{
  "meta_info": {
    "case_id": "case13_09",
    "bench_subset": "NeuroBench-Multi",
    "difficult": 1.5,
    "original_dataset": "SEED-VIG"
  },
  "agent_input": {
    "data_path": "data/multi/SEEDVIG_03_session-20151024_noon.edf",
    "instruction": "Analyze EEG channel `POZ` and EOG channel `eog_h` over [160, 240).\n\nDivide both streams into aligned, non-overlapping 8-second windows and pair the windows by index.\n\nFor each window, compute `theta_alpha_ratio` from the EEG signal as theta-to-alpha power ratio using [4, 8) and [8, 13) Hz. Compute `blink_amplitude_mean` from the EOG signal as mean absolute amplitude at detected blinks.\n\nKeep windows with finite values for both features. If at least 6 valid pairs remain and both feature sequences have non-zero variance, compute their Pearson correlation. Otherwise, the sign cannot be determined.\n\nReport correlation_sign as `positive` when the correlation is greater than zero and `negative` when it is less than zero. Report null when the correlation cannot be determined or is exactly zero.\n\nReport only correlation_sign."
  },
  "eval_config": {
    "parser_prompt": "### ROLE\nYou are a precise JSON extraction engine for multimodal physiological correlation reports.\n\n### TASK\nExtract `correlation_sign`, the reported sign of the correlation between the two\nwindow-level feature sequences.\n\n### STRICT CONSTRAINTS\n1. Output ONLY one valid JSON object.\n2. Do not output Markdown, explanations, units, numeric correlation values, or extra fields.\n3. The only key must be `correlation_sign`.\n4. The value must be exactly `positive` or `negative`, or null if the sign is\n   missing, zero, or unclear.\n\n### OUTPUT TEMPLATE\n{\"correlation_sign\": \"<positive|negative|null>\"}",
    "metrics": [
      {
        "metric_id": "correlation_sign",
        "type": "categorical_check",
        "target_key": "correlation_sign",
        "weight": 100,
        "params": {
          "gt_value": "positive"
        }
      }
    ]
  }
}