{
  "baseline_label": "B0",
  "bootstrap": {
    "interpretation": "descriptive paired benchmark interval; rows are not a sampled query population",
    "samples": 10000,
    "seed": "v24.3-joint-lexeme-dose29-retention",
    "unit": "row"
  },
  "control_label": "B0",
  "created_at": "2026-07-15T17:46:49.877265+00:00",
  "decision": {
    "assessments": {
      "J69_s0900": {
        "checks": {
          "all_outputs_nonempty": true,
          "elder_diagnostic_chrf_noninferior": true,
          "elder_diagnostic_empty_zero": true,
          "elder_diagnostic_loops_not_increased": true,
          "lexical_exact_gain_over_control": true,
          "natural_dev_text36_chrf_noninferior": true,
          "natural_dev_text36_empty_zero": true,
          "natural_dev_text36_loops_not_increased": true,
          "synthetic_dev_chrf_noninferior": true,
          "synthetic_dev_empty_zero": true,
          "synthetic_dev_loops_not_increased": true,
          "usage_diagnostic_chrf_noninferior": true,
          "usage_diagnostic_empty_zero": true,
          "usage_diagnostic_loops_not_increased": true
        },
        "dose": 1000000,
        "eligible": true,
        "lexical_exact_gain_over_control": 20,
        "lexical_exact_paired_difference": {
          "descriptive_row_bootstrap_95_high": 11.447811447811448,
          "descriptive_row_bootstrap_95_low": 2.356902356902357,
          "percentage_points": 6.7340067340067336
        },
        "mean_retention_chrf_delta": 2.231445078930668,
        "minimum_retention_chrf_delta": 0.439048848140132,
        "retention_chrf_delta_from_stronger_control": {
          "elder_diagnostic": 0.439048848140132,
          "natural_dev_text36": 3.1980733149968437,
          "synthetic_dev": 4.512079964857236,
          "usage_diagnostic": 0.776578187728461
        }
      },
      "J69_s1800": {
        "checks": {
          "all_outputs_nonempty": true,
          "elder_diagnostic_chrf_noninferior": true,
          "elder_diagnostic_empty_zero": true,
          "elder_diagnostic_loops_not_increased": true,
          "lexical_exact_gain_over_control": true,
          "natural_dev_text36_chrf_noninferior": true,
          "natural_dev_text36_empty_zero": true,
          "natural_dev_text36_loops_not_increased": true,
          "synthetic_dev_chrf_noninferior": true,
          "synthetic_dev_empty_zero": true,
          "synthetic_dev_loops_not_increased": true,
          "usage_diagnostic_chrf_noninferior": true,
          "usage_diagnostic_empty_zero": true,
          "usage_diagnostic_loops_not_increased": true
        },
        "dose": 1000000,
        "eligible": true,
        "lexical_exact_gain_over_control": 40,
        "lexical_exact_paired_difference": {
          "descriptive_row_bootstrap_95_high": 18.855218855218855,
          "descriptive_row_bootstrap_95_low": 8.417508417508419,
          "percentage_points": 13.468013468013467
        },
        "mean_retention_chrf_delta": 4.905562167460687,
        "minimum_retention_chrf_delta": 0.945519864198527,
        "retention_chrf_delta_from_stronger_control": {
          "elder_diagnostic": 0.945519864198527,
          "natural_dev_text36": 5.100579522145747,
          "synthetic_dev": 8.62012268644063,
          "usage_diagnostic": 4.956026597057843
        }
      },
      "J69_s2700": {
        "checks": {
          "all_outputs_nonempty": true,
          "elder_diagnostic_chrf_noninferior": true,
          "elder_diagnostic_empty_zero": true,
          "elder_diagnostic_loops_not_increased": true,
          "lexical_exact_gain_over_control": true,
          "natural_dev_text36_chrf_noninferior": true,
          "natural_dev_text36_empty_zero": true,
          "natural_dev_text36_loops_not_increased": true,
          "synthetic_dev_chrf_noninferior": true,
          "synthetic_dev_empty_zero": true,
          "synthetic_dev_loops_not_increased": true,
          "usage_diagnostic_chrf_noninferior": true,
          "usage_diagnostic_empty_zero": false,
          "usage_diagnostic_loops_not_increased": true
        },
        "dose": 1000000,
        "eligible": false,
        "lexical_exact_gain_over_control": 45,
