{
  "seed": 42,
  "num_samples": 999,
  "samples": [
    {
      "features": {
        "hrm_confidence": 0.932186814566789,
        "trm_confidence": 0.36522885075930284,
        "mcts_value": 0.7145138679647244,
        "consensus_score": 0.7101167835754781,
        "last_agent": "none",
        "iteration": 1,
        "query_length": 2637,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9283419105971059,
        "trm_confidence": 0.6511300084852987,
        "mcts_value": 0.10612189126399749,
        "consensus_score": 0.5519417910279142,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 921,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8931595360241993,
        "trm_confidence": 0.6525812194404133,
        "mcts_value": 0.351698200208428,
        "consensus_score": 0.5779273962479688,
        "last_agent": "none",
        "iteration": 6,
        "query_length": 4441,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9482893515977746,
        "trm_confidence": 0.5358341835586543,
        "mcts_value": 0.6430777574912441,
        "consensus_score": 0.679972291175198,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 2234,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9335150491221286,
        "trm_confidence": 0.1622342921363004,
        "mcts_value": 0.3890189803478681,
        "consensus_score": 0.4036835270262114,
        "last_agent": "trm",
        "iteration": 1,
        "query_length": 3720,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9234286467723452,
        "trm_confidence": 0.7966752297173753,
        "mcts_value": 0.2682939337358132,
        "consensus_score": 0.6368912112821515,
        "last_agent": "none",
        "iteration": 5,
        "query_length": 3981,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7389764516006414,
        "trm_confidence": 0.30396424576879205,
        "mcts_value": 0.1449897306915453,
        "consensus_score": 0.42993960829016165,
        "last_agent": "mcts",
        "iteration": 4,
        "query_length": 811,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9100795306006747,
        "trm_confidence": 0.25304182222607313,
        "mcts_value": 0.6741966292520353,
        "consensus_score": 0.6733921988589547,
        "last_agent": "mcts",
        "iteration": 4,
        "query_length": 4492,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9047486511924926,
        "trm_confidence": 0.112465622685393,
        "mcts_value": 0.16087585630415913,
        "consensus_score": 0.29416916401088267,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 3904,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9115496135879005,
        "trm_confidence": 0.6336003434427341,
        "mcts_value": 0.3724329203075315,
        "consensus_score": 0.6529425316366341,
        "last_agent": "none",
        "iteration": 1,
        "query_length": 1235,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9005208885371415,
        "trm_confidence": 0.37712235352817686,
        "mcts_value": 0.4524833101935956,
        "consensus_score": 0.6297086222361764,
        "last_agent": "trm",
        "iteration": 6,
        "query_length": 2831,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8677621482236241,
        "trm_confidence": 0.23336138024133235,
        "mcts_value": 0.02366076687148142,
        "consensus_score": 0.3622715762919518,
        "last_agent": "mcts",
        "iteration": 2,
        "query_length": 1390,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9560209219804499,
        "trm_confidence": 0.2002570943780814,
        "mcts_value": 0.0499083666946623,
        "consensus_score": 0.35833890608879715,
        "last_agent": "mcts",
        "iteration": 3,
        "query_length": 2177,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8671096457023835,
        "trm_confidence": 0.6013358774543538,
        "mcts_value": 0.5096013245558383,
        "consensus_score": 0.6406263215255393,
        "last_agent": "hrm",
        "iteration": 8,
        "query_length": 1613,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7068136219401581,
        "trm_confidence": 0.054642268545230264,
        "mcts_value": 0.4383374938777725,
        "consensus_score": 0.3923065741713311,
        "last_agent": "trm",
        "iteration": 1,
        "query_length": 4505,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7456936308139505,
        "trm_confidence": 0.4496096311936752,
        "mcts_value": 0.2880802654858253,
        "consensus_score": 0.47066542105044684,
        "last_agent": "none",
        "iteration": 3,
        "query_length": 3422,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8085437831660172,
        "trm_confidence": 0.062103805426426364,
        "mcts_value": 0.08361234832438866,
        "consensus_score": 0.410466178548847,
        "last_agent": "hrm",
        "iteration": 9,
        "query_length": 2480,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7797609884378558,
        "trm_confidence": 0.6588082922365092,
        "mcts_value": 0.5293644842266623,
        "consensus_score": 0.6993559594654748,
        "last_agent": "mcts",
        "iteration": 4,
        "query_length": 3687,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7289172886460498,
        "trm_confidence": 0.5676622519597796,
        "mcts_value": 0.2866455884313108,
        "consensus_score": 0.46821438263809284,
        "last_agent": "trm",
        "iteration": 3,
        "query_length": 4046,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7530318348817695,
        "trm_confidence": 0.5593963977267774,
        "mcts_value": 0.49533740036194435,
        "consensus_score": 0.6464811355136844,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 1550,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8752293906738207,
        "trm_confidence": 0.5037801849547963,
        "mcts_value": 0.06546371962314065,
        "consensus_score": 0.46465257885137445,
        "last_agent": "mcts",
        "iteration": 0,
        "query_length": 905,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7989583636998355,
        "trm_confidence": 0.10101639056485813,
        "mcts_value": 0.07227436912106161,
        "consensus_score": 0.3416119555641275,
        "last_agent": "mcts",
        "iteration": 1,
        "query_length": 1507,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8743183419101185,
        "trm_confidence": 0.26858765190610195,
        "mcts_value": 0.45755670357289335,
        "consensus_score": 0.43804834000231074,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 2097,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9348205681750859,
