A scoring method that ranks cache entries by how well they reconstruct the full model's attention outputs.