        "lexical_exact_paired_difference": {
          "descriptive_row_bootstrap_95_high": 20.2020202020202,
          "descriptive_row_bootstrap_95_low": 10.1010101010101,
          "percentage_points": 15.151515151515152
        },
        "mean_retention_chrf_delta": 5.776385216810932,
        "minimum_retention_chrf_delta": 1.0722849961428835,
        "retention_chrf_delta_from_stronger_control": {
          "elder_diagnostic": 1.0722849961428835,
          "natural_dev_text36": 5.999971872055813,
          "synthetic_dev": 9.702506484001347,
          "usage_diagnostic": 6.330777515043685
        }
      },
      "J69_s3598": {
        "checks": {
          "all_outputs_nonempty": true,
          "elder_diagnostic_chrf_noninferior": true,
          "elder_diagnostic_empty_zero": true,
          "elder_diagnostic_loops_not_increased": true,
          "lexical_exact_gain_over_control": true,
          "natural_dev_text36_chrf_noninferior": true,
          "natural_dev_text36_empty_zero": true,
          "natural_dev_text36_loops_not_increased": true,
          "synthetic_dev_chrf_noninferior": true,
          "synthetic_dev_empty_zero": true,
          "synthetic_dev_loops_not_increased": true,
          "usage_diagnostic_chrf_noninferior": true,
          "usage_diagnostic_empty_zero": true,
          "usage_diagnostic_loops_not_increased": true
        },
        "dose": 1000000,
        "eligible": true,
        "lexical_exact_gain_over_control": 43,
        "lexical_exact_paired_difference": {
          "descriptive_row_bootstrap_95_high": 19.865319865319865,
          "descriptive_row_bootstrap_95_low": 9.090909090909092,
          "percentage_points": 14.47811447811448
        },
        "mean_retention_chrf_delta": 5.370022937107108,
        "minimum_retention_chrf_delta": 0.6656234052750634,
        "retention_chrf_delta_from_stronger_control": {
          "elder_diagnostic": 0.6656234052750634,
          "natural_dev_text36": 5.594655686933461,
          "synthetic_dev": 10.462328446429353,
          "usage_diagnostic": 4.7574842097905545
        }
      }
    },
    "selected_label": "J69_s3598",
    "selection_order": [
      "all hard checks pass",
      "highest closed-set lexical exact count",
      "best worst-case retention chrF++ delta",
      "best mean retention chrF++ delta",
      "lowest lexical dose"
    ],
    "status": "ADVANCE"
  },
  "lexical_endpoint": "lexicon_closed_set",
  "models": {
    "B0": {
      "lexical": {
        "empty_outputs": 0,
        "interpretation": "Closed-set governed reconstruction. Wilson bounds are a conservative frozen-benchmark description, not a future-query population reliability interval.",
        "normalized_exact_count": 48,
        "normalized_exact_percent": 16.161616161616163,
        "rows": 297,
        "source_copy_outputs": 10,
        "strata": {
          "target_not_seen_in_documented_v21_2": {
            "exact_count": 1,
            "exact_percent": 0.9523809523809524,
            "rows": 105
          },
          "target_seen_in_documented_v21_2": {
            "exact_count": 47,
            "exact_percent": 24.479166666666664,
            "rows": 192
          }
        },
        "token_contains_accepted_percent": 16.161616161616163,
        "wilson_95": {
          "high": 0.20775600744370654,
          "low": 0.1241179942715992
        }
      },
      "prediction_file": "/workspace/v24.3-joint-lexeme-dose29-rerun1/predictions/B0.retention.json",
      "prediction_sha256": "5dc094cb5298f09220af65d7cfeffba4dc38680862c95e888b11b297013e8f4b",
      "sentence_endpoints": {
        "elder_diagnostic": {
          "corpus_chrf": 24.616066875984558,
          "empty_outputs": 0,
          "exact_count": 0,
          "mean_sentence_chrf": 26.623232014816452,
          "mean_token_length_ratio": 0.5880877668028823,
          "repeated_segment_rows_at_least_10": 0,
          "rows": 43,
          "source_copy_outputs": 0,
          "underlength_rows_below_0_5": 13
        },
        "natural_dev_text36": {
          "corpus_chrf": 30.63483112608454,
          "empty_outputs": 0,
          "exact_count": 0,
          "mean_sentence_chrf": 29.469202880107375,