        "trm_confidence": 0.06906470554037497,
        "mcts_value": 0.40627238425745194,
        "consensus_score": 0.4681939515285418,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 813,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8420468203170861,
        "trm_confidence": 0.19810844189938528,
        "mcts_value": 0.24603972019376524,
        "consensus_score": 0.4328661412977197,
        "last_agent": "mcts",
        "iteration": 4,
        "query_length": 1430,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9478875772583073,
        "trm_confidence": 0.7598435856691546,
        "mcts_value": 0.11891546028371681,
        "consensus_score": 0.6196894364448695,
        "last_agent": "hrm",
        "iteration": 1,
        "query_length": 3333,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7843701351517025,
        "trm_confidence": 0.4512891576261886,
        "mcts_value": 0.4975334024340737,
        "consensus_score": 0.6314603967875196,
        "last_agent": "hrm",
        "iteration": 1,
        "query_length": 3077,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7690641972684642,
        "trm_confidence": 0.025031401865777046,
        "mcts_value": 0.37123192203583794,
        "consensus_score": 0.3626269638291808,
        "last_agent": "hrm",
        "iteration": 9,
        "query_length": 2306,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7951416678468146,
        "trm_confidence": 0.6624000747790032,
        "mcts_value": 0.20222911121111173,
        "consensus_score": 0.5562683771253194,
        "last_agent": "mcts",
        "iteration": 2,
        "query_length": 4957,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7493823452746055,
        "trm_confidence": 0.029164163348725774,
        "mcts_value": 0.2825443492468585,
        "consensus_score": 0.452172065434564,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 463,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9672377472635574,
        "trm_confidence": 0.7748306511117701,
        "mcts_value": 0.4499735556104071,
        "consensus_score": 0.6938664616947368,
        "last_agent": "none",
        "iteration": 8,
        "query_length": 1684,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.812097318662113,
        "trm_confidence": 0.0672694610295489,
        "mcts_value": 0.5317868798463629,
        "consensus_score": 0.4228766563639145,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 4740,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7368273797234458,
        "trm_confidence": 0.5292753052442554,
        "mcts_value": 0.09761564970867558,
        "consensus_score": 0.39042643985700676,
        "last_agent": "hrm",
        "iteration": 6,
        "query_length": 3603,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7589303997143719,
        "trm_confidence": 0.20448170182489145,
        "mcts_value": 0.5122556808021458,
        "consensus_score": 0.5862545459926631,
        "last_agent": "hrm",
        "iteration": 5,
        "query_length": 2174,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7041808863124605,
        "trm_confidence": 0.13875378375170938,
        "mcts_value": 0.07964446437793447,
        "consensus_score": 0.3430581128699396,
        "last_agent": "mcts",
        "iteration": 1,
        "query_length": 1490,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9082787306928659,
        "trm_confidence": 0.46970419528561314,
        "mcts_value": 0.16147441014713151,
        "consensus_score": 0.5739773506116561,
        "last_agent": "none",
        "iteration": 7,
        "query_length": 431,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.739317325467194,
        "trm_confidence": 0.07911795076613058,
        "mcts_value": 0.5930068093130082,
        "consensus_score": 0.44999633394709904,
        "last_agent": "mcts",
        "iteration": 3,
        "query_length": 669,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8988592638290747,
        "trm_confidence": 0.763408491622485,
        "mcts_value": 0.22883022193361457,
        "consensus_score": 0.7153276783241302,
        "last_agent": "hrm",
        "iteration": 0,
        "query_length": 2989,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8901925335043255,
        "trm_confidence": 0.08367933776133821,
        "mcts_value": 0.11089530175490933,
        "consensus_score": 0.3454119215367851,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 838,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9799069664800634,
        "trm_confidence": 0.7077627732121353,
        "mcts_value": 0.4112523272131102,
        "consensus_score": 0.7565933788188737,
        "last_agent": "mcts",
        "iteration": 0,
        "query_length": 3321,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9488285844648209,
        "trm_confidence": 0.6763611211604272,
        "mcts_value": 0.19747211169221357,
        "consensus_score": 0.6137078572256313,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 4667,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8809321472016177,
        "trm_confidence": 0.32219040146723615,
        "mcts_value": 0.29221234846655453,
        "consensus_score": 0.4836213829865048,
        "last_agent": "none",
        "iteration": 7,
        "query_length": 4814,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8361690646228899,
        "trm_confidence": 0.18245181923444714,
        "mcts_value": 0.17422351039937214,
        "consensus_score": 0.446817654134259,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 4797,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7199676570865534,
        "trm_confidence": 0.3685296458124635,
        "mcts_value": 0.0906226837308878,
        "consensus_score": 0.4579728336345698,
        "last_agent": "hrm",
        "iteration": 3,
        "query_length": 4618,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.976291141746235,
        "trm_confidence": 0.14505398231091812,
        "mcts_value": 0.2494976860299916,
        "consensus_score": 0.38767028240079326,
        "last_agent": "hrm",
        "iteration": 1,
        "query_length": 2706,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.716618622749278,
        "trm_confidence": 0.10768718328337355,
        "mcts_value": 0.032916293776263636,
        "consensus_score": 0.3039694631584993,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 4018,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7953973290856148,
        "trm_confidence": 0.35084619767887465,
        "mcts_value": 0.6084760997666595,
