The divide-by-zero test is the part that stuck with me. The first prompt produced a correct-looking margin, then a single zero-revenue row took the whole report down with ORA-01476, which is exactly the kind of failure that only shows up in real data. Putting the guard in Report Context instead of relying on every user to phrase the prompt correctly feels like the right call, since the CASE WHEN F=0 THEN NULL version also makes the intent explicit: a zero-revenue margin is undefined, not zero. I also liked the discipline in the conclusion: use the natural-language calculation to explore, then lock the agreed definition into SQL or a view once the team has settled what margin means. Did the Luna model ever suggest a different fallback, such as returning 0 instead of NULL, or was NULL its consistent choice across your runs?