          "mean_token_length_ratio": 1.4235849056603773,
          "repeated_segment_rows_at_least_10": 0,
          "rows": 53,
          "source_copy_outputs": 0,
          "underlength_rows_below_0_5": 0
        },
        "synthetic_dev": {
          "corpus_chrf": 47.72283585854748,
          "empty_outputs": 0,
          "exact_count": 396,
          "mean_sentence_chrf": 62.014714148558014,
          "mean_token_length_ratio": 0.9367398394820023,
          "repeated_segment_rows_at_least_10": 0,
          "rows": 1609,
          "source_copy_outputs": 0,
          "underlength_rows_below_0_5": 83
        },
        "usage_diagnostic": {
          "corpus_chrf": 48.412271340338926,
          "empty_outputs": 0,
          "exact_count": 3,
          "mean_sentence_chrf": 49.17943516814359,
          "mean_token_length_ratio": 0.9473450491307635,
          "repeated_segment_rows_at_least_10": 0,
          "rows": 84,
          "source_copy_outputs": 1,
          "underlength_rows_below_0_5": 1
        }
      }
    },
    "J69_s0900": {
      "lexical": {
        "empty_outputs": 0,
        "interpretation": "Closed-set governed reconstruction. Wilson bounds are a conservative frozen-benchmark description, not a future-query population reliability interval.",
        "normalized_exact_count": 68,
        "normalized_exact_percent": 22.895622895622896,
        "rows": 297,
        "source_copy_outputs": 3,
        "strata": {
          "target_not_seen_in_documented_v21_2": {
            "exact_count": 2,
            "exact_percent": 1.9047619047619049,
            "rows": 105
          },
          "target_seen_in_documented_v21_2": {
            "exact_count": 66,
            "exact_percent": 34.375,
            "rows": 192
          }
        },
        "token_contains_accepted_percent": 22.895622895622896,
        "wilson_95": {
          "high": 0.2800214598628667,
          "low": 0.1848129395210349
        }
      },
      "prediction_file": "/workspace/v24.3-joint-lexeme-dose29-rerun1/predictions/J69_s0900.retention.json",
      "prediction_sha256": "15c356a55124be01f8a1963dda8d5ac205ad3c82f74079c6ecfe861a1cce9c05",
      "sentence_endpoints": {
        "elder_diagnostic": {
          "corpus_chrf": 25.05511572412469,
          "empty_outputs": 0,
          "exact_count": 0,
          "mean_sentence_chrf": 27.262812679521662,
          "mean_token_length_ratio": 0.6041727544970414,
          "repeated_segment_rows_at_least_10": 0,
          "rows": 43,
          "source_copy_outputs": 0,
          "underlength_rows_below_0_5": 10
        },
        "natural_dev_text36": {
          "corpus_chrf": 33.83290444108138,
          "empty_outputs": 0,
          "exact_count": 0,
          "mean_sentence_chrf": 31.56488108068914,
          "mean_token_length_ratio": 1.2352201257861635,
          "repeated_segment_rows_at_least_10": 0,
          "rows": 53,
          "source_copy_outputs": 0,
          "underlength_rows_below_0_5": 3
        },
        "synthetic_dev": {
          "corpus_chrf": 52.23491582340471,
          "empty_outputs": 0,
          "exact_count": 595,
          "mean_sentence_chrf": 68.7948655450125,
          "mean_token_length_ratio": 0.9402293252550047,
          "repeated_segment_rows_at_least_10": 0,
          "rows": 1609,
          "source_copy_outputs": 0,
          "underlength_rows_below_0_5": 76
        },
        "usage_diagnostic": {
          "corpus_chrf": 49.18884952806739,
          "empty_outputs": 0,
          "exact_count": 6,
          "mean_sentence_chrf": 50.270612547337805,
          "mean_token_length_ratio": 1.0122590702947845,
          "repeated_segment_rows_at_least_10": 0,
          "rows": 84,
          "source_copy_outputs": 0,
          "underlength_rows_below_0_5": 2
        }
      }
    },
    "J69_s1800": {
      "lexical": {
        "empty_outputs": 0,
        "interpretation": "Closed-set governed reconstruction. Wilson bounds are a conservative frozen-benchmark description, not a future-query population reliability interval.",
        "normalized_exact_count": 88,
        "normalized_exact_percent": 29.629629629629626,
        "rows": 297,
        "source_copy_outputs": 6,
        "strata": {
