Why conformal prediction intervals blow up under log transform

A developer reports that after log-transforming target data, fitting a model, and applying conformal prediction, the back-transformed upper interval reaches 350x the point forecast. The back-transform is correct; the explosion stems from lognormal skew and inflated log-space scores due to small calibration sets. Coverage remains valid because monotone transforms preserve interval coverage exactly.