{
  "schema_version": 1,
  "cutoff": "2026-11-14 Asia/Shanghai",
  "purpose": "Published retrospective evidence; no model weights or private/real payment records included. Exported report fields and original synthetic prediction records allow metric recalculation, not independent retraining.",
  "source_snapshot_commit": "6647c1bc30dc10c1e0d27b38f8f34df94a57a98a",
  "base_models": {
    "e5": {
      "id": "intfloat/multilingual-e5-small",
      "revision": "614241f622f53c4eeff9890bdc4f31cfecc418b3",
      "license": "MIT"
    },
    "native": {
      "id": "Qwen/Qwen3-1.7B",
      "revision": "70d244cc86ccca08cf5af4e1e306ecf908b1ad5e",
      "license": "Apache-2.0"
    }
  },
  "e5_v1": {
    "evaluation_role": "Exploratory reused synthetic test; not independent payment validation",
    "test_examples": 960,
    "methods": {
      "uniform": {
        "count": 960,
        "accuracy": 0.303125,
        "nll": 1.3647815207406997,
        "brier_multiclass_sum": 0.7111111111111115,
        "ece_10_bins": 0.014236111111111106
      },
      "train_majority": {
        "count": 960,
        "accuracy": 0.3770833333333333,
        "nll": 1.2755177621275253,
        "brier_multiclass_sum": 0.6769675925925926,
        "ece_10_bins": 0.01180555555555555
      },
      "tfidf_fixed_vocabulary": {
        "count": 960,
        "accuracy": 0.7072916666666667,
        "nll": 0.8820875666223112,
        "brier_multiclass_sum": 0.47863065716211317,
        "ece_10_bins": 0.21764808179921885
      },
      "trained_scorer_raw": {
        "count": 960,
        "accuracy": 0.640625,
        "nll": 0.8708030674238718,
        "brier_multiclass_sum": 0.46396099367184135,
        "ece_10_bins": 0.030501947810964013
      },
      "trained_scorer_calibrated": {
        "count": 960,
        "accuracy": 0.640625,
        "nll": 0.9527929958309654,
        "brier_multiclass_sum": 0.4985226465375998,
        "ece_10_bins": 0.1450815223437118
      },
      "finetuned_encoder_cosine_uncalibrated": {
        "count": 960,
        "accuracy": 0.30416666666666664,
        "nll": 1.3644097266732063,
        "brier_multiclass_sum": 0.7104773246450304,
        "ece_10_bins": 0.08114191507187926
      }
    },
    "frozen_head": {
      "count": 960,
      "accuracy": 0.45625,
      "nll": 1.2302640580849968,
      "brier_multiclass_sum": 0.6524428410228916,
      "ece_10_bins": 0.07487606699734495
    },
    "frozen_head_trainable_parameters": 393729,
    "frozen_encoder_parameters": 117653760,
    "joint_trainable_parameters": 118047489,
    "joint_config": {
      "epochs": 3,
      "batch_size": 8,
      "gradient_accumulation": 4,
      "encoder_learning_rate": 2e-05,
      "head_learning_rate": 0.0001,
      "weight_decay": 0.01,
      "patience": 2,
      "seed": 20261113
    },
    "selected_epoch": 1,
    "calibration": {
      "evidence_sufficient": {
        "temperature": 2.039512949617754,
        "count": 320,
        "nll_before": 0.6468701470470063,
        "nll_after": 0.5818012432366432
      },
      "next_action": {
        "temperature": 2.4090812176262415,
        "count": 320,
        "nll_before": 1.7475032883968118,
        "nll_after": 1.4363403125700476
      },
