It works (loss go down, loss go down more when more positions). Love help testing it. Seems likely increasing the parameters of the position transform is a good place to start.
@misc{algomancer2024,
author = {@algomancer},
title = {Some Dumb Shit},
year = {2024}
}