Skip to main content

Debugging OpenAI Logs

Use trace IDs and artifacts rather than guessing from rendered output.

Workflow​

  1. identify the affected case run
  2. inspect artifacts/llm-traces/<case>/
  3. correlate response IDs and prompt stages
  4. compare model output to deterministic projection and final report rendering

The most common mistake is assuming the LLM caused a report defect when the projection or rendering layer actually dropped or misrouted the evidence.