Given this program specification:
"{specification}"

{examples_context}

Generate {num_examples} test INPUTS that will be run through the compiled neural program.
DO NOT generate outputs - the neural program will generate those when we run the tests.

Your task:
1. Analyze the specification and examples to determine the expected input format
2. Generate straightforward, representative test inputs in the correct format

IMPORTANT: Generate typical, realistic inputs (not edge cases or adversarial examples).
The goal is to demonstrate the program works correctly, not to test its limits.

Input Format Rules:
- If the program expects TEXT only: return input as a string
- If the program expects IMAGES: return input as an array with image placeholders
- If the program expects MIXED (text + images): return input as an array with both

For images, use placeholders like:
- {{"type": "image", "placeholder": "descriptive_name"}}
Examples: "fruit_basket", "person_photo", "document_scan", "cat_photo", etc.

Format your response as JSON:
{{
  "input_format_description": "Brief description of expected input format (e.g., 'single image', 'text then image', 'text only')",
  "examples": [
    {{
      "input": "text input"  // For text-only
    }},
    {{
      "input": [{{"type": "image", "placeholder": "example_image"}}]  // For image-only
    }},
    {{
      "input": ["some text", {{"type": "image", "placeholder": "example_photo"}}]  // For mixed
    }}
  ]
}}

Make the test inputs realistic, diverse, and cover different edge cases.
REMEMBER: Only generate INPUTS, not outputs. The neural program will produce outputs when we run it.
