This is an earlier experiment. Read the amendments below before treating its findings as current.
The question
The last free-flight collisions were fast pairs whose predicted miss was already inside the required spacing. But avoiding took priority over other steering only once craft were already too close.
What was tested
Avoiding now also took priority on a predicted near-miss, for free-flying craft only. Launch clearance now looked 4 s ahead.
What was found then
The full test set (every scenario family and variant, about 400 cases) fell from 7 collisions to 2. The two pad results were unchanged, and all 56 worse-link cases stayed at zero. The work score (trips and cargo delivered) rose 0.12 %. Given to every craft, the rule raised one-at-a-time pad collisions from 3 to 10 on one layout. A rule holding lane craft at their height was tried and left out: it traded pad collisions for as many in wind.
What we learned later
Rerunning the slow-link cases on 2026-09-04 showed this rule let hovering grid craft act on out-of-date pictures (collisions 37 → 311). Craft holding station were then exempted.
README.md:481; git aa0a0a4
What changed
Priority on predicted near-misses, for free-flying craft only; launch clearance looks 4 s ahead.
Numbers and source references
- 4 s
- launch look-aheadREADME.md:387
- 7 → 2
- full-matrix collisions (run 16)README.md:387; git 5709428
- 56
- degraded-link rows still at zeroREADME.md:387
- +0.12 %
- workREADME.md:387
- 3 → 10
- metered pad collisions if applied to everyone, one seedREADME.md:387