AletheionAGI is an independent research organization building causal language architectures around persistent state.
Our research program, Aletheion State Models (ASM), asks a direct question: can language generation be modeled as the evolution of a compact latent state, instead of repeated attention over the entire token history?
The current answer is experimental—not a product claim. We publish the architecture, protocols, negative results, corrections, and limitations needed to examine it.
token xₜ ──→ persistent state zₜ ──→ learned transition ──→ zₜ₊₁ ──→ logits
│ │
└── selective memory ────┘
| Layer | Name | Role |
|---|---|---|
| Organization | AletheionAGI | Independent research, engineering, and open scientific collaboration |
| Program | Aletheion State Models | A family of attention-free causal state models evaluated under shared protocols |
| Theory | Directional Relational Manifolds | The geometric framework that originated the research direction |
| Architectures | ASM-R · ASM-S · ASM-F · ASM-X | Comparable variants that isolate transition, memory, and relational geometry |
Current ASM variants use no Transformer blocks, self-attention, Q/K/V attention, nn.MultiheadAttention, or KV cache. Their memory is an evolving causal state.
The promoted architecture, ASM-R, combines a direct contextual transition, relational metric naturalization, a causal local mixer, token-to-state residuals, and selective forget/write memory. ASM-S remains the efficiency-oriented candidate, while ASM-X preserves the explicit DRM geometric path.
| ASM-R confirmation | Result |
|---|---|
| Training scale | 100M tokens |
| Independent runs | 3 seeds |
| Frozen-validation targets | 4,834,787 |
| Mean cross-entropy | 1.344538 |
| Population standard deviation | 0.000561 |
These results promote ASM-R within the current ASM experimental protocol. They do not establish general superiority over Transformers, Mamba, or production language models.
Inspect the 100M multiseed evidence →
| Model | Research role | Current status |
|---|---|---|
| ASM-R | Relational state model | Promoted for quality per token |
| ASM-S | Selective state model | Candidate for efficiency per time |
| ASM-F | Relational frame model | Generation 1 requires a stabilized rerun |
| ASM-X | Explicit DRM state model | Geometric reference architecture |
The architecture follows evidence, not naming. Variants are promoted, revised, or retired as ablations and scaling results accumulate.
Evidence before narrative. Claims travel with their protocol, scale, artifacts, and limitations.
- Reproduce. Configurations, seeds, frozen evaluations, and benchmark artifacts remain inspectable.
- Compare. Variants share controlled protocols designed to isolate architectural contributions.
- Correct. Invalid comparisons are retracted, documented, and kept available for audit.
- Scale carefully. We move from replication to scaling, external comparison, and publication.
The historical GPT-2 comparisons affected by a target double-shift bug have been explicitly retracted as comparative evidence. Keeping that correction visible is part of the research record.
| Explore | Read |
|---|---|
| Run the code | ASM repository and quickstart |
| Understand the family | Model taxonomy and promotion criteria |
| Study the architecture | Causal state paths and implementation |
| Audit the results | Benchmark artifacts |
| Read the caveats | Known limitations |
| Trace the theory | DRM paper and mathematical notes |
git clone https://github.com/gnai-creator/ASM.git
cd ASM
python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
python scripts/train_tiny.py --config configs/tiny.yaml --text data/tiny.txtASM is research infrastructure. It is not currently a production model, and cross-entropy alone does not measure usefulness, safety, reasoning, or conversational quality. The current evidence uses a principal dataset and a byte-level tokenizer; broader datasets, external baselines, and independent reproductions remain part of the roadmap.
The source is available under AGPL-3.0 for open work, with commercial licensing available for proprietary use. See the repository license files for the terms that apply to a specific component.
Reproduce an experiment, challenge an assumption, propose an ablation, or help take causal state models to their next honest result.
Explore ASM · Visit the website · Email the research team
AletheionAGI · independent research in causal state models