—Rapid RF triage must support not just signal
classification but reliable command delivery under poor link
quality. We extend a normalized 1024-pt FFT with light
post-filters — previously shown to outperform a tiny CNN
(AUROC 0.754 vs 0.671, 48× fewer FLOPs, 1.5 ms p99 at 0
dB) — into a live control plane. Using fleet telemetry, we map
connection quality (signal/100) to command success rate,
retransmits, and p95 latency. Results show: Q1 (worst 20%)
59.5% success, 1.93 retransmits, 3722 ms p95; Q5 (best 20%)
99.9% success, 0.19 retransmits, 406 ms p95. A hybrid gate using
FFT confidence as a proxy for link quality cuts average compute
11×whilepushingsuccessabove99.7%.F ullreproducibleharnessreleased.
I. INTRODUCTION
RF control planes demand rapid spectrum sensing that
not only classifies signals but actively improves end-to-end
command reliability. Traditional approaches treat signal processing and transport reliability as separate domains, missing
opportunities for cross-layer optimization. We argue that
spectrum sensing should drive transport policy — using spectral
confidence as a predictor of link quality to enable adaptive
compute allocation and robust command delivery.
Modern tactical operations require sub-second command
latency with near-100% reliability across heterogeneous RF
environments. Pure CNN-based approaches achieve high
classification accuracy but impose prohibitive computational
overhead for edge deployment. Conversely, lightweight FFTonly methods sacrifice accuracy for speed. We bridge this gap
through a hybrid gate that dynamically escalates low-confidence
classifications to heavier models while maintaining fast-path
performance for high-quality links.