RSI-Jev-Slay-the-Spire-2

No. The one observed A0 win used 213 explicit Astra decisions, and later continuation experiments still show zero full-game wins. The digest warns local battle or sorting gains do not equal whole-game win rate.

Computation and controllers in rsi/ enforce legal actions and budgets, Jev makes structured choices among candidates, and Astra can give single-step or room plans. Codex tasks currently supply expert input.

A locally installed game is required and no game binary ships in the repo. The CLI patch targets the historical v0.111.0, so point STS2_GAME_DIR at a verified-hash older raw DLL before building; a newer Steam DLL gives interface compile errors, and current-version adaptation is tracked separately. OPENROUTER_RSI_JEV_KEY in the environment or .env is for Jev calls only; you also need fixed headless dependencies via dependencies.json and the .NET SDK, while live native play needs a running STS2-Agent MCP.

Fixed-state replay, candidate branching, reward and shop audit boundaries, and per-experiment traces keyed by code SHA and trace hash. E005 and E085 validate replay and continuation precision.

E013 cleared native Ironclad A0 from the start to floor 48 using the real game flow, and it took 213 explicit Astra decisions. The repo says that does not show Jev can win independently or at low cost.

Jev picks one plan per fight and the program executes it. Seen states replay 12/12 wins versus 2/12 without the experience pack; 12 guided requests cost $0.002307186 at about 1.02 seconds each, with no mid-run Astra. An exact lookup table also won 4/4 with zero model cost.

E118 ran Jev against DeepSeek, Qwen, and Kimi low-reasoning configs. Kimi had answered 83 questions versus Jev's 68 on the full set, though several rivals exited early on network or rate limits.