From f671409644de086e203893ff587b63b71dc1f2d0 Mon Sep 17 00:00:00 2001 From: Cursor Agent Date: Fri, 4 Sep 2026 20:27:17 +0000 Subject: [PATCH] docs: add ELI5 blurb at top of README Place a short, honest public-proof summary after the title so GitHub visitors see the 11-model Crashlanded spread (felix, scoring 1.1, N=1) before the technical lead. Co-authored-by: Jason --- README.md | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/README.md b/README.md index 06a18d1..df14ac6 100644 --- a/README.md +++ b/README.md @@ -1,5 +1,9 @@ # RLE — RimWorld Learning Environment +## ELI5 + +RLE is a public research benchmark — not a product — where AI agent setups try to keep a RimWorld colony alive. The published leaderboard is one 11-model Crashlanded run (2026-06-11, N=1, seed 42, 10 ticks) on the `felix` harness at scoring 1.1 — Grok 4.3 leads at mean composite 0.836; it is content-first and not statistically valid. Winners advance to N=4; N=4 is not published. Those rows are not comparable to scoring 1.2. Harness × model and scoring 1.2 are how the code works now; that matrix is not a published result yet. Full table: [Benchmark Results](#benchmark-results). + A **harness × model** benchmark: swappable agent harnesses manage a RimWorld colony under uncertainty and are scored on the same footing against an unmanaged baseline. Think [FLE](https://github.com/chenhao-wang/FLE) (Factorio Learning Environment) but stochastic, multi-agent-capable, and with the *harness* — not just the model — as a first-class variable. ## What makes this different