Question for anyone who's run the hierarchical classification cookbook on real data — how does the beam search behave under domain shift?

Question for anyone who's run the hierarchical classification cookbook on real data — how does the beam search behave under domain shift? My case: same 641-leaf taxonomy everywhere, but the tag vocabulary differs a lot from site to site across ~560 industrial sites. For my current model, per-class shift correction was worth +46pp over CORAL, so the shift is severe and heterogeneous rather than a single global offset. What I can't tell from the docs: does Jev's calibration hold when the state looks unlike anything in training, or does confidence degrade silently? A miscalibrated gate is worse than no gate, since everything downstream branches on it. Happy to report back with numbers once I'm off the waitlist.