Research

Published, reproducible, verified. Every result in our papers is backed by real model experiments.

Published Papers

Paper 1 Zenodo · June 2026
DOI: 10.5281/zenodo.15623044

Holosynthics: Neuron-Level Decomposition of Language Models into 8 Universal Atom Types

We present Holosynthics and demonstrate that transformer language model neurons fall into 8 fundamental semantic types: Boolean, Number, Symbol, Identity, Space, Time, Relation, and Entropy. We show these 8 types appear universally across architecturally distinct models. Every model tested — regardless of size, training data, or architecture — contains all 8 types in consistent proportions. This is the first empirical demonstration of universal neuron-type organization across AI language models.

Key Findings

  • 8 universal atom types confirmed across all tested architectures
  • Symbol neurons are the most numerous — dominant in every model
  • Entropy neurons are the rarest — and the most diagnostically important
  • Relation neuron proportion scales with model capability

Models Tested

Gemma 2BGPT-2 SmallGemma 9BLlama 3.1 8BQwen3 1.7BOLMo 3 7B
Read Paper →
Paper 2 Zenodo · June 2026
DOI: 10.5281/zenodo.15640300

The Periodic Table of AI: 56 Universal Neuron Elements Across 11 Language Models

We extend the 8-type Holosynthics framework to 56 sub-type elements (8 types × 7 sub-types each) and validate universality across 11 language models. The same 56 elements appear in 10 of 11 models tested. We characterize each element by its functional role, typical network position, and behavioral signature. We introduce Entropy.Confidence — a late-layer neuron cluster whose presence or absence predicts whether a model will hallucinate on a given prompt.

Key Findings

  • 56 elements validated across 11 distinct architectures
  • 10 of 11 models confirm all 8 types and all 56 sub-types
  • Entropy.Confidence neuron cluster correlates with hallucination absence
  • Full element profiles with behavioral characterization published

Models Tested

GPT-2 SmallGPT-2 MediumGPT-Neo 2.7BPythia 1.4BPythia 2.8BDeepSeek 1.3BFalcon 7BMistral 7BOLMo 3 7BPhi-3 MiniQwen3 1.7B
Read Paper →

Experiment Log

Summaries only — full datasets and methodology are available in the published papers and on request.

Experiment Model(s) Date Key Result
Hallucination Trace Experiment GPT-2 Small June 2026 Entropy.Confidence neurons absent in all hallucinated responses; present (weakly) in correct responses — first neuron-level hallucination signature
Gemma 2B Full Layer-by-Layer Analysis Gemma 2B June 2026 All 26 layers analyzed; complete periodic table element distribution mapped; Conductor GI identified
Qwen3 1.7B Cross-Architecture Validation Qwen3 1.7B June 2026 All 8 types confirmed; anomaly cluster detected; results independently validated against Paper I findings
Six-Model Periodic Table Comparison Gemma 2B/9B, GPT-2, Llama, Qwen3, OLMo June 2026 Symbol is most variable across models; Relation grows monotonically with model scale; type proportions are model-fingerprints
Llama 3.1 8B Deep Classification Llama 3.1 8B June 2026 Full 32-layer classification complete; Symbol dominant, Relation second at all layers, Entropy exclusively late-layer