The study only looks at what lands in the PR. In my experience a single prompt can trigger 20+ tool calls, most of them reads and greps. The final diff is a tin

The study only looks at what lands in the PR. In my experience a single prompt can trigger 20+ tool calls, most of them reads and greps. The final diff is a tiny fraction of what actually happened. Hard to judge quality without seeing the process.