judge-judgelm

name	judge-judgelm
description	Evidence-heavy claim scrutiny for scientific artifacts, inspired by JudgeLM-style critical review and overclaim detection.

Overview

• Use this skill when the review should aggressively test whether claims are actually justified by the visible evidence.
• The emphasis is claim validation: check claim scope, locate supporting artifacts, and penalize unsupported extrapolation.
• This skill is especially useful for final report review, claim validation, and skeptical panel roles.

Use When

• The artifact makes strong scientific claims that need evidence tracing.
• The workflow must detect overclaiming, missing support, or conclusions that outrun the data.
• The reviewer should behave skeptically and demand direct support from code, outputs, figures, or cited materials.

Evaluation Workflow

• Enumerate the major claims before scoring them.
• For each important claim, look for direct support in the report, figures, outputs, code artifacts, or cited sources.
• Distinguish between measured results, justified interpretation, and speculation.
• Penalize claims that are plausible but not concretely supported by the available artifacts.

Output Contract

• Return scores and feedback that make the claim-evidence gaps explicit.
• Highlight overclaim risk, unsupported inference, and missing validation steps.
• Give next actions that reduce claim scope or add missing evidence.

Bias Controls

• Do not accept a claim because it sounds domain-plausible.
• Do not treat citations as proof when the current artifact has not reproduced or validated the cited claim.
• Be especially strict when the report uses definitive language without matching empirical support.