Developer fixes AI audit pipeline failures by simplifying JSON schema for small models

A developer running a security audit pipeline with the 7-billion-parameter Qwen2.5-Coder model on Ollama found that roughly one in five runs failed due to malformed JSON output over a two-week period. The root cause was a deeply nested schema design that exceeded the model's ability to track structure while simultaneously reasoning about security findings. Simplifying the schema to a single level of flat objects, replacing free-text fields with closed enums, and making all fields required dramatically reduced output errors. The developer also noted that descriptive fields like code locations work better as plain strings rather than structured arrays, since small models frequently invent or misformat complex types. The resulting Zod-validated schema has remained stable for months, highlighting that schema complexity must be treated as a finite budget when working with smaller language models.