A course helper took a weekend to build and weeks to teach the words "I don't know"

Refusal costs more than answering because you are working against the model's default. According to the dev.to writeup, the training gradient pushes toward being helpful, so a question outside the knowledge base gets answered anyway, and answering without knowledge is guessing ^[10]. Instructing it not to hallucinate was ignored ^[12]. What moved behaviour was a concrete script, "I do not have that, here is what to do next", plus worked examples of acceptable refusals and explicit permission not to know ^[11][12].

The second lever was length. The team stopped treating long answers as good answers, on the theory that a model which believes it must fill space will invent space to fill; once brevity and "ask a mentor about this one" were both acceptable outcomes, the inventing dropped ^[15].

Scoping did two jobs, and the second is the one operators will care about. Every chunk carries its origin and an authority rank, a published lesson outranks a forum reply, and random forum replies were cut from the corpus entirely ^[6]. That gives a debugging path: a bad answer traces back to the document that caused it, and the author reports that half the fixes were deleting one stale file rather than editing a prompt ^[9]. The maintenance object here is the corpus, not the prompt ^[19]. Before the tagging, the blob problem was epistemic: a throwaway comment in an old draft carried the same confidence as core material, because nothing in the input distinguished canon from noise ^[5].

The cheapest fix was also the least technical. Until the system prompt named the covered subjects and declared everything else out of scope, the model assumed its scope was anything a helpful assistant might discuss ^[7]. The author rates writing that fence down as more effective than any retrieval technique they tried ^[8].

Then the test suite, which is the part with the clearest transfer to other projects. Out-of-scope questions are the easy half; the near misses, questions that sound in scope and are not, are where the money is ^[13]. Before that suite existed, every improvement on the answering side quietly degraded refusal, and the trade was invisible until a student hit it ^[14]. That is the usual shape of an untested invariant, and it is why refusal rate belongs in CI rather than in a launch review.

The handoff is engineering, not etiquette. The helper forwards what was asked and what it could not cover, so the human does not start cold, and the author's argument for spending that effort is that a handoff which drops everything the student said feels worse than having no helper at all ^[17].

Arithmetic on the calendar: if "weeks" means even two, the guardrail work ran about seven times longer than the weekend that produced the demo ^[20]. The author's own summary is blunter, that answering is a demo and the product is refusing, scoping and handing off ^[18].