Run the models yourself
Three of the tiny models, quantized to int8 and running on ONNX Runtime Web — tap a phrase, pick a world, and hear where each one takes it. Nothing leaves your machine.
The three worlds are the same model, raised differently. Normal saw ordinary tonal music. Leading tone erased never saw a leading tone (major collapsed into Mixolydian). No dominant pull had the leading tone flattened only inside dominant chords. Give them the same phrase and listen for where the harmony wants to resolve — or refuses to.
Models: d384 / 6 layers, exported with a fixed causal mask at opset 17, then dynamically quantized to int8 — 11 MB each. The browser recomputes the full context every step (no KV cache), samples at top-p 0.95, and decodes through a JavaScript port of the same tokenizer. The two matched pilots behind the demo are on the downloads page.