One state vector, five skeletons, three spectral bands
The model exists to make a subject addressable by something other than time. Each layer holds a different kind of invariance, so a change can be applied at the level it actually belongs to.
K0–K4: the nested skeleton
Layered so that a change to motion cannot silently alter identity. K0 is deliberately immutable during a session — the runtime exposes a lock for it.
Identity and topology
What makes this subject this subject. It does not move.
Motion and deformation
Everything that moves, layered on top of K0 rather than replacing it.
Contours, masks, boundaries
Where one region ends and the next begins.
Semantic regions, coarse visuals
Regions that own their own detail budget.
Spectral, residual, high frequency
The detail that gets regenerated rather than stored.
S0–S2: spectral decomposition
The split is a budgeting decision, not an aesthetic one. Under pressure the runtime drops S2 first, because S2 is the band whose absence costs the least that cannot be recovered.
Coarse blocks and low frequency
The structure that survives any quality level.
Signed residual texture
The correction that makes S0 match the source. This is what must be preserved.
Frequency detail modulation
Detail that can be synthesized. Dropping it costs sharpness, not identity.
Nested regions
Regions are not a flat list. A face contains eyes and a mouth; each owns a detail budget and a dirty state. When one axis changes, the dirty graph resolves which regions are affected and their downstream closure — and only those are rebuilt.
This is what makes partial reconstruction possible at all. Without region ownership, any change is a full-frame change, and the structure buys you nothing over a video file.
Dirty propagation
Every axis declares which nodes it invalidates. The engine resolves the transitive closure and rebuilds exactly that set. The inspector in the live runtime shows this happening per frame: which nodes went dirty, which surfaces were rebuilt, and how long it took.