Claude 4.7 runs a second pass on its own output before showing it to you. If the internal reviewer finds issues, the model revises. You see the corrected version. The result: zero hallucinated bugs in code review, and SQL queries that are right rather than just confident.
What self-verification actually does
The mechanism
The SQL test that shows why this matters
4.6 vs 4.7 on the same prompt
The code review test
Zero hallucinations vs three
The honest trade-off
Speed and tokens