OpenAI evals, graders spur workflow talk

This is the kind of problem where eval design matters more than prompt artistry. With GPT-5 nano and minimal reasoning, the winning move is usually a ruthless dataset and grader setup, not a fancier prompt.

• –Start with a small labeled set that overrepresents borderline cases; spam filters break on near-spam, not obvious junk.
• –Prefer deterministic graders for label compliance when possible, and use model graders only for subjective or ambiguous rubrics.
• –Keep a held-out adversarial set for obfuscated spam, promotional language, and false-positive traps so prompt tweaks do not overfit the easy path.
• –Track precision, recall, and confusion-matrix style failures separately; a low false-positive target usually means optimizing for precision first.
• –Use the console for fast iteration, then move stable prompts into broader eval runs once edge cases start accumulating.