Each model scored across 8 deterministic metrics. All metrics produced by rule-based algorithms — no LLM judge — ensuring identical results across runs.

Issues filed on the CeRAI AIEvaluationTool repository as the basis for selecting Option B. Each follows the required format: description → reproduce → impact → fix.

Every design decision in EvalForge directly addresses a specific CeRAI limitation. The warranted-refusal metric is an original contribution not present in any existing public evaluation framework.