The previous aerolog-scaling step found why a crossing-path formula undercounts conflicts between delivery drones from different operators by about 7%. The formula counts paths that cross into the separation disc while both drones are flying. It misses the conflicts that begin inside the disc, because a drone takes off from its hub, or turns for home at a delivery point, right next to another operator's drone. That step estimated the missing "start term" by sampling pairs of flights. This step computes it exactly.
Why no pairs are needed. Under the semicircular rule a flight's outbound leg and its return leg follow the same segment at the same speed, in opposite altitude layers. So wherever a drone starts, and in whichever layer, exactly one leg of any other flight is in that layer, and it spends the same time near the start point either way. The start term does not depend on layers. It factorises into one operator's expected airborne time within R of another operator's hub and delivery points. That is a finite sum over a destination list, evaluated by visiting only the legs whose bearing passes within R.
On 24 eight-operator layouts over New York (2014 pickup locations standing in for deliveries, R = 50 m, 240 flights an hour), five of the six pre-registered predictions held:
- the exact term matches the sampled one (pooled ratio 1.0002; 23 of 24 layouts within three standard errors);
- 60% of it comes from drones turning for home at delivery points inside another operator's traffic, and 40% from hub launches under another operator's routes. Points that two operators both serve add 0.6%;
- it grows as πR². It is 3.93 times larger at 100 m than at 50 m, and 0.257 times as large at 25 m;
- added to the crossing-path formula, it brings the formula within 2% of the exact estimator in all 24 layouts (mean 1.0025), and it picks the same altitude-layer split in 23 (the 24th is a tie).
The prediction that failed: I expected an exact start term to make the whole estimate less noisy. It doesn't (median spread ratio 1.02). The sampled start chords share flight pairs with the exact count, so their noise partly cancels in the ratio, and the crossing term's own noise dominates anyway. The stdlib implementation is also about five times slower than simply sampling.
What a planner can take from it: the correction depends only on where hubs are, where each operator delivers, and how fast its drones fly. It is the same whatever layer plan you are comparing, and you can scale it by R² when you change the separation minimum. The model has straight legs, Poisson launches, no avoidance and no wind, and covers one city. Code, tests and results: research/aerolog-scaling, milestone 11t.