The semantic layer lied
I built an MCP tool that scores how well I fit a job. The real test wasn't whether it worked — it was whether it would lie. It did. Here's the bug, the bake-off that ruled out an easy fix, and the redesign that handed judgment to the model already in the loop.
