Statistics Summary for op agent (mean ± std):
==================================================
accuracy: 0.0000 ± nan
total_entries: 126.0000 ± nan
correct_matches: 0.0000 ± nan
average_execution_time: 1.0636 ± nan
metrics_by_complexity:
  complex:
    count: 2.0000 ± nan
    correct_matches: 0.0000 ± nan
    pv_matched: 1.0000 ± nan
    pv_mismatched: 5.0000 ± nan
    pv_match_rate: 16.6667 ± nan
    timing_matched: 0.0000 ± nan
    timing_mismatched: 6.0000 ± nan
    timing_match_rate: 0.0000 ± nan
    temp_matched: 1.0000 ± nan
    temp_mismatched: 1.0000 ± nan
    temp_match_rate: 50.0000 ± nan
    full_matched: 0.0000 ± nan
    full_mismatched: 6.0000 ± nan
    full_match_rate: 0.0000 ± nan
    accuracy: 0.0000 ± nan
    average_best_codebleu:
      codebleu: 0.3514 ± nan
      ngram_match_score: 0.0545 ± nan
      weighted_ngram_match_score: 0.0641 ± nan
      syntax_match_score: 0.7154 ± nan
      dataflow_match_score: 0.5714 ± nan
    average_best_levenshtein: 142.0000 ± nan
    average_best_normalized_levenshtein: 0.4703 ± nan
    average_execution_time: 1.2419 ± nan
==================================================
