ShinsaEvolve LunarLander recipe evolution: exploratory and prospective search evidence

Abstract

Clean-room LunarLander data and code supporting When Training Fitness Becomes the Target: Auditing LLM-Driven Recipe Evolution in LunarLander. Includes 12 exploratory searches and 20 prospectively planned searches (10 pairs), 480 candidate recipes and saved policies, experimental protocols, seed manifests, 12,288 role-level final-test episodes, resource accounting, and offline numerical verification scripts. The cohorts have separate inference designs and must not be pooled for the primary comparison. The prospective mean paired difference is 95.3458 with an approximate 95% Student-t interval of [-59.3786, 250.0703]. Mixed and unfavorable results are retained. This package does not establish general superiority or equivalence of the selection methods, isolated reward-scale causality, peer review, or journal submission. No manuscript source or manuscript PDF is included. This is a disclosed public derivative: machine-specific paths in 32 configurations were replaced with package-relative paths, historical explanatory text was cleaned up, and dependent hash references were updated. Candidate programs, policy parameters, raw episode outcomes, and call/resource logs are unchanged. PUBLIC_DERIVATION.json records original and public hashes; public derivative hashes are not evidence of original freezing or external preregistration. Read START_HERE.txt first. Software retains Apache-2.0; author-owned data and explanatory documentation use CC BY 4.0. The retained numerical plot Figure_1.eps receives no additional reuse license. Existing licenses and third-party rights remain effective. See LICENSE_SCOPE.txt and FILE_LICENSES.json for per-file scope.