Pick your LLM-judge shape by whether you need hard gates: DAG for deterministic branches, G-Eval for subjective criteria

DeepEval·medium signal

Use a DAG-based judge when evaluation needs deterministic branches, hard pass/fail gates, or multi-step scoring logic (e.g., 'if no citation, fail regardless of fluency'), and use G-Eval when you're scoring fuzzy, natural-language criteria like helpfulness or tone. Choosing G-Eval where you actually needed a gate is why judges 'pass' outputs that violate a non-negotiable rule. For builders: model your rubric's structure first — branching requirements want a DAG, holistic quality wants G-Eval — instead of writing one prose prompt and hoping the judge respects your hard constraints.