3rd Place
Method
Using Astra to beat Astra.
Core trick
Model iteratively rewrites a game-playing procedure, runs it against a fixed final score, and keeps versions that improve the result.
Description
Using Astra to beat Astra. Continuous policy search for long-horizon tasks. I used Astra to design a strategy to play factory that runs faster and more accurately on repeated iterations than just prompting Astra.
View results here https://method-ai-hq.github.io/method-factorio/
My goal is to use this to make vertical ai (AI for X) faster and more accurate, especially in tough to verify domains. that's what my startup is doing (hill climbing in hard to verify domains, withmethod.ai).
README excerpt
# Method × Factorio [Read the experiment story online](https://method-ai-hq.github.io/method-factorio/). [Download the illustrated experiment story](story/complete.html) and open it in a browser. This single HTML file includes the result charts and three reviewed game replay videos. See the [story guide](story/README.md). Can Astra invent and improve executable Methods that finish long tasks when the final result is clear, but the value of each action is uncertain? Our task is to build a factory and launch a rocket in base Factorio. Astra will design the playing procedure, inspect its results, and test changes. The procedure can use plain-English instructions, agents, model calls, bounded loops, and code. It must choose where reasoning and verification are worth their time and cost. **Status: the supplied-kit policy search is finished.** The [final report](evidence/policy-search-final-2026-09-10/README.md) covers direct Astra, deterministic scripting, and 16 new Methods made from game feedback. The earlier selected code Method completed two unseen maps in a mean 9.06 seconds, versus 51.78 seconds for the initial agent Method. The later hill climb reduced actions from 13 to 10 and coal loaded from 130 to 9, but did not show a further speed gain. No advantage over the deterministic reference script was established. This is a small production test, not a rocket launch or a fresh-world win. ## Harder task: automatic science The [red-and-green science task](docs/science-tas