A Streaming Market Monitor with CNet Born Attention

CNet's Born attention, a complex-valued attention with an exact finite-state recurrence, runs as a fixed-memory streaming model on one-minute bars of US stocks. One causal tokeniser builds the model's input from ten years of consolidated history and from the live feed. Over 530,932 tokens (454 sessions, 130 times the training window) read without reset, the CPU scorer's memory stays flat at 183 MiB, and its scores match the GPU's chunkwise evaluation to 2.3e-6 nats per byte. On held-out GOOGL data the model reduces next-byte loss by 13.8% against a no-context baseline, mostly on range and volume rather than return direction, and it transfers to seven other liquid symbols. A one-layer GRU does slightly better on the same tokens (14.9%), so the contribution is the streaming system, not an advantage of the mixer. A live monitor runs eight symbols on one CPU. In its first two sessions under a protocol frozen in advance it flagged 0.13% of minutes against a nominal 0.1%, with no drift. An event study finds no trading edge in the alerts: they mark unusual minutes, not profitable ones. Nothing here forecasts prices.