World Labs: expert evals for AI-made games, by game developers who ship.

AI has made it easy to generate a game. It still can't tell you whether the game is any good, or how to make it great.

Fun doesn't show up in a unit test. Game feel, pacing, and coherence decide whether anyone keeps playing, and they only register in the judgment of people who have shipped games and watched players bounce off them.

We put AI-made games in front of exactly those people, under blind, structured conditions, and capture their verdicts as clean data.

So that models and agents can build games that aren't just playable, but worth playing.

Evals for training models

We work with labs building world models, game-playing agents, and generative game systems. Expert judgment becomes rubric scores, written critiques, and preference pairs, ready for RLHF, reward models, and benchmarking, with every verdict traceable to a rater's specialty and years of experience.

Get in touch

Evals for game tools & agents

We work with prompt-to-game tools, NPC and agent builders, and studios shipping AI features. Blind expert evals of your output, scored on anchored rubrics and tracked release over release, so you know exactly what got better and what broke before your players find out.

Get in touch

Join our mission to keep games good

Bring your shipped-title judgment in level design, systems, UX, QA, or game feel, and get paid to decide what AI gets right and wrong about games.

Level design Systems design UX & onboarding QA & playtest leads Game feel

Seventeen questions. One harness.

From gameplay decisions to game economies: the full taxonomy of expert evals we're building toward.

The playtest is the eval.