The shared-model transceiver from Latent Radio, retrained and stress-tested against what
deep-space and jammed military links actually face: burst jamming (chunks of the signal wiped out),
deep fades, and extreme low SNR. The question — does "meaning survives damage" hold up when the channel is
actively hostile, not just noisy?
Why this technology fits space & defense. These are closed systems — one
operator controls both the transmitter and the receiver — so the shared-model requirement that sinks consumer
use is a natural fit (and doubles as implicit keying). They are bandwidth- and power-starved,
with no retransmission (a Mars round-trip is tens of minutes; you can't ask again under
jamming). That is exactly the regime where graceful degradation beats a digital cliff.
Trained through the threat, the link barely flinches as a jammer wipes out up to 70% of
the signal. At 70% jamming it holds 17.0 dB — versus 12.9 dB for the clean-noise model
(which collapses below even the classical baseline) and 14.3 dB for idealized digital.
contested-trained (threat-aware)AWGN-only model (ablation)classical digital, idealized
The honest lesson is the red line: a transceiver tuned for clean noise — stock Latent Radio —
degrades badly under jamming, worse than classical digital. Robustness isn't automatic; it comes from
training through the actual threat model. Do that, and the shared-model system is remarkably jam-resistant.
See it — images under jamming
Top: originals. Next three rows: contested-trained at 0%, 25%, 45% jamming — still clearly
readable. Then the clean-noise model and classical digital at 45% jamming, both visibly worse.
Extreme low SNR, with jamming
Holding a 25% jammer on, sweeping the noise floor. Below −4 dB the digital
link runs out of capacity and drops to the blank-image floor; the shared-model system keeps delivering a
recognizable picture.
contested-trainedclassical digital, idealized
What this is / isn't
Is: honest evidence that the shared-model / joint-coding idea holds up under contested conditions — burst
jamming, fades, extreme SNR — and a demonstration that you must train for the threat to get there. The
closed-system nature of space and defense makes the approach a natural fit.
Isn't: a deployable system. This is a simulation with idealized channel models (real RF adds Doppler,
timing, hardware limits, adaptive jammers). Two more honest cautions specific to these domains: a neural decoder
can hallucinate plausible-but-wrong detail — dangerous for science or targeting data, so you'd send
verifiable residuals — and a shared model is not cryptography (layer real crypto on top). The digital
baseline here is deliberately generous (ideal interleaving), so the true classical cliff under bursts is harder
than shown; the neural advantage is if anything understated.