        "consensus_score": 0.6551328675414908,
        "last_agent": "trm",
        "iteration": 0,
        "query_length": 2470,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7710234613313187,
        "trm_confidence": 0.1673449143479575,
        "mcts_value": 0.3833105111979031,
        "consensus_score": 0.42381211409969827,
        "last_agent": "hrm",
        "iteration": 0,
        "query_length": 45,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8571258846152912,
        "trm_confidence": 0.07697842942649595,
        "mcts_value": 0.6310330448273446,
        "consensus_score": 0.4321048262493981,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 3450,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9530724854793258,
        "trm_confidence": 0.7700285610160599,
        "mcts_value": 0.8356447951906794,
        "consensus_score": 0.9133204566167403,
        "last_agent": "trm",
        "iteration": 8,
        "query_length": 3143,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9336989063734387,
        "trm_confidence": 0.11217602900417172,
        "mcts_value": 0.44691933531130246,
        "consensus_score": 0.5004426642677205,
        "last_agent": "mcts",
        "iteration": 9,
        "query_length": 4170,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.815526848830106,
        "trm_confidence": 0.45762469209576373,
        "mcts_value": 0.1906616579001964,
        "consensus_score": 0.41589141513281347,
        "last_agent": "hrm",
        "iteration": 5,
        "query_length": 834,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.769770982158662,
        "trm_confidence": 0.24614874601571698,
        "mcts_value": 0.24539903097564403,
        "consensus_score": 0.3859386992658501,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 922,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7890629423713855,
        "trm_confidence": 0.653822834874885,
        "mcts_value": 0.6314214625791688,
        "consensus_score": 0.6876178322684955,
        "last_agent": "mcts",
        "iteration": 3,
        "query_length": 2308,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9545681466320873,
        "trm_confidence": 0.5576803541380092,
        "mcts_value": 0.6874076335747309,
        "consensus_score": 0.7397631666614238,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 371,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9204679486930872,
        "trm_confidence": 0.16606648132791707,
        "mcts_value": 0.5700595955075086,
        "consensus_score": 0.6243418221746831,
        "last_agent": "mcts",
        "iteration": 1,
        "query_length": 4763,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7285287244689803,
        "trm_confidence": 0.45613311816464464,
        "mcts_value": 0.05310641505342561,
        "consensus_score": 0.49977738376914194,
        "last_agent": "trm",
        "iteration": 1,
        "query_length": 4801,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.940265252825529,
        "trm_confidence": 0.49885035958088786,
        "mcts_value": 0.6576118411438995,
        "consensus_score": 0.757932118980056,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 459,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8770227686074927,
        "trm_confidence": 0.07385539064541159,
        "mcts_value": 0.4787747781537166,
        "consensus_score": 0.410809239954098,
        "last_agent": "none",
        "iteration": 6,
        "query_length": 49,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8397955458943682,
        "trm_confidence": 0.3866406596674736,
        "mcts_value": 0.5651471213314001,
        "consensus_score": 0.6570433856001606,
        "last_agent": "trm",
        "iteration": 5,
        "query_length": 1186,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9793708707070305,
        "trm_confidence": 0.10529022997310968,
        "mcts_value": 0.10297746470847831,
        "consensus_score": 0.31342132417835983,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 3647,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9322964248425529,
        "trm_confidence": 0.5586635055387557,
        "mcts_value": 0.2776855134490763,
        "consensus_score": 0.669221790749102,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 3176,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8092576053316162,
        "trm_confidence": 0.2230189473801963,
        "mcts_value": 0.11178726047613538,
        "consensus_score": 0.31091127890442094,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 2780,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.814995946833223,
        "trm_confidence": 0.5217223241664531,
        "mcts_value": 0.3953878002827034,
        "consensus_score": 0.6645966877951524,
        "last_agent": "hrm",
        "iteration": 8,
        "query_length": 3067,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8129078420442478,
        "trm_confidence": 0.7033773655847808,
        "mcts_value": 0.5116969009545275,
        "consensus_score": 0.766232968195287,
        "last_agent": "hrm",
        "iteration": 1,
        "query_length": 1362,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8911221651973622,
        "trm_confidence": 0.0964549421489706,
        "mcts_value": 0.4653839426303197,
        "consensus_score": 0.5215396230133073,
        "last_agent": "trm",
        "iteration": 0,
        "query_length": 487,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9476198533588963,
        "trm_confidence": 0.250352173750384,
        "mcts_value": 0.3886744578520911,
        "consensus_score": 0.5173449870673691,
        "last_agent": "none",
        "iteration": 3,
        "query_length": 2583,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9343882106294548,
        "trm_confidence": 0.09227366636739937,
        "mcts_value": 0.8319139640886433,
        "consensus_score": 0.6953652852233815,
        "last_agent": "trm",
        "iteration": 3,
        "query_length": 3892,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7319258595614149,
        "trm_confidence": 0.631360115596809,
        "mcts_value": 0.42066339906103695,
        "consensus_score": 0.6246747945181796,
        "last_agent": "trm",
        "iteration": 0,
        "query_length": 3407,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7086998509613408,
        "trm_confidence": 0.14659200303695535,
        "mcts_value": 0.08705739409814221,
        "consensus_score": 0.3694700041705898,
        "last_agent": "mcts",
        "iteration": 2,
        "query_length": 82,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.896880711798247,
