-
0. Overview
A warning, the dragons game, why it reminded us of NYPC, and a map of this postmortem series.
-
1. RL
Our first plan was reinforcement learning: the model, how we trained it, and why a scripted bot ended it.
-
2. Mass code generation
How we replaced one trained bot with hundreds of LLM-written variants, a local league to rank them, and what that league got wrong.
-
3. The arena and vast.ai workers
tokenz-arena, the team's shared league on the official judge: how its pieces fit together, the life of a submission, the workers it rented on vast.ai, and the C++ judge.
-
4. Metrics engine and probability analysis
How the arena turned every replay into per-round metrics and a P(A wins) curve, how well that curve predicted real games, and what it taught us about the top teams.
-
5. The LLM loop up close
Who did what in our agent loop (orchestrator, analysts, strategists, workers), what a brief looked like, how verdicts were judged, and the rules we added each time it went wrong.
-
6. What the factory made
The bot lines we bred between 09-28 and 10-09, what each one did, and which of them moved the server rating.
-
7. Postmortem and results
Where UnderTheC ended up on the ladder and in the sprint tournament, the numbers behind it, and what we would change.