Keep selected factual correctness separate from nominal workflow equivalence. Record all completed MCP attempts, including failed/undecodable replies, and require the planned ordered calls, arguments and completion status. Extra inventory, repeated or reordered attempts cannot pass workflow acceptance; retain their actual cost/errors and supported cited facts. Validation: 31 targeted context/grading regressions and factual plus workflow regrading of all four original model trials without reruns. SDK and comparator code are unchanged and are not retested. Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com> Co-authored-by: Kaoru0822 <208281322+kaoru0822-kitauji@users.noreply.github.com>