        "trm_confidence": 0.02881736647956746,
        "mcts_value": 0.004326930085425239,
        "consensus_score": 0.2203399195239597,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 4901,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7715658461882166,
        "trm_confidence": 0.5704339683573875,
        "mcts_value": 0.03843501631801768,
        "consensus_score": 0.5203377143495425,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 610,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9094362351935912,
        "trm_confidence": 0.6782915533551156,
        "mcts_value": 0.032499916743507586,
        "consensus_score": 0.4804323239078826,
        "last_agent": "none",
        "iteration": 1,
        "query_length": 3490,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9235564384986533,
        "trm_confidence": 0.5188503109313061,
        "mcts_value": 0.7009544972153812,
        "consensus_score": 0.6454963473699984,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 274,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7812276253953401,
        "trm_confidence": 0.4836066911859404,
        "mcts_value": 0.6677425780176791,
        "consensus_score": 0.6665010193950671,
        "last_agent": "trm",
        "iteration": 0,
        "query_length": 360,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.71270516547405,
        "trm_confidence": 0.5417206443861774,
        "mcts_value": 0.4347622805893685,
        "consensus_score": 0.4976882661010326,
        "last_agent": "hrm",
        "iteration": 1,
        "query_length": 3854,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9940081538662577,
        "trm_confidence": 0.40995695340003957,
        "mcts_value": 0.7009747611084067,
        "consensus_score": 0.7289282912131234,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 2238,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9838073360674285,
        "trm_confidence": 0.26632882928637897,
        "mcts_value": 0.5108558557039758,
        "consensus_score": 0.6269525292704077,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 2219,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7445316969760174,
        "trm_confidence": 0.3276494531736131,
        "mcts_value": 0.2604129715008172,
        "consensus_score": 0.4390317864466729,
        "last_agent": "none",
        "iteration": 1,
        "query_length": 4868,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7834226637343772,
        "trm_confidence": 0.20824203195435526,
        "mcts_value": 0.2924387541146086,
        "consensus_score": 0.45023199267510605,
        "last_agent": "hrm",
        "iteration": 6,
        "query_length": 4858,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8226349328286091,
        "trm_confidence": 0.1572659757712483,
        "mcts_value": 0.4251306463016947,
        "consensus_score": 0.4317520338972982,
        "last_agent": "mcts",
        "iteration": 0,
        "query_length": 1500,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8422398025270511,
        "trm_confidence": 0.16744400594740463,
        "mcts_value": 0.424901063235603,
        "consensus_score": 0.4913493373230704,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 2740,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8957299169563094,
        "trm_confidence": 0.25162106077135565,
        "mcts_value": 0.6265833766096053,
        "consensus_score": 0.6011403281656927,
        "last_agent": "none",
        "iteration": 4,
        "query_length": 2653,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8081972003350762,
        "trm_confidence": 0.36312049753242803,
        "mcts_value": 0.5217329060540332,
        "consensus_score": 0.6416307786519208,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 4253,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8560825344221707,
        "trm_confidence": 0.6047680473796243,
        "mcts_value": 0.23775067596476046,
        "consensus_score": 0.6336768917245321,
        "last_agent": "hrm",
        "iteration": 5,
        "query_length": 217,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7216177441218831,
        "trm_confidence": 0.5233979204122917,
        "mcts_value": 0.034542003153205686,
        "consensus_score": 0.3826415097891913,
        "last_agent": "trm",
        "iteration": 3,
        "query_length": 2443,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7941680109405967,
        "trm_confidence": 0.5155534217839192,
        "mcts_value": 0.010192360309046276,
        "consensus_score": 0.505405949248658,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 2082,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7460838697181904,
        "trm_confidence": 0.388193295848214,
        "mcts_value": 0.07731850799225726,
        "consensus_score": 0.3768490967352044,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 392,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9316314674063757,
        "trm_confidence": 0.25860537680230905,
        "mcts_value": 0.5718838939220751,
        "consensus_score": 0.6284548524737226,
        "last_agent": "mcts",
        "iteration": 4,
        "query_length": 4629,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7032182934927028,
        "trm_confidence": 0.12610740406362417,
        "mcts_value": 0.3167428700488906,
        "consensus_score": 0.314773116720065,
        "last_agent": "trm",
        "iteration": 1,
        "query_length": 2095,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9967399008050817,
        "trm_confidence": 0.49855996974088884,
        "mcts_value": 0.7524273072394143,
        "consensus_score": 0.8473067254603828,
        "last_agent": "trm",
        "iteration": 1,
        "query_length": 2519,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8177718147525759,
        "trm_confidence": 0.05745711959236592,
        "mcts_value": 0.5421547088642604,
        "consensus_score": 0.459217019867586,
        "last_agent": "hrm",
        "iteration": 5,
        "query_length": 4596,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7542779956692605,
        "trm_confidence": 0.5934979397686945,
        "mcts_value": 0.029212916442549666,
        "consensus_score": 0.4055667409489095,
        "last_agent": "none",
        "iteration": 3,
        "query_length": 980,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8759335519044388,
        "trm_confidence": 0.3827602430806626,
        "mcts_value": 0.06526791031235166,
        "consensus_score": 0.39005405924362707,
        "last_agent": "mcts",
