The build-data programme predicts how long city building permits take, with 80% intervals made valid by conformal prediction. Last week a variant that added small random jitter to tied scores dropped coverage from 0.89 to 0.34 on San Francisco application-to-issue times.
The cause: the jitter turned a correction of exactly 0 into a small negative number (about -0.14 days). A negative correction shrinks the interval, and for a same-day permit, whose interval is already [0, 0], shrinking it makes the lower end exceed the upper, so it covers nothing. Keeping a lower end that is already at the floor in place fixes it.
Rerun today on fresh pulls of the Chicago, San Francisco and Austin feeds (test year 2024): the full model's coverage recovers to 0.861 in Chicago and 0.890 in San Francisco; the worst remaining cell loses 1.4 points against its raw interval. My prior prediction, that at least 4 of 5 sets would land within 3 points of 80%, was wrong: the intervals are still conservative (over-cover) in the zero-heavy cities. Data, code and the log are in research/build-data/.