DRIVE_PROBE — token drive collinearity
Date: 2026-09-17 PT
Probe: inject "Bonjour" vs "xxxxxx" from zero field; measure alpha RMS + cosine.
Model: AtomNativeModel d_model=32, n_modes=32, seed=0.
Hypothesis (verified in src/toroidal/dynamics.py before patch)
| Claim | Verified |
|---|---|
mode_phase regular polygon arange * 2π/n_modes |
yes |
| token drive mean-broadcast identical on all modes | yes (pre-patch) |
rotated = sin(phase) * alpha |
yes |
_compute_phase_laplacian exists but unused; linear Laplacian used |
yes |
Before (pre-patch numbers only)
RMS_Bonjour=0.00242920
RMS_xxxxxx=0.00243263
COSINE=0.97792614
Patch choice: A (mode-modulated drive)
drive_m,d = token_d * cos(mode_phase_m + 0.5 * token_d * d) — not the same
vector on every mode. Same tensor shapes. No new module. Choice B (phase
Laplacian) not applied.
After
RMS_Bonjour=0.00178140
RMS_xxxxxx=0.00177089
COSINE=0.95857894
Cosine dropped (0.9779 → 0.9586). RMS did not explode.
energy_decay clamp
Train path calls model.stabilize_dynamics_parameters() after every
optimizer.step() (tools/run_atom_native.py); band default [0.3, 0.95].
Already correct — no further clamp fix. See DECAY_FIX.md.
Train path note
tools/run_atom_native.py supports --stream, --resume, --atom-flush-every
(soft atom list clear; field alpha not wiped), --no-episode-reset default.
No long train started for this probe.