I Fine-Tuned a Compliance Judge and Beat the Stock Model by +29.6pp F1

I Fine-Tuned a Compliance Judge and Beat the Stock Model by +29.6pp F1 The problem: if your LLM-powered product touches personal information in South Africa, POPIA sits over it. The regulator doesn't ask "is your model good?" — they ask "can you demonstrate the output was validated against the clause, and can you show me the validation?" The uncomfortable answer most teams give today: "we call GPT-4 as a judge with a prompt that mentions POPIA." That's not a defence. It's non-deterministic, s...