      "payment_state": {
        "temperature": 2.2003868410416927,
        "count": 320,
        "nll_before": 1.5023290080809963,
        "nll_after": 1.2724034980932613
      }
    },
    "source_report_sha256": {
      "frozen": "f9ca6d47ab631528822fda6c4438468ee80e3881f5f2c25a57006168acc5a4a5",
      "joint": "1d0447cabefa5624a74a68608945431437f255316c29945ee2ccbd7b418312b3"
    },
    "joint_training_history": [
      {
        "epoch": 1,
        "train_nll": 1.135139419619615,
        "validation_nll": 0.9508499378642598
      },
      {
        "epoch": 2,
        "train_nll": 0.43224510593063314,
        "validation_nll": 2.359482401872052
      },
      {
        "epoch": 3,
        "train_nll": 0.1873129366537493,
        "validation_nll": 2.770654012202626
      }
    ],
    "tfidf_configuration": {
      "source": "src/paydecision/baselines.py:32 TextBaseline",
      "input": "context + newline + question; no candidate descriptions",
      "per_task_models": true,
      "training_records": 4608,
      "vectorizer": {
        "analyzer": "char",
        "ngram_range": [
          2,
          4
        ],
        "max_features": 20000,
        "min_df": 2
      },
      "logistic_regression": {
        "C": 4,
        "max_iter": 500,
        "random_state": 20261113
      }
    }
  },
  "external_e5": {
    "benchmark": "BenchCompass public L1-normal MCQ open-book subset",
    "source": "https://github.com/ant-intl/BenchCompass",
    "revision": "606a32d84c947518244d486369d13b07fbfcc0f9",
    "data_sha256": "c6eba28b1b87da12375dc2b28aac37f080f237739b300df3930eced179387265",
    "total": 95,
    "evaluated": 69,
    "coverage": 0.7263157894736842,
    "calibration": "No synthetic-task temperatures applied to this out-of-distribution task",
    "limitations": [
      "Coverage-limited subset, not the full benchmark",
      "Token-window filtering can introduce selection bias",
      "Public benchmark scores do not establish real-world safety",
      "This benchmark is never used as training data or to choose checkpoints"
    ],
    "scorer_raw": {
      "count": 69,
      "accuracy": 0.21739130434782608,
      "nll": 1.4929933468320322,
      "brier_multiclass_sum": 0.8105241129068305,
      "ece_10_bins": 0.19124492075760233
    },
    "finetuned_encoder_cosine_raw": {
      "count": 69,
      "accuracy": 0.463768115942029,
      "nll": 1.3840603869622712,
      "brier_multiclass_sum": 0.7488812031761336,
      "ece_10_bins": 0.21255635556158664
    }
  },
  "native001": {
    "status": "interrupted_before_calibration_and_test",
    "history": [
      {
        "step": 256,
        "epoch": 1,
        "validation_nll": 1.709574171393078,
        "validation_accuracy": 0.2708333333333333
      },
      {
        "step": 512,
        "epoch": 1,
        "validation_nll": 1.6242749759397173,
        "validation_accuracy": 0.22916666666666666
      }
    ],
    "selection": {
      "epoch": 1,
      "step": 512,
      "validation_accuracy": 0.22916666666666666,
      "validation_nll": 1.6242749759397173
    },
    "config": {
      "alpha": 32,
      "batch_size": 2,
      "dropout": 0.05,
      "epochs": 2,
      "eval_batch_size": 4,
      "eval_every_steps": 256,
