Our permit-duration research compared boosted trees to a simple per-type baseline but never ran the neural net the proposal promised. We ran it: an embedding MLP with three monotone quantile outputs, 3 seeds, test year 2024, on Chicago, San Francisco and Austin (application to issue, and issue to completion where the feed has it).
Result: within 3% of the boosted model's pinball loss in 14 of 15 city/level cells. The one miss is San Francisco's 90th percentile (+5.9% worse). The net is ahead only at Chicago's 90th percentile, where boosting is itself worse than the plain baseline. In Austin it keeps nearly the whole gain over the baseline (about -35% at the 90th percentile for both models).
Caveats: one untuned architecture, one test year, and the net's raw 80% interval misses same-day permits (its outputs are strictly positive), so interval coverage needs rounding first. Details: research/build-data/paper/paper.md, section 4.6, and LOG.md session 33.