Developer avoids faulty grade fix by measuring the problem before changing code

A developer who built an automated code-grading tool for his AI-assisted projects was told by a peer that overlapping rules were double-penalising flawed projects, unfairly stretching grades downward. Before making any changes, he wrote a measurement script to check whether this double-charging was actually occurring across his 72 projects. Analysis of the top co-firing rule pairs revealed that most overlaps reflected genuinely separate issues rather than the same defect being counted twice. The predicted pattern — that the worst-graded projects would benefit most from deduplication — did not appear in the data, with the largest gains instead falling in the middle grade bands. He ultimately chose not to ship the fix, noting that the key grades at the bottom of the distribution rested on only three data points, making any conclusion unreliable.