Compiler-verified Faust DSP coding model
BlackDog
A coding model for Faust audio code, built so that what it writes has to pass a compiler and signal tests before anyone trusts it.
Specification-first · spec 0.3.11 · no tagged release yet · formerly FaustCoder
BLACKDOGa coding model for Faust DSPgenerate → compile → test the signalWhat it is
BlackDog is a locally deployable model specialized in generating, repairing, explaining and optimizing programs in Faust, a language for audio signal processing. Correctness is meant to rest on evidence, not on the model’s confidence: output is checked by a pinned compiler and independent signal tests, and a successful compile is not treated as proof the audio is right.
It was called FaustCoder until October 17, 2026 (Pacific time). The Python package and the specification folder still use the faustcoder name.
Where it stands
The project is specification-first. What exists so far is infrastructure: a content-addressed artifact store, a sandboxed Faust compiler harness (faustwasm 0.19.0), a sandboxed signal-testing harness, dataset rights transitions, and a normalization and de-duplication snapshot. The first milestone is infrastructure, contracts, provenance and baseline evaluation, not a trained model.
No model-quality claim is established. A Qwen smoke run on an NVIDIA DGX Spark has recorded bounded training, prefill and decode throughput; repeated throughput qualification, data and split gates and baseline gates are still pending. There is no tagged release yet: the first, v0.1.0, waits for the foundation milestone in the spec.
How it is kept honest
Training data is tracked with its source rights, evaluation holdouts are kept apart from training and model selection, and every dataset, run and receipt carries an immutable content digest. Release needs explicit human approval. The specification and docs are Apache-2.0; the code, data and model weights are under the Zero State Proprietary License 1.0.
Built with
- Faustthe audio language the model writes
- faustwasm 0.19.0pinned Faust compiler that checks every program
- Node.jsruns the sandboxed compiler worker
- Pythonfile store, test harnesses and dataset tools
- JSON Schemavalidates every dataset, run and receipt
- NVIDIA DGX Sparkone recorded Qwen smoke run for training and response speed; not a trained model
Changelog
Milestones in plain English, newest first, to Nov 7, 2026
- Dataset snapshots now require reviewed source ranges for every function, with offline tooling to review them.
- A Qwen smoke run on an NVIDIA DGX Spark recorded bounded training, prefill and decode throughput. Repeated qualification is still pending, and this is not a trained model.
- Renamed from FaustCoder to BlackDog, the same day a sandboxed signal-testing harness landed.
- Project began as FaustCoder; a fingerprinted file store, sandboxed Faust compiler and data-rights tracking followed.
Roadmap
What comes next. Plans, not promises, and no dates.
- In progressFreeze the split between training and test data, then measure baseline models before any real training.
- PlannedFirst full specialist training on the DGX Spark waits until the data and baseline checks pass.
- PlannedTurn generated Faust code into audio plugins (CLAP, VST3, LV2) and code for small embedded audio boards.
- PlannedServe the model locally through a programming interface and connect it to code editors.