The target application distribution must be specified before pooling these error rates. Variant type, genomic context, depth, and species can alter both calibration and the frequency of difficult calls. Estimate performance within prespecified strata, then standardize those estimates using weights from the intended application population, not the incidental validation sample. Otherwise, an apparently portable workflow may simply benefit from a favorable mixture of species, platforms, depths, or callable regions.
The target deployment distribution also needs to be named. Variant-type and depth-specific error rates can look favorable while the aggregate calibration fails after transport to a species, sequencing platform, or genomic-context mix unlike the validation sample. I would report stratum-specific false-positive and false-negative rates, then standardize them using prespecified weights from each intended application population. Weights estimated from the incidental validation mix would only describe that sample, not portability.
The standardization distribution needs to be named before evaluation. Weights based on the external sample would make the summary depend on its incidental referral mix. Prefer prespecified weights from the intended referral population, with stratum-specific estimates reported alongside the standardized result. If those target weights are unavailable, the analysis should remain stratified rather than presenting a pooled transportability claim.
Batch controls establish analytical stability, but they do not establish transportability to a referral population. The evaluation should also prespecify the target distribution of ancestry, ascertainment, prior evidence, and variant class, then report calibration and reportable-call yield within those strata. Otherwise stable batch performance can coexist with clinically important miscalibration after the referral mix changes.
Site and time strata still need a target distribution. Report stratum-specific calibration and threshold performance first, then standardize them to the intended referral population. Otherwise the pooled estimate remains contingent on the validation sample’s case mix, even when the mapping is locked.
The deployment population should be specified before batch stability is interpreted: diagnostic referrals, cascade testing, and population screening can have different prior probabilities and evidence profiles. Stable assay behavior would not by itself preserve likelihood calibration or classification thresholds across those settings. For the JAG1 work, external validation should stratify calibration and decision consequences by referral pathway, ancestry, and reporting policy, with any setting-specific recalibration made explicit.
