Resonant Cortex — live grokking instrument

This is not a recording. The complex-valued network (RCNet) is training in this tab, full-batch AdamW, on (a + b) mod P — the same architecture as resonant_cortex.py: tanh amplitude, learnable phase bias, skip-connected columns, neurogenesis on frustration. Watch the embedding phases rotate until the hue gradient locks into rotational order.

Phase wheel · embedding phases of the 3 loudest neurons · hue = token value · angle = learned phase

Each ring is one neuron. Each dot is one of the P tokens, placed at the phase that neuron assigns it. Scrambled hue order = lookup-table territory. A clean hue sweep around the ring = the network found the circle: token n living at phase k·2πn/P, so addition becomes rotation.

Training traces · accuracy + log₁₀ loss · dashes mark neurogenesis

Phase field · wave-state phase, first 64 equations × 48 neurons · cyclic colormap

Interference probe · query the live weights

0 + 0 ≡ ? (mod 29)

b12

Distribution over answers. Early in training it's mush; after phase-lock it's a spike.

Phase wheel · embedding phases of the 3 loudest neurons · hue = token value · angle = learned phase

Training traces · accuracy + log₁₀ loss · dashes mark neurogenesis

Phase field · wave-state phase, first 64 equations × 48 neurons · cyclic colormap

Interference probe · query the live weights

Run log · honest ledger