How Estragon-9B went from 38% to 83% on an engine-verified GDScript benchmark for ~$150: the dataset, the RL cycles that failed, the one that worked, and honest quant numbers.
Measured benchmarks, the KV-cache math that makes 8 GB fit, and a Windows 11 setup kit that takes an RTX 3060 Ti from blank slate to chained tool calls against a locally-hosted 9B model in one PowerShell command.
I built a complete RTS game in Godot using an experimental workflow: multiple AI sessions working autonomously on separate tasks, with automatic testing and merging. Here's how it works and what I learned.