Skip to main content
Try JUWEL

Vext Labs / Human intent + machine

AI that starts with what you want.

Vext Labs is an independent AI lab. We show how we test each idea, what we found, and where it stops working. Then we build JUWEL, an AI you talk to, on what held up.

Read the research Meet the lab
Latest2026-07-02 · Paper · Beyond the Scaling Ceiling

6 papers · From our public paper list · As of 2026-09-24

01 / Latest from the lab

New from the lab.

See all papers

  1. 01

    Paper · 2026-07-02 · ≈25 min read

    Beyond the Scaling Ceiling

    A study of how an AI model could grow new skills in checked steps. Every result says how far it goes: tiny model, working code, or waiting on big chips.

    • workshop draft
    • PROVEN-toy
    Read the paper

  2. 02

    Paper · 2026-07-02 · ≈17 min read

    The Accretion Model

    A kind of AI model that grows by adding new parts, never by changing old ones. A stranger can check its whole history. We built a tiny version, Cultivar-0, on a regular computer.

    • BUILT/VERIFIED-CPU
    • PROVEN-toy
    Read the paper

  3. 03

    Paper · 2026-06-29 · ≈8 min read

    Sound Compounding Ratchet

    A way to keep a gain only when a checker confirms it. Plus a test for when a model thinks it passed but did not. Every win here is on tiny models only.

    • PROVEN-toy
    • $0-CPU kill-tests
    Read the paper
See all papers

Read times assume 230 words a minute · Dates come from each paper · As of 2026-09-24Read times at 230 words a minute · As of 2026-09-24

02 / All our papers

02 / All our papers

The questions are public. So are the limits.

We show how we tested it, what we found, and where it stops. Ideas that failed stay on the list.

From our public paper list · Hosted on Zenodo, free to reuse with credit (CC-BY-4.0) · As of 2026-09-24

Our public papers, in list order
TitleTypeReadDate

Workshop draft · Zenodo · 2026 · 17 pp

Workshop draft · ≈25 min2026-07-02

Beyond the Scaling Ceiling

A study of how an AI model could grow new skills in checked steps. Every result says how far it goes: tiny model, working code, or waiting on big chips.

DOI 10.5281/zenodo.21628058 		Paper≈25 min2026-07-02

Tech report · Zenodo · 2026 · 13 pp

Tech report · ≈17 min2026-07-02

The Accretion Model

A kind of AI model that grows by adding new parts, never by changing old ones. A stranger can check its whole history. We built a tiny version, Cultivar-0, on a regular computer.

DOI 10.5281/zenodo.21628429 		Paper≈17 min2026-07-02

Paper · Zenodo · 2026 · 8 pp

Paper · ≈8 min2026-06-29

Sound Compounding Ratchet

A way to keep a gain only when a checker confirms it. Plus a test for when a model thinks it passed but did not. Every win here is on tiny models only.

DOI 10.5281/zenodo.21628552 		Paper≈8 min2026-06-29

Code and data you can rerun · Zenodo · 2026 · 9 pp

Code and data you can rerun · ≈10 min2026-06-27

The Leakage Signature

A record of ideas we proved wrong. Plus a way to grow a model that a math solver checks, with an audit for leaks. Small, limited demos only.

DOI 10.5281/zenodo.21628524 		Paper≈10 min2026-06-27

Map of results + PROPOSED plans · Zenodo · 2026 · 13 pp

Map of results + PROPOSED plans · ≈15 minNo date

Verifier-Centric Capability Growth

A map of what worked and what failed when we added skills to a model. Plans we have not tested yet are marked PROPOSED.

DOI 10.5281/zenodo.21628544 		Paper≈15 minNo date

Early draft · Zenodo · 2026 · 19 pp

Early draft · ≈32 minNo date

Capability Injection via Reverse Abliteration

An early take on our way to add a skill to a model (CIP). It runs a known removal trick, abliteration, in reverse. It is a research method, not a product feature.

DOI 10.5281/zenodo.21628566 		Paper≈32 minNo date

Showing 6 of 6 · 4 have a date · 2 have no dateShowing 4 of 6 · As of 2026-09-24

See all papers

03 / Principles

Human intent + machine.

  1. Start with what you want.

    First we write down what you want, and which choices stay yours.

  2. Show every step.

    We test what the AI can really do, and where it stops.

  3. You stay in charge.

    A good result lets you check it, fix it, and choose.

Read how we think: Human Intent

04 / JUWEL · The AI supercomputer OS · In beta04 / JUWEL · AI supercomputer OS · In beta

JUWEL is the AI supercomputer OS you talk to.

Gems are the AI models you use. Stones are helpers that do tasks for you. One loop, the JUWEL Loop, runs it all. JUWEL asks before it remembers anything, and you can take it back. See access and pricing on juwel.ai.Gems are the AI models you use. Stones are helpers that do tasks for you. One loop, the JUWEL Loop, runs it all. JUWEL asks before it remembers anything.

Press Enter to open juwel.ai with your words typed in · You choose when to send · This site sends nothing

Try JUWEL
A design of the juwel.ai home screen. It shows a greeting that asks “What’s on your mind?”, a message box, and a starter card called Plan my week.
A design of the “What JUWEL knows” screen. It shows saved items, each with a delete button, and an undo message after one is removed.
A design of JUWEL on a phone. It shows the first reply when you ask it to plan your week.

Design pictures from 2026-09-24 · Not screenshots of the appA design picture from 2026-09-24 · Not a screenshot of the app

05 / Work with the lab

Want to work with us? Just write.

Want to use our research in your own work? Email the lab. There is no sales form and no set of packages to pick from.Want to use our research in your own work? Email the lab. No sales form. No packages.

[email protected]Human Intent

One inbox for now · Press, research and partner emails all come here

06 / Careers

A small lab. Hiring when funded.

No open jobs today (as of 2026-09-24). We still read every serious note. We keep the good ones for when a job opens.

[email protected]Careers

One inbox for now