C. Moreno

For insertions in the linear versus graph comparison, would those fixed matching criteria include inserted-sequence similarity, or use only breakpoint distance and length?

Your held-out validation proposal gives the regional comparison a concrete target. I'd also select the validation events independently of the linear and graph calls. For structural variants, validating only graph-rescued calls would leave linear-only events outside the comparison; an independently selected event set would let us compare which validated events each representation misses in the withheld region.

The downsampling test should stay SNP-specific unless a separate structural-variant set has independent support. For an extension to structural variants, I would require breakpoint evidence matched to event type, then compare linear and graph calls only within that validated set. Otherwise, apparent geographic persistence could reflect caller-specific breakpoint placement rather than reduced reference bias.

Breakpoint evidence first: require concordant split long-read support and junction sequence for graph-rescued events, then test whether event type remains stable across callers. Treat physical mapping as orthogonal support where available. Only validated events should enter the callability-matched geographic comparison; otherwise breakpoint error and reference bias remain confounded.

Structural variants are not yet a clean probe: graph-derived calls may change missingness, allele frequency, and linkage structure simultaneously. Match variants by callable samples and frequency, prune within each representation, then compare geography-conditioned structure against permuted geographic labels. Otherwise a PCA shift could reflect altered ascertainment rather than reduced reference bias.