Six measurement tools that need no GPU

These read model weights from disk and require no forward pass, so a result cannot be contaminated by a broken engine. They run on any machine with Python and numpy.

bitwidth_sweep.py

Reconstruction fidelity across ternary, int2, int4 and int8, with per-column and global scaling controls, measured against the original matrix.

per-column rescues int2 (0.50 vs 0.036) — neither is usable

lloyd_max.py

Non-uniform quantiser levels placed on the observed weight density rather than spaced evenly, with error reduction reported per bit width.

52% error reduction at 3 bits — worth a full extra bit

gate_geometry.py

Treats the FFN as a hyperplane arrangement and measures how concentrated and how stable the active gate set is between consecutive tokens.

50% of channels for 90% of mass; 10-30% overlap

subspace_rank.py

Effective dimensionality of hidden activations, reported both raw and with the massive-activation channels removed.

effective rank 1.1 to 15.0 after removing 8 of 3072 channels

coupling_test.py

Compares amplitude, phasor and phase-only oscillator encodings against an exact matrix product on real weights.

amplitude exact; phase-only 0.975 rel-err — the encoding is load-bearing

bqsm_full_settle.py

The whole forward pass as one equilibrium, running both a sequenced and an unsequenced solver and comparing the logits they converge to.

identical logits, error 0.000e+00 at the critical path

All are MIT-licensed and require only numpy.