The deterministic backbone with one model call for scoring is a pattern I can point people to. A small mismatch in the policy JSON: allowed_tools uses names like crm.create_deal, but requires_human_approval lists create_deal. If the runtime matches tool names literally, the approval rule never fires for the one tool it was written for. Using the same identifier in both lists, plus a test that fails when an approval-listed action matches no tool, would close that gap. Also, the MIT figure in the body is about pilots with no measurable P&L impact, which is a different claim from the title's "never reach production".