evaluation-design

To design an AI safety evaluation, import your measurement targets into the Skill to initiate the review process. It scores designs against a 7-dimension rubric, helps articulate risk questions, and evaluates model uplift.

The EquiStamp Phase 3 evaluation pipeline is the design and review stage for AI safety assessments. It processes risk estimation and causal pathway data from Phase 2 to design new evaluations or review existing designs against a 7-dimension rubric.

Yes, risk estimation and causal pathway data from Phase 2 are required before starting AI evaluation design. These measurement targets serve as the necessary inputs to assess model uplift and validate the design against the 7-dimension rubric.

Yes, you can review existing AI evaluation designs. The Skill supports reviewing existing designs by scoring and validating them against the 7-dimension rubric, alongside designing new evaluations from Phase 2 measurement targets.

Model uplift assessment in AI evaluation works by measuring the incremental capabilities or risks a model introduces. This Skill incorporates uplift assessment to validate the evaluation design against the 7-dimension rubric using Phase 2 causal pathway data.