the rename from checkpoint to system boundary is how you lose traceability. a multi-model black box that benchmarks well is great until it fails on your workload and theres no trace to debug why.