glen-muscle-router

Glen Muscle Router

A 2.17M parameter ternary router for the HEIDES-VOLT harness. Given an intent sentence it predicts one of 16 harness actions: check, staged, verify, remember, recall, curious, webget, websearch, export, import, skills, stats, delegate, peers, sync, chat.

This is the "muscle" model: a reflex layer that lets familiar work skip the brain. It is trained from scratch with quantisation aware training (QAT): every linear weight is ternary {-1, 0, +1} with one f32 scale per group of 128 weights, the same shape as Bonsai Q2_0 g128.

Results

Accuracy is measured on held out templates per class, so it tests unseen phrasing rather than memorised sentences. The synthetic template data is the limit here; real verified harness traces will raise it. QAT itself is free: the ternary model matches the FP32 baseline within noise.

Files

• router_ternary.bin packed ternary weights, format below
• metrics.json full metrics including per class accuracy

Binary format

Little endian, magic GMUS, version 4:

"GMUS" u32 version=4
u32 vocab, classes, max_len, d, layers, heads, ff, group
class names: classes x (u16 len, utf8 bytes)
tokens: vocab x (u16 len, utf8 bytes)
token embedding: vocab f32 row scales, vocab*d int8 values
position embedding: max_len*d f32
final layernorm: d f32 weight, d f32 bias
per layer: ln1 weight, ln1 bias, ln2 weight, ln2 bias (d each)
ternary tensors in order (per layer q,k,v,o,up,down, then head):
  u32 group, u32 rows, u32 cols
  rows*cols/group f32 scales
  rows*cols/4 bytes codes, 4 codes per byte, low bits first
  code = weight/scale + 1 in {0,1,2}, value = (code - 1) * scale

The token embedding is int8 with per row scales; every other weight matrix is ternary. Codes pack like GGUF Q2_0 g128, so the same decoder works.

Runtime and verification

The harness runs this model in process with zero dependencies: src/muscle/gmus.js implements the forward pass (pre-norm blocks, exact erf GELU, causal attention, mean pool). It is verified against torch: the test test/gmus.test.js replays 8 held out inputs and matches the exact logits within 0.1 with identical argmax. Set HARNESS_ROUTER_MODEL to the binary and the harness fast path routes real turns through it.

Training

Code lives in the HEIDES-VOLT harness repo under trainers/ternary/. Kaggle kernel jamalkt/glen-muscle-router runs it on a free GPU:

python3 train.py --out out --steps-fp 1800 --steps-qat 900

Contract: prints one JSON line with {ok, evals_pass, acc_fp, acc_ternary, retention, params, bytes_ternary, out} and writes router.pt, router_ternary.bin, metrics.json.

License

MIT. Created for the HEIDES-VOLT resident project.