Abstract
Handwritten text recognition systems that serve real users must be more than accurate—they must remain reliable under noise, device constraints, and occasional compute faults. We study reliability-centered Kazakh handwritten text recognition by combining error-correcting output codes with redundant lightweight ensembles and a simple watchdog for selective prediction. Each character is mapped to a structured codeword and the network predicts code bits; inference uses nearest-neighbor Hamming decoding with optional parity checks, converting small logit perturbations into correctable events. We evaluate robustness under controlled corruptions (blur, noise, occlusion), quantization, and bit-flip fault injection. In bit-flip simulations, error-correcting-code decoding reduces misclassification from 78.73% to 2.27% at p = 0.10, and from 90.69% to 23.43% at p = 0.20, while correctly recovering 86.66% of faulted trials at p = 0.10. With three diverse replicas, bit-wise ensemble fusion further lowers error to 0.01% at p = 0.10, outperforming label voting. Finally, the watchdog reduces failure-without-warning to 0.01% at 64.95% coverage, enabling practical coverage–risk trade-offs. Together, these results show that classical redundancy and coding concepts translate naturally to modern deep handwritten text recognition, providing a practical path to robust and deployable recognition for Kazakh and other under-resourced scripts.