{
  "meta_info": {
    "case_id": "case36_02",
    "bench_subset": "NeuroBench-Multi",
    "difficult": 2,
    "original_dataset": "DEAP"
  },
  "agent_input": {
    "data_path": "data/multi/DEAP_01.bdf",
    "label_path": "data/multi/DEAP_01.npy",
    "instruction": "Select all DEAP trials with both arousal and dominance ratings at least `6`, regardless of valence or liking.\n\nMatch rating row `i` to trial `i` in chronological order, using `Status = 4` as stimulus onset and its following `Status = 5` marker as stimulus end.\n\nFor each selected trial, use the 3 seconds before onset as baseline and divide the following 60 seconds into six consecutive 10-second windows. Exclude a trial only if these intervals are not fully available.\n\nFor EEG channels `P3`, `P4`, `PO3`, `PO4`, `O1`, and `O2`, compute alpha power in `[8, 13)` Hz using a 2-second Hann-window Welch PSD with 50% overlap. Average alpha power across the six channels first, then compute each window's suppression as `log(baseline power + 1e-12) - log(window power + 1e-12)`. Select the window with the largest suppression.\n\nFor `GSR1`, subtract the baseline mean from each window mean and select the window with the largest increase. If maxima are tied, use the earliest window.\n\nFor each trial, define `peak lag = GSR peak-window index - EEG peak-window index`. Report the median lag across trials as `median_peak_lag_windows` and the fraction of trials with positive lag as `gsr_after_eeg_fraction`.\n\nReport only `median_peak_lag_windows` and `gsr_after_eeg_fraction`."
  },
  "eval_config": {
    "parser_prompt": "### ROLE\nYou are a precise JSON extraction engine for a DEAP cortical--autonomic timing analysis.\n\n### TASK\nExtract the median GSR-minus-EEG peak-window lag and the fraction of selected trials whose GSR peak occurs later than the EEG peak.\n\n### REQUIRED FIELDS\n- `median_peak_lag_windows`: a finite number, or `null`\n- `gsr_after_eeg_fraction`: a finite number between 0 and 1, or `null`\n\n### STRICT CONSTRAINTS\n1. Output only one valid JSON object.\n2. Use exactly the two required keys.\n3. Do not include Markdown, explanations, units, labels, or extra fields.\n4. Preserve the lag direction: positive values mean that the GSR peak is later.\n5. Do not calculate or infer a missing value.\n\n### OUTPUT TEMPLATE\n{\"median_peak_lag_windows\": <number or null>, \"gsr_after_eeg_fraction\": <number or null>}",
    "metrics": [
      {
        "metric_id": "median_peak_lag_windows",
        "type": "numeric_check",
        "target_key": "median_peak_lag_windows",
        "weight": 50,
        "params": {
          "gt_value": -1.5,
          "tolerance": 0.01
        }
      },
      {
        "metric_id": "gsr_after_eeg_fraction",
        "type": "numeric_check",
        "target_key": "gsr_after_eeg_fraction",
        "weight": 50,
        "params": {
          "gt_value": 0.15,
          "tolerance": 0.025
        }
      }
    ]
  }
}