You are an expert in computer vision and explainable artificial intelligence.
Your task is to analyze visual examples and discover interpretable, human-understandable
features that can differentiate between {n_classes} visual categories:
{class_list}

You must reason purely from the visual evidence itself, without assuming any predefined
feature types, and describe the distinguishing properties as if explaining them to a
human researcher.

I am providing you with example images. Each image belongs to one of the {n_classes} categories
listed above, but you are NOT told which image belongs to which category.

Do not use the category names or paraphrases of them as feature names or
values. Describe the property itself instead, in general terms rather than
names or phrases taken from the examples.

Your task is to:
  1. Examine all images carefully and propose key visual features that appear to
     systematically vary across subsets of the images and together distinguish all
     {n_classes} categories in a classification model.
  2. Focus on characteristics meaningful to a human observer: differences in structure,
     shape, proportion, colour, texture, material, markings, or the way parts relate
     to each other.
  3. Prefer intrinsic properties of the depicted subject over properties of the
     photograph. Background, lighting, framing, or image quality are only worth
     proposing when they genuinely distinguish the categories rather than the
     circumstances of the shot.
  4. Each feature must capture a different property. Do not propose two features
     that would almost always take related values.
  5. Some categories will resemble each other much more closely than others.
     Make sure the feature set also separates the closest pairs, not only the
     obvious groupings.
  6. Express each feature as a short snake_case attribute name.
  7. For each feature, provide a brief explanation of why it could be discriminative
     across the {n_classes} hidden groups.
  8. For each feature, provide 3-6 possible values: short snake_case strings
     without filler words, mutually exclusive, together covering the realistic
     range. If a feature cannot apply to some of the categories, include an
     explicit value for that case (for example not_applicable) in the list.
     These exact values will later be used to label new images, so choose
     values you can apply consistently.
  9. Provide at least {min_features} distinct features.
  10. Respond only with valid JSON. Do not wrap it in markdown code fences and do not include any explanation outside the JSON.

Output format:
{{
  "proposed_features": [
    {{
      "feature": "feature_name_1",
      "description": "why it separates the {n_classes} groups",
      "possible_values": ["value1", "value2", "value3"]
    }}
  ]
}}