Anthropic maps reasoning for AI safety

Anthropic and Redwood Research introduced the Conceptual Reasoning Index, combining three benchmarks that test argument judgment, consistency, and decision theory; its LMCA component contains 560 position texts and 1,461 counterarguments. The research release says the index weights LMCA at 60%, ACCoRD at 20%, and DTBench at 20%, with scores rising roughly linearly since late 2024 and DTBench already near its ceiling. For safety teams, the useful artifact is not a leaderboard rank but a missing deployment check: a model that writes persuasive safety analysis still needs separate tests for whether its beliefs stay consistent when the answer cannot be empirically verified; the site’s reproducibility gate for Claude’s math work makes the same distinction.