Why the check exists

Three properties of language models are why the pipeline is shaped the way it is:

• Fluency is independent of correctness. A wrong answer is written with the same vocabulary, structure and confidence as a right one, so you cannot tell them apart by reading.
• Self-assessment is unreliable. A model asked whether its own work is correct will usually say yes, and will produce plausible reasoning for that answer if you ask why.
• Output varies between runs. The same input can produce different results, so a run that came out right once does not tell you the next one will.

A model answering road law from memory is a model that will, some fraction of the time, tell a sixteen-year-old the wrong distance with complete assurance, and the fraction does not have to be large for that to be unacceptable. The pipeline answers the three properties directly. The model is given the law and told to write from it, so there is something to check against. The check is code, not a model, so it cannot be talked round. And because output varies, the check runs on every answer, every time.