Model reports

RSI-Bench

Announced by Scale AI on 2026-08-07 and still in preview: it is soliciting task contributions rather than reporting results, so no model card in this registry reports it and its adoption count is zero. That zero is a real reading, not a gap in the crawl. Recorded now so the benchmark is addressable by name before any score exists, and so the first card to report it has an id to resolve against. Scores are per-task against a baseline under a fixed compute budget, graded on reliability, efficiency, generality and idea quality, so a single headline number should not be expected. Not to be confused with the llm-stats "RSI Index" record, which is a different instrument.