RLE

介绍

RLE：在实时 RimWorld 殖民地上进行 × 模型基准测试。可交换的代理 harness（Felix 多代理、通过 MCP 的编码代理，或自定义插件）在未管理基线下以相同方式评分。FLE 用于不确定性下的多代理协同。

AI 翻译

取自仓库 README

README

正在加载完整 README。 以下为已保存的摘录。

RLE is a public research benchmark — not a product — where AI agent setups try to keep a RimWorld colony alive. The live published board is spread-2026-09-07 on rle.appsprout.dev: scoring 1.2, harness × model, Crashlanded, N=1, seed 42, 10 ticks, 12 rows (Sep 6 grok-4.6 four + Flash 8-cell). Grok Build (ACP) leads at mean composite 0.811 (final 0.833). N=1 is content-first and not statistically valid. Winners advance to N=4; N=4 is not published. spread-2026-09-06 is a prior 4-row snapshot. The June 2026 scoring 1.1 felix 11-model table is historical and not comparable (different scoring era).