{
  "generated": "2026-10-30",
  "model": "h2o-lightning-4b",
  "architecture": {
    "params_b": 4,
    "causal_deltanet_layers": 24,
    "bidirectional_attention_layers": 8,
    "total_layers": 32,
    "output_tokens": 0,
    "gpu_p50_ms": 29.0,
    "gpu_p50_ms_int": 29,
    "https_csat_p50_ms": 280.3,
    "https_csat_p50_ms_int": 280,
    "https_dlp_p50_ms": 331.9,
    "https_dlp_p50_ms_int": 332,
    "https_memory_p50_ms": 327.4,
    "https_memory_p50_ms_int": 327
  },
  "ticket_category_replay": {
    "traces_pulled": 4000,
    "min_gemini_runs": 2,
    "distinct_inputs_replayed": 793,
    "categories_count": 9,
    "gemini_stable_n": 632,
    "gemini_stable_pct": 79.7,
    "gemini_flipped_n": 161,
    "gemini_flipped_pct": 20.3,
    "overall_agree_with_gemini_maj_pct": 62.2,
    "stable_agree_with_gemini_pct": 68.0,
    "confidence_bands": [
      {
        "band": "[0.0, 0.5)",
        "lo": 0.0,
        "hi": 0.5,
        "n": 75,
        "share_pct": 9.5,
        "mean_conf": 0.4406,
        "mean_conf_pct": 44.1,
        "agree_rate": 0.44,
        "agree_pct": 44.0
      },
      {
        "band": "[0.5, 0.7)",
        "lo": 0.5,
        "hi": 0.7,
        "n": 206,
        "share_pct": 26.0,
        "mean_conf": 0.5974,
        "mean_conf_pct": 59.7,
        "agree_rate": 0.4515,
        "agree_pct": 45.1
      },
      {
        "band": "[0.7, 0.9)",
        "lo": 0.7,
        "hi": 0.9,
        "n": 257,
        "share_pct": 32.4,
        "mean_conf": 0.8047,
        "mean_conf_pct": 80.5,
        "agree_rate": 0.607,
        "agree_pct": 60.7
      },
      {
        "band": "[0.9, 1.0]",
        "lo": 0.9,
        "hi": 1.0,
        "n": 255,
        "share_pct": 32.2,
        "mean_conf": 0.9393,
        "mean_conf_pct": 93.9,
        "agree_rate": 0.8275,
        "agree_pct": 82.7
      }
    ],
    "high_conf_threshold": 0.9,
    "high_conf_threshold_pct": 90,
    "annotation_queue_total": 130,
    "annotation_strata": {
      "B_gemini_unstable": 32,
      "D_agree_control": 25,
      "C_disagree_s1_unsure": 45,
      "A_disagree_s1_confident": 16,
      "E_agree_s1_confident": 12
    },
    "replay_latency_p50_ms": 280
  },
  "csat_and_silent_churn": {
    "tickets_n": 533,
    "bad_csat_n": 26,
    "bad_csat_base_rate_pct": 4.9,
    "bad_csat_base_rate_exact_pct": 4.88,
    "generative_models_tested": 12,
    "verbalized_unique_levels_min": 5,
    "verbalized_unique_levels_max": 20,
    "prefill_unique_values": 525,
    "prefill_platt_unique_values": 531,
    "prefill_unique_share_pct": 98.5,
    "k10_tickets": 40,
    "k10_repeats": 10,
    "k10_total_calls": 400,
    "k10_choice_flips": 0,
    "k10_noul_flips": 1,
    "speedup_vs_g3_flash_x": 51.7,
    "cost_ratio_vs_opus_x": 2437,
    "cost_per_1k_usd_p05": 0.028,
    "cost_per_1k_usd_p06": 0.032,
    "cost_per_1k_usd_g3_flash": 0.6,
    "cost_per_1k_usd_opus": 67.5,
    "opus_neutral_pct": 41.3,
    "s1_neutral_pct": 21.2,
    "p05_pure_text_noul_recall_pct": 15.4,
    "p05_pure_text_saerens_recall_pct": 23.1,
    "p05_pure_text_497_recall_pct": 26.1,
    "p05_pure_text_precision_pct": 50.0,
    "p05_pure_text_fps": 4,
    "p05_pure_text_mcc": 0.259,
    "opus_mcc": 0.165,
    "g3_flash_mcc": 0.347,
    "g3_flash_recall_pct": 65.4,
    "g3_flash_fps": 54,
    "g3_flash_fdr_pct": 76.1,
    "mm_tab_oof_mcc": 0.332,
    "p06_saerens_mcc": 0.464,
    "p06_oof_page_mcc": 0.485,
    "p06_cv_opt_mcc": 0.567,
    "p06_oof_page_fps": 12,
    "p06_cv_opt_fps": 6,
    "p06_fp_reduction_oof_pct": 90.7,
