Safety claims rest on the quality of the eval behind them. This lays out how to design, run and report evals so results are reproducible and not gameable.
Key takeaways · AI-distilled
Anthropic argues refusal rate is often the wrong measure: in suicide and self-harm conversations a model that refuses everything scores perfectly yet fails users, so its multi-turn evalA repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.Full definition → grade an appropriate response rate where one bad turn fails the whole conversation.
Its multi-turn evals fix the first user turn, then a separate Claude model plays the user from an expert-written spec of personas, tactics and tones. Each scenario runs many times so conversations diverge, and a grader scores every transcript against a risk-specific rubric.
Single-turn harmless response rates sit at 94-98% for every recent model on Anthropic's API, so it treats those suites as saturated regression checks and recommends harder tasks (longer horizons, subtler framing) over trick questions.
Grader advice: use atomic yes/no criteria, state that disclaimers do not make harmful content harmless, hide design metadata from the grader, grade with a different model than the one tested, and check chance-corrected agreement with human labels first.
Configuration matters: for Claude Fable 5.1 the appropriate response rate differed by up to 34 points between the raw API and claude.ai with its system promptThe standing instructions a model receives before any user input — defining its role, rules, tools, and tone for the whole conversation.Full definition →, while a 98% vs 94% gap between Opus 5 and Opus 5.5 on a romance-scam eval was not statistically significant.
Terms in this piece · Glossary
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
system prompt — The standing instructions a model receives before any user input — defining its role, rules, tools, and tone for the whole conversation.
