Two labels, never mixed. The pregame role for a week uses only games that kicked off earlier (recent games count more; last season counts half) and the latest depth chart published before kickoff. The game role describes what he did in that game and is used only for history. Projections, matchup notes and prop tests read the pregame role.
Features are shrunk toward the position average when the sample is small: share of targets short over the middle, aDOT, deep share, screens and throws behind the line, yards after catch over expected, contested and end-zone targets, third-down and red-zone share, Next Gen Stats cushion, and for backs carry, early-down, goal-line, target and third-down shares. Slot is a logistic model fitted to depth-chart slot labels; other traits are percentiles among 2023–25 regulars.
Role-vs-scheme tests group players by their pregame role and use per-play coverage (man/zone, shells, pressure) from 2023–25 participation data and FTN blitz charting, with routes measured as on-field dropbacks. Standard errors come from resampling players. Player splits on player pages use only data available before the week and are shrunk toward the role by how much player splits repeat (they mostly don’t).