The data retention example is the most interesting part of this. The ECL didn't just learn the rule, it learned why the rule exists - reps kept getting it wrong. That's a different thing entirely. Most knowledge systems store the conclusion and quietly lose the reasoning that produced it.

Which makes me wonder: how does the maintenance agent know when to revisit a rule like that? "Feature X ships in Q3" is easy - facts go stale and you can detect it. But "don't let reps answer data retention questions" - that rule could still look valid in the ECL long after the original reasons for it stopped applying. Does it track enough of its own provenance to catch that kind of drift?