{
  "claim_boundary": {
    "mode": "B_reverse_loop_small_must_separate",
    "pipeline": [
      "veto",
      "complement",
      "verify"
    ],
    "not_claimed": [
      "Mode_A_residual_stream_as_field",
      "layer_Jacobian",
      "in_silico_gene_perturbs",
      "fusion_audit",
      "clinical"
    ],
    "true_adt_events_note": "complement (b) admits true ADT as typed evidence for this probe only; production path keeps holdout ADT verifier-only"
  },
  "inputs": {
    "pairs_path": "/Users/tianchichen/Documents/GitHub/teddy_mm/outputs/anm_cite_bridge/reverse_step1/must_separate_pairs.jsonl",
    "n_pairs": 5000,
    "n_unique_cells": 2872,
    "z_path": "/Users/tianchichen/Documents/GitHub/teddy_mm/outputs/anm_cite_bridge/z_rna_512.npy",
    "panel": [
      "CD19",
      "CD72",
      "CD22",
      "CD3",
      "CD2",
      "CD5",
      "CD16",
      "CD11c",
      "CD36"
    ]
  },
  "veto": {
    "claim": "z_512 (mean-pool last-layer @ ctx 1024, L2 512-D) is INSUFFICIENT to separate these must-pairs: near-identical z carries disagreeing true ADT lineage / protein labels. Downstream readout must abstain, rewrite observer, or admit a complement evidence channel.",
    "n_pairs_written": 5000,
    "flag_counts": {
      "diff_adt_lineage": 2059,
      "diff_protein_profile": 4639,
      "diff_key_marker_lineage": 3287,
      "diff_coarse_cell_type": 790
    },
    "lineage_confusion": {
      "myeloid|t_lineage": 1490,
      "b_lineage|t_lineage": 375,
      "abstain|t_lineage": 213,
      "b_lineage|myeloid": 194,
      "abstain|myeloid": 168,
      "abstain|b_lineage": 24
    },
    "n_myeloid_t_pairs": 1490,
    "frac_myeloid_t": 0.298,
    "myeloid_t_cos_sim_quantiles": {
      "min": 0.9800012707710266,
      "q25": 0.9810009449720383,
      "q50": 0.9822191596031189,
      "q75": 0.983788788318634,
      "max": 0.990542471408844
    },
    "myeloid_t_focus_protein_deltas": {
      "CD5": {
        "n": 1490,
        "mean_abs_delta": 0.7155161401769459,
        "median_abs_delta": 0.40774770081043243,
        "q75_abs_delta": 1.0865748226642609
      },
      "CD3": {
        "n": 1490,
        "mean_abs_delta": 0.6981261888506428,
        "median_abs_delta": 0.3616558760404587,
        "q75_abs_delta": 0.8664438873529434
      },
      "CD16": {
        "n": 1490,
        "mean_abs_delta": 1.0650999203649343,
        "median_abs_delta": 0.907050609588623,
        "q75_abs_delta": 1.776064932346344
      }
    },
    "veto": true,
    "mode": "B_sufficiency_for_readout",
    "not_mode_a": true,
    "baseline_pred_O0_myeloid_t": {
      "false_agreement_rate_among_both_decided": 0.8536754507628294,
      "soft_separation_rate": 0.1738255033557047,
      "Q_among_decided": 0.7409695817490495
    }
  },
  "complements": {
    "a_observer": [
      "complement_a_O1_abstain",
      "complement_a_O2_observer"
    ],
    "b_true_adt": [
      "complement_b_true_ADT_O0",
      "complement_b_true_ADT_O1"
    ],
    "c_zplus": {
      "n_myeloid_t_pairs": 1490,
      "held_lineage": "myeloid",
      "held_proteins": [
        "CD16",
        "CD11c",
        "CD36"
      ],
      "zplus_proteins": [
        "CD19",
        "CD72",
        "CD22",
        "CD3",
        "CD2",
        "CD5"
      ],
      "cos_z512": {
        "mean": 0.9826505275780723,
        "median": 0.9822191596031189,
        "frac_still_ge_0_98": 1.0,
        "frac_dropped_below_0_95": 0.0
      },
      "cos_zplus_true_complement": {
        "mean": 0.8255820943005133,
        "median": 0.8865448236465454,
        "frac_still_ge_0_98": 0.09395973154362416,
        "frac_dropped_below_0_95": 0.7208053691275168
      },
      "cos_zplus_pred_complement": {
        "mean": 0.9877049229289061,
        "median": 0.9969751238822937,
        "frac_still_ge_0_98": 0.8348993288590604,
        "frac_dropped_below_0_95": 0.07583892617449664
      },
      "prior_global_relative_must_reduction_vs_z512": {
        "zplus_adt_true_complement": -0.27272456550677143,
        "zplus_adt_pred_complement": -0.27353544257798595,
        "zplus_adt_pred_all9_leakcheck": -0.2508802113539737
      },
      "prior_note": "Prior global z\u207a probe (held myeloid) relative_must_reduction primary=-0.27272456550677143; NEGATIVE => z\u207a alone worsens must-rate among near.",
      "interpretation": "On FIXED myeloid\u2194T must-pairs: true-ADT z\u207a lowers cosine (frac still \u22650.98 = 0.09395973154362416; frac <0.95 = 0.7208053691275168), so geometry can separate this fixed set. Pred z\u207a does NOT (frac still \u22650.98 = 0.8348993288590604). Prior GLOBAL probe still found z\u207a worsens must-rate among its own near neighborhoods (relative_must_reduction primary=-0.27272456550677143). Decision winner remains true-ADT typed evidence + O1, not z\u207a alone.",
      "decision_vs_geometry_note": "z\u207a true geometry helps on fixed pairs but is not the winning decision complement; prior z\u207a alone worsened global must-rate. Pred z\u207a fails fixed-pair separation."
