<p>Progress in offline Arabic handwriting recognition is difficult to assess objectively because published systems often differ simultaneously in visual encoders, training procedures, and decoding strategies, making the source of reported performance gains unclear. This study isolates the effect of decoder design through a controlled evaluation in which a common encoder architecture, training protocol, and experimental setup are held constant across five decoding configurations. The evaluated systems include a Connectionist Temporal Classification (CTC) decoder, an attention-based encoder-decoder, two CTC-attention combinations employing different coupling strategies, and a CTC decoder enhanced with a character-level n-gram language model. Experiments are conducted on the KHATT and Muharaf benchmark datasets using identical training and evaluation conditions. The results show that language-model-assisted CTC decoding achieves the best recognition accuracy on both datasets, reaching character error rates of 7.76% on KHATT and 11.67% on Muharaf, while outperforming all neural decoder combinations without increasing the number of trainable neural parameters. Further error analysis reveals that the dominant failure categories, particularly dot-related character confusions and word-boundary errors, remain largely consistent across all five systems. These findings indicate that decoder design alone offers limited gains, and that future advances are more likely to arise from stronger visual representations and higher-quality training data. The study provides a reproducible benchmark and practical guidance for selecting decoding strategies in offline Arabic handwriting recognition systems.</p>