This is an earlier experiment. Read the amendments below before treating its findings as current.
The question
The four classic scenarios looked clean. A harder test was needed to find where the rules fail.
What was tested
An automated harness ran the JavaScript simulator without graphics: 14 scenario families, 25 variants each (later 29, with worse radio links), 1,200 s per run. It logged both craft's recent history at every collision, starting with 48 cases from the families without landing pads.
What was found then
The fleet families broke on physics the rules had skipped. A package of fixes took the 48 cases from 198 collisions to 2. It included a minimum push strong enough to open the gap in time, steering on ground speed in gusts, avoidance first in line for acceleration, and the ground as an obstacle. Start, launch and speed-limit changes too. The work score (trips and cargo delivered) rose about 2 %. On the first run of all cases, 321 of 329 collisions were at the landing pads.
What we learned later
The work score added trips and cargo units together; the 2026-09-08 review had them split into separate measures.
README.md:1244 (Two measurement fixes); git a17e672
What changed
Adopted: a minimum push when avoiding, pair geometry on speed over the ground, avoidance first in the acceleration budget, launch clearance, and the ground as an obstacle.
Numbers and source references
- 14
- scenario familiesREADME.md (Stress campaign)
- 25
- variants before the link variants (350 cases)git 59cdf31 commit message
- 29
- variants with the four degraded links (406 cases)README.md (Stress campaign)
- 1,200 s
- length of each runREADME.md (Stress campaign)
- 48
- cases in the fleet subsetREADME.md (Stress campaign)
- 198 collisions to 2
- fleet subset before and after the fixesREADME.md (Stress campaign); git 284f2e8
- 2 %
- more work doneREADME.md (Stress campaign)
- 321 of 329
- collisions at the pads on the first full matrixREADME.md (Stress campaign) (funnel families)