Two-Stage Product Category Assignment

Stage one request, top level of the taxonomy:

{
  "model": "jev-latest",
  "state": {
    "product": {
      "title": "{{product-title}}",
      "description": "{{product-description}}",
      "attributes": "{{product-attributes}}"
    }
  },
  "questions": {
    "department": {
      "type": "choice",
      "instructions": "Which top-level department best matches this product? Read the subtree contents in each option before choosing.",
      "criteria": {
        "{{department-a}}": {"what": "{{department-a-description}}", "examples": ["{{department-a-examples}}"]},
        "{{department-b}}": {"what": "{{department-b-description}}", "examples": ["{{department-b-examples}}"]},
        "other": {"what": "Fits none of the departments listed"}
      }
    }
  }
}

Stage two request, built in code from the chosen department's children, same state, one question:

"subcategory": {
  "type": "choice",
  "instructions": "Within the chosen department, which subcategory best matches this product?",
  "criteria": { one entry per child, each with what and examples, plus "other" }
}

Traversal rule: keep the top two departments from stage one when the leader's probability is below 0.6, run stage two for both, and pick the path with the higher geometric mean of edge probabilities. That lets a clear subcategory match repair an ambiguous department call. Send anything whose final confidence is below 0.7 to a human with both candidate paths shown.