Design an eval suite that simultaneously detects regressions in safety and helpfulness after any data update.

A robust eval suite should include a fixed red-team set of approximately 500 harmful prompts to measure safety recall and a fixed borderline set of 500 benign-but-risky prompts to measure false-positive rates. To ensure capabilities are maintained, include MMLU and a chat-style benchmark. Model releases should be gated on passing defined thresholds for all four metrics, and these should be tracked on a single dashboard to ensure that no individual regression is masked by an aggregate score.