LABEL RELIABILITY BEFORE LABEL ACQUISITION: A COST-ACCOUNTED STUDY OF NEURAL OFDM ADAPTATION

Objective: Labels recovered through forward error correction (FEC) can adapt a neural orthogonal frequency-division multiplexing (OFDM) receiver without dedicated training symbols, but an oracle-label gap does not by itself justify transmitting new training frames. We ask whether the gap reflects label correctness, observation selection or label quantity, and whether paid known-payload probes remain worthwhile after full radio costs are charged. 
Methodology: A coded 64-subcarrier OFDM link with a pretrained recurrent receiver was simulated. A development experiment (32 episodes) separated the three factors at matched update effort. A frozen confirmation (24 episodes at each of two nominal Eb/N0 values, 7.25 and 8 dB) compared a codeword-level zero-syndrome mask, batch-level FEC adaptation, known-payload probes and conventional receivers by error-free user bits per transmitted sample, charging pilots, cyclic prefixes, probe frames and signaling. 
Key findings: The oracle gap was chiefly a label-correctness effect. The mask, needing no new transmissions, improved efficiency over matched-update FEC adaptation by 0.0419 and 0.0352 bit/sample (95% intervals 0.0395–0.0443 and 0.0326–0.0378). Probes slightly lowered block error rate but lost net efficiency over the registered horizon. Charging measured serialized processing time put the mask below the frozen receiver and a colored-noise conventional receiver in throughput. 
Conclusion: Ordinary labels should be filtered at codeword level before new training transmissions are purchased. This empirical control result is conditional on one checkpoint and a stable macro channel model, and does not imply that probes are unnecessary at longer horizons or under other impairments.