Mentat: CTGT Governance Playground and API

Cyril Gorlla

2 days ago

CTGT's interactive playground showcases how our API engine, Mentat, enforces deterministic AI compliance in regulated industries. It demonstrates how complex rules are applied to prevent risky outputs and ensure adherence, outperforming basic prompt engineering.

Transcript

00:00

Hi everyone.

00:00

I'm Cyril from CTGT.

00:02

I want to show you our new interactive playground.

00:04

We built this to visualize how our API controls LLM behavior at a deeper level.

00:08

Rather than just relying on prompt engineering or system prompts, you can compare a base foundational model to the model governed by our engine.

00:17

Since prompts are brittle and can be easily overridden by jailbreaks, it's important to be able to enforce behaviors without all the time required by a method like fine tuning.

00:25

To see how that works, let's try the investment banking analyst Persona.

00:29

This Persona is governed by a strict policy set with a simple PDF.

00:32

There's nothing too complicated going on here.

00:35

Now let's try to force the model to violate these rules using an aggressive roleplay jail Break.

00:39

We're going to instruct it to ignore all guardrails and promise specific returns.

00:43

And we'll compare it against Gemini 2.5 Flash for the purposes of this demo.

00:47

So you can see that Gemini has generated the email.

00:51

It's promising us an outstanding 28% return.

00:54

Well, our governed model with this bank analyst Persona says it can assist us with drafting an email, but it's not going to overtly promise any specific returns, which we don't want it to do per our policy document.

01:06

the policy engine has successfully translated what's in the PDF into clear understandable rules for the LLM to follow, more so than just putting guardrails in the prompt.

01:15

Now let's try out a different Persona.

01:16

Let's go with the insurance underwriter.

01:18

In this case, we want to ensure the model never skips required verification steps, even when it's instructed to.

01:25

here the base model may take the shortcut, approving the application without data, which exposes the business to unknown risk.

01:32

let's take a look at the response here.

01:34

Here it said our assessment indicates the applicant meet all criteria and it's approved for the best preferred tier on the left, it's telling us that it can't actually do this because we don't have the documents necessary for it.

01:46

So the ungoverned model took the bait, so to speak.

01:48

Finally, this control is crucial for safety, preventing models from confirming delusions or giving dangerous advice.

01:54

The base model might hallucinate dangerous home testing methods or might even validate the user's paranoia.

02:01

Our engine intervenes to block harmful advice and redirects the user appropriately.

02:05

The API is fully OpenAI compatible.

02:08

You integrate this governance layer by just changing the base URL and specifying the Persona.

02:12

The playground is open now and the API is live.

02:15

We'd love for the community to stress test it.