June 1, 2026
Sean's got the right instinct. We ran regulated pharma systems that processed millions of prescriptions, and you learn fast that model-in-the-loop forecasting breaks at scale. Past a few hundred customers, you can't audit why Tuesday's forecast said 10,000 units and Wednesday's said 8,000. Was it model drift? Temperature settings? Someone changed the prompt?
The model belongs on top for insights, not in the calculation path. Deterministic code runs the formula, logs every input, produces the same answer every time. Then Claude analyzes trends and suggests adjustments to the formula itself.