Training procedure that simulates low‑precision arithmetic during training to preserve accuracy after quantization.