Publications & Papers¶
All publications documenting the MoE Sovereign project — available directly from this repository.
Technical Whitepaper (v2.8 · August 2026)¶
The full technical whitepaper covers the complete system architecture: formal logic layer, GraphRAG pipeline, self-correction loop, MCP precision tools, evaluation, and sovereignty principles.
| Language | Download | Pages |
|---|---|---|
| English | moe-sovereign-whitepaper-en-v2.8.pdf | 116 |
| Deutsch | moe-sovereign-whitepaper-de-v2.8.pdf | 116 |
v2.8 changes (August 2026):
- §8 Attribution table: corrected de Vries (2007) author name (Andreas, not Madelon); added Belnap (1977) four-valued semantics {T, F, B, N} as separate entry
- §10.1 Judge node: activated Constitutional AI citation (Bai et al. 2022)
- §10.7 Agentic loop: activated Self-Refine citation (Madaan et al. 2023)
- §10.8 Paraconsistent resolution: separated algebraic motivation (de Vries 2007) from the formal semantics (Belnap 1977)
- MCP tool count corrected: 28 → 51 deterministic tools
- Full DE section mirroring all EN corrections
LUMI-G Training Paper (August 2026)¶
Empirical practitioner's report on the EuroHPC LUMI-G training campaign for the Judge and Planner models. Documents 20+ SLURM job iterations, OOM root causes, resume bugs, and the distillation pipeline.
| Format | Link | Pages |
|---|---|---|
| HTML (this site) | system/lumi_training_paper/ | — |
| moe-sovereign-lumi-paper.pdf | 23 | |
| DE | system/de/lumi_training_paper/ | — |
Key findings: Judge deployed 2026-07-19; max_seq_len 8192→4096 yields
only 1.2× speedup with --no-packing; expandable_segments silently
inactive on LUMI-G ROCm build (July 2026).
arXiv Paper (IEEE format)¶
Short paper in IEEE two-column format covering the formal routing and Optimal Transport layer — suitable for submission.
| Format | Link |
|---|---|
| arxiv-paper.pdf |