An engineering manager tested using Claude to audit a team's codebase over four months, aiming to objectively measure what was delivered, its complexity, and development time. Traditional metrics like story points often incentivize quantity over quality, so this experiment sought a more impartial assessment by analyzing actual code changes rather than commit messages.

The audit, guided by a detailed six-phase prompt, produced a comprehensive report. It found most deliverables were rated MODERATE or lower complexity, with timelines 2-10x longer than senior-engineer benchmarks. Rework consumed a significant portion of effort, and the overall efficiency calculation suggested room for improvement, though the assessment may underestimate real-world integration challenges.

This experiment raises critical questions about AI's role in performance assessment. While LLMs can cut through human bias and identify patterns, they may miss hidden production complexities like legacy system debt or coordination overhead. The results prompt debate on whether such tools should complement existing metrics like DORA or influence compensation and promotions.