{
  "schema_version": "zils.model-comparison/v1",
  "experiment": "abcd-002-jev",
  "completed_date": "2026-10-06",
  "title": "ABCD-trained JevK5 versus TypeSafe Jev",
  "status": "completed",
  "data_mode": "recorded_experiment",
  "task": "Predict the next recorded support tool action when an action is due, among all 30 choices.",
  "dataset": {
    "name": "ABCD",
    "url": "https://github.com/asappresearch/abcd",
    "source_commit": "6b8700ce67c6b37b062dd7a60abc76d7ef832a97",
    "license": "MIT",
    "description": "Public human-roleplay customer-support conversations"
  },
  "splits": {
    "training_conversations": 1024,
    "development_conversations": 500,
    "test_conversations": 500,
    "action_choices": 30,
    "unit": "One selected action per distinct conversation",
    "final_test_used_for_selection": false
  },
  "model": {
    "name": "JevK5 4B",
    "base": "alibiserikbay/JevK5",
    "base_revision": "c4f7fdb3aeab5582336406e78d3bef11bf98833d",
    "adapter_sha256": "298752aa8f12cf9ce0ad2c0de964cf899f4ae8dff3bb74d03a6a2cfa78c848e9",
    "native_temperature": 1.22,
    "knockout_temperature": 0.93
  },
  "comparator": {
    "name": "TypeSafe Jev 1.13.0",
    "model_id": "jev-1.13.0",
    "provider": "TypeSafe",
    "access": "Hosted API",
    "docs": "https://docs.typesafe.ai/models"
  },
  "method": {
    "same_inputs_and_question": true,
    "case_count": 500,
    "new_jev_inference": true,
    "reused_verified_jevk5_predictions": true,
    "test_used_for_training_or_selection": false,
    "jev_prompt_optimized_separately": false,
    "prediction_rule": "Use returned choice, including ties and one probability-ranking mismatch.",
    "probability_rule": "47 Jev distributions sum to 0.99; normalize by the sum for Brier. Retain original probabilities for confidence thresholds.",
    "confidence_rule": "Use the probability assigned to the selected action, not the separate TypeSafe confidence field.",
    "paired_bootstrap": {
      "resamples": 2000,
      "seed": 553,
      "unit": "conversation"
    },
    "timing_comparable": false
  },
  "test": {
    "experiment": "jev-vs-abcd-002-20261006",
    "all_500_responses_received": true,
    "http_failures": 0,
    "api_attempts": 500,
    "pinned_jev_model": "jev-1.13.0",
    "raw_validation_flags": 48,
    "validation_resolution": "All flagged answers are complete; normalize bounded probability-sum deviations for secondary probability scores. Preserve returned choices, including one ranking mismatch, and original returned-choice probability thresholds. Report canonical argmax accuracy as sensitivity.",
    "jev_api_cost_usd": 0.067028598,
    "metrics": {
      "jev": {
        "n": 500,
        "correct": 353,
        "accuracy": 0.706,
        "macro_f1": 0.6107575060796364,
        "multiclass_brier": 0.40353605411692683,
        "raw_brier_before_normalization": 0.40348860000000003,
        "nll_clipped_1e15": 1.3068122594803144,
        "mean_top_probability": 0.75468,
        "ece_10_bins": 0.051119999999999985,
        "distributions_normalized": 47,
        "probability_sum_range": [
          0.99,
          1.0000000000000002
        ],
        "exact_top_probability_ties": 5,
        "returned_choice_not_maximal": 1,
        "canonical_argmax_correct": 354,
        "coverage": [
          {
            "threshold": 0.5,
            "automated": 433,
            "coverage": 0.866,
            "wrong": 105,
            "accuracy_among_automated": 0.7575057736720554
          },
          {
            "threshold": 0.7,
            "automated": 317,
            "coverage": 0.634,
            "wrong": 45,
            "accuracy_among_automated": 0.8580441640378549
          },
          {
            "threshold": 0.8,
            "automated": 262,
            "coverage": 0.524,
            "wrong": 24,
            "accuracy_among_automated": 0.9083969465648855
          },
          {
            "threshold": 0.9,
            "automated": 176,
            "coverage": 0.352,
            "wrong": 8,
            "accuracy_among_automated": 0.9545454545454546
          },
          {
            "threshold": 0.95,
            "automated": 116,
            "coverage": 0.232,
            "wrong": 4,
            "accuracy_among_automated": 0.9655172413793104
          },
          {
            "threshold": 0.99,
            "automated": 32,
            "coverage": 0.064,
            "wrong": 0,
            "accuracy_among_automated": 1.0
          }
        ],
        "per_action": [
          {
            "action": "ask-the-oracle",
            "support": 22,
            "correct": 10,
