UTF-8 text + source code, 13.2 MB mixed corpus. No vocabulary, no pre-training, no GPU.
Order-1 byte Markov surprisal; greedy segmentation where each patch carries ≈15.9 bits of accumulated information. Content-addressed by FNV-1a-64: the same substring is always the same token.
Each patch id projects to 12 of 1024 input neurons via splitmix64 — no embedding table, works for unseen patches.
512 LIF neurons, Dale's law (50% inhibitory), homeostatic thresholds, refractory periods. Recurrent synapses follow reward-modulated STDP.
256 LIF neurons reading the pooled fast spike volley every 8 tokens (multi-scale). Its readout contributes a top-down prior fused with gain β (variance-matched). A counterfactual trust rule self-calibrates β.
Local Rescorla–Wagner delta rule on active units, gated by a dopaminergic RPE broadcast to both areas. No backprop, no gradients, no autodiff.
Hybrid addressing: primary full-key hash (pattern separation) + leave-one-out lanes (CA3-style completion). Writes associations in a single pass and keeps the episodic trace.
Offline "sleep": episodic traces are re-run through the cortex in training mode, consolidating one-shot memories into slow cortical weights. Cortex-only recall rises 0.3% → 46.9% while the hippocampus stays intact.
Readout scores fused with hippocampal votes; strong recall overrides the vocabulary. Closed loop: the emitted patch is fed back as spikes.
Multi-byte patches emitted at ≈2,000 chars/s on 2 vCPU. One-shot teaching of new text in 0.4 s. Readout vocabulary 16,384 (64% held-out coverage).