Proposed example record

examples/baseball_swing.json describes the current artifact example. It is marked split: case-study, not assigned to train or test. There is only one source sequence here.

Required release fields:

schema_version, sequence_id, source_group_id, character_id, motion_id, split.
timing: frame count, fps and first/last stored pose. The 52-frame display lasts 52/30 seconds; the span from the first sample to the last is 51/30 seconds.
artifacts: repository-relative paths for input presentation video, final video, comparison, FBX, Blender scene and validation.
representation: deformation type, topology continuity and the explicit closing-surface handoff.
supervision_status: what is actually bundled and which training labels remain unprepared.

Future training records should add camera matrices with coordinate conventions, original uncropped inputs, canonical/reference geometry, stable per-sequence vertex or surface correspondence, VMR control transforms and influence weights, accepted target geometry, masks and confidence, contact labels, correction history, solver settings and provenance. Preserve units and transformations rather than assuming every character has identical scale or bone ordering. Vertex indices must not be compared across unrelated meshes without a correspondence map.

Known or synthetic ground truth must be distinguishable from reconstructed estimates and manual corrections. Keep a closing topology change as an explicit event until continuous correspondence is solved. Do not train it as if it were a standard continuous deformation.

All camera variants, adjacent frames, corrections and augmentations of a source belong to the same source_group_id. Before training, create splits that hold out character identities and motions. Assess missing fields and uncertainty rather than filling unavailable supervision with invented values.