Why the format looks the way it does
One resource
A whole scene — geometry, appearance, motion, audio and camera — is one file, one URL, one cache entry.
Continuous time
Each gaussian carries its own birth time, temporal extent, motion and validity window, so the number of live gaussians varies over time with no frame machinery.
Seekable without a sidecar
A byte-range index maps time to chunks. Displaying an arbitrary instant reads the file's own index and then only the ranges that instant needs.
Bounded memory
Every read path is streamable. No conforming reader ever needs the whole file resident, on any path.
Audio in the file
A scene with a soundtrack carries it inline, and the audio track is the scene clock's master. A scene without one carries nothing at all.
Forward compatible by construction
Every top-level structure is a length-prefixed, opcode-tagged record, so a reader skips what it does not recognize instead of failing.
Stated error bounds
Lossy encodings declare, per attribute, the maximum deviation a decoder may observe, and the encoder is expected to verify it.
Renderer-agnostic
The format defines how to reconstruct splat state at a given time and nothing beyond that. How that state is drawn is out of scope.
SDKs
Support is stated per feature, not per language: the feature support matrix records only what the conformance suite proves, and a partial SDK is a documented state rather than a defect.