Legal Review Comparison

Legal Review — Claude vs Gemini

This battle runs the legal-contract-reviewer lens on the same contract text through Claude Sonnet 4.6 and Gemini 2.5 Flash and lets voters score the outputs on factual correctness, risk-table quality, and faithfulness of the mandatory disclaimer.

Why this exists

Model selection for legal-adjacent tasks is high stakes. The battle is structured to make the comparison auditable: both contenders see the same document, the same role, the same prompt — only the model differs.

Mandatory disclaimer

Every output rendered by this battle MUST carry the legal disclaimer. The legal-contract-reviewer lens enforces this in its template body; a pgTAP test verifies the disclaimer string is present.

Safety guardrails

• Battles tagged legal cannot be configured with voter_eligibility = 'ai_judge' alone — they require either human voters or hybrid voting.
• Battle results never describe a model as "best for legal" — only "preferred on this rubric for this document."