            "predicted": 11,
            "f1": 0.6060606060606061
          },
          {
            "action": "enter-details",
            "support": 27,
            "correct": 13,
            "predicted": 22,
            "f1": 0.5306122448979592
          },
          {
            "action": "instructions",
            "support": 8,
            "correct": 4,
            "predicted": 6,
            "f1": 0.5714285714285714
          },
          {
            "action": "log-out-in",
            "support": 20,
            "correct": 14,
            "predicted": 15,
            "f1": 0.8
          },
          {
            "action": "make-password",
            "support": 7,
            "correct": 1,
            "predicted": 2,
            "f1": 0.2222222222222222
          },
          {
            "action": "make-purchase",
            "support": 8,
            "correct": 7,
            "predicted": 13,
            "f1": 0.6666666666666666
          },
          {
            "action": "membership",
            "support": 16,
            "correct": 3,
            "predicted": 4,
            "f1": 0.3
          },
          {
            "action": "notify-team",
            "support": 11,
            "correct": 9,
            "predicted": 12,
            "f1": 0.782608695652174
          },
          {
            "action": "offer-refund",
            "support": 6,
            "correct": 5,
            "predicted": 10,
            "f1": 0.625
          },
          {
            "action": "promo-code",
            "support": 9,
            "correct": 8,
            "predicted": 13,
            "f1": 0.7272727272727273
          },
          {
            "action": "pull-up-account",
            "support": 99,
            "correct": 96,
            "predicted": 110,
            "f1": 0.9186602870813397
          },
          {
            "action": "record-reason",
            "support": 23,
            "correct": 10,
            "predicted": 20,
            "f1": 0.46511627906976744
          },
          {
            "action": "search-boots",
            "support": 1,
            "correct": 0,
            "predicted": 3,
            "f1": 0.0
          },
          {
            "action": "search-faq",
            "support": 33,
            "correct": 2,
            "predicted": 2,
            "f1": 0.11428571428571428
          },
          {
            "action": "search-jacket",
            "support": 7,
            "correct": 6,
            "predicted": 8,
            "f1": 0.8
          },
          {
            "action": "search-jeans",
            "support": 3,
            "correct": 3,
            "predicted": 7,
            "f1": 0.6
          },
          {
            "action": "search-membership",
            "support": 7,
            "correct": 6,
            "predicted": 11,
            "f1": 0.6666666666666666
          },
          {
            "action": "search-policy",
            "support": 12,
            "correct": 7,
            "predicted": 7,
            "f1": 0.7368421052631579
          },
          {
            "action": "search-pricing",
            "support": 7,
            "correct": 4,
            "predicted": 7,
            "f1": 0.5714285714285714
          },
          {
            "action": "search-shirt",
            "support": 8,
            "correct": 7,
            "predicted": 17,
            "f1": 0.56
          },
          {
            "action": "search-timing",
            "support": 5,
            "correct": 5,
            "predicted": 11,
            "f1": 0.625
          },
          {
            "action": "select-faq",
            "support": 30,
            "correct": 30,
            "predicted": 40,
            "f1": 0.8571428571428571
          },
          {
            "action": "send-link",
            "support": 6,
            "correct": 6,
            "predicted": 6,
            "f1": 1.0
          },
          {
            "action": "shipping-status",
            "support": 9,
            "correct": 5,
            "predicted": 7,
            "f1": 0.625
          },
          {
            "action": "subscription-status",
            "support": 4,
            "correct": 4,
            "predicted": 8,
            "f1": 0.6666666666666666
          },
          {
            "action": "try-again",
            "support": 12,
            "correct": 5,
            "predicted": 11,
            "f1": 0.43478260869565216
          },
          {
            "action": "update-account",
            "support": 8,
            "correct": 5,
            "predicted": 9,
            "f1": 0.5882352941176471
