BrickedUp!

Full disclosure Claude wrote all this below- please go watch the video I explained it better there :D

How do you train models on outcomes that aren't immediately verifiable and depend on luck? Math and code are easy to verify. The hard, valuable domains are semi-verifiable: the final outcome is concrete, but the intermediate decisions are luck-dependent and hard to judge. Catan is the perfect test case. Chess and Go are solved, but Catan isn't, because correct play depends on the unpredictable dice rolls that follow, so a "good move" is barely verifiable in the moment.

Our insight: you can make a semi-verifiable domain trainable by grounding the judgment in a verified expert. I happen to be the all-time ELO world record holder in 1v1 Catan on colonist.io, so the ground truth is my own play. I pulled my 100 most recent games through the replay feature and compiled my notes into a rubric that encodes my playstyle.

We built a fully autonomous loop. Qwen 3.5 8B plays itself 100 times in 1v1 Catan on Fireworks with reasoning on. Claude and GPT sub-agents cross-grade the transcripts against the rubric to surface failure modes, and I adjudicated the disputed ones myself. We took the top weaknesses, synthesized training environments by augmenting snapshots of my real games, and hill-climbed each with GRPO on Hud. We ran the full loop twice overnight.

The result: targeted placement training drove the biggest gain. The post-placements checkpoint beats the untrained baseline 88% on paired boards where we switch sides to cancel dice variance (all evaluated under a 300-turn cap), and it frequently wins both sides of the mirror.

On a stronger base model, this loop runs recursively until the model plays at my level. The framework applies to any semi-verifiable task with concrete outcomes but hard to judge intermediate steps, Catan is just the domain we proved it on.