Evaluation Suite

Every number on this page carries the same three qualifiers: internal evaluation, a deliberately small held-out corpus, self-supervised training. They bound what the results mean — strong evidence of what the architecture can learn, not benchmark claims. The protocol is strict about leakage: validation splits are drawn at the level of whole recordings, never individual windows, so scores are never earned on movement the model has already seen — a discipline that becomes more binding as the corpus grows.

How capability is expected to grow with that corpus is mapped on the data-scale ladder.

Reconstruction error is measured as MPJPE — mean per-joint position error — under progressively harder occlusions: scattered missing joints, whole missing time spans, and body regions missing over time. Difficulty scales exactly as it should: scattered gaps are easy, contiguous space-time holes are hard. How the model learns to fill them is described in the training documentation.

A linear probe asks a hard question gently: can a straight line through the frozen embedding recover a physical quantity the model was never told about? For body-mechanics concepts the answer is emphatic — speed, trunk lean, and left-right knee asymmetry are all linearly readable. That is the intended shape of the representation: the model learned how the body moves without a single label.

Beyond point metrics, the suite checks that the embedding space is organized: unsupervised clustering over movement windows recovers coherent movement archetypes without any event feed, and transitions between clusters align with visible changes on video. Archetype counts and quality scores stay internal until the corpus is large enough to publish them responsibly.

What the numbers do not show — tactical context, opponent pressure, match state — is deliberately out of scope: the model sees the body, not the pitch.