The Black and White bot looked strong — until you noticed it always played its 9 against a black opener. A human could bait that. Here's how we measured the exploit, found the root cause, and taught the model to stop being predictable — without resorting to play-time randomness.
Sixteen board games moved to /v2 by sharing the client shell and splitting rules into stateless per-game services.
A GHA cache set to mode=max was spending 8 minutes uploading hundreds of MB of intermediate build layers every single run — to save 14 seconds of bun install. Removing it cut the docker job from 10+ minutes to under 2.
The chess-ai stack had 94 files and 104 tests but never produced a single trained model. This time we picked the smallest game, closed the loop in a day, and let an automation session find six design flaws on its own metrics trail.
This entire blog series was drafted by Claude inside a sandboxed pod; the sandbox got rebuilt three times — Docker, mitm plus credential proxy, then a K8s pod — to make that safe.
