HF-TINYSTORIES-215

A recurrent bio-inspired core trained with cycling plasticity phases (hebbian → scaling → pruning) converges stably on narrative text.

Recurrent core with gated state blocks + proto-neurons, GPT-2 vocab, TinyStories corpus. Phase schedule 50/30/20 (hebbian/scaling/pruning), adaptive neuromodulator, LR 3e-4 constant.

• ◆Plasticity phase schedule
• ◆Neuromodulator strength (adaptive)

• →21,476 total steps across 215 cycles: loss 10.91 → ~0.86, PPL 54,728 → ~2.1–2.9.
• →Neuromodulator strength self-adjusted 1.00 → 1.05 across the run.

• ·Phase transitions visible in the loss trace but never destabilized training.
• ·1,211-page JSON step log preserved in the patent package.

Plasticity-phase training is stable end-to-end on this domain. Domain PPL is not comparable to WikiText-class numbers.

Cycle-150 behavioral evaluation (HF-CYCLE150-EVAL).