v2.0.0 — RWKV-v7 Goose + Pip Suite + correct backward pass

The backward pass is real. The state learns.
Three names on the file.

Complete browser-native LLM training runtime in a single HTML file.

Architecture: RWKV-v6 + BitNet b1.58 ternary quantization
Training: OOMB chunk-recurrent backpropagation
Optimizer: AdamW (default) + Muon Newton-Schulz quintic
Alignment: GRPO critic-free RL
Tokenizer: BPE via WebWorker
Storage: IndexedDB + JSON export/import

Zero dependencies. Zero server. Runs offline.