Understand the idea first.
Inspect the evidence next.

Simple version

Sometimes the AI your product uses changes, even when your code does not.

In plain terms

If your product depends on an AI provider, small behaviour changes can affect answers, refusals, formats, and workflows. VertRule reruns the same tests over time so teams can see when behaviour has shifted, instead of finding out from a customer.

Simple version

Run a small AI model right in your browser and watch what it does, step by step.

In plain terms

This is an interactive demo that runs a small transformer model entirely on your machine — nothing leaves the page. As it generates each word, it records what every layer did, and you can step through the result afterward.

Simple version

We ran a very large AI model on a single machine and kept a careful record of how each piece of every answer was formed.

In plain terms

Kimi K2.5 is far too large to fit in normal memory, so it is loaded from storage piece by piece during a run. We captured five tokens of output formation against a frozen starting state, verified the capture itself did not change behaviour, and had the resulting bundle reviewed adversarially.

Simple version

Same careful map of how a model works, but for a model so large we had to stream it from a network drive.

In plain terms

Grok-1 has 314 billion parameters and uses a routing design that picks different sub-networks for different inputs. We ran it from network-attached storage, captured a bounded slice of its routing and attention structure, and ran adversarial checks against the result.

Simple version

A careful map of how a much bigger, more complex AI model routes its work internally.

In plain terms

Nemotron-3 Super has 120 billion parameters and a mixture-of-experts design that sends different parts of an answer to different sub-networks. We captured both the expert-routing decisions and the attention edges, and showed the map reproduces exactly across independent runs on the same machine.

Simple version

We carefully recorded a small map of how an AI model worked while it answered a question on an ordinary computer.

In plain terms

When a model produces an answer, lots of internal connections are exercised. We captured a bounded slice of those connections on a normal CPU, confirmed the recording itself did not change the answer, and saved the map so anyone else can reproduce it from the same inputs.