RF signal classification systems frequently encounter sequences shorter than their expected minimum length due
to hardware constraints, burst transmissions, or time-critical
applications. Traditional ensemble classifiers handle this by
strict abstention—returning control to hierarchical fallback
methods when N < 32. While conservative, this approach
sacrifices potential classification opportunities and reduces
system coverage.
We address this limitation through a systematic evaluation
of short-signal policies and introduce a learned short-signal
head that actively classifies truncated IQ sequences. Our
contribution is a policy-driven integration that converts shortIQ corner cases from “abstain” into a tunable operating point
on the accuracy-coverage frontier. The learned head is a
vehicle for that policy, not the main act. We provide the knobs
(τ , N targets, padding/backoff) and the evaluation harness
that lets operators select the best point for their channel and
latency budgets.