The key technical constraints follow from how the system was developed and validated. The reference populations used to train and validate HIrisPlex-S are modern populations. The allele frequencies, the relationships between genotype and phenotype, and the prediction accuracy of the system have been established from studies of people alive today. Applying the system to ancient DNA involves an assumption: that the same genetic variants predict the same traits in an ancient individual, with the same relationship between genotype and phenotype, as they do in modern individuals.
This assumption is not unreasonable — the basic biochemical mechanisms of melanin production in skin are ancient and conserved — but it is an assumption, and it introduces uncertainty that cannot be quantified from the genomic data alone. The ancient genome might contain additional pigmentation-related variants not captured by the HIrisPlex-S model — either variants not yet known to researchers or variants in regulatory regions that affect gene expression without changing the protein-coding sequence. These additional sources of variation, undetected by the current model, could push the actual phenotype in either direction relative to the prediction.
Furthermore, ancient DNA from specimens ten thousand years old is by definition damaged and incomplete. The sequencing of any ancient genome produces coverage that is typically lower and patchier than what would be obtained from modern DNA — not every position in the genome is read equally well, and positions important to a predictive model may be among those with lower coverage or greater damage-induced uncertainty. The research team would have been working with the variants they could read reliably, excluding positions where the data were ambiguous.