          "target_not_seen_in_documented_v21_2": {
            "exact_count": 6,
            "exact_percent": 5.714285714285714,
            "rows": 105
          },
          "target_seen_in_documented_v21_2": {
            "exact_count": 82,
            "exact_percent": 42.70833333333333,
            "rows": 192
          }
        },
        "token_contains_accepted_percent": 29.96632996632997,
        "wilson_95": {
          "high": 0.3505614421400262,
          "low": 0.24723335491247128
        }
      },
      "prediction_file": "/workspace/v24.3-joint-lexeme-dose29-rerun1/predictions/J69_s1800.retention.json",
      "prediction_sha256": "65a71808bab03c90b518fea581a23668bc86cdac30ca72bc6b42cb78135c9dc0",
      "sentence_endpoints": {
        "elder_diagnostic": {
          "corpus_chrf": 25.561586740183085,
          "empty_outputs": 0,
          "exact_count": 0,
          "mean_sentence_chrf": 27.437130846992417,
          "mean_token_length_ratio": 0.5970282845047321,
          "repeated_segment_rows_at_least_10": 0,
          "rows": 43,
          "source_copy_outputs": 0,
          "underlength_rows_below_0_5": 12
        },
        "natural_dev_text36": {
          "corpus_chrf": 35.735410648230285,
          "empty_outputs": 0,
          "exact_count": 0,
          "mean_sentence_chrf": 33.21525527204322,
          "mean_token_length_ratio": 1.3169811320754716,
          "repeated_segment_rows_at_least_10": 0,
          "rows": 53,
          "source_copy_outputs": 0,
          "underlength_rows_below_0_5": 1
        },
        "synthetic_dev": {
          "corpus_chrf": 56.34295854498811,
          "empty_outputs": 0,
          "exact_count": 779,
          "mean_sentence_chrf": 74.5980101726973,
          "mean_token_length_ratio": 0.9505425446562038,
          "repeated_segment_rows_at_least_10": 0,
          "rows": 1609,
          "source_copy_outputs": 0,
          "underlength_rows_below_0_5": 73
        },
        "usage_diagnostic": {
          "corpus_chrf": 53.36829793739677,
          "empty_outputs": 0,
          "exact_count": 8,
          "mean_sentence_chrf": 53.9335576092068,
          "mean_token_length_ratio": 1.0063303099017384,
          "repeated_segment_rows_at_least_10": 0,
          "rows": 84,
          "source_copy_outputs": 0,
          "underlength_rows_below_0_5": 2
        }
      }
    },
    "J69_s2700": {
      "lexical": {
        "empty_outputs": 0,
        "interpretation": "Closed-set governed reconstruction. Wilson bounds are a conservative frozen-benchmark description, not a future-query population reliability interval.",
        "normalized_exact_count": 93,
        "normalized_exact_percent": 31.313131313131315,
        "rows": 297,
        "source_copy_outputs": 6,
        "strata": {
          "target_not_seen_in_documented_v21_2": {
            "exact_count": 9,
            "exact_percent": 8.571428571428571,
            "rows": 105
          },
          "target_seen_in_documented_v21_2": {
            "exact_count": 84,
            "exact_percent": 43.75,
            "rows": 192
          }
        },
        "token_contains_accepted_percent": 31.64983164983165,
        "wilson_95": {
          "high": 0.3679775429251391,
          "low": 0.2630573535445074
        }
      },
      "prediction_file": "/workspace/v24.3-joint-lexeme-dose29-rerun1/predictions/J69_s2700.retention.json",
      "prediction_sha256": "8d485ba0e3c2d8c0438d1a3c6def7ac5a66c49f016c3df4b474638ee0ac31e07",
      "sentence_endpoints": {
        "elder_diagnostic": {
          "corpus_chrf": 25.68835187212744,
          "empty_outputs": 0,
          "exact_count": 0,
          "mean_sentence_chrf": 27.472106802479413,
          "mean_token_length_ratio": 0.594930807433617,
          "repeated_segment_rows_at_least_10": 0,
          "rows": 43,
          "source_copy_outputs": 0,
          "underlength_rows_below_0_5": 12
        },
        "natural_dev_text36": {
          "corpus_chrf": 36.63480299814035,
          "empty_outputs": 0,
          "exact_count": 0,
          "mean_sentence_chrf": 34.071486017964524,
          "mean_token_length_ratio": 1.2968553459119496,
          "repeated_segment_rows_at_least_10": 0,
