The Metabolic Therma Neuron: A Complete, Reproducible Account of Homeostatic, Metabolic, and Consolidative Neuron Architectures

A complete, reproducible account of the metabolic Therma Neuron: a metabolic activation whose thermal state gates gradient flow, conserves total gradient mass via ThermaGradient, and stores blocked signal in a Hebbian associative overlay.

Validated on Permuted MNIST (3 seeds, 0.0089 mean forgetting), CIFAR-100 (FC-only 0.0731, full conv+FC vault 0.0863, tuned champion 0.0543), and DeepSeek-Coder 1.3B across Python, JavaScript, and C++ (negative forgetting: Python perplexity improved by 1.85 points after training on later tasks; JavaScript improved by 12.91).

The document extends the core neuron with a Complex-STDP memory manifold where magnitude encodes factual association and phase encodes causal direction. The complex outer-product update yields LTP for consistent temporal orderings and LTD for reversed orderings through constructive and destructive interference, with no hand-crafted rules.

Inference-time mechanisms include a corrected warmth gate (applied before low-rank projection on the full intermediate-size vector), orthogonal STDP that forces new updates into the null space of crystallised columns, a column-norm mask that freezes trained memories on the first inference step, selective phi decay that eliminates silent memory drift, prefill-only STDP updates, post-memory trace updates, and an energy-tracking adaptive fast-trace decay that prevents Fire saturation.

Empirical telemetry shows Hot Norm unchanged at 0.00000000 across all turns while Cold Norm grows, demonstrating that the consolidated vault is permanently sealed while remaining open to new knowledge.

Every number is from an actual run. Every bug is documented with root cause and fix. Reproducibility is the whole point: every result should be regenerable from the code included.