Ruling out non-LLMs seems harder though. A possible strategy could be to generate a random set of 20 words, then ask an LLM to write a long story about them. Then challenge the user to quickly summarize it, check that the response is short enough and use another LLM to check that the response is indeed a summary and grammatically and ortographycally correct. Repeat 100 times in parallel. You can also maybe generate random Leetcode problems and require a correct solution.