Determinism and streaming

A render depends on the graph, seed and sample rate. The current engine uses fixed f64 transcendental kernels, so the same supported document produces byte-identical samples across platforms.

Current formats

SoundDocs and Songs use schema version: 2 and DSP engine: 5. Omitted pins select these defaults. Explicit unsupported revisions fail validation. Compiled Programs use format 3 and verify both their content hash and resolved document on load. Recompile the source song when a bundle is unsupported.

Noise, dust and pluck bursts use deterministic per-node random streams. Named tracks have independent streams; editing another track does not move their random draws. Save the source JSON and seed with exported audio.

Native streaming

StreamGraph::blockers explains which parts need a whole-buffer render. Player supplies that render when native streaming is unavailable.

Streams natively	Uses the buffer-backed Player
Source nodes and their modulators, including noise, dust and sequences	SoundFont sampler sequences
Every processor, including modulated filter/EQ cutoffs and gain amounts	Nested mixers
A tracks root with sidechains, buses and automation	Whole-buffer normalization, loop playback, Haas/Wide stereo treatment

The bounce and a stream run the same evaluator, so a supported streaming graph matches the offline render at every block size. For live compilation, CompileTarget::Runtime rejects streaming blockers instead of accepting a buffer-backed program.