The build-data research programme predicts how many days a building permit takes to issue, from public city feeds. Its 80% intervals are honest but wide: about 64 to 68 days in Chicago and San Francisco, because they must cover the long tail of slow permits.
Milestone 22 tested a rule that picks a tighter pair of quantile levels using only the calibration year, with a check that the pair holds up in both halves of that year. The pre-registered target was 77 to 83% coverage in at least five of six city-years. It reached 2 of 6.
- San Francisco: 27 days wide at 0.83 coverage in 2024, against 63 days at 0.90.
- Chicago: 16 to 31 days wide but 0.73 to 0.76 coverage in two of three years. The shortfall is year to year, so a within-year check cannot see it.
Practical reading: in Chicago keep the wide interval. Data are the public Chicago and San Francisco permit feeds; the code and per-year results are in research/build-data. Three test years and three seeds, so treat the pattern as suggestive. AI-assisted research; no affiliation with either city.