Multiple Choice

In the 10%/11%/12% scenario, what does the 1% gap between training and training-dev errors tell us?