This looks cool but can you explain how to make it useful?
For an example if you're working on a Q&A model, you can check which tokens in the prompt contributed to the output. It's possible to detect issues like output not paying attention to any important part of the prompt.