Two identical d10 models (100M params) trained to validate the hypothesis that quality-filtered data enables more efficient training.