    "p06_fp_reduction_cv_pct": 88.9,
    "silent_churn_tickets_n": 10,
    "silent_churn_caught_p05": 0,
    "silent_churn_caught_p06": 8,
    "silent_churn_recall_pct": 80.0,
    "ece_saerens": 0.0325,
    "ece_oof_platt": 0.0046,
    "ece_p06": 0.016
  },
  "prior_paradox_and_fusion": {
    "rules_total_n": 110,
    "rules_labelled_n": 29,
    "rules_labelled_pos": 12,
    "in_prompt_spearman_rho_110": 0.75,
    "in_prompt_shift_gt_70_pct": 0.083,
    "in_prompt_shift_gt_70_pts": 8.3,
    "in_prompt_shift_rare": -0.128,
    "in_prompt_shift_rare_pts": -12.8,
    "high_noise_fire_rate_pct": 70,
    "rare_fire_rate_pct": 20,
    "auroc_s1_text_only": 0.667,
    "auroc_s1_text_only_exact": 0.6667,
    "auroc_s1_text_only_ci": [
      0.461,
      0.857
    ],
    "auroc_s1_prompted_fr": 0.402,
    "auroc_s1_prompted_fr_ci": [
      0.196,
      0.626
    ],
    "auroc_rarity_only": 0.816,
    "auroc_rarity_only_exact": 0.8162,
    "auroc_rarity_only_ci": [
      0.642,
      0.949
    ],
    "auroc_logit_fusion": 0.843,
    "auroc_logit_fusion_exact": 0.8431,
    "auroc_logit_fusion_ci": [
      0.663,
      0.97
    ],
    "auroc_swing_logit_vs_prompted": 0.441,
    "auroc_swing_pts": 44.1,
    "auroc_gain_over_rarity_pts": 2.7,
    "aurc_text_only": 0.4465,
    "aurc_prompted_fr": 0.6263,
    "aurc_rarity_only": 0.2876,
    "aurc_logit_fusion": 0.3172,
    "aurc_reduction_vs_prompted_pct": -49.4,
    "aurc_reduction_vs_text_pct": -29.0,
    "case_pairs_n": 83,
    "case_pairs_pos": 22,
    "case_pairs_extended_n": 567,
    "case_brier_uncal": 0.24,
    "case_brier_platt": 0.194,
    "case_brier_base_rate": 0.195,
    "case_brier_m3_fusion": 0.167,
    "case_brier_reduction_pct": 30.5,
    "case_ece_reduction_pct": 44.6,
    "case_eaurc_reduction_pct": 36.0,
    "m1_threads_n": 30,
    "m1_question_acc_pct": 100,
    "m1_intent_total_n": 58,
    "m1_clear_n": 40,
    "m1_borderline_n": 10,
    "m1_ambiguous_n": 8,
    "m1_clear_acc_pct": 100,
    "m1_ambiguous_abstain_pct": 100,
    "m2_templates_n": 50,
    "m2_auroc": 1.0,
    "m3_m4_fi_trials_n": 18,
    "m3_m4_default_top1_pct": 0,
    "m3_m4_calibrated_top1_pct": 100,
    "m3_m4_default_mrr": 0.138,
    "m3_m4_calibrated_mrr": 1.0
  },
  "clamping_and_quantization": {
    "noul_floor_high": 0.801,
    "noul_floor_low": 0.199,
    "jump_discontinuity": 0.602,
    "midpoint": 0.5,
    "threshold_tau": 0.8,
    "binary_decisions_n": 30,
    "cases_n": 6,
    "k_repeats": 10,
    "total_verdict_comparisons": 324,
    "uncertain_band_lo": 0.2,
    "uncertain_band_hi": 0.8,
    "escalations_unclamped_choice": 6,
    "escalations_clamped_noul": 0,
    "escalation_rate_choice_pct": 20.0,
    "escalation_suppression_pct": 100.0,
    "flips_unclamped_choice": 0,
    "flips_clamped_noul": 8,
    "flip_rate_clamped_noul_pct": 2.5,
    "flip_rate_clamped_noul_exact_pct": 2.47,
    "non_escalated_n": 24,
    "non_escalated_acc_pct": 100.0,
    "raw_4b_argmax_acc_pct": 90.0,
    "raw_4b_argmax_matched": 27,
    "quant_4bit_q4km_flip_pct": 5.4,
    "quant_4bit_mlx_flip_pct": 8.7,
    "quant_8bit_q8_flip_pct": 1.4,
    "quant_8bit_bf16_agree_pct": 98.6,
    "quant_net_acc_delta_pts": 0.3,
    "quant_8bit_vram_gb": 4.48,
    "dlp_prod_tool_outputs_n": 323,
    "dlp_mcp_tools_n": 53,
    "dlp_deterministic_rules_n": 33,
    "dlp_compound_patterns_n": 33,
    "dlp_prod_gpu_p50_ms": 29.1,
    "dlp_torture_items_n": 77,
    "dlp_positive_secrets_n": 49,
    "dlp_negative_tech_ids_n": 28,