        "iteration": 9,
        "query_length": 4272,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8947447150402823,
        "trm_confidence": 0.5326404951932822,
        "mcts_value": 0.6063131424699416,
        "consensus_score": 0.5895211472454582,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 3721,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8015369449891978,
        "trm_confidence": 0.5924331288738127,
        "mcts_value": 0.33854244341010364,
        "consensus_score": 0.6312296903162472,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 1856,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9728656731614049,
        "trm_confidence": 0.5124803335859882,
        "mcts_value": 0.7421752482233381,
        "consensus_score": 0.7106252441085485,
        "last_agent": "trm",
        "iteration": 5,
        "query_length": 4242,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7314939147688211,
        "trm_confidence": 0.2516834826978827,
        "mcts_value": 0.5792931580045768,
        "consensus_score": 0.5469899665633753,
        "last_agent": "hrm",
        "iteration": 1,
        "query_length": 52,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7574809028988099,
        "trm_confidence": 0.016320735107701632,
        "mcts_value": 0.6097875396601716,
        "consensus_score": 0.45083785820693534,
        "last_agent": "hrm",
        "iteration": 3,
        "query_length": 720,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7021943368879047,
        "trm_confidence": 0.16742333812465543,
        "mcts_value": 0.4233627716865501,
        "consensus_score": 0.45774743684688934,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 4016,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8432517620883198,
        "trm_confidence": 0.5659360942979396,
        "mcts_value": 0.6714000325807745,
        "consensus_score": 0.737668485666015,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 3121,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.960040431478578,
        "trm_confidence": 0.09813411461423771,
        "mcts_value": 0.629905225135514,
        "consensus_score": 0.5507109969815835,
        "last_agent": "trm",
        "iteration": 6,
        "query_length": 4808,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9909445350116708,
        "trm_confidence": 0.8772044609437818,
        "mcts_value": 0.25679538015345144,
        "consensus_score": 0.7550654915238845,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 3189,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7371609324592663,
        "trm_confidence": 0.026089803397021747,
        "mcts_value": 0.49529267197729376,
        "consensus_score": 0.41745441764124175,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 4301,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.993375093722152,
        "trm_confidence": 0.36769174953390693,
        "mcts_value": 0.7090558656421148,
        "consensus_score": 0.6070047574284242,
        "last_agent": "none",
        "iteration": 6,
        "query_length": 2707,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9774105001659665,
        "trm_confidence": 0.7217430409667073,
        "mcts_value": 0.03243850424508178,
        "consensus_score": 0.5517378167205987,
        "last_agent": "mcts",
        "iteration": 0,
        "query_length": 1851,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9025916886002452,
        "trm_confidence": 0.5724551001857711,
        "mcts_value": 0.6209817892745624,
        "consensus_score": 0.7717675022979922,
        "last_agent": "trm",
        "iteration": 8,
        "query_length": 1973,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7146891129469838,
        "trm_confidence": 0.1441662035796492,
        "mcts_value": 0.38227379065973416,
        "consensus_score": 0.48533476332042036,
        "last_agent": "hrm",
        "iteration": 0,
        "query_length": 4760,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9031862198093191,
        "trm_confidence": 0.023780166634012397,
        "mcts_value": 0.3223632530707673,
        "consensus_score": 0.49557018944810566,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 1280,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9558343388895738,
        "trm_confidence": 0.29785724316647905,
        "mcts_value": 0.7303216724735169,
        "consensus_score": 0.6211264817336376,
        "last_agent": "trm",
        "iteration": 6,
        "query_length": 2250,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7824475152265201,
        "trm_confidence": 0.6050290054094732,
        "mcts_value": 0.12802284211179998,
        "consensus_score": 0.42212877249641706,
        "last_agent": "trm",
        "iteration": 3,
        "query_length": 2255,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9422294817928845,
        "trm_confidence": 0.8411711110154104,
        "mcts_value": 0.2496048616995867,
        "consensus_score": 0.6592568750622169,
        "last_agent": "trm",
        "iteration": 1,
        "query_length": 605,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9992740111788752,
        "trm_confidence": 0.6302840773303063,
        "mcts_value": 0.5352594385037045,
        "consensus_score": 0.7000796609219875,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 1105,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7403100739939185,
        "trm_confidence": 0.23395551080682137,
        "mcts_value": 0.04300749469983864,
        "consensus_score": 0.2794868340379491,
        "last_agent": "trm",
        "iteration": 0,
        "query_length": 1516,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8903620794855391,
        "trm_confidence": 0.2713253444375048,
        "mcts_value": 0.33225381110472557,
        "consensus_score": 0.5898222662448384,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 2521,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7853622625335476,
        "trm_confidence": 0.6147677554924093,
        "mcts_value": 0.16112669194044266,
        "consensus_score": 0.48548744942546856,
        "last_agent": "mcts",
        "iteration": 9,
        "query_length": 1537,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9226953848253929,
        "trm_confidence": 0.48600301576815713,
        "mcts_value": 0.537581421516067,
        "consensus_score": 0.6086365990061708,
        "last_agent": "mcts",
        "iteration": 2,
        "query_length": 1505,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7466324692250267,
        "trm_confidence": 0.5653600569069068,
