A difficult-region percentage divides by the whole gene span — including bases the reference never sequenced. Where a gene sits over an assembly gap, that dilutes the number, and the gene reads cleaner than it is.
GIAB's difficult-region tracks cover sequence that exists. An assembly gap — a run of N bases — is in neither the difficult track nor its complement, but it is still in the denominator when we compute what share of a gene is difficult. The result is arithmetically correct and practically misleading: GRK1 on GRCh37 reports 8.8% difficult when 83.9% of its span is not sequenced at all. Of the bases that do exist, 54.7% are difficult.
This page shows all three framings side by side for every gene affected, with all three assemblies, so the contrast is visible. Most of these gaps are GRCh37 artifacts that GRCh38 or T2T-CHM13 resolves — which is itself the finding.