Inspect a measured reference build
Run the baseline, read its traces, then switch on the missing-order check and run the same tasks again. The numbers below are computed by the workflow in your browser. The inputs and recorded decisions are invented teaching data. They are not measurements of a live AI model.
Recorded-decision workflow: sample question → choose action → read sample order → reply or hand over → stop.
No network calls, model charges, customer data, or write tools. This reference does not test live language understanding.
The baseline treats a missing order as delivered. That produces an unsupported answer. The repaired build hands that case to a person instead. Inspect the successful order-status trace and the missing-order trace to see the difference. Refund requests and timeouts also hand over, and no path writes to an order.
Each safeguard has a reason: read-only lookup prevents order changes; the missing-order check prevents invented status; one lookup and no retries bound the reference calls. A live loop also needs an enforced stopping limit. This list is your harness inventory: the surrounding checks and tools that make the model's choices usable. Keep its reasons beside the design.