terrizoaguimor/tinymars: v1.2.0 — Native (from-scratch) result + paper

The from-scratch experiment: a 110M decoder born with proprioceptive channels from layer 1.

• Perpendicular force replicates from scratch — the channel decides the preferred output on 88.8% of 455 held-out counterfactual pairs (chance 25%).
• Relief valve (new) — with channels zeroed, the native predicts held-out targets better than two channel-less baselines that bracket its parameter count (4.825 vs 5.252 / 5.253 nats), attributing the gain to the channel pathway, not size.
• Paper: docs/paper/proprioceptive-channels.{md,tex,pdf} — two experiments (adapter + native), 15 verified citations.
• New: training/native/, eval/native/native_eval.py.

Honest scope: toy scale (110M / 1B tokens), logprob on in-distribution held-out data, one iteration. Built on nanochat (A. Karpathy, MIT).