        "mcts_value": 0.18315666384793758,
        "consensus_score": 0.5106809422095007,
        "last_agent": "hrm",
        "iteration": 8,
        "query_length": 2929,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8315158775752578,
        "trm_confidence": 0.348389783042449,
        "mcts_value": 0.7276401227978485,
        "consensus_score": 0.6707680898676567,
        "last_agent": "hrm",
        "iteration": 8,
        "query_length": 3361,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9362769508688018,
        "trm_confidence": 0.15486121796165153,
        "mcts_value": 0.47013042707773395,
        "consensus_score": 0.4408016967656381,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 1601,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8538196192900334,
        "trm_confidence": 0.32638315973663695,
        "mcts_value": 0.02701896922834707,
        "consensus_score": 0.49436215388418314,
        "last_agent": "trm",
        "iteration": 1,
        "query_length": 1475,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7738199834514815,
        "trm_confidence": 0.04415571589845285,
        "mcts_value": 0.3066674970423677,
        "consensus_score": 0.3780986257155231,
        "last_agent": "mcts",
        "iteration": 3,
        "query_length": 247,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.733480110474768,
        "trm_confidence": 0.24357587584709162,
        "mcts_value": 0.03834356971646067,
        "consensus_score": 0.3781044638014227,
        "last_agent": "hrm",
        "iteration": 2,
        "query_length": 4886,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8182377331534292,
        "trm_confidence": 0.2992247584654667,
        "mcts_value": 0.0011927829157792366,
        "consensus_score": 0.2952991657112979,
        "last_agent": "hrm",
        "iteration": 9,
        "query_length": 821,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8524502067438625,
        "trm_confidence": 0.3683258859978843,
        "mcts_value": 0.2506367703176279,
        "consensus_score": 0.4767363583824148,
        "last_agent": "trm",
        "iteration": 8,
        "query_length": 2341,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.778104563024575,
        "trm_confidence": 0.21868223955202493,
        "mcts_value": 0.1644287290937298,
        "consensus_score": 0.38304452400032296,
        "last_agent": "mcts",
        "iteration": 7,
        "query_length": 1673,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7992207226164599,
        "trm_confidence": 0.6505442046219245,
        "mcts_value": 0.033960653943758315,
        "consensus_score": 0.4867291147752941,
        "last_agent": "mcts",
        "iteration": 7,
        "query_length": 965,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7142122060016445,
        "trm_confidence": 0.08488742619537909,
        "mcts_value": 0.5643524206239566,
        "consensus_score": 0.356335973971218,
        "last_agent": "mcts",
        "iteration": 2,
        "query_length": 4085,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7331888402686609,
        "trm_confidence": 0.3926716075936083,
        "mcts_value": 0.15300306038828654,
        "consensus_score": 0.440130246169654,
        "last_agent": "none",
        "iteration": 6,
        "query_length": 3441,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7014224178735622,
        "trm_confidence": 0.5132355235771408,
        "mcts_value": 0.37235209791058166,
        "consensus_score": 0.4615520436747531,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 4976,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7762790451262471,
        "trm_confidence": 0.6213609131888949,
        "mcts_value": 0.3073400370129043,
        "consensus_score": 0.589011555976013,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 406,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9439478037460338,
        "trm_confidence": 0.26871223620600315,
        "mcts_value": 0.6744944968198383,
        "consensus_score": 0.6491982960298203,
        "last_agent": "hrm",
        "iteration": 2,
        "query_length": 2582,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7952906903090685,
        "trm_confidence": 0.054308038543778275,
        "mcts_value": 0.02074383789783102,
        "consensus_score": 0.2594099199983288,
        "last_agent": "hrm",
        "iteration": 0,
        "query_length": 3514,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9175549962552606,
        "trm_confidence": 0.5789035275376786,
        "mcts_value": 0.6038300850843744,
        "consensus_score": 0.6635268371851003,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 748,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7378079025239879,
        "trm_confidence": 0.09166489801650933,
        "mcts_value": 0.4420617404990008,
        "consensus_score": 0.3584336343015816,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 615,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7012022814200018,
        "trm_confidence": 0.009967814726302898,
        "mcts_value": 0.5970447162996493,
        "consensus_score": 0.4529993709754845,
        "last_agent": "hrm",
        "iteration": 1,
        "query_length": 2868,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9641140900383076,
        "trm_confidence": 0.4633383320539006,
        "mcts_value": 0.5372992540761998,
        "consensus_score": 0.6095546636896446,
        "last_agent": "trm",
        "iteration": 0,
        "query_length": 2241,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7884265579608288,
        "trm_confidence": 0.4568543529427487,
        "mcts_value": 0.5756933887050966,
        "consensus_score": 0.5106235419857642,
        "last_agent": "none",
        "iteration": 6,
        "query_length": 789,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9621665493711404,
        "trm_confidence": 0.21870105232034467,
        "mcts_value": 0.5265855782305323,
        "consensus_score": 0.5798592279919281,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 3761,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9177308904146302,
        "trm_confidence": 0.46360254817622953,
        "mcts_value": 0.6201852913233449,
        "consensus_score": 0.763856325169989,
        "last_agent": "none",
        "iteration": 4,
        "query_length": 4622,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7037409633823639,
        "trm_confidence": 0.4805950474667002,
        "mcts_value": 0.3140842018588201,
        "consensus_score": 0.48106569030807445,
        "last_agent": "mcts",
