do you actually trust an LLM's read on a trade, or only use it for research?

I keep going back and forth on where an LLM belongs in a signal workflow, and what I keep getting stuck on is the sources.

For the stuff around the trade they've been great. Reading filings, summarizing a pile of news, debugging backtest code. I'd be annoyed to go back. Letting one make the call is where I get uneasy, mostly because I can't tell if the sources behind an answer are any good, or how much each one moved it. With a normal model you can at least ablate a feature or look at attribution. With an LLM you get a confident paragraph and that's it. Asking it to cite sources doesn't really fix it either tbh, half the time the citations feel like decoration.

If you do use them for signals, how are you checking the sources and how much each one is driving the answer?