The AI judge that called a half-finished audit 'exhaustive'

Short summary

An AI judge without ground truth conflates confidence and correctness. The author discovered this flaw when benchmarking coding agents: the judge rated a 44%-complete audit as 'exhaustive' because it was well-written. The lesson applies everywhere LLMs grade LLMs—code review, ticket scoring, RAG verification—and the fix is reference-aware grading with verifiable answer keys.

• •An AI judge without reference data scores based on confidence and fluency, not correctness—a polished wrong answer scores like a right one
• •The author's benchmark nearly published false results when an LLM judge called a 44%-complete audit 'exhaustive' because it was well-structured
• •Fix: give judges ground truth (answer keys) and require precise matching (file:line citations), not vague credit for approximations

Generated with AI, which can make mistakes.