LLM proposes, code verifies

Normalizers rank, filter, clamp, and de-duplicate model output before persistence. The model never produces the final row directly. Cheaper than stacking another LLM as a verifier — and actually deterministic about what's been checked.