      "gradient_accumulation": 8,
      "learning_rate": 0.0001,
      "max_length": 8192,
      "max_train_examples": 1536,
      "rank": 16,
      "seed": 20261114,
      "validation_examples": 192,
      "weight_decay": 0.01
    },
    "trainable_parameters": 3211264,
    "total_parameters_including_adapter": 1723786240,
    "data_hashes": {
      "calibration.jsonl": "762ae764276a5172a68f28b68050061373c9a1065151c28a1d923699a5d29612",
      "source_notes.json": "cd139f8316f90f3ffdef74ec4129f38644dd08271aee9a0ccb8a45c70dbeeb6a",
      "test.jsonl": "744ef68e407502ec4d7e141fa4caf2d4a642fcb1f018dd4509fad2640320746b",
      "train.jsonl": "e3c25967f07ea760df318486a88d3ab67db79aa05fd2846fc0551267cf02375c",
      "validation.jsonl": "4ac74357026c7d94f3d53428312990520dbc82cb43f01825958c30a535dab66f"
    },
    "best_adapter_manifest_sha256": "834bd4a40e3b028802401b61499ceadf1db4b5c8830e0e8e7286936e2ab5de62"
  },
  "native002": {
    "status": "interrupted_before_calibration_and_test",
    "history": [
      {
        "step": 256,
        "epoch": 1,
        "validation_nll": 1.826032241503059,
        "validation_accuracy": 0.20833333333333334
      },
      {
        "step": 512,
        "epoch": 1,
        "validation_nll": 1.6439579984103478,
        "validation_accuracy": 0.22395833333333334
      },
      {
        "step": 768,
        "epoch": 1,
        "validation_nll": 1.5698205569232735,
        "validation_accuracy": 0.2864583333333333
      }
    ],
    "selection": {
      "epoch": 1,
      "step": 768,
      "validation_accuracy": 0.2864583333333333,
      "validation_nll": 1.5698205569232735
    },
    "config": {
      "alpha": 32,
      "batch_size": 1,
      "dropout": 0.05,
      "epochs": 2,
      "eval_batch_size": 2,
      "eval_every_steps": 256,
      "gradient_accumulation": 1,
      "last_trainable_layers": 8,
      "learning_rate": 0.0003,
      "max_length": 8192,
      "max_train_examples": 1536,
      "rank": 16,
      "seed": 20261115,
      "validation_examples": 192,
      "weight_decay": 0.01
    },
    "trainable_parameters": 917504,
    "total_parameters_including_adapter": 1721492480,
    "data_hashes": {
      "calibration.jsonl": "2a0c9ca95a186ef7ce438e11a94a7fddb9c9b070b82450575531b8336b02498f",
      "contracts.jsonl": "fd6bcf6a27693584c216aea584f042546f93de6d479e0af9f703e5c991fb492c",
      "source_notes.json": "322fb5d5b8bf0bd3405a75c3eb974b449c07f7c4a9cdadc9cb48e2046303ce37",
      "test.jsonl": "460466faa9b6eb4a72c5615697cd2e89710030fe3b917d0b1cbf0e7f19bc3c63",
      "train.jsonl": "256f95bc89c9a54fc996ae9c700f8010ee82339c17d85d4b819b8bbd801dece1",
      "validation.jsonl": "d0e0f2c475d51dcabccb63362846a97bbed875921517d63a8d152ac2e8158ff1"
    },
    "best_adapter_manifest_sha256": "6cb6771b690a531c15498b6d37a499f8e121322abd9bdcdbb0b7ba09d3dd0dd0"
  },
  "native_base_validation_console_capture": {
    "native001": {
      "count": 192,
      "accuracy": 0.265625,
      "correct": 51
    },
    "native002": {
      "count": 192,
      "accuracy": 0.2604166666666667,
      "correct": 50
    },
    "provenance": "Values captured in training console and reports/PROGRESS.md; interrupted runs did not produce final report.json."
