Numerical likelihood ratios outputted by LR systems are often based on extrapolation: When to stop extrapolating?