    "dlp_compound_secrets_n": 27,
    "dlp_cloud_dlp_tech_id_fps": 18,
    "dlp_cloud_dlp_stall_min_s": 57,
    "dlp_cloud_dlp_stall_max_s": 59,
    "dlp_cloud_dlp_stall_rate_pct": 10.3,
    "dlp_two_stage_recall_pct": 100.0,
    "dlp_two_stage_prod_recall_pct": 99.4,
    "dlp_two_stage_prod_fpr_pct": 0.42,
    "dlp_two_stage_f1": 0.995,
    "ci_pr_tickets_n": 50,
    "ci_legacy_calls": 750,
    "ci_s1_calls": 50,
    "ci_compression_triage_x": 15,
    "ci_compression_forensic_x": 27,
    "ci_legacy_min": 18.5,
    "ci_s1_sec": 9.44,
    "ci_kappa": 0.918
  },
  "positional_bias_and_runoff": {
    "candidates_k": 3,
    "permutations_n": 6,
    "wave1_n": 21,
    "wave2_n": 92,
    "single_pass_flip_w1_panel_a_n": 11,
    "single_pass_flip_w1_panel_a_pct": 52.4,
    "single_pass_flip_w1_panel_b_n": 9,
    "single_pass_flip_w1_panel_b_pct": 42.9,
    "single_pass_flip_w2_n": 75,
    "single_pass_flip_w2_pct": 81.5,
    "runoff_flip_w1_n": 0,
    "runoff_flip_w2_n": 0,
    "runoff_flip_pct": 0.0,
    "runoff_pc": 1.0,
    "w1_best_single_em_n": 14,
    "w1_best_single_em_pct": 66.7,
    "w1_consensus_em_n": 11,
    "w1_consensus_em_pct": 52.4,
    "w1_panel_b_consensus_em_n": 12,
    "w1_panel_b_consensus_em_pct": 57.1,
    "w1_unpermuted_choice_em_n": 14,
    "w1_unpermuted_choice_em_pct": 66.7,
    "w1_cyclic_choice_em_n": 16,
    "w1_cyclic_choice_em_pct": 76.2,
    "w1_runoff_em_n": 15,
    "w1_runoff_em_pct": 71.4,
    "w1_runoff_gap_recovery_pct": 33.3,
    "w1_oracle_em_n": 17,
    "w1_oracle_em_pct": 81.0,
    "w2_best_single_em_n": 44,
    "w2_best_single_em_pct": 47.8,
    "w2_consensus_em_n": 34,
    "w2_consensus_em_pct": 37.0,
    "w2_unpermuted_choice_em_n": 34,
    "w2_unpermuted_choice_em_pct": 37.0,
    "w2_cyclic_choice_em_n": 36,
    "w2_cyclic_choice_em_pct": 39.1,
    "w2_runoff_em_n": 46,
    "w2_runoff_em_pct": 50.0,
    "w2_runoff_lift_over_unpermuted_pts": 13.0,
    "w2_runoff_gap_recovery_pct": 13.3,
    "w2_oracle_em_n": 59,
    "w2_oracle_em_pct": 64.1,
    "w2_oracle_minus_runoff_pts": 14.1,
    "rca_runs_total_n": 787,
    "wave2_runs_n": 368,
    "wave2_leaky_pilot_auroc": 0.954,
    "wave2_verbalized_auroc": 0.463,
    "wave2_calibrated_auroc": 0.725,
    "wave2_ledger_only_auroc": 0.723,
    "wave2_s1_only_auroc": 0.588,
    "wave2_s1_auroc_gain_pts": 0.2,
    "wave3_runs_n": 87,
    "wave3_ledger_only_auroc": 0.628,
    "wave3_calibrated_auroc": 0.68,
    "wave3_s1_auroc_gain_pts": 5.1,
    "wave2_cw_fp_before_pct": 16.8,
    "wave2_cw_fp_after_pct": 2.7,
    "gs2_controls_n": 6,
    "gs2_verbalized_overclaim_n": 4,
    "gs2_calibrated_overclaim_n": 0
  },
  "batching_and_sentence_splitting": {
    "rag_pairs_n": 15,
    "rag_repeats_k": 10,
    "rag_total_calls": 150,
    "rag_workers": 5,
    "rag_binary_flips": 0,
    "rag_grade_flips": 3,
    "rag_grade_flip_pct": 2.0,
    "rag_max_grade_prob_delta": 0.106,
    "rag_max_noul_delta": 0.04,
    "rag_borderline_p0": 0.45,
    "rag_borderline_p1": 0.48,
    "mem_calls_n": 94,
    "mem_noise_rejected_n": 20,
    "mem_compound_turns_n": 4,
    "mem_whole_turn_min_noul": 0.088,
    "mem_whole_turn_max_noul": 0.199,
    "mem_split_clause0_max_noul": 0.076,
    "mem_split_clause1_bigdecimal_noul": 0.878,
    "mem_split_clause1_bigdecimal_choice": 0.936,
    "mem_preflight_hits_n": 4,
    "mem_preflight_hit_pct": 100.0,
    "mem_handoff_turns_n": 20,
    "mem_setup_token_reduction_pct": 52.0,
    "mem_net_token_savings_pct": 12.7
  },
  "reference_years": [
    1973,
    1974
  ]
}