Short-Signal Resilience: Learned Heads and Policy Boundaries for N < 32 IQ Classification

RF ensemble classifiers often abstain on very short
IQ sequences (N < 32), forfeiting coverage under burst or bandwidth constraints. We present a policy-driven framework with a
lightweight learned short-signal head and show it improves the
accuracy-coverage frontier versus strict abstention and padding
baselines. On synthetic truncations of four modulations, the head
attains higher utility across N ∈ [12, 28) and remains robust
down to 0 dB SNR. The framework integrates via configuration
switches (including a confidence threshold τ ) and exposes a clear
trade space to operators. Utility is computed as (accuracy ×
coverage) with both terms in [0,1]; we assert and clamp to enforce
bounds.