Developer Guide

DantinoX is designed to be extended without touching the core library. Every major seam — core layers, training paradigms, benchmark tasks, and visualization charts — has a clean plugin interface.

---

Extension points

What to add	Interface	Guide
New attention variant, MLP type, or normalization layer	core/ NNX module + Config field	Adding a Core Layer
New training objective (e.g. contrastive, RLHF)	Paradigm subclass	Custom Paradigm
New evaluation metric or dataset	BenchmarkTask subclass	Custom Benchmark Task
New chart type	Chart subclass + @Visualizer.register	Custom Chart

---

General principles

• Extend, don’t modify. New functionality belongs in new files, not edits to existing ones (unless fixing a bug).

• Typed contracts. Use ClassVar[str] for name, typed signatures for run() and loss_fn() — the type checker will catch mismatches.

• No global state. Paradigms and tasks are stateless; all mutable state lives in the model or optimizer.

• Test your addition. Add a test in tests/ before opening a PR. See Contributing.