        "iteration": 1,
        "query_length": 3864,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8185021720121669,
        "trm_confidence": 0.49043636088886106,
        "mcts_value": 0.10732565579300138,
        "consensus_score": 0.5643979888062163,
        "last_agent": "hrm",
        "iteration": 1,
        "query_length": 3413,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9576948581673141,
        "trm_confidence": 0.782717527302752,
        "mcts_value": 0.1819577406062614,
        "consensus_score": 0.6347502196604081,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 3226,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8137158085111915,
        "trm_confidence": 0.36900498107278173,
        "mcts_value": 0.5293747276201115,
        "consensus_score": 0.616925701065254,
        "last_agent": "trm",
        "iteration": 8,
        "query_length": 4750,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7313845290414325,
        "trm_confidence": 0.5707547505270651,
        "mcts_value": 0.5465091100904041,
        "consensus_score": 0.6758356170499672,
        "last_agent": "trm",
        "iteration": 1,
        "query_length": 26,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9230801095322956,
        "trm_confidence": 0.019046151583096152,
        "mcts_value": 0.8059803187685097,
        "consensus_score": 0.5581311293657699,
        "last_agent": "mcts",
        "iteration": 7,
        "query_length": 4657,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8183889664450247,
        "trm_confidence": 0.2292612586441427,
        "mcts_value": 0.437312910421789,
        "consensus_score": 0.5111855865865955,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 60,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9806144108699484,
        "trm_confidence": 0.41181074908830173,
        "mcts_value": 0.1732539995949861,
        "consensus_score": 0.4973342497610128,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 3847,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7488614501663664,
        "trm_confidence": 0.44418593967111597,
        "mcts_value": 0.22027402789520992,
        "consensus_score": 0.5620868176153923,
        "last_agent": "trm",
        "iteration": 2,
        "query_length": 999,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9260513689000773,
        "trm_confidence": 0.7277810367370762,
        "mcts_value": 0.22973591915308827,
        "consensus_score": 0.5682737285925096,
        "last_agent": "mcts",
        "iteration": 2,
        "query_length": 2400,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8405249923740312,
        "trm_confidence": 0.19212168542979885,
        "mcts_value": 0.033451822114106196,
        "consensus_score": 0.3516644347045683,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 1231,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8486519669885855,
        "trm_confidence": 0.08286983895231823,
        "mcts_value": 0.18898692545925058,
        "consensus_score": 0.3324577003816266,
        "last_agent": "hrm",
        "iteration": 8,
        "query_length": 1515,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9704925560699316,
        "trm_confidence": 0.8570596444224141,
        "mcts_value": 0.8551428849736049,
        "consensus_score": 0.9848315217089132,
        "last_agent": "trm",
        "iteration": 0,
        "query_length": 2056,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7913399579428099,
        "trm_confidence": 0.38224065571157784,
        "mcts_value": 0.06704468949902856,
        "consensus_score": 0.4827011592782339,
        "last_agent": "trm",
        "iteration": 6,
        "query_length": 2019,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7495933970101873,
        "trm_confidence": 0.16451308392705075,
        "mcts_value": 0.10413763288980703,
        "consensus_score": 0.40980133616797476,
        "last_agent": "hrm",
        "iteration": 6,
        "query_length": 2625,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7888098588730077,
        "trm_confidence": 0.25572247655900815,
        "mcts_value": 0.2788680909106773,
        "consensus_score": 0.4931364697963714,
        "last_agent": "none",
        "iteration": 8,
        "query_length": 225,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9824616570214297,
        "trm_confidence": 0.10647370900903526,
        "mcts_value": 0.7907933777490511,
        "consensus_score": 0.5466553770564673,
        "last_agent": "hrm",
        "iteration": 2,
        "query_length": 897,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7538476217038262,
        "trm_confidence": 0.27041040829999513,
        "mcts_value": 0.29413943336981513,
        "consensus_score": 0.3885176157291874,
        "last_agent": "mcts",
        "iteration": 7,
        "query_length": 679,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9623725419566715,
        "trm_confidence": 0.29262210136103023,
        "mcts_value": 0.4577908039206839,
        "consensus_score": 0.5206097631859256,
        "last_agent": "hrm",
        "iteration": 2,
        "query_length": 1488,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9820057032422544,
        "trm_confidence": 0.7832112112455303,
        "mcts_value": 0.6856348186200852,
        "consensus_score": 0.8204828819601692,
        "last_agent": "hrm",
        "iteration": 5,
        "query_length": 2263,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8609704040731267,
        "trm_confidence": 0.33068964615766583,
        "mcts_value": 0.10026152842446387,
        "consensus_score": 0.355773296792126,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 456,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9859616574715566,
        "trm_confidence": 0.14492284170007116,
        "mcts_value": 0.4911918073749651,
        "consensus_score": 0.4822355528525055,
        "last_agent": "none",
        "iteration": 2,
        "query_length": 4373,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7356859932220634,
        "trm_confidence": 0.5828276906378326,
        "mcts_value": 0.2043940190441073,
        "consensus_score": 0.5292617433509195,
        "last_agent": "none",
        "iteration": 5,
        "query_length": 4855,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8595658274038205,
        "trm_confidence": 0.14222030388897,
        "mcts_value": 0.7509097381536648,
        "consensus_score": 0.6478935302229244,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 1861,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7458631366379007,
        "trm_confidence": 0.5946241390199914,
        "mcts_value": 0.220489322146777,
        "consensus_score": 0.4303468705409431,
        "last_agent": "mcts",
        "iteration": 3,
        "query_length": 294,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8868126234964624,
        "trm_confidence": 0.590769276685474,
        "mcts_value": 0.6244423384763792,
        "consensus_score": 0.643065668435547,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 2660,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8917040947792494,
        "trm_confidence": 0.0018278647108377233,
        "mcts_value": 0.786453907816844,
        "consensus_score": 0.5161745662951414,
        "last_agent": "mcts",
        "iteration": 0,
        "query_length": 1994,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7387090171546125,
        "trm_confidence": 0.0972924433311407,
        "mcts_value": 0.4038447340742791,
        "consensus_score": 0.3918675431673908,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 3242,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9178540385151703,
        "trm_confidence": 0.3771120311092349,
        "mcts_value": 0.539743836851922,
        "consensus_score": 0.6315090401686404,
        "last_agent": "none",
        "iteration": 5,
        "query_length": 3074,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8028140928264867,
        "trm_confidence": 0.12456360802715516,
        "mcts_value": 0.47657406224828497,
        "consensus_score": 0.5371854332955573,
        "last_agent": "trm",
        "iteration": 0,
        "query_length": 4940,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9677463069061979,
        "trm_confidence": 0.6504787881584428,
        "mcts_value": 0.8606374831518863,
        "consensus_score": 0.8325702883916245,
        "last_agent": "mcts",
        "iteration": 7,
        "query_length": 869,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9842606798997977,
        "trm_confidence": 0.3239360681824274,
        "mcts_value": 0.6530359849135624,
        "consensus_score": 0.6343722851557455,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 756,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8526802455254062,
        "trm_confidence": 0.6827459592509878,
        "mcts_value": 0.3155995705425496,
        "consensus_score": 0.6464229514016462,
        "last_agent": "hrm",
        "iteration": 3,
        "query_length": 3486,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8320150324700657,
        "trm_confidence": 0.09209262859482964,
        "mcts_value": 0.06714311190465266,
        "consensus_score": 0.3639149089525342,
        "last_agent": "none",
        "iteration": 7,
        "query_length": 3668,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7059299336028074,
        "trm_confidence": 0.197864506698318,
        "mcts_value": 0.11985890761574668,
        "consensus_score": 0.39684806714775983,
        "last_agent": "mcts",
        "iteration": 9,
        "query_length": 3759,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9003547476543958,
        "trm_confidence": 0.10353613391771245,
        "mcts_value": 0.22021963019442897,
        "consensus_score": 0.3357677365021964,
        "last_agent": "hrm",
        "iteration": 3,
        "query_length": 1412,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9730980730555757,
        "trm_confidence": 0.17682192410918202,
        "mcts_value": 0.47328816085929903,
        "consensus_score": 0.5813476062567055,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 2916,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7765320793082142,
        "trm_confidence": 0.4231560380588548,
        "mcts_value": 0.21682255200993333,
        "consensus_score": 0.5673959012640091,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 2207,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7069682747040275,
        "trm_confidence": 0.2428246057360107,
        "mcts_value": 0.531162558238724,
        "consensus_score": 0.51569730177957,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 890,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9561953128441687,
        "trm_confidence": 0.27587954965212524,
        "mcts_value": 0.4662397894445769,
        "consensus_score": 0.5640697351987828,
        "last_agent": "hrm",
        "iteration": 6,
        "query_length": 4067,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7732615002297665,
        "trm_confidence": 0.6144668090215966,
        "mcts_value": 0.622009131556389,
        "consensus_score": 0.7385770574216465,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 114,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9294001685685999,
        "trm_confidence": 0.6324923508225583,
        "mcts_value": 0.39861605011890444,
        "consensus_score": 0.6447932954062084,
        "last_agent": "trm",
        "iteration": 2,
        "query_length": 3035,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7867960408307003,
        "trm_confidence": 0.17706179507643613,
        "mcts_value": 0.4024359518088873,
        "consensus_score": 0.4374277402170165,
        "last_agent": "hrm",
        "iteration": 2,
        "query_length": 3291,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9581709757883128,
        "trm_confidence": 0.14391486044111265,
        "mcts_value": 0.006442317663250758,
        "consensus_score": 0.3367353783367668,
        "last_agent": "hrm",
        "iteration": 8,
        "query_length": 680,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7414987210348147,
        "trm_confidence": 0.49673549026026803,
        "mcts_value": 0.07422407534739722,
        "consensus_score": 0.4154986135787992,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 1439,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7461890747732879,
        "trm_confidence": 0.27320895570225096,
        "mcts_value": 0.4004230314625919,
        "consensus_score": 0.5598538437953609,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 3167,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7413992552043716,
        "trm_confidence": 0.5071787475187043,
        "mcts_value": 0.17358326090088597,
        "consensus_score": 0.5510235664425928,
        "last_agent": "none",
        "iteration": 7,
        "query_length": 2268,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9497587854772022,
        "trm_confidence": 0.14986038328681847,
        "mcts_value": 0.359868932238171,
        "consensus_score": 0.49693301937926315,
        "last_agent": "none",
        "iteration": 5,
        "query_length