Ask: “Validate this model and run a simulation.” Claude uses run_simulation on a saved model version and saves the run. If you do not give a number of draws, the engine chooses and keeps going until the ranking of the options is stable. Each run records its seed, number of draws, engine version and numerical diagnostics, so it can be reproduced.To read a run later, ask for it by its run ID (get_run) or list all runs of a model, newest first (list_runs).
Ask: “Run again, but assume setup cost is 50% higher.” Claude passes a declared stress change. It is recorded with the run, so it never silently replaces the base case. See Stress tests.
Ask: “At what value of adoption does this option stop paying off?” Claude uses break_even: the value of one input at which expected economic value reaches zero, with every other input keeping its uncertainty. It holds only under those other assumptions.
Ask Claude to quote the run ID next to every number. Figures should come from a saved run, not from the AI’s own arithmetic. If the model has changed since a run, get_run tells you the run is out of date.