Skip to content- Sources: primary, discussion
- Summary: The post, dated 2026-08-12 with four Redwood Research authors and two from Anthropic, describes three benchmarks and a weighted aggregate. LMCA carries 560 position texts and 1,461 arguments with 2,140 human ratings and measures how closely a model's ratings match the researchers', ACCoRD measures logical consistency across separately asked probability and preference questions using 567 of about 14,000 generated constraints, and DTBench capabilities is 407 handcrafted decision-theory questions, weighted 60, 20, and 20 percent. Scores are as of 2026-08-10, the authors put the practical ceiling near 91, the top score is Opus 5 at 73.6 with a 95 percent confidence interval of plus or minus 2.1, and the methodology section states that Fable 5's score used Opus 5 as a fallback where Fable 5 refused, which is a caveat on that one figure.
- Why it matters: The benchmarks target the class of task where no verifiable answer exists and training feedback loops cannot be built, which is the gap the authors argue standard evaluation misses.