In Firmament, each of four real-time audio neural networks is trained on a distinct dataset of recordings of para-scientific experiments, either funded or ostensibly perpetrated by clandestine State agencies, in using sound to interface with an Other. The networks reconstruct each other in live feedback.