The branching threshold is calibrated as a minimum alternative token confidence score—defaulting to 10%—to isolate meaningful divergent entity associations from low-probability language model noise.