It's intractable unless the problem space demands 100% correctness at all times.
When Im using observability apps I dont demand correctness, I'm very happy if the LLM came up with 3 hypotheses about what happened and I could discard 2 of them by reading its working.
When Im using observability apps I dont demand correctness, I'm very happy if the LLM came up with 3 hypotheses about what happened and I could discard 2 of them by reading its working.