  },
  "training_stopped": true,
  "stop_verified_at_utc": "2026-11-13T16:18:34.943995+00:00",
  "diagnostics": {
    "native-diagnostic-001": {
      "initial": {
        "train_subset_accuracy": 0.07692307692307693,
        "train_subset_nll": 1.63241821261636
      },
      "final": {
        "train_subset_accuracy": 0.46153846153846156,
        "train_subset_nll": 1.340232495170549
      },
      "restored": {
        "train_subset_accuracy": 0.46153846153846156,
        "train_subset_nll": 1.340232495170549
      },
      "elapsed_seconds": 26.49862283305265,
      "reload_prediction_agreement": true,
      "reload_max_absolute_logit_difference": 0.0,
      "claim_limit": "Only demonstrates learning on this repeated training subset and adapter serialization; not held-out ability",
      "training_rows": 13,
      "steps": 64,
      "learning_rate": 0.0003
    },
    "native-diagnostic-002": {
      "initial": {
        "train_subset_accuracy": 0.46153846153846156,
        "train_subset_nll": 1.340232495170549
      },
      "final": {
        "train_subset_accuracy": 1.0,
        "train_subset_nll": 1.644862469828931e-05
      },
      "restored": {
        "train_subset_accuracy": 1.0,
        "train_subset_nll": 1.644862469828931e-05
      },
      "elapsed_seconds": 79.73531320900656,
      "reload_prediction_agreement": true,
      "reload_max_absolute_logit_difference": 0.0,
      "claim_limit": "Only demonstrates learning on this repeated training subset and adapter serialization; not held-out ability",
      "training_rows": 13,
      "steps": 256,
      "learning_rate": 0.001
    }
  },
  "v21_dataset": {
    "row_counts": {
      "train": 4608,
      "validation": 1152,
      "calibration": 960,
      "test": 960
    },
    "groups_per_split": {
      "train": 192,
      "validation": 48,
      "calibration": 40,
      "test": 40
    },
    "contracts": 640,
    "manifest": {
      "calibration.jsonl": "2a0c9ca95a186ef7ce438e11a94a7fddb9c9b070b82450575531b8336b02498f",
      "contracts.jsonl": "fd6bcf6a27693584c216aea584f042546f93de6d479e0af9f703e5c991fb492c",
      "source_notes.json": "322fb5d5b8bf0bd3405a75c3eb974b449c07f7c4a9cdadc9cb48e2046303ce37",
      "test.jsonl": "460466faa9b6eb4a72c5615697cd2e89710030fe3b917d0b1cbf0e7f19bc3c63",
      "train.jsonl": "256f95bc89c9a54fc996ae9c700f8010ee82339c17d85d4b819b8bbd801dece1",
      "validation.jsonl": "d0e0f2c475d51dcabccb63362846a97bbed875921517d63a8d152ac2e8158ff1"
    },
    "languages": [
      "en",
      "zh",
      "th",
      "id"
    ],
    "tasks": [
      "payment_state",
      "next_action",
      "evidence_sufficient"
    ],
    "independence_unit": "Group: two cases, four language variants, three tasks. Ordinary renderers and semantic families overlap across partitions."
  },
  "published_prediction_sha256": {
    "e5-frozen-predictions.jsonl": "87d2c59591b0f0c7a2f5994c51e7bc43c095faf0db1590c60244ac13a5587acd",
    "e5-joint-predictions.jsonl": "233b23df3265ac83914910d8a35f6c7bcae02b69b48b54f353250f876151efaa"
  },
  "prediction_data_license": {
    "license": "CC-BY-4.0",
    "attribution": "Original synthetic PayDecision experiment data, 2026; published with this Lei blog retrospective. No real payment records."
  },
  "illustrative_math": {
    "purpose": "Explanatory examples only; not model measurements",
    "softmax_scores": [
      1,
      2
    ],
    "softmax_probabilities": [
      0.2689414213699951,
      0.7310585786300049
    ],
    "temperature2_probabilities": [
      0.3775406687981454,
      0.6224593312018545
    ],
    "negative_log_probabilities": {
      "0.8": 0.2231435513142097,
      "0.2": 1.6094379124341003
    }
  },
  "native_prompt_example": {
    "example_id": "v21-validation-7-0-zh-payment_state",
    "source": "Frozen native002 validation example, used only for tokenizer/prompt export; no model inference or training performed",
    "language": "zh",
    "task": "payment_state",
    "letter_token_ids": {
      "A": [
        32
      ],
      "B": [
        33
      ],
      "C": [
        34
      ],
      "D": [
        35
      ],
      "E": [
        36
      ]
    },
    "file": "native-prompt-example.json",
    "sha256": "aeee23e46ea3c7dfc323b4e49cfd0abf239dd36e689e26cce2c9eabe13c5623f",
    "prompt_text_sha256": "4ef546693735c7b5687693a5e5e0d780df8e9015dc73c77e3de80e06327e2adf"
  }
}
