Skip to content

Trace and re-judge history in IRIS

Trace. Each cached LLM result (Basic, tag llm-cache) has a second extension llm-trace: system prompt, input, tool steps with hits, reasoning summary, raw output and rejected items. GET /api/triggers/{id}/llm-trace returns them. One write and one read, so the trace cannot drift from its result.

Re-judge. The Collector asks for a new verdict on one prior, with an optional note. All priors go as context. Each result is a Basic tagged llm-override. The newest one is active. Restoring an old one adds it again as the newest. History is append-only.

  • In-memory log: empty after every deploy.
  • Separate trace resource: two writes that can drift.
  • Overwrite the cached result on re-judge: loses history.
  • Re-judge the whole case: changes verdicts the user did not ask about.
  • Raw chain of thought is not available. “Reasoning” is the model’s reasoning summary.
  • A failed re-judge is not stored. The API answers 502.
  • A Collector decision always wins over any AI verdict.