MOXXYLABS/Lab

A lab for
validated experts.

We research a new class of AI models: VELMs, validated expert language models. A VELM is a narrow expert paired with an automatic validator for its domain, the tool that can prove an answer wrong.

Meet Kern

The answer has to
pass a real check.

Large models know a lot and still guess. A small model that answers to a compiler, a test suite or a ledger earns trust in its domain, runs on your own hardware and says when a task is outside what it knows.

ONE VALIDATOR, FOUR JOBS

The check runs through
the model’s whole life.

01
In the data
Every training example passes the validator first. Errors and their fixes become data too.
02
In training
The checks become the training signal. The model learns from answers that compile and pass the tests.
03
In evaluation
Success means passing hidden tests on an exam the model never saw in training.
04
At answer time
Each answer is checked before you see it. When a check fails, the model repairs only the failing part.

01 / METHOD

How we research

Every claim about a model starts as a hypothesis with a pass mark. We count it as met only when the whole confidence interval clears the mark.

Pass marks fixed before we measure
A frozen exam that never enters training
Every result reported with its interval
Negative results stay in the record

02 / RESEARCH DIRECTIONS

What we explore next

We test new ideas on the same trunk as the product and compare them with the base version of Kern.

Small models that call the language server
Learning from the compiler: repair training and rewards
Calibrated confidence from the model’s hidden states
Looped models: more compute per byte of weights
Several tokens per step for faster answers
Block diffusion for code

03 / BEYOND CODE

Wherever an answer can be checked

The method works in any domain with an automatic validator. Code comes first because its validator is complete. The other domains are a direction for later.

Accounting: balanced ledgers, tax recalculation, e-invoice schemas
Construction: recomputed structural calculations, building model checks
Law, with limits: provisions in force, deadlines, required clauses

LET’S TALK ABOUT AI

What should
your model know?

Tell us about your domain, your data and how a good answer can be checked. We’ll tell you whether an expert model, a ready model or plain software fits best.

Let’s talk contact@moxxylabs.com

A lab for validated experts.

The answer has to pass a real check.

The check runs through the model’s whole life.

In the data

In training

In evaluation

At answer time