Constant acceleration was checked after the primary results, with a fixed formula and no model change. Its RMSE was 0.347 yards versus 0.303 for the joint model: a 0.044-yard absolute advantage. The estimated squared-error gain was 23.9%, with a 95% player interval of 6.9%–36.8%. This is a post-hoc sensitivity check.

Only two tight ends contributed 20 evaluation plays; no dependable TE-wide claim.

One training group failed the registered gauntlet repeatability check. Added motion-scalar, learned Combine-embedding and shuffled-profile forecast arms were withheld. Their NFL value was not tested.

The fitted baselines include conventional tests and recording controls, including drill duration and path length. Their contribution was not isolated. Trees use history summaries; the joint model uses the full history sequence. There is no matched independent-player neural ablation, so the gain cannot be credited specifically to attention or interactions.

Xs are offensive players, Os are defenders and Q marks the quarterback. The selected receiver is emphasized. Only continuously observed actors are drawn; the source has partial coverage/QB scenes, not all 22 players. There is no invented ball or blocking.

Both panels share 0.9 seconds of recorded history. At 0 s the right panel becomes a frozen forecast; the left stays observed. The actual throw occurs at +1.1 s, just after this clip ends. Zero is the forecast origin, not the snap.

A geometric opening does not establish catchability or a coverage assignment. A nearest-defender change does not prove a handoff.

Four outer receiver folds and a fixed held-game set. A held receiver is excluded from training even when appearing as another actor. Two inner training-game folds select each model’s settings; all choices freeze before evaluation.

Forty saved fits: bounded nonlinear trees and a compact temporal GRU plus player-attention model. Each predicts without later paths, future target labels, ball landing or output length. Small coordinate perturbations changed the geometric opening classification in about 2% of checks.

The original single split failed support; it was replaced before any model scores by the declared nested evaluation. Both 2023 and the broader project have prior exposure. Player/game intervals hold fitted models fixed.

The failure case is deliberately error-selected. Examples are individual plays, not player grades. Tight-end examples do not establish generalization to tight ends.