This is an earlier experiment. Read the amendments below before treating its findings as current.
The question
The failed version did two things: it removed slow craft's extra room and added early reaction (steering away sooner from closing neighbours). Which one broke the storm results?
What was tested
Each change was flown alone on the storm test set (storms plus a poor radio link) for wing and multirotor, and on the wing's mixed test set (145 standard scenarios), paired scenario by scenario with the old controller.
What was found then
Removing the room alone was clearly worse: wing storm collisions rose from 308 to 963. Early reaction with the room kept was clearly better on every test flown: wing storm collisions fell from 308 to 217, wing mixed-set collisions from 544 to 158, and the multirotor's fell too. With gusts and outdated position messages, the extra room is what protects. The crossing version had always combined both changes. Why the room protects was not measured.
What we learned later
Reviewers noted that the earlier 'no clear change' on the mixed set was not a simple sum of the two effects; the two changes interact.
research/reviews/auditor-c134.md (F-TAU-SPLIT-ATTRIBUTION)
What changed
nothing
Numbers and source references
- 308 to 963
- wing storm-set collisions, bubble removed without chargeresearch/data/oos/tau_stormsplit.md
- 544 to 1 416
- wing matrix collisions, bubble removed without chargeresearch/data/oos/tau_stormsplit.md
- 308 to 217
- wing storm-set collisions, charge with bubble keptresearch/data/oos/tau_stormsplit.md
- 583 to 417
- wing storm-set craft lost, charge with bubble keptresearch/data/oos/tau_stormsplit.md
- 544 to 158
- wing matrix collisions, charge with bubble keptresearch/data/oos/tau_stormsplit.md
- 67 to 48
- multirotor storm-set collisions, charge with bubble keptresearch/data/oos/tau_stormsplit.md