@mfowler Doesn't the fact that in legacy migration, the unpredictable nature of LLM responses would mean you could confidently glean a requirement from the legacy code that isn't discovered correctly? And, because nobody remembers the details, you would be unable to verify it? Similarly, missing requirements, context entirely, how would using and LLM improve on that?
Deliberation would instead consider the original problem and how it might be solved.