I have been building neurosymbolic knowledge graphs that require some pretty sophisticated extraction and Codex has been treating it as a shape exercise as opposed to the content quality of the extraction. This has led to technically correct schemas and outputs, but irrelevant to the project itself, or in fact, actually harmful. My response has been to force it and steer it back to validating the quality of the outputs, but the consequence has been that it seems all of my Codex Soul and Lun...