Research¶
Paper¶
End-to-End Prime Factorization in a Generative Language Model: Emergent Algebraic Semantics from Joint Training
J. Arturo Ornelas Brand (2026). Independent Researcher.
Key Results¶
| Metric | Value | Notes |
|---|---|---|
| Language cost | Zero | PPL 7.69 vs 7.56 ablation (+1.7%) |
| Analogy verification | 98% | 51 analogies on XL model |
| Subsumption accuracy | 98.3% | Held-out, 158 supervised anchors |
| Domain separation | 1.21 mean | 12 domains, sentence-level |
| GPT-2 transfer gap | 48% closure | Toward Engine PCA upper bound |
| Signature uniqueness | 100% | All evaluated concepts |
| Scale crossover | ~20M params | Semantic ordering emergence |
Pre-trained Models¶
| Model | Params | Description | Link |
|---|---|---|---|
| triadic-gpt-40m | 40M | Production model (Run 15, v1.4-strongalign) | HuggingFace |
| triadic-gpt2-medium | 355M | GPT-2 medium with triadic head (Experiment 10) | HuggingFace |
Benchmark Suite¶
12 evaluation scripts covering scaling, analogies, subsumption, entropy, language quality, interpretability, topology, and engine comparison.
| Script | What it measures |
|---|---|
scaling_study.py |
4-point scaling study (1.3M--40M), PPL + triadic quality |
analogy_benchmark.py |
Analogy verification accuracy (51 analogies) |
subsumption_benchmark.py |
Taxonomic subsumption accuracy |
bit_entropy.py |
Per-bit entropy analysis (dead bit detection) |
language_quality.py |
Perplexity, Distinct-n, repetition, MAUVE |
engine_comparison.py |
TriadicGPT vs Triadic Engine (5 projection methods) |
Citation¶
Paper¶
@article{ornelas2026triadicgpt,
author = {Ornelas Brand, J. Arturo},
title = {End-to-End Prime Factorization in a Generative Language
Model: Emergent Algebraic Semantics from Joint Training},
year = 2026,
month = mar,
doi = {10.5281/zenodo.19206545},
url = {https://doi.org/10.5281/zenodo.19206545}
}
Software¶
@software{ornelas2026triadicgpt_repo,
author = {Ornelas Brand, J. Arturo},
title = {TriadicGPT: End-to-End Prime Factorization in a
Generative Language Model (Repository)},
year = 2026,
month = mar,
doi = {10.5281/zenodo.19207845},
url = {https://doi.org/10.5281/zenodo.19207845}
}
Related Work¶
| Project | Role | Link |
|---|---|---|
| Triadic Engine | Mathematical foundation (post-hoc projection) | PyPI |
| reptimeline | Training analysis tools | Docs |
License¶
BUSL-1.1 (Business Source License 1.1). Same terms as the Engine.