{
    "meta_info": {
        "case_id": "ISRUC_02.edf",
        "bench_subset": "NeuroBench-Sleep",
        "difficult": 1.5,
        "original_dataset": "ISRUC"
    },
    "agent_input": {
        "label_path": "data/sleep/ISRUC_02.npy",
        "instruction": "Please use the provided sleep-stage label file to calculate the light-sleep to deep-sleep ratio for the whole night. Then interpret what this ratio suggests about the sleep architecture, including whether sleep appears balanced, relatively restorative, too shallow, or atypical. Please report the final ratio and give a brief explanation of the sleep quality implication."
    },
    "eval_config": {
        "parser_prompt": "### ROLE\nYou are a precise JSON Extraction Engine for sleep medicine reports. Your sole task is to convert natural language reports into structured JSON data.\n\n### TASK\nExtract the light-sleep to deep-sleep ratio from the agent's report.\n\n### STRICT CONSTRAINTS (MANDATORY)\n1. Output ONLY a valid JSON object.\n2. DO NOT include Markdown code blocks.\n3. DO NOT include conversational text, explanations, or units.\n4. The key must be EXACTLY \"light_deep_sleep_ratio\".\n5. The value must be a FLOAT or null.\n6. If the value is missing or cannot be found, return {\"light_deep_sleep_ratio\": null}.\n\n### OUTPUT TEMPLATE\n{\"light_deep_sleep_ratio\": <float|null>}",
        "metrics": [
            {
                "metric_id": "light_deep_ratio_accuracy",
                "type": "numeric_check",
                "target_key": "light_deep_sleep_ratio",
                "weight": 40,
                "params": {
                    "gt_value": 2.929936305732484,
                    "tolerance": 0.1464968152866242
                }
            },
            {
                "metric_id": "sleep_quality_semantic_consistency",
                "type": "semantic_consistency",
                "target_key": "full_text",
                "weight": 60,
                "params": {
                    "judge_prompt": "You are judging semantic consistency for a sleep-medicine benchmark. The sample is ISRUC_02.edf. The ground-truth light/deep sleep ratio is 2.929936, where the ratio is defined as (N1+N2)/N3. The correct response must explicitly conclude that the sleep structure is balanced or normal. Mark the response as correct only if its final conclusion matches that rule. If the response gives the opposite conclusion or contradicts that rule, mark it as incorrect. Output only a valid JSON object with exactly one key named status whose value is either True or False. Do not output Markdown, explanations, or additional fields."
                }
            }
        ]
    }
}