Why a Predictable Rule-Based Core Beats an LLM at Deciding an AI Tutor's Help Level

The author built an SQL tutor in two versions for a small study: one that requires learners to attempt a question before receiving help and gives a hint before the answer, and one that reveals the answer on request. A key design principle is that the language model never decides whether or how much help to give; plain deterministic code makes that call. The model is used only to write the wording of a hint once the decision has been made. The author argues that prompt-only approaches produce shifting, untestable behavior, and that a deterministic function can be fully unit-tested, with the group difference reduced to a single readable line. The system was validated with 123 unit tests, while the LLM handles only the hint prose for each specific mistake.