Compiler-verified Faust DSP coding model, developed on NVIDIA DGX Spark.
Status: specification-first. The V0.1 content-addressed artifact store (T02), sandboxed Faust compiler harness (T03, faustwasm 0.19.0), and dataset rights transitions (T05) are implemented. No model-quality claim is established. For v0.2, R-001 (amended 0.3.1) uses cited external DGX Spark benchmarks plus a per-machine fingerprint; FaustCoder throughput on the Spark is not measured yet.
- Canonical specification
- Source rights inventory
- Implementation backlog
- Architecture decisions: ADR-0001, ADR-0002, ADR-0003, ADR-0004
- Constitution
- Contribution guidelines and security policy
- Changelog and versioning
- DGX Spark feasibility study and the env_verified runbook
- Validate generated Faust through a pinned compiler and independent signal tests.
- Track all datasets, runs, receipts and releases with immutable content digests and parent references.
- Separate an acyclic evidence DAG from stateful/recursive Faust DSP semantics.
- Keep frozen evaluation holdouts isolated from training and model selection.
- Benchmark model quality and training throughput before choosing a student model.
The v0.1 milestone is infrastructure, contracts, provenance and baseline evaluation, not a trained model.
This repository is split-licensed (ADR-0003):
| Path | License |
|---|---|
specs/ and docs/ (including ADRs) |
Apache-2.0 |
| Everything else: code, scripts, CI, model weights, adapters, checkpoints, datasets and evaluation data | Zero State Proprietary License 1.0 |
The proprietary license is the default for any path not listed above. See NOTICE. Third-party material keeps its own license; GPL code used for training is never stored in this repository or in released data.