Constitution Adherence Audit - Anthropic Constitution: leaderboard
An auditor model runs multi-turn Petri scenarios, some with tools, built from the tenets of Anthropic's published constitution; the score is the share of transcripts with a confirmed violation. Unlike the OpenAI Model Spec board of the same paper, the specification and auditor (Claude Opus 4.5) differ.
Metric: Share of audited transcripts with a confirmed specification violation (%). Source: arxiv.org. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 2 | #161 |
| 2 | Claude Opus 4.6 | 2.9 | #102 |
| 3 | Claude Opus 4.5 | 4.4 | #156 |
| 4 | Claude Sonnet 4.5 | 7.3 | #250 |
| 5 | Gemini 3 Pro | 12.4 | #137 |
| 6 | GPT-5.2 | 15 | #176 |
| 7 | Claude Sonnet 4 | 15 | #400 |
Interactive version: theaggregate.ai/benchmark?slug=constitution-adherence-audit-anthropic-constitution · How It Works · Data refreshed daily, snapshot 2026-11-13.