An architecture made of observable roles.
The current branch combines topology-guided multi-field encoding, shared latent integration, sparse expert computation, and a stress-aware projection path. Select a component to inspect its computational role.
Signal path
Each stage has a narrower responsibility. The names are analogies; the implementation is standard PyTorch tensor computation.
A learning signal, not yet a comparative result.
The uploaded v2 run trained for 2,500 steps on the FineWeb-derived setup encoded in the script. Validation bits per byte fell from 4.1189 to 1.4971. That demonstrates end-to-end learning; it does not establish superiority over a matched transformer baseline.
Validation BPB across training
Lower is betterRecorded points: steps 0, 250, 500, …, 2500. No smoothing and no baseline substitution. CSV ↗ · JSON ↗
Run readout
What the current evidence supports
- The architecture executes and trains end to end in the recorded setup.
- Validation BPB decreases consistently across the 2,500-step v2 trace.
- Internal architecture signals — resonance state, perception gate, and axon stress — are instrumented and logged.
- The model can export a compressed state artifact and a separately distilled resonance snapshot.
What remains unproven
- No matched parameter-count baseline is included in the uploaded run.
- No ablation isolates the contribution of each biological analogy.
- No result demonstrates consciousness, subjective experience, or biological equivalence.
- The v3.12 branch is still changing; its final efficiency and quality are unresolved.
One project, two experimental branches.
Dopamine has evolved through competing implementations rather than a single frozen story. The page keeps those branches distinct.
Resonance and perception branch
Tests whether memory can emerge from similarity between current representations and a non-parametric buffer of previous latent flow.
- ResonanceBuffer with 64 EMA slots
- PerceptionAgent cross-attention over current and past flow
- Soma treated as a shared medium rather than a central processor
- Recorded 2,500-step run shown above
Topology-guided morphology
Tests whether structural divergence measured from multiple latent views can modulate how local, mid-range, and global fields are mixed.
- SemanticTopologyMap and plastic prototype recall
- Shared Soma nucleus with prototype cross-attention
- Top-2 mixture-of-experts blocks
- Semantic loss, QAT, checkpointing, and H100/RTX execution paths
Read the architecture, not the mythology.
The repository exposes the training scripts, model components, and version history. The useful question is not whether the names sound biological; it is whether the separated mechanisms survive controlled comparison.