Raw 1090 MHz Aircraft Transponder Signals as a Testbed for Complex-Valued Streaming Models

Radio signals are genuinely complex-valued, which makes them a more natural testbed for complex neural networks than text or market data. With a HackRF Pro in the UK we recorded aircraft transponder replies at 1090 MHz: three hours at 2 Msps and 96 minutes at 10 Msps, plus later days and a second receiver. We kept only the bursts (673,609 and 249,690, gated without losing a decodable message), quantised the raw I/Q samples to bytes, and labelled each burst with a Mode S decoder. Streaming: read as a byte stream, CNet's complex gated-linear-attention model reaches 0.731 nats per byte on a held-out hour once per-block normalisation removes a training divergence. That beats a one-layer real GRU (0.752) but trails a two-layer GRU (0.723). Fingerprinting: identifying which of 93 aircraft sent a burst from 17 microseconds of signal with no address bits depends almost entirely on phase. Real and complex networks plateau together at 45-46% top-1 at 2 Msps and 60-61% at 10 Msps, so the sample rate, not the network algebra, sets the ceiling. A learned embedding verifies a claimed address within one pass at 89% true accepts for 1% false accepts over 20 bursts, and runs live at the antenna. Across 36 hours it falls to 10%. A prospective second-day test, with models and thresholds frozen in advance, replicated every ranking, gap and alert rate.