arXiv:2608.12918v2 Announce Type: replace Abstract: While learned wireless receivers are typically studied using synthetic data, the impact of over-the-air (OTA) measurements for training remains unclear. We conducted a 5.88 GHz measurement campaign with a 5G/6G-like orthogonal frequency-division multiplexing (OFDM) system across diverse environments and mobility conditions, and trained a neural channel estimator and a capacity-matched end-to-end neural receiver using mixtures of measured and synthetic data. Increasing the OTA fraction revealed a fundamental asymmetry: measured data consistently improved the end-to-end receiver, whereas the channel estimator peaked at an intermediate fraction and degraded with fully measured training. We identified supervision quality as a key contributor: OTA channel labels are derived from noisy received signals and therefore contain supervision errors correlated with the receiver input, whereas decoded bits validated by a cyclic redundancy check (CRC) provide effectively error-free supervision. Denoising the channel labels mitigated the degradation, and a controlled simulation showed that label noise of equal power degrades the estimator far more when it is correlated with the input. For OTA training, end-to-end receivers thus benefit from fully measured data, whereas channel estimators benefit from moderate OTA fractions but require improved label quality.