Document expected output scale and extrapolation behaviour

README.md

@@ -80,6 +80,14 @@ Three values per sequence, in the order `["K562", "HepG2", "SKNSH"]`, on the sca
 MPRA activity log2 fold-change. They are not probabilities and are not calibrated
 across cell types.
 
+For calibration, across the 754,588 real library sequences the predictions have
+median ~0.2 and a 99th percentile of 3.9-4.7 depending on cell type, with a
+maximum around 11-17. The output layer is unbounded, so out-of-distribution input
+(homopolymers, short tandem repeats, synthetic sequence unlike the library) can
+extrapolate well past that range -- a plain `"ACGT" * 50` repeat scores above 20.
+Treat values far outside the percentiles above as a signal that the input is
+outside what the model was trained on, not as a strong prediction.
+
 ## Verified performance
 
 Predicted vs empirical MPRA activity across all 754,588 length-200 sequences on