Six stages, one broadcast feed, and no target-match labels at inference.

There is no court-line detector. Every detection is projected through one homography calibrated by hand from four outer corners — and the corners sit on the outer edge of a painted line, so reading them off a gridded crop beats every automatic pick tried.

Pose gives ankles, but ankles float above the floor and bias depth. The ground point blends the ankle down toward the box bottom (0.65 of the way), which is what pulls far-court error from metres to centimetres.

The frame offset between our video and the labels is not one number. Player feet are insensitive — their error curve is flat over ±10 frames — but the shuttle moves 20+ px per frame and pins the true contact sharply, and the right offset differs per match.

Two offsets are stored per match: a global +6 for the (flat) player tracks, and a fitted per-match shuttle offset used everywhere contact timing matters.

A smash off a descending lift has no 2D direction change at all. Detecting contacts by looking for the shuttle to turn around misses the single most important shot in the sport.

Two per-frame signals are unioned over ±4-frame velocities: an acceleration term (|Δv| ≥ 30 px/f) that catches smashes, and a reversal term weighted by speed that catches slow net play. Serves can’t kink — they start from rest — so they get a dedicated motion-onset pass.

Drives and pushes look nearly identical from geometry alone. Position, landing and timing get you to the mid-80s on easy classes and stall on the ones that decide rallies — you need the body, not the trajectory.

BST-0 reads our YOLO pose and TrackNetV3 shuttle windows. A stacked classifier combines its probabilities with court geometry and estimated contact height. Training uses detected contacts from other matches, so it sees the timing noise present when the system is deployed.

Shuttle visibility is useless as a rally signal. TrackNetV3 fires happily on replays and warm-ups — 83–94% visible during rallies versus 77–82% during the gaps. The two distributions barely separate.

Rallies are camera runs where both players hold in-court tracks (98% of rally frames, 7–18% of gaps), with bounding-box height bands self-calibrated off the dominant cluster to throw out close-ups and zoomed replays, then split at dead-shuttle restarts.

The BWF score digits are about 12 pixels tall. OCR engines return noise. And the overlay updates two to five seconds late — sometimes fifteen-plus seconds behind, when the broadcast cuts to a replay.

Digit templates are bootstrapped from a labeled match. A rule-constrained decoder repairs score sequences; delayed events are associated with already-ended rallies. The next server and service-court parity provide independent checks. Correct final scores do not make every inferred rally winner exact.