Part of my research responsibilities is building and maintaining a fire simulation program for an area of Eastern Oregon. The idea is to simulate fire and growth under different suppression and fire treatment policies so we can generate an optimal policy. Everything was proceeding smoothly until the closed-source fire spread model refused to simulate any landscape past 40 years. To make things worse, any of 1 million pixels could cause the error, and each pixel could reach its error state via any combination of annual growth and fire transitions.