    },
    "variants_run": [
      "baseline_pred_O0",
      "complement_a_O1_abstain",
      "complement_a_O2_observer",
      "complement_b_true_ADT_O0",
      "complement_b_true_ADT_O1"
    ]
  },
  "verify": {
    "all_written_pairs": {
      "baseline_pred_O0": {
        "n_pairs": 5000,
        "both_decided": 4666,
        "same_action_among_both_decided": 4165,
        "differ_action_among_both_decided": 501,
        "false_agreement_rate_among_both_decided": 0.8926275182168881,
        "any_abstain": 334,
        "both_abstain": 142,
        "abstain_pair_frac": 0.0668,
        "soft_separation_rate": 0.167,
        "correct_separation_rate": 0.0416,
        "n_decided_cell_slots": 9524,
        "mean_Q_among_decided_defined": 0.8081765087605451,
        "n_Q_defined": 7705,
        "match_expected_among_decided_with_expected": null,
        "n_decided_with_expected": 7705,
        "Q_among_decided": 0.8081765087605451,
        "mean_Q_recomputed": 0.8081765087605451
      },
      "complement_a_O1_abstain": {
        "n_pairs": 5000,
        "both_decided": 4462,
        "same_action_among_both_decided": 4031,
        "differ_action_among_both_decided": 431,
        "false_agreement_rate_among_both_decided": 0.9034065441506051,
        "any_abstain": 538,
        "both_abstain": 496,
        "abstain_pair_frac": 0.1076,
        "soft_separation_rate": 0.1938,
        "correct_separation_rate": 0.0114,
        "n_decided_cell_slots": 8966,
        "mean_Q_among_decided_defined": 0.8665363426337302,
        "n_Q_defined": 5627,
        "match_expected_among_decided_with_expected": null,
        "n_decided_with_expected": 5627,
        "Q_among_decided": 0.8665363426337302,
        "mean_Q_recomputed": 0.8665363426337302
      },
      "complement_a_O2_observer": {
        "n_pairs": 5000,
        "both_decided": 2582,
        "same_action_among_both_decided": 2027,
        "differ_action_among_both_decided": 555,
        "false_agreement_rate_among_both_decided": 0.7850503485670023,
        "any_abstain": 2418,
        "both_abstain": 1969,
        "abstain_pair_frac": 0.4836,
        "soft_separation_rate": 0.5946,
        "correct_separation_rate": 0.087,
        "n_decided_cell_slots": 5613,
        "mean_Q_among_decided_defined": 0.8514386376981797,
        "n_Q_defined": 5109,
        "match_expected_among_decided_with_expected": null,
        "n_decided_with_expected": 5109,
        "Q_among_decided": 0.8514386376981797,
        "mean_Q_recomputed": 0.8514386376981797
      },
      "complement_b_true_ADT_O0": {
        "n_pairs": 5000,
        "both_decided": 4905,
        "same_action_among_both_decided": 2963,
        "differ_action_among_both_decided": 1942,
        "false_agreement_rate_among_both_decided": 0.6040774719673803,
        "any_abstain": 95,
        "both_abstain": 0,
        "abstain_pair_frac": 0.019,
        "soft_separation_rate": 0.4074,
        "correct_separation_rate": 0.2226,
        "n_decided_cell_slots": 9905,
        "mean_Q_among_decided_defined": 0.9567061759201497,
        "n_Q_defined": 8015,
        "match_expected_among_decided_with_expected": null,
        "n_decided_with_expected": 8015,
        "Q_among_decided": 0.9567061759201497,
        "mean_Q_recomputed": 0.9567061759201497
      },
      "complement_b_true_ADT_O1": {
        "n_pairs": 5000,
        "both_decided": 4529,
        "same_action_among_both_decided": 2653,
        "differ_action_among_both_decided": 1876,
        "false_agreement_rate_among_both_decided": 0.5857805255023184,
        "any_abstain": 471,
        "both_abstain": 116,
        "abstain_pair_frac": 0.0942,
        "soft_separation_rate": 0.4694,
        "correct_separation_rate": 0.1092,
        "n_decided_cell_slots": 9413,
        "mean_Q_among_decided_defined": 0.9899108794350092,
