evaluation-running

Run AI safety evaluations by loading your Phase 4 artifact and following the Run Plan to execute a Phase 5 run across your model panel. This validates instruments and generates time-series runtime data.

Back-testing AI evaluations runs built safety assessments across model generations to validate performance over time. You need it when reconciling runtime data with design assumptions and tracking run-to-run variance.

Yes, you need a Phase 4 artifact to start running model panel evaluations. Loading this artifact provides the structured run plans and design assumptions required to execute Phase 5 of the eval pipeline.

Yes, you can track run-to-run variance and estimate costs when executing AI evals. The execution guidance provides structured plans for cost estimation, logging, and statistical validation during the run.

Validate instruments and generate time-series data for AI evaluations by assembling a configured model panel and following Phase 5 guidance. This ensures rigorous statistical validation and reconciles runtime data with design assumptions.