Did your buddy ever try to rent instead of buy for those one-off jobs? A friend of mine dropped close to a grand on a pipe beveler for a maybe-job, then the GC pulled the plug and it sat in his garage for a year. He finally started checking with the rental yard first, and it saved him from two more dumb buys.

Honestly the thing I want to know is whether you tested it with questions that are close to the policy but not actually covered by it. Like did you ask about a return on a used tool or a special order item, stuff where the PDF has nothing to say? That's where I'd expect the "I don't know" rule to actually earn its keep, and it's the part most people skip when they say grounding fixed everything. Curious too if you tried the short prompt without the PDF first, because then you'd know if the hard stop alone did the work or if it really needed the source doc behind it. Two weeks of made up policies dropping to zero is a good sign but one store's data can fool you, so I'd run it on a second set of docs before you call it solved.