Prove
Set a falsifiable boundary before declaring growth. Kill-tests and controls belong in the record.
Independent American AI lab · Est. 2026
An independent American AI research lab, studying how a model gains new skills without losing what it already knows.
Mechanisms, artifacts, and explicit scope labels.
A killed mechanism stays in the record. That is the credibility.
A practical AI product, built by the lab.
Research index
The archive holds both the mechanism and the limits. Continual learning (CIP) stays toy-scoped unless a paper says otherwise, and the scope sits beside every title.
Set a falsifiable boundary before declaring growth. Kill-tests and controls belong in the record.
Toy stays toy. CPU-built stays CPU-built. Proposed work never presents itself as a measured result.
Growth is additive. Existing work stays byte-frozen or is kept behind a regression gate.
Failures are part of the boundary. Source, tools, and the offline verifier make a claim testable. We publish either way.
How rule four stays honest: signed receipts ship in JUWEL today. Check one yourself, offline, with our servers unplugged.
The research program is public. JUWEL OS gives people an AI with its own computer for real work today, and signs what it does onto a record you can check.
We began in security, where an uncheckable claim is not enough. Today we build tools, publish research, and read every serious note.