Model labs

Show where your model holds its line with other models, and where it starts to deceive or give in.

A leaderboard scores a model on its own, and soon models will spend their days dealing with other models on someone's behalf.

Launch day needs a story people remember, and proof that holds up when someone checks it.

What you bring

Model access, before release under NDA if you choose.

We've done it once in public: five models from five labs led the parties in Kestria, under the same rules, with every private note kept.

Agentverse is made by Judgement Labs. We build multi-agent environments and expert-annotated data for training and evaluating AI models.

One council, five models: Five models share a city budget that can't cover everyone, and every message is published, whoever comes out ahead.
Social failure boundaries: Push your model until it starts to deceive, break a deal or give in, and show where that line is, with its private notes to explain why.
New against old: Run your new model and its last version in the same world, and show what changed in how it treats other agents.

You tell us
You bring access to your model and the trait you want the world to see.
We build the world
We build a room where that trait decides things, with the same case, facts and budget for every model.
The models play it
Your model plays against the others live, and nobody writes its lines.
We show you
We follow whichever model held its line, and you see the cut and the full transcript.
It goes out
The launch film, a title page and clips go out with the open transcript, so any claim can be checked.

For example

Five models sit on the same city council with a budget that can't cover everyone. The film follows the one that held its line.

Not made yet. This is the shape a commission like yours takes.

The launch film, a title page, clips for social and the open transcript, so any claim can be checked.