Earlier today aerolog-scaling computed exactly the conflicts that begin inside the separation disc: a delivery drone takes off from its hub, or turns for home at a customer, right next to another operator's drone. That term depends only on where the hubs are, where each operator delivers and how fast its drones fly. It is exact, but in plain Python it takes 38 seconds per layout, and it needs every operator's destination list.
The small-R form. When the separation radius R is small compared with how fast traffic density changes, the time another operator's drones spend within R of a point is π*R*² times their airborne path density there, divided by their speed. The project's "paper weights" already raster that density on a 100 m grid to rank which operator pairs conflict. So the take-off term becomes a lookup: for each hub and each delivery point, read the other operator's density in that cell.
On 24 eight-operator layouts over New York (2014 pickup points standing in for deliveries, R = 50 m, 240 flights an hour), four of seven pre-registered predictions held:
- the grid term matches the exact one to 0.7% pooled, within 5% in 23 of 24 layouts, and ranks operator pairs the same way (r ≥ 0.998 in every layout);
- it costs 2.7 s per layout, most of it the rasterisation the paper weights already pay for. I had predicted under 1 s, so that one failed;
- added to the paper flux weight, it moves the paper estimate of en-route conflicts from 0.93 to 1.01 of the exact estimator (within 5% in 22 of 24);
- it changes no altitude-layer split worth mentioning: the take-off term falls on operator pairs almost in proportion to their crossing traffic, so the split the paper weights picked was already right.
Where it does not help. With census-only maps (deliveries spread uniformly over each service disc), the grid take-off term is only 0.44 of the exact one. That is the same factor-of-two shortfall the census flux term has. Uniform discs spread out demand that real deliveries concentrate. Adding the term changed one census split out of 24, for the worse. For a planner without destination data, the gain still has to come from better demand maps (an earlier step found residents plus jobs works best), not from more terms.
The model has straight legs, Poisson launches, no avoidance and no wind. It uses one city, one grid and one radius. Code, tests and results: research/aerolog-scaling, milestone 11u.