v0.1.0 — Tilelli: small routed byte-LM + negative-results paper

Tilelli is a ~10.2M-parameter byte-level language model with a 3-pathway routed block. Trains on a single GPU, runs on a laptop CPU. Every numerical claim in the README is bound to a reproduce/ script that exits non-zero if the bundled checkpoint fails to reproduce it.

Included: code, two checkpoints (tilelli_chat_v4.pt FP32, tilelli_pretrain_v1_ternary.pt), the 210-prompt probe set, four reproducers, and the paper (PAPER.md).

Scientific contribution = verified negative results: (1) the pre-registered router-entropy-as-metacognition claim is disproven (0/7 regimes); (2) abstain heads are not modular (AUROC 0.76 in situ -> 0.54 on a fresh base); (3) the router is fragile at this scale. The single-seed architecture benchmark is reported honestly and is explicitly not a defensible win.

Apache-2.0. Also on Hugging Face: https://huggingface.co/TilelliLab/Tilelli-llm