          "rows": 53,
          "source_copy_outputs": 0,
          "underlength_rows_below_0_5": 2
        },
        "synthetic_dev": {
          "corpus_chrf": 57.425342342548824,
          "empty_outputs": 0,
          "exact_count": 829,
          "mean_sentence_chrf": 76.38321985134282,
          "mean_token_length_ratio": 0.9449671209151963,
          "repeated_segment_rows_at_least_10": 0,
          "rows": 1609,
          "source_copy_outputs": 0,
          "underlength_rows_below_0_5": 74
        },
        "usage_diagnostic": {
          "corpus_chrf": 54.74304885538261,
          "empty_outputs": 1,
          "exact_count": 9,
          "mean_sentence_chrf": 55.64473526044015,
          "mean_token_length_ratio": 0.9803571428571428,
          "repeated_segment_rows_at_least_10": 0,
          "rows": 84,
          "source_copy_outputs": 0,
          "underlength_rows_below_0_5": 2
        }
      }
    },
    "J69_s3598": {
      "lexical": {
        "empty_outputs": 0,
        "interpretation": "Closed-set governed reconstruction. Wilson bounds are a conservative frozen-benchmark description, not a future-query population reliability interval.",
        "normalized_exact_count": 91,
        "normalized_exact_percent": 30.63973063973064,
        "rows": 297,
        "source_copy_outputs": 5,
        "strata": {
          "target_not_seen_in_documented_v21_2": {
            "exact_count": 8,
            "exact_percent": 7.6190476190476195,
            "rows": 105
          },
          "target_seen_in_documented_v21_2": {
            "exact_count": 83,
            "exact_percent": 43.22916666666667,
            "rows": 192
          }
        },
        "token_contains_accepted_percent": 30.976430976430976,
        "wilson_95": {
          "high": 0.36102078255623826,
          "low": 0.25671807414654857
        }
      },
      "prediction_file": "/workspace/v24.3-joint-lexeme-dose29-rerun1/predictions/J69_s3598.retention.json",
      "prediction_sha256": "a39927380b2cad5b3c5a2146183a26b20e4379e9f165a73721182b7528102542",
      "sentence_endpoints": {
        "elder_diagnostic": {
          "corpus_chrf": 25.28169028125962,
          "empty_outputs": 0,
          "exact_count": 0,
          "mean_sentence_chrf": 26.80680366972627,
          "mean_token_length_ratio": 0.5925365142051283,
          "repeated_segment_rows_at_least_10": 0,
          "rows": 43,
          "source_copy_outputs": 0,
          "underlength_rows_below_0_5": 12
        },
        "natural_dev_text36": {
          "corpus_chrf": 36.229486813018,
          "empty_outputs": 0,
          "exact_count": 0,
          "mean_sentence_chrf": 33.894810776286725,
          "mean_token_length_ratio": 1.3088050314465407,
          "repeated_segment_rows_at_least_10": 0,
          "rows": 53,
          "source_copy_outputs": 0,
          "underlength_rows_below_0_5": 2
        },
        "synthetic_dev": {
          "corpus_chrf": 58.18516430497683,
          "empty_outputs": 0,
          "exact_count": 857,
          "mean_sentence_chrf": 77.13974052582428,
          "mean_token_length_ratio": 0.9497983546661658,
          "repeated_segment_rows_at_least_10": 0,
          "rows": 1609,
          "source_copy_outputs": 0,
          "underlength_rows_below_0_5": 67
        },
        "usage_diagnostic": {
          "corpus_chrf": 53.16975555012948,
          "empty_outputs": 0,
          "exact_count": 6,
          "mean_sentence_chrf": 54.033433933973,
          "mean_token_length_ratio": 1.0061507936507936,
          "repeated_segment_rows_at_least_10": 0,
          "rows": 84,
          "source_copy_outputs": 0,
          "underlength_rows_below_0_5": 1
        }
      }
    }
  },
  "promotion_effect": "none; ADVANCE only authorizes the next research stage",
  "retention_endpoints": [
    "synthetic_dev",
    "natural_dev_text36",
    "usage_diagnostic",
    "elder_diagnostic"
  ],
  "schema_version": 1,
  "stage": "one-seed-development-only-screen",
  "thresholds": {
    "blank_outputs": 0,
    "chrf_noninferiority_margin": 1.0,
    "minimum_exact_gain_over_control": 1,
    "repeated_segment_rows": "must not exceed stronger of B0 and C0"
  }
}