          },
          {
            "action": "update-order",
            "support": 17,
            "correct": 12,
            "predicted": 19,
            "f1": 0.6666666666666666
          },
          {
            "action": "validate-purchase",
            "support": 25,
            "correct": 22,
            "predicted": 32,
            "f1": 0.7719298245614035
          },
          {
            "action": "verify-identity",
            "support": 50,
            "correct": 44,
            "predicted": 57,
            "f1": 0.822429906542056
          }
        ],
        "latency_median_seconds": 0.31784439549664967,
        "latency_p95_seconds": 0.7131867916148621,
        "total_input_tokens": 1595919
      },
      "trained_jevk5": {
        "n": 500,
        "correct": 396,
        "accuracy": 0.792,
        "macro_f1": 0.7583245392546261,
        "multiclass_brier": 0.33754536386605183,
        "raw_brier_before_normalization": 0.33754536386605183,
        "nll_clipped_1e15": 0.9160290729773535,
        "mean_top_probability": 0.8887458870369801,
        "ece_10_bins": 0.09674588703698009,
        "distributions_normalized": 0,
        "probability_sum_range": [
          0.9999999999999996,
          1.0000000000000002
        ],
        "exact_top_probability_ties": 2,
        "returned_choice_not_maximal": 0,
        "canonical_argmax_correct": 396,
        "coverage": [
          {
            "threshold": 0.5,
            "automated": 480,
            "coverage": 0.96,
            "wrong": 88,
            "accuracy_among_automated": 0.8166666666666667
          },
          {
            "threshold": 0.7,
            "automated": 457,
            "coverage": 0.914,
            "wrong": 74,
            "accuracy_among_automated": 0.838074398249453
          },
          {
            "threshold": 0.8,
            "automated": 441,
            "coverage": 0.882,
            "wrong": 68,
            "accuracy_among_automated": 0.8458049886621315
          },
          {
            "threshold": 0.9,
            "automated": 381,
            "coverage": 0.762,
            "wrong": 44,
            "accuracy_among_automated": 0.8845144356955381
          },
          {
            "threshold": 0.95,
            "automated": 125,
            "coverage": 0.25,
            "wrong": 4,
            "accuracy_among_automated": 0.968
          },
          {
            "threshold": 0.99,
            "automated": 0,
            "coverage": 0.0,
            "wrong": 0,
            "accuracy_among_automated": null
          }
        ],
        "per_action": [
          {
            "action": "ask-the-oracle",
            "support": 22,
            "correct": 15,
            "predicted": 18,
            "f1": 0.75
          },
          {
            "action": "enter-details",
            "support": 27,
            "correct": 13,
            "predicted": 17,
            "f1": 0.5909090909090909
          },
          {
            "action": "instructions",
            "support": 8,
            "correct": 6,
            "predicted": 14,
            "f1": 0.5454545454545454
          },
          {
            "action": "log-out-in",
            "support": 20,
            "correct": 13,
            "predicted": 13,
            "f1": 0.7878787878787878
          },
          {
            "action": "make-password",
            "support": 7,
            "correct": 3,
            "predicted": 4,
            "f1": 0.5454545454545454
          },
          {
            "action": "make-purchase",
            "support": 8,
            "correct": 5,
            "predicted": 6,
            "f1": 0.7142857142857143
          },
          {
            "action": "membership",
            "support": 16,
            "correct": 12,
            "predicted": 18,
            "f1": 0.7058823529411765
          },
          {
            "action": "notify-team",
            "support": 11,
            "correct": 10,
            "predicted": 12,
            "f1": 0.8695652173913043
          },
          {
            "action": "offer-refund",
            "support": 6,
            "correct": 3,
            "predicted": 5,
            "f1": 0.5454545454545454
          },
          {
            "action": "promo-code",
            "support": 9,
            "correct": 9,
            "predicted": 10,
            "f1": 0.9473684210526315
          },
          {
            "action": "pull-up-account",
            "support": 99,
            "correct": 97,
            "predicted": 109,
            "f1": 0.9326923076923077
          },
          {
