Why the Inference Chain Is Ordered the Way It Is

A junior engineer proposes simplifying the inference pipeline: run harm detection once, right after prompt safety and before generation, by having the harm-detection classifier assess the (enhanced) prompt instead of the generated image — "if the prompt is safe, the output will be safe, and we save a pipeline stage."

1. Explain why harm detection on the generated image is a distinct stage from prompt safety, not a redundant one.
2. Where does the super-resolution cascade belong in the chain, and why does that placement matter for cost?
3. This subject says the two safety stages defer their depth to guardrails-and-prompt-injection-defense. What is this subject's own responsibility regarding those stages, if not re-deriving classifier design?