Passes (passes, handoffs, inbounds) and drives are checked against the NBA's own per-game counts: about 98% / 93% and 92% / 91% (found / right). A drive starts 20+ ft from the rim, gets within 18 ft and isn't a fast break.
Ball screens: an on-ball screen, how the defense played it (over, under, switch, trap) and what the screener did after (roll, pop, short roll). Only screens that led to baskets can be checked against the NBA (about 80%).
Shot types: each shot's play type (spot up, P&R ball handler / roll man, handoff, off-ball screen, cut, iso, post up, transition, putback) and contest (open / light / heavy / blocked), in Shax's vocabulary: agreement with his Knicks charting (9 games) ~64% for play type, ~62% for contest (open vs contested ~73%); the contest is the NBA feed's closest defender joint to the ball.
Help (experimental): a defender who leaves his man toward a driver or a roller (HELP: gets to him or in front of him; TAG: meets the roller; STUNT: a fake at him), the dashed line pointing to the man he left. Checked so far against the analysts' posts. Ball-screen coverage uses Shax's vocabulary: the handler's defender (over / under) | the big (drop / level / hedge-blitz / switch), tuned on his Knicks charting (9 games: over / under 86%, the big 67%). Topside: on a side screen the handler's defender set up 2+ ft on the top side, the ICE setup; only about 1 in 4 is an ICE in his charting (a defender who then goes over looks the same beforehand), so it's a flag, not a call.
Off-ball screens (experimental): two teammates without the ball meeting, the slower one the screener. Set: the screener stood within 3 ft of the line the cutter's defender ran and that defender got within 4 ft of him (a legal screen puts the body in the defender's way). Unconfirmed: either failed, so the tracking doesn't confirm the screener was in the way; some are still real screens (a defender who goes wide around one never reaches it), others are teammates passing near each other. About 63 a team-game, 41 of them set. Against the NBA's screen assists, keeping only set ones is right a bit more often (81% vs 80%) but misses more (76% vs 78% found): the NBA credits only screens that lead straight to a basket and no public count covers the rest, so ✓ / ✗ reviews here are the real check.
Handoffs: a dribble handoff when the giver had nearly stopped (3 ft/s or less) and then screened the receiver's defender (within 4 ft in the next 1.2 s), else a plain handoff.
Experimental layers aren't checked against anything yet: named sets come from matching templates. Reviewing them with ✓ / ✗ is how they get checked.
Players: the strip under the court shows who's on the floor (headshot, jersey number, name). Hover a headshot or a dot to find that player and their path; click or tap to keep them highlighted (several at once is fine). Names labels every dot.
3D is the Discord /replay3d view: pins as tall as the listed height that rise on jumps; contested shots show the defender's measured hand height in purple.
All games builds a queue across every tracked game (newest first). For detected actions each clip centers on one detection, and Y / N marks it and moves on: the fastest way to review a layer.
Credits: the shot play types, contest levels and ball-screen coverage words follow
Shax (@ShaxNBA)'s charting vocabulary,
and the agreement numbers above are measured against his hand-charted Knicks games.