An AI research lab working on how a model can grow new capability — and prove that it grew.
Most of the field makes a model better by making it bigger, or by retraining it. We work on a third path: add capacity to a model that already exists, train only the new capacity, freeze everything that came before, and refuse the result unless it passes a gate. Growth that cannot be verified does not ship.
| Product | juwel.ai — JUWEL, an agent workspace |
| Lab | vextlabs.ai |
| Papers | vextlabs.ai/research · Vext-Labs-Inc/research |
| Models | huggingface.co/VextLabsinc |
| Author | Annalea Layton — ORCID 0009-0001-3695-9569 |
Six preprints, each with a DOI, and each published together with its LaTeX source,
figure code, and underlying experiment data in
Vext-Labs-Inc/research.
| Paper | DOI |
|---|---|
| Beyond the Scaling Ceiling: Verified, Non-Forgetting, Provable Capability Accretion | 10.5281/zenodo.21628058 |
| The Accretion Model: A Growable, Non-Forgetting, Externally-Fueled Model Type with a Cryptographically Verifiable Growth History | 10.5281/zenodo.21628429 |
| The Leakage Signature: Falsification and Solver-Certified Capability Enlargement in Sequence Models | 10.5281/zenodo.21628524 |
| Verifier-Centric Capability Growth: A Measured Mechanism Map and a Research Program | 10.5281/zenodo.21628544 |
| Sound Compounding: A Verifier-Frontier Ratchet for Capability Acquisition and a Believed-versus-True Test for Self-Improvement Leakage | 10.5281/zenodo.21628552 |
| Capability Injection via Reverse Abliteration | 10.5281/zenodo.21628566 |
All six are CC BY 4.0 and collected in the Vext Labs community on Zenodo.
On scope. These papers describe a mechanism, not a frontier system. Most positive results are measured at CPU scale or on small models, and every paper states which. We make no state-of-the-art or frontier-parity claim anywhere. The negative results are the majority of the program — the published ledger accounts for 81 pre-registered experiments, 43 of which were killed, nulled, or found to be leakage — and those are published too, because a growth claim is worth only as much as the tests it survived.
Open weights on Hugging Face: three JUWEL bases and sixteen GEM specialists.
- JUWELS —
juwel-garnet·juwel-emerald·juwel-sapphire(vision-language) - GEMS — sixteen domain specialists composed on a shared base: alexandrite, aquamarine, bloodstone, citrine, jade, lapis, opal, pearl, peridot, ruby, sardonyx, spinel, tanzanite, topaz, tourmaline, turquoise
pentest-7b— security-testing model from our earlier work
The bases are derived from Qwen models under Apache-2.0, and each model card says so. We do not claim from-scratch weights or our own tokenizer.
Agent SDK — theron-agent-sdk.
Open-core client for building agents against our fleet. Any model, any provider. MIT.
Stoa — an agent-native memory specification and runtime: memory an agent can write to, query, and carry between sessions, with provenance attached to each fact. Spec CC BY 4.0, runtime Apache-2.0. These repositories are archived and read-only; the design is published for reference and the ideas continue inside JUWEL.
stoa-spec ·
stoa-sdk ·
stoa-bus ·
stoa-edge ·
stoa-graph ·
stoa-identity ·
stoa-conformance
Vext Labs began as a security company, building agents that ran real tools against authorized bug bounty targets. That work is where the verification discipline in the research above comes from: a finding you cannot reproduce is not a finding. These repos date from that period and are kept for reference.
vext-pentest-7b— 7B model for autonomous penetration testing. Parses 25+ security tools, plans, and classifies findings.vext-shield— detection suite for prompt injection and related attacks on AI systems.sentry-benchmarks— audited benchmark results with raw responses, test cases, and a SHA-256 receipt per score.
- Research and papers: vextlabs.ai/research
- Hiring: vextlabs.ai/careers
- Annalea Layton, founder — alayton@tryvext.com
Vext Labs, Inc. Incorporated March 2026. Built in Maryland.