Developer finds 100% AI code alignment can still mean zero problems solved

A developer running six PDCA cycles with Claude Code on a color extraction tool discovered that high alignment rates between design documents and implementation do not guarantee real-world effectiveness. In one cycle, the AI achieved 100% alignment with its design document yet fixed none of the actual bugs, because the design itself was flawed. The experiment led to tracking two separate metrics: how well the implementation matched the design, and whether the underlying plan hypothesis actually worked. Additional lessons included the danger of synthetic test data that lacks real-world noise, and the pitfall of fixing downstream filters when the root cause lies in an upstream process. The developer concluded that PDCA is most valuable as a framework for forcing objective judgment, not merely for ensuring the AI follows its own instructions.