HF-GOLDEN-RATIO

Regularizing the ratio of the wavelet-stream and SDM-stream activation magnitudes toward the golden ratio (φ ≈ 1.618), with a PID-tuned loss weight, improves language-model training.

HibbieFormer V3 Metal (143.8M params: wavelet state block, Kanerva SDM, neuromodulator gating, predictive coding — no attention), 919M-token C4+WikiText-103 corpus, Colab A100, AdamW 3e-4 cosine, φ-ratio PID (Kp 0.05/Ki 0.01/Kd 0.02) plus a temperature-duality KL term.

• ◆φ-ratio regularizer weight (PID-controlled)
• ◆Duality KL weight

• →The control mechanism worked: stream ratio driven from 8.64 to ~1.51–1.66 (target 1.618) by step ~240 and held.
• →The hypothesis didn't: per-batch loss sat at ~7.5–7.9 (PPL ~1,900–2,700) for the entire ~20-hour run.
• →Run died at step 1,560 of a planned 1,000,000 (0.16%); one epoch-1 checkpoint saved (running-avg PPL 2,287).

• ·The Python per-timestep loop in the wavelet block held throughput to ~45 s/step (~360 tok/s on an A100) — the architecture starved the GPU.
• ·No baseline, no validation split; the logged 'PPL' was a running average that overstates progress.

Steering an activation ratio to φ is achievable and apparently useless for LM loss at this scale. Filed as inconclusive-leaning-negative; the fused-kernel lesson fed directly into the mlx-recurrence work.

Any revival requires a fused scan kernel and a matched transformer control first.