Abstract Powered Research

An independent research cooperative building small, modular AI systems that are trained on commodity hardware, measured honestly, and released with the evidence behind them.

---

What we work on

Most models grow by retraining everything at once. We build them from parts that can be attached, measured, detached and attached again, so each experiment adds to the last one instead of replacing it.

Four ideas run through the work:

• Detachable parts. Behaviour lives in small add-on modules ("arms") on a frozen core. Switching an arm off returns the core exactly, bit for bit, so every arm's contribution can be measured on its own.
• Geometric addressing. Our addressing read (the aleph read) combines every anchor's signed contribution instead of picking a winner. We avoid winner-take-all routing because it kept failing in our measurements.
• Byte-level language models. The Beatrix models read and write raw UTF-8 bytes, with no tokenizer to retrain or translate.
• Compact training. Runs are sized for a few rented cards, not a cluster. That keeps experiments cheap enough to repeat.

How we measure

• Two seeds before a result is called settled; one seed is labelled as such.
• A control that could fail runs beside every claim.
• Numbers are scored on data the model and any fitted map never saw.
• Negative and mixed results are published with the same care as positive ones.

Mission

Our longer-term research question is machine introspection and rationalization: can a model's reasoning about itself be studied as measurable, repeatable behaviour? We approach it through careful protocols and instruments rather than claims. The modular, measurable parts above are the tools for that study.

---

What ships here

This organization holds official releases only. The research record behind each release (every checkpoint, every experiment, including the ones that did not work) stays public at AbstractPhil, and the code lives on GitHub at AbstractEyes.

Releases

Status labels. Official: passed every release check below. Research preview: passed the checks that apply, released as an early stage, with what was not evaluated listed on its card.

Families

---

What an official release means

Every release passes these checks before it is published, and the results are kept in its manifest:

1. Licence chain. Every input (base model, training data, teacher models, code) is listed with its terms, read from its primary source and dated. A release never carries a more permissive licence than its inputs allow.
2. File census. Exactly the declared files ship, in safe formats (safetensors), each with its hash.
3. Numeric parity. The released files compute what the training code computed, checked to exact equality where the math allows it.
4. Clean install. The quick start on the model card runs, as written, in a fresh environment with no cached files and no access token.
5. Held-out evaluation. Every number on the card is measured on data that nothing was fitted to, and traces to a public raw result.
6. Plain-text review. The card says what the model is, what it is for, and where it falls short.
7. Post-publish check. After upload, the hub copy is re-checked against the gated hashes and reloaded from scratch.

How to verify a release yourself: open its folder in the AbstractPowered/manifests dataset (published with the first release). It lists every file's hash, the environment each check ran in, the exact code version, and where each number on the card comes from.

---

Licences

Each release carries its base licence forward. Non-commercial releases (for example LoRAs on non-commercial base models) are marked at the top of their card. Data-derived releases (for example feature banks) list each source's terms. Our own code is MIT unless its card says otherwise.

Responsible release

• Behaviour lives in modules that can be audited, switched off and measured.
• Every release is reproducible from its manifest.
• Official chat-grade releases of the Beatrix line will carry an ethics component trained into the model, a tamper-evidence evaluation, and the published cost of removing it. Early checkpoints are released as research previews and say so.

How to cite

Every repository carries a CITATION.cff. For a release:

@misc{abstractpowered_<release>_<version>,
  title  = {<release name>},
  author = {AbstractPhil and Claude (Anthropic)},
  year   = {2026},
  url    = {https://huggingface.co/AbstractPowered/<release>}
}

Working with us

All releases and demos here are public and open to everyone.

• Research groups: reproduce or extend a release from its manifest; joint studies on modular training and measurement are welcome.
• Labs and companies: trial the detachable-module approach on your own models, or the feature banks for distillation.
• Sponsors and foundations: support open reports on compact, modular training and its costs.

Partnerships and talks: on request.

---

Earlier notes on this page described an "Infinity-CFG" guidance module; it is retired. Our current guidance work is the CFG Megapack, a ComfyUI bench for comparing published guidance methods on one engine.