Eric Evans explores a core challenge: integrating AI components into conventional deterministic software. The flexibility of AI input is useful, but its non-deterministic outputs make further processing difficult. Using a code analysis example, he shows how LLM responses vary, preventing reliable comparison across different files or modules.

The key insight is separating tasks. Modeling—creating a categorization scheme—is fundamentally different from classification. An LLM can generate a useful taxonomy, but it will differ each time. For consistent results, developers must first create a canonical category list and then use it as a fixed prompt for classifying individual code modules.

This approach allows applications to aggregate and compare results hierarchically. However, it requires accepting that the classification is only consistent within a single, frozen model. For ongoing projects, an incremental method can update the taxonomy as new code is analyzed, but the core principle remains: impose a deterministic structure on AI outputs to make them usable.