A staff handbook that knows when not to answer.

This is a synthetic handbook with deliberate gaps (no parental-leave or pension policy). The point: the assistant must decline and refer when the answer isn't in the handbook — not invent one. Watch it follow the relevant policy, ground its answer, and admit what it doesn't know.

What it can and can't guarantee — and how to get the best out of it. It does the verifiable part with certainty: every turn shows which policy it used, whether the answer is grounded in that policy, and it declines and refers when the handbook doesn't cover the question — auditable end to end. It cannot vouch for the rules themselves (a wrong policy, perfectly followed, is still wrong) — it only knows the handbook it is given. To get the best out of it: keep the rule-set small and non-competing, retrieve the relevant rule rather than dumping the whole book, and size the model to how many rules actually compete per turn — the "optimal, not perfect" discipline (more rules → lower adherence, including on the safety-critical ones).