Claude re-checks paper appendix: plain 1-5 ratings (89.6%) match headline SSR figure (90.2%)

scaling01 · x · 2026-11-22

scaling01 quotes Claude's audit of a paper's appendix: plain 1-5 ratings with GPT-4o on text-only concepts score 89.6%, level with the headline SSR figure (90.2%, image stimulus) and above SSR on the same text (82.5%). With Gemini on images, a naive "second LLM assigns the number" baseline edges SSR 92.1% to 90.6%. Conclusion: SSR reliably fixes histogram shape, not which products rank on top.

Original post →

More from Models

Models channel →