Anthropic构建了一个30万查询的行为审计工具，因为模型行为会发生变化。以下是生产版本。

Anthropic has developed an internal tool named Petri, an automated behavioral auditing system designed to track how model behavior shifts across different versions and training runs. The system executed over 300,000 test queries, revealing 'thousands of direct contradictions and interpretive ambiguities' among Claude, GPT-4o, Gemini, and Grok. This development coincided with the Pentagon's CTO labeling Claude as a supply chain risk, noting that Anthropic's training constitution is 'baked into the model' and 'directly shapes Claude's behavior.' Anthropic confirmed that the 2026 constitution plays a crucial role in this process. This situation implies that developers using these APIs must exercise greater caution when shipping their products.