back.

That band assumption breaks the moment you check a real back four. A right-sided centre-back in a back four doesn't sit at the exact midpoint of the pitch — he skews toward his side, same as the left-sided one skews the other way. When we audited the classifier against real players, three regular starting centre-backs failed it outright: Saliba (average y=32.2), Van Dijk (63.4) and Gabriel Magalhães (64.5) all fell just outside the "central" band and were quietly relabelled as full-backs — which meant their reports were built from the wrong module set: carrying and crossing sections instead of build-up and defensive-action sections.

The actual signal was hiding in a different axis. We compared average depth (x-coordinate) across known centre-backs and known full-backs — including at high-possession clubs, where centre-backs push further up than usual. Every genuine centre-back in the sample, from Saliba to Rúben Dias to Pau Cubarsí, stayed inside a tight depth band (roughly 33–39), regardless of which side they skewed toward. Every genuine full-back — Trent Alexander-Arnold, Marc Cucurella, Kieran Trippier — sat both further forward (≥42) and hugging a touchline. Depth, not width, is what actually separates the two roles. The classifier now uses that instead, and the three misclassified centre-backs are centre-backs again.

The bug survived as long as it did because it wasn't a crash — it produced a plausible-looking report for the wrong position. That's the harder class of data bug to catch, and the reason every position family in the model gets checked against real players before it ships.