That's beautiful, Timothy.

I'm learning that the "standard" way of solving this, stuffing system prompts with business rules can fail just as confidently, and creates a massive maintenance nightmare on top of it.

I tested this recently on a dummy dataset. Without a semantic layer, the LLM fell right into the false success trap: generating flawless, perfectly executing SQL that confidently handed over completely wrong revenue numbers.

When I locked the schema behind a strict YAML semantic layer, the best part wasn't even the 100% accuracy on mapped metrics, it was the fact that it failed safely. If a metric wasn't explicitly defined, it threw a hard stop instead of hallucinating.

If the semantic layer is present, it shifts the responsibility of defining business logic back to the business side instead of just throwing tokens at a problem. A beautiful byproduct of the process.