        "n_Q_defined": 5947,
        "match_expected_among_decided_with_expected": null,
        "n_decided_with_expected": 5947,
        "Q_among_decided": 0.9899108794350092,
        "mean_Q_recomputed": 0.9899108794350092
      }
    },
    "myeloid_t_focus": {
      "baseline_pred_O0": {
        "n_pairs": 1490,
        "both_decided": 1442,
        "same_action_among_both_decided": 1231,
        "differ_action_among_both_decided": 211,
        "false_agreement_rate_among_both_decided": 0.8536754507628294,
        "any_abstain": 48,
        "both_abstain": 17,
        "abstain_pair_frac": 0.032214765100671144,
        "soft_separation_rate": 0.1738255033557047,
        "correct_separation_rate": 0.08389261744966443,
        "n_decided_cell_slots": 2915,
        "mean_Q_among_decided_defined": 0.7409695817490495,
        "n_Q_defined": 2104,
        "match_expected_among_decided_with_expected": null,
        "n_decided_with_expected": 2104,
        "Q_among_decided": 0.7409695817490495,
        "mean_Q_recomputed": 0.7409695817490495
      },
      "complement_a_O1_abstain": {
        "n_pairs": 1490,
        "both_decided": 1391,
        "same_action_among_both_decided": 1156,
        "differ_action_among_both_decided": 235,
        "false_agreement_rate_among_both_decided": 0.8310567936736161,
        "any_abstain": 99,
        "both_abstain": 83,
        "abstain_pair_frac": 0.06644295302013423,
        "soft_separation_rate": 0.22416107382550335,
        "correct_separation_rate": 0.032214765100671144,
        "n_decided_cell_slots": 2798,
        "mean_Q_among_decided_defined": 0.7777777777777778,
        "n_Q_defined": 1332,
        "match_expected_among_decided_with_expected": null,
        "n_decided_with_expected": 1332,
        "Q_among_decided": 0.7777777777777778,
        "mean_Q_recomputed": 0.7777777777777778
      },
      "complement_a_O2_observer": {
        "n_pairs": 1490,
        "both_decided": 1109,
        "same_action_among_both_decided": 935,
        "differ_action_among_both_decided": 174,
        "false_agreement_rate_among_both_decided": 0.8431018935978359,
        "any_abstain": 381,
        "both_abstain": 330,
        "abstain_pair_frac": 0.2557046979865772,
        "soft_separation_rate": 0.3724832214765101,
        "correct_separation_rate": 0.08389261744966443,
        "n_decided_cell_slots": 2269,
        "mean_Q_among_decided_defined": 0.8824110671936759,
        "n_Q_defined": 2024,
        "match_expected_among_decided_with_expected": null,
        "n_decided_with_expected": 2024,
        "Q_among_decided": 0.8824110671936759,
        "mean_Q_recomputed": 0.8824110671936759
      },
      "complement_b_true_ADT_O0": {
        "n_pairs": 1490,
        "both_decided": 1475,
        "same_action_among_both_decided": 674,
        "differ_action_among_both_decided": 801,
        "false_agreement_rate_among_both_decided": 0.4569491525423729,
        "any_abstain": 15,
        "both_abstain": 0,
        "abstain_pair_frac": 0.010067114093959731,
        "soft_separation_rate": 0.5476510067114094,
        "correct_separation_rate": 0.36845637583892615,
        "n_decided_cell_slots": 2965,
        "mean_Q_among_decided_defined": 0.9716981132075472,
        "n_Q_defined": 2120,
        "match_expected_among_decided_with_expected": null,
        "n_decided_with_expected": 2120,
        "Q_among_decided": 0.9716981132075472,
        "mean_Q_recomputed": 0.9716981132075472
      },
      "complement_b_true_ADT_O1": {
        "n_pairs": 1490,
        "both_decided": 1407,
        "same_action_among_both_decided": 558,
        "differ_action_among_both_decided": 849,
        "false_agreement_rate_among_both_decided": 0.39658848614072495,
        "any_abstain": 83,
        "both_abstain": 17,
        "abstain_pair_frac": 0.05570469798657718,
