This is an earlier experiment. Read the amendments below before treating its findings as current.
The question
An outside review of the testbed arrived on 2026-09-08, and testing was paused to check it.
What was tested
Every concrete claim was checked against the code. The defects in the simulated world and in reporting were fixed the same day.
What was found then
Every concrete claim that could be checked against the code was correct. The gust field did not wrap around at the edges of looping test worlds. The neighbour search was limited by airspeed, while collisions depend on speed over the ground. The work score mixed cargo units with trips. The reference-run check asserted nothing. Two Python reports were wrong. The criticism of method stands: the rules were tuned and tested on the same scenarios, so results describe this simulator, not collision avoidance in general.
What we learned later
Fixed the same day: gusts wrap around the edges, the neighbour search uses ground speed, reference checks now fail loudly, every result records time spent in close calls, and the work score was split. Losses from storms on a slow link then fell about a sixth (one layout 62 → 52).
README.md:1244, :1377; git a17e672, 7b00f6c
What changed
Nothing at the time (testing paused).