            "action": "record-reason",
            "support": 23,
            "correct": 13,
            "predicted": 15,
            "f1": 0.6842105263157895
          },
          {
            "action": "search-boots",
            "support": 1,
            "correct": 1,
            "predicted": 1,
            "f1": 1.0
          },
          {
            "action": "search-faq",
            "support": 33,
            "correct": 24,
            "predicted": 30,
            "f1": 0.7619047619047619
          },
          {
            "action": "search-jacket",
            "support": 7,
            "correct": 6,
            "predicted": 6,
            "f1": 0.9230769230769231
          },
          {
            "action": "search-jeans",
            "support": 3,
            "correct": 3,
            "predicted": 3,
            "f1": 1.0
          },
          {
            "action": "search-membership",
            "support": 7,
            "correct": 5,
            "predicted": 6,
            "f1": 0.7692307692307693
          },
          {
            "action": "search-policy",
            "support": 12,
            "correct": 6,
            "predicted": 6,
            "f1": 0.6666666666666666
          },
          {
            "action": "search-pricing",
            "support": 7,
            "correct": 3,
            "predicted": 3,
            "f1": 0.6
          },
          {
            "action": "search-shirt",
            "support": 8,
            "correct": 7,
            "predicted": 9,
            "f1": 0.8235294117647058
          },
          {
            "action": "search-timing",
            "support": 5,
            "correct": 5,
            "predicted": 7,
            "f1": 0.8333333333333334
          },
          {
            "action": "select-faq",
            "support": 30,
            "correct": 30,
            "predicted": 39,
            "f1": 0.8695652173913043
          },
          {
            "action": "send-link",
            "support": 6,
            "correct": 6,
            "predicted": 7,
            "f1": 0.9230769230769231
          },
          {
            "action": "shipping-status",
            "support": 9,
            "correct": 9,
            "predicted": 12,
            "f1": 0.8571428571428571
          },
          {
            "action": "subscription-status",
            "support": 4,
            "correct": 4,
            "predicted": 8,
            "f1": 0.6666666666666666
          },
          {
            "action": "try-again",
            "support": 12,
            "correct": 6,
            "predicted": 10,
            "f1": 0.5454545454545454
          },
          {
            "action": "update-account",
            "support": 8,
            "correct": 6,
            "predicted": 13,
            "f1": 0.5714285714285714
          },
          {
            "action": "update-order",
            "support": 17,
            "correct": 12,
            "predicted": 21,
            "f1": 0.631578947368421
          },
          {
            "action": "validate-purchase",
            "support": 25,
            "correct": 23,
            "predicted": 28,
            "f1": 0.8679245283018868
          },
          {
            "action": "verify-identity",
            "support": 50,
            "correct": 41,
            "predicted": 50,
            "f1": 0.82
          }
        ],
        "latency_median_seconds": 0.873066412499611,
        "latency_p95_seconds": 1.0209821118489344,
        "total_input_tokens": 3176942
      },
      "shared_jevk5": {
        "n": 500,
        "correct": 289,
        "accuracy": 0.578,
        "macro_f1": 0.49643748291373646,
        "multiclass_brier": 0.6169341384932315,
        "raw_brier_before_normalization": 0.6169341384932315,
        "nll_clipped_1e15": 1.5107449360579095,
        "mean_top_probability": 0.6744458796737537,
        "ece_10_bins": 0.10175795422630485,
        "distributions_normalized": 0,
        "probability_sum_range": [
          0.9999999999999996,
          1.0000000000000002
        ],
        "exact_top_probability_ties": 9,
        "returned_choice_not_maximal": 0,
        "canonical_argmax_correct": 289,
        "coverage": [
          {
            "threshold": 0.5,
            "automated": 383,
            "coverage": 0.766,
            "wrong": 140,
            "accuracy_among_automated": 0.6344647519582245
          },
          {
            "threshold": 0.7,
            "automated": 244,
            "coverage": 0.488,
            "wrong": 70,
            "accuracy_among_automated": 0.7131147540983607
          },
          {
            "threshold": 0.8,
            "automated": 165,
            "coverage": 0.33,
            "wrong": 44,
            "accuracy_among_automated": 0.7333333333333334
          },
          {
            "threshold": 0.9,
            "automated": 65,
            "coverage": 0.13,
            "wrong": 16,
            "accuracy_among_automated": 0.7538461538461538
          },
          {
            "threshold": 0.95,
            "automated": 22,
            "coverage": 0.044,
            "wrong": 5,
            "accuracy_among_automated": 0.7727272727272727
          },
          {
            "threshold": 0.99,
            "automated": 0,
            "coverage": 0.0,
            "wrong": 0,
            "accuracy_among_automated": null
          }
        ],
        "per_action": [
          {
            "action": "ask-the-oracle",
            "support": 22,
            "correct": 16,
            "predicted": 27,
            "f1": 0.6530612244897959
          },
          {
            "action": "enter-details",
            "support": 27,
            "correct": 6,
            "predicted": 12,
            "f1": 0.3076923076923077
          },
          {
            "action": "instructions",
            "support": 8,
            "correct": 3,
            "predicted": 4,
            "f1": 0.5
          },
          {
            "action": "log-out-in",
            "support": 20,
            "correct": 14,
            "predicted": 15,
            "f1": 0.8
          },
          {
            "action": "make-password",
            "support": 7,
            "correct": 6,
            "predicted": 7,
            "f1": 0.8571428571428571
          },
          {
            "action": "make-purchase",
            "support": 8,
            "correct": 5,
            "predicted": 7,
            "f1": 0.6666666666666666
          },
          {
            "action": "membership",
            "support": 16,
            "correct": 0,
            "predicted": 0,
            "f1": 0.0
          },
          {
            "action": "notify-team",
            "support": 11,
            "correct": 10,
            "predicted": 11,
            "f1": 0.9090909090909091
          },
          {
            "action": "offer-refund",
            "support": 6,
            "correct": 5,
            "predicted": 13,
            "f1": 0.5263157894736842
          },
          {
            "action": "promo-code",
            "support": 9,
            "correct": 6,
            "predicted": 11,
            "f1": 0.6
          },
          {
            "action": "pull-up-account",
            "support": 99,
            "correct": 90,
            "predicted": 105,
            "f1": 0.8823529411764706
          },
          {
            "action": "record-reason",
            "support": 23,
            "correct": 2,
            "predicted": 4,
            "f1": 0.14814814814814814
          },
          {
            "action": "search-boots",
            "support": 1,
            "correct": 1,
            "predicted": 7,
            "f1": 0.25
          },
          {
            "action": "search-faq",
            "support": 33,
            "correct": 0,
            "predicted": 4,
            "f1": 0.0
          },
          {
            "action": "search-jacket",
            "support": 7,
            "correct": 6,
            "predicted": 14,
            "f1": 0.5714285714285714
          },
          {
            "action": "search-jeans",
            "support": 3,
            "correct": 3,
            "predicted": 13,
            "f1": 0.375
          },
          {
            "action": "search-membership",
            "support": 7,
            "correct": 5,
            "predicted": 13,
            "f1": 0.5
          },
          {
            "action": "search-policy",
            "support": 12,
            "correct": 6,
            "predicted": 8,
            "f1": 0.6
          },
          {
            "action": "search-pricing",
            "support": 7,
            "correct": 4,
            "predicted": 11,
            "f1": 0.4444444444444444
          },
          {
            "action": "search-shirt",
            "support": 8,
            "correct": 7,
            "predicted": 23,
            "f1": 0.45161290322580644
          },
          {
            "action": "search-timing",
            "support": 5,
            "correct": 5,
            "predicted": 16,
            "f1": 0.47619047619047616
          },
          {
            "action": "select-faq",
            "support": 30,
            "correct": 3,
            "predicted": 4,
            "f1": 0.17647058823529413
          },
          {
            "action": "send-link",
            "support": 6,
            "correct": 6,
            "predicted": 10,
            "f1": 0.75
          },
          {
            "action": "shipping-status",
            "support": 9,
            "correct": 2,
            "predicted": 4,
            "f1": 0.3076923076923077
          },
          {
            "action": "subscription-status",
            "support": 4,
            "correct": 4,
            "predicted": 20,
            "f1": 0.3333333333333333
          },
          {
            "action": "try-again",
            "support": 12,
            "correct": 8,
            "predicted": 16,
            "f1": 0.5714285714285714
          },
          {
            "action": "update-account",
            "support": 8,
            "correct": 7,
            "predicted": 24,
            "f1": 0.4375
          },
          {
            "action": "update-order",
            "support": 17,
            "correct": 9,
            "predicted": 27,
            "f1": 0.4090909090909091
          },
          {
            "action": "validate-purchase",
            "support": 25,
            "correct": 24,
            "predicted": 40,
            "f1": 0.7384615384615385
          },
          {
            "action": "verify-identity",
            "support": 50,
            "correct": 26,
            "predicted": 30,
            "f1": 0.65
          }
        ],
        "latency_median_seconds": 0.7662553700029093,
        "latency_p95_seconds": 0.899243401798958,
        "total_input_tokens": 3178609
      }
    },
    "comparisons": {
      "trained_vs_jev": {
        "accuracy_gain_percentage_points": 8.6,
        "accuracy_gain_95_ci_percentage_points": [
          4.8,
          12.2
        ],
        "brier_improvement": 0.06599069025087502,
        "brier_improvement_95_ci": [
          0.015399874212914205,
          0.11475712522423882
        ],
        "trained_correct_other_wrong": 72,
        "other_correct_trained_wrong": 29,
        "both_correct": 324,
        "both_wrong": 75
      },
      "trained_vs_shared_jevk5": {
        "accuracy_gain_percentage_points": 21.4,
        "accuracy_gain_95_ci_percentage_points": [
          16.8,
          25.8
        ],
        "brier_improvement": 0.27938877462717987,
        "brier_improvement_95_ci": [
          0.2213867081272137,
          0.33584692554146667
        ],
        "trained_correct_other_wrong": 128,
        "other_correct_trained_wrong": 21,
        "both_correct": 268,
        "both_wrong": 83
      }
    },
    "limitations": [
      "Same prompt and full candidate set; no separate provider prompt optimization.",
      "Frozen test already evaluated for abcd-002; no tuning or model selection in this new comparison.",
      "Agreement with recorded action is not full conversation success or guaranteed policy compliance.",
      "Public-data overlap with either models pretraining cannot be excluded.",
      "One training seed; no across-seed uncertainty measured.",
      "Jev elapsed time includes API network transport; historical JevK5 timing is local GPU inference. Not an apples-to-apples speed benchmark."
    ],
    "evidence": {
      "protocol.json": "e475f2f29764700bed7eff54958a9ce357deb51f0e8ab147bed7fffbd6d37e7e",
      "response-format-amendment.json": "72217cbaa1b6e836087fdadedbb2e9a432a0d3b1d5aa9fa0cbcc4a991bd8079d",
      "run_jev.py": "86841e0969e21ee1ba04afa466c289c5b4d6aeae6c3354b39714b5ecf77f0104",
      "jev-responses.jsonl": "84019aebe79e011def498b2f21aad0028e836a4ee2e795f728d0c685e3f8e3aa",
      "analyze.py": "b3c530f0fbc10bf78817facb54f7d72f3e8689e4cb4195d9f6fcbcdaabdce5a1"
    }
  },
  "deployment_status": "Research adapter; not deployed",
  "artifact_hashes": {
    "frozen_inputs_sha256": "d5b81470e9d9b7ae4c2bf2c42e7d9f06533c635cc8b1da2a8f719269ed7ae193",
    "frozen_labels_sha256": "cefc5660bc04b5d1dc5f303d5794bbbddbc48b742b5853c24f3fd30b897bfbc1",
    "original_comparison_protocol_sha256": "e475f2f29764700bed7eff54958a9ce357deb51f0e8ab147bed7fffbd6d37e7e",
    "response_format_amendment_sha256": "72217cbaa1b6e836087fdadedbb2e9a432a0d3b1d5aa9fa0cbcc4a991bd8079d",
    "independent_verification_sha256": "b6933c7e83e9f8e076905fa2eefb7799edd2a9c135e6243a2763689379e71099"
  },
  "sources": {
    "abcd-002-jev-results.json": "5451f7369a650ce448e917e0c4e5e4032ae3e24686bc69c2fda0bad8da001330",
    "abcd-002-jev-predictions.json": "b31138747c0daf1676ea788f4c95780ab989e0b4724d8655745e305060667561",
    "abcd-002-jev.md": "3f64234a19ea928e17e13a617f3042ba6326680e508724cb3f41c82d2ccf18c0"
  }
}