        "soft_separation_rate": 0.625503355704698,
        "correct_separation_rate": 0.17248322147651007,
        "n_decided_cell_slots": 2880,
        "mean_Q_among_decided_defined": 0.9868421052631579,
        "n_Q_defined": 1368,
        "match_expected_among_decided_with_expected": null,
        "n_decided_with_expected": 1368,
        "Q_among_decided": 0.9868421052631579,
        "mean_Q_recomputed": 0.9868421052631579
      },
      "complement_c_zplus_geometry": {
        "n_pairs": 1490,
        "soft_separation_rate": 0.7208053691275168,
        "correct_separation_rate": null,
        "false_agreement_rate_among_both_decided": 0.09395973154362416,
        "abstain_pair_frac": null,
        "Q_among_decided": null,
        "note": "geometry only: soft_sep\u2248frac cos dropped <0.95; false_agree\u2248frac still \u22650.98"
      }
    }
  },
  "winner": {
    "winner": "complement_b_true_ADT_O1",
    "winner_row": {
      "complement": "complement_b_true_ADT_O1",
      "score": 2.8412603305318287,
      "soft_separation_rate": 0.625503355704698,
      "Q_among_decided": 0.9868421052631579,
      "false_agreement_rate_among_both_decided": 0.39658848614072495,
      "correct_separation_rate": 0.17248322147651007,
      "abstain_pair_frac": 0.05570469798657718
    },
    "ranking": [
      {
        "complement": "complement_b_true_ADT_O1",
        "score": 2.8412603305318287,
        "soft_separation_rate": 0.625503355704698,
        "Q_among_decided": 0.9868421052631579,
        "false_agreement_rate_among_both_decided": 0.39658848614072495,
        "correct_separation_rate": 0.17248322147651007,
        "abstain_pair_frac": 0.05570469798657718
      },
      {
        "complement": "complement_b_true_ADT_O0",
        "score": 2.610050974087993,
        "soft_separation_rate": 0.5476510067114094,
        "Q_among_decided": 0.9716981132075472,
        "false_agreement_rate_among_both_decided": 0.4569491525423729,
        "correct_separation_rate": 0.36845637583892615,
        "abstain_pair_frac": 0.010067114093959731
      },
      {
        "complement": "complement_c_zplus_geometry",
        "score": 2.3476510067114096,
        "soft_separation_rate": 0.7208053691275168,
        "Q_among_decided": null,
        "false_agreement_rate_among_both_decided": 0.09395973154362416,
        "correct_separation_rate": null,
        "abstain_pair_frac": null
      },
      {
        "complement": "complement_a_O2_observer",
        "score": 1.78427561654886,
        "soft_separation_rate": 0.3724832214765101,
        "Q_among_decided": 0.8824110671936759,
        "false_agreement_rate_among_both_decided": 0.8431018935978359,
        "correct_separation_rate": 0.08389261744966443,
        "abstain_pair_frac": 0.2557046979865772
      },
      {
        "complement": "complement_a_O1_abstain",
        "score": 1.3950431317551684,
        "soft_separation_rate": 0.22416107382550335,
        "Q_among_decided": 0.7777777777777778,
        "false_agreement_rate_among_both_decided": 0.8310567936736161,
        "correct_separation_rate": 0.032214765100671144,
        "abstain_pair_frac": 0.06644295302013423
      },
      {
        "complement": "baseline_pred_O0",
        "score": 1.2349451376976295,
        "soft_separation_rate": 0.1738255033557047,
        "Q_among_decided": 0.7409695817490495,
        "false_agreement_rate_among_both_decided": 0.8536754507628294,
        "correct_separation_rate": 0.08389261744966443,
        "abstain_pair_frac": 0.032214765100671144
      }
    ],
    "note": "Winner = best Mode B decision complement on myeloid\u2194T must-pairs (soft_separation + Q_among_decided \u2212 false_agreement). z\u207a geometry reported separately; prior showed it worsens global must-rate."
  },
  "wall_seconds": 2.169034004211426
}