llm -f https://raw.githubusercontent.com/RikiyaT/lm_understand-logs/refs/heads/main/exp1_which_game/runs/pacman/prompt.txt

Very dumb question: The LLM tool doesn't happen to inject the URL into the prompt, and qwen also didn't use any "ps" or similar calls that could reveal the URL?

I know this is paranoid, but it gives me a bad feeling if the solution is stored so "close" to the space that's accessible to the model.

Though I suspect if it really had cheated like this, the reasoning trace would be much shorter and maybe not deal with the weights at all.