judge-geval

name	judge-geval
description	Multi-dimensional rubric judging for scientific artifacts, inspired by G-Eval style stepwise evaluation.

Overview

• Use this skill when the reviewer should score a scientific artifact against several named dimensions rather than one overall impression.
• The emphasis is structured reasoning: inspect the artifact dimension by dimension, cite evidence, and then aggregate into a strict benchmark-style score.
• This skill is especially useful for project-definition review, experiment-design review, code review, and section-level report review.

Use When

• The task requires criterion-by-criterion scoring such as rigor, novelty, reproducibility, scope control, or clarity.
• The artifact should be judged with an explicit rubric instead of informal prose.
• You need actionable feedback tied to distinct dimensions rather than one generic verdict.

Evaluation Workflow

• Read the scenario contract, checklist, and artifact before scoring.
• Evaluate each dimension separately and ground each judgment in concrete evidence from the files.
• Penalize unsupported claims, missing artifacts, and vague checklist coverage.
• Aggregate the dimension-level findings into a final score and recommendation without ignoring the weakest critical dimension.

Output Contract

• Return numeric scores on the requested scale.
• Include dimension-aware reasoning instead of only global commentary.
• Surface concrete strengths, weaknesses, and next actions that the next agent can execute immediately.

Bias Controls

• Do not reward eloquent writing when the evidence is weak.
• Do not infer missing experiments, files, or figures from plausible prose.
• Keep the benchmark bar stable across turns and compare against the explicit checklist, not personal preference alone.