When 90% Means 75%: Conformal Crop Yield Intervals for Pakistan's Districts under Temporal Shift

Year-ahead yield forecasts for districts and crops can support food-security planning, but only if their stated uncertainty can be trusted. We study split-conformal prediction intervals for one-year-ahead district-crop yields in Pakistan, using official district crop statistics from 1981-82 to 2024-25 (70,225 district-crop-year rows, 98 crops, after cleaning). With a leakage-safe temporal design (training up to 2015-16, calibration on 2016-17 to 2019-20, test on 2020-21 to 2024-25), a pooled ridge model has a lower RMSE than persistence and a three-year median, but its nominal 90% intervals cover only 0.751 of 17,896 test yields. The shortfall is present from the first test year, is concentrated in Punjab (0.600) and Balochistan (0.658), and is not caused by shifts in province proportions or in how many years of history each series has: reweighting the test results to the calibration-period mix still gives 0.737 and 0.723. Recalibrating each year on the previous four years raises coverage to 0.839 but nearly doubles interval width; an adaptive conformal update reaches 0.884 only with intervals too wide to use in some years. In an exploratory follow-up, province-wise rolling calibration reaches 0.849 with narrower intervals than pooled rolling calibration. No method reached the nominal level at a usable width. We release code, cleaning rules, a pre-specified protocol and all results.