TL;DR
HibbieFormer trains in explicit phases — hebbian, synaptic scaling, pruning, on a 50/30/20 schedule — with an adaptive neuromodulator. The TinyStories run completed 215 cycles and 21,476 steps without a single divergence, ending at PPL ~2.1–2.9.
The schedule
Each cycle allocates 50% of steps to hebbian learning, 30% to synaptic scaling, and 20% to pruning. The neuromodulator strength is adaptive; across the full run it drifted from 1.00 to 1.05 — the system leaning slightly harder into plasticity as training matured.
Honest caveats
TinyStories is a narrow domain: PPL ~2 here is not comparable to WikiText-class numbers, and the cycle-150 behavioral eval showed total domain lock-in — the model answers factual questions as children's stories. The result is stability of the training regime, not general capability.
The Stage-0 audit later flagged one discrepancy in the 2025 archive: a checkpoint named 'best_verified' matches the PPL-12.2 file, not the PPL-10.2 one. That flag is now part of the permanent record.