Deterministic Code as a Safety Net

The DGG project made contact with the outside world this week.

Agentic DGG conversion

I integrated a collaborator's FoxFlow transpiler service, which exposes an OpenAI-compatible endpoint. Around it I built an extractor for FoxFlow code blocks, a lightweight local validator with three syntax checks, deterministic repairs for known problems, and a validate-and-retry loop.

The bigger addition is a deterministic emitter that turns the typed DGG model straight into FoxFlow, with no LLM involved. I wrote a golden corpus of six hand-built models (particle loop, birth-death, diffusion along an edge, microtubule growth, boundary interaction, observable counting) and documented how our representation maps onto FoxFlow and the DGGML C++ library.

The first smoke test against the live service timed out, which is exactly why a deterministic path matters. When the LLM path fails, the system can still produce something checkable instead of nothing.

Lost in the Middle

I prepared the reproduction for release: a mocked test for the Ollama adapter, an MIT license, a preflight document with 23 passing tests and every citation checked against ACL Anthology, arXiv or OpenReview, and a tagged v0.1 release.

RAG saturation

I added proper packaging with pinned dependencies, made metric saving idempotent so re-runs replace rows instead of appending duplicates, and wrote a release manifest that maps every artifact to the run that produced it.