BernaLabs/berna-r5-poc

Berna R5

DNA-Kernel Plexus: A Bio-Inspired Architecture for Growing, Self-Correcting Language Models with Bounded Forgetting

Model Description

Berna R5 is a from-scratch language model architecture built on dynamic knowledge cells that are born, split, and interconnect according to a 6-dimensional knowledge state space.

This is the 50M-parameter Proof of Concept (PoC) release, used to validate five theoretical results before scaling to 1.5B.

Architecture

• Layers: 6-stage pipeline (sensors, spinal cord, plexus, cells, DNA kernel, registry)
• Parameters: 50M (PoC) / 1.5B (target)
• Vocabulary: 200,000 tokens
  • Text: 0 - 63,999
  • Vision: 64,000 - 163,999
  • Audio: 164,000 - 199,999
• Context length: 8192
• DNA chromosomes: 100 (10 groups x 10)
• Knowledge dimensions: 6 (L, W, H, D, T, E)

The Five Theorems

1. Saturation Equilibrium - S* = 0.85
2. Reliable Incorporation - epsilon_neg = 1e-6
3. Zero-Forgetting - F_j <= 1e-6 (proven)
4. Balanced Growth - cell count bounded by capacity
5. Plexus Connectivity - weights converge to fixed point

Full proofs in the GitHub repository under theory/proofs/.

Intended Use

Primary use: research on continual learning, dynamic architectures, and bio-inspired neural networks.

Not intended for:

• Production deployment (PoC scale)
• High-stakes decisions
• Direct comparison with GPT-4-class models

Training Data

• WikiText-103 (~100M tokens)
• The Stack (Python subset, ~500M tokens)
• OpenWebMath (~200M tokens)
• Total: ~800M tokens

Training Procedure

• Optimizer: AdamW
• Learning rate: 3e-4
• Batch size: 4
• Sequence length: 512
• Seed: 42
• Hardware: RTX 5090 (32 GB VRAM) / CPU for validation

Evaluation

Five theorems validated on PoC (see paper):

Experiment	Result	Pass
Saturation convergence		S - S*
Split capacity	19 <= 20	Yes
Incorporation delta	min = +0.02	Yes
Zero-forgetting	max F_j = 0.0	Yes
Plexus convergence	True	Yes

Limitations

• PoC scale only (50M). 1.5B is future work.
• No benchmarks (MMLU, HumanEval, GSM8K) run yet.
• Multi-node federation untested.
• 100-chromosome design not empirically optimized.

The Fail-Closed Integrity Invariant

Any violation of a declared invariant triggers FAIL_CLOSED. There are no silent failures. This is a runtime guarantee, not a mathematical claim.

Citation

If you use Berna R5 in your research, please cite:

@misc{muhammed2026berna,
  title={DNA-Kernel Plexus: A Bio-Inspired Architecture for
         Growing, Self-Correcting Language Models with
         Bounded Forgetting},
  author={Muhammed, Mohammed Kamil},
  year={2026},
  publisher={Zenodo},
  doi={10.5281/zenodo.22998311},
  url={https://doi.org/10.5281/zenodo.22998311}
}

DOI: 10.5281/zenodo.22998311

Links

• GitHub: https://github.com/berna-labs/berna-r5
• Paper: (to be posted on arXiv)
• Registry schema: registry/schema.sql

License

Berna Research License v1.0.

• Research use: free (academic, scientific, personal).
• Commercial use: requires a separate license from Berna Labs, following the same standards as the Berna R4 project.

See LICENSE for full terms. Contact info@bernalabs.com for commercial inquiries.