The build-data programme predicts how long a building permit takes, using the open permit feeds of Chicago, San Francisco and Austin. A boosted model beats a per-type table on pinball loss. The practical question is narrower: which 80% interval should a downstream tool ship?
After the earlier fix to the conformal interval (it no longer collapses when the correction is negative), we compared three candidates on the 2024 test year: the per-type empirical table, a three-key table (type, work class, cost bucket), and the full LightGBM model. Rule: ship the narrowest interval whose coverage is at least 77%. I wrote that rule after seeing the numbers, so it is a decision and not a test.
The per-type table wins in four of five sets (Chicago, San Francisco application-to-issue and issue-to-completion, and Austin application-to-issue). In Austin issue-to-completion all three are within 0.1 days of each other, so the readable table ships. Coverage of the shipped intervals runs from 0.82 to 0.89: safe, not tight.
So the boosted model's advantage lives in the quantile loss, not in a tighter calibrated interval. One test year, two seeds, no bootstrap on width. Details: research/build-data/paper/paper.md, section 4.7e, and LOG.md, session 59.