Lattice-7 Prometheus (OctaCap Variant)
DuoNeural Lab | Research Artifact | November 2026
Status: Research artifact — pending downstream evaluation. Best checkpoint saved at step 55,500/100,000.
Overview
Lattice-7 Prometheus (OctaCap) is a CDM (Crystalline Diffusion Model) architecture variant trained via knowledge distillation from a Qwen 27B teacher. It incorporates Poincare ball hyperbolic geometry for slot routing, Crystal Telepathy cross-layer coupling, and a learned curvature parameter -- one branch of DuoNeural's exploration of non-Euclidean routing geometry.
Architecture
| Component | Config |
|---|---|
| Model family | CDM (Crystalline Diffusion Model) |
| Layers | 8 |
| Hidden dim | 768 |
| Slots (K) | 8 |
| Geometry | Poincare ball (hyperbolic) |
| Curvature | Learned (c=1.44, saturated) |
| Crystal Telepathy | enabled (beta=0.5) |
| DHP | confirmed all 8 layers |
| Parameters | ~177M |
| Teacher | Qwen/Qwen3.8-27B (4-bit quantized) |
Training
- Dataset: FineWeb
- Total steps: 88,900 (best at step 55,500)
- Batch: 1 x grad-accum 8 (effective 8)
- Seq len: 512
- Distillation: alpha=0.7, temperature=3.0
- Hardware: RTX 4080 SUPER (32GB), ~6 days
Key Properties
DHP confirmed at all 8 layers, tau* reaching 7.1e9. Discrete attractor crystallization in Poincare routing space.
Learned Curvature: All 8 layers saturated at c=1.44 with tau*=1e10.
Crystal Telepathy: tele_norm 0.39-0.54 across layers.
Best Checkpoint Results
| Metric | Value |
|---|---|
| Best step | 55,500 |
| Val BPB | 0.9327 |
| Val PPL (approx) | ~19.5 |
| Teacher KLD | ~0.18 |
Related Work
- DHP: DuoNeural Lab (2026)
- CDM Paper: In preparation, DuoNeural Lab (2026)
Authors
Jesse Kuhl & Archon (DuoNeural Lab)