How good can prime-editing efficiency predictors get? A reproducibility ceiling shows HEK293T is nearly saturated and locates the remaining K562 gap

Sequence-based models that predict prime-editing efficiency are now compared on differences of a few hundredths of a correlation point, yet none of these comparisons measure how reproducible the underlying assay is, so none can say how much of the remaining error is addressable. Using the three biological replicates per pegRNA in the PRIDICT2 Library Diverse dataset, I estimate that ceiling directly. The assay is highly reproducible: mean inter-replicate Pearson r is 0.99 in HEK293T and 0.96 in K562 (Spearman 0.97 and 0.84). Against this ceiling, HEK293T prediction is nearly saturated, with the stronger of two sequence-only models reaching Pearson 0.95 and Spearman 0.95, within 0.02 to 0.03 on both. K562 requires separating the two metrics. In rank terms the models are already close to the ceiling across the full library (Spearman about 0.80 against 0.84), and the large K562 Pearson gap (models at 0.68 to 0.70 against 0.96) is a magnitude effect amplified by the library's zero-inflation. A gap remains only on high-efficiency pegRNAs, where the Spearman shortfall reaches about 0.17. Whether chromatin features would close it cannot be tested on this library, because it has no per-target chromatin context, which is itself why epigenetics-aware models have never been measured against a reproducibility ceiling. I recommend reporting the ceiling, in both Pearson and Spearman, alongside model performance.