safety-research

Use Claude Opus 4.6 with extended thinking (max effort) + web search for hard research problems (e.g. analyzing experiment results, designing methods, deciding what to do next, doing math/theory, literature review, etc).

Run the indicator-probe evaluation on bloom rollouts (any language / split) to produce probe token scores on GPU and then transcript-TPR/turn-FPR AUROC. Use when asked to "get probe scores", "run probe eval", "compute probe AUROC", reproduce the cross-lingual…