Digital Whiteboard — drag chemistry into a real play

Each card on the right names a chemistry mechanism from the dossier — Third-Man Triangle, Decoy Run, Pin the Fullback, Rest-Defense Anchoring — and shows what happens to P(score) and P(concede) when you apply that mechanism as a counterfactual on the play below. Click a card to see the shift animate on the pitch and the delta curves over time. Reset to return to what actually happened.

Play

Goal-leadup focus (10 s) Linked movement View:

How to use this

Use a Suggested move (panel on the right →). Each card is a real tactical concept — Third-Man Triangle, Decoy Run, Pin — applied to this play. Click a card to apply it: the named players slide to their counterfactual positions on the pitch, the Δ P(score) and Δ P(concede) chips show the model's verdict, and the chart below the pitch shows the per-frame swing. Click the same card again or hit "reset to original" to return to what actually happened.
Or go freestyle. Hit "✋ freestyle: drag any player" once (~5 MB ONNX runtime loads in your browser). Then drag any player dot to a new spot. The model re-runs across the whole focus window and an orange "counterfactual" curve overlays the faded baseline live as you drag — you see both the original trajectory and the one that would have unfolded.
Linked movement (on by default): dragging a player pulls their nearby teammates slightly along to maintain shape. Dragging an LB nudges the CBs; dragging a striker tugs the CMs forward. Same-team only — opponents stay put. Toggle it off to drag a single player in isolation.
Goal-leadup focus (on by default) zooms the scrubber to the 10 s before the goal — the window where a different decision would have mattered. A small ⌃ key decision marker on the scrubber pinpoints the steepest P(score) rise in that window; click it to jump there. Outside the focus window the slider is dimmed.

frame 0/0

How to read this

Δ P(score) is the model's swing in scoring probability over the next 10 s for the team in possession, if you applied that single move. Positive = the move would have helped them score; negative = it would have hurt.
Δ P(concede) is the symmetric defensive swing. A move that's "good for chemistry" tends to push score up *and* concede down — both arrows pointing the same way.
Mechanism tags link the move to the named chemistry concept it tests. The bigger the |Δ net| the more the model thinks that specific geometry was load-bearing in the real frame.
Cards with no mechanism tag are free-form sweeps over the most-attended off-ball players — surfaced when they produce a larger swing than the curated mechanisms, as a sanity check.

Frame-VAEP transformer (P_score / P_concede). 22 players + ball tokens, pre-norm encoder, 2 layers × 4 heads. CPU inference ≈ 0.2 ms/frame. Mechanism dossier at research/site/data/chemistry_concepts.json. Underlying counterfactuals pre-computed by research/scripts/compute_whiteboard_counterfactuals.py.