(i) increase number of refinement cycles to 30 (or more)
(ii) drop convergence criterion from 0.1 (both A and %) to a
lower value e.g. 0.02
(ii) repeat a couple of times on randomly oriented copies.
(iv) if you know where the cavity is give xyz coords for center
and search for this cavity
(v) average the converged volumes for 10 different orientations (if you want to strictly follow the manual) or just 2-3 orientations with very strict convergence criteria (I found that between 20 and 25 cycles, the volume varies only about <5 A^3, or ~1-2% for me). Also, in the latter case, on the 25th cycle, there is <0.5 A^3 difference between different orientations. So, there is little point really in calculating cavity volumes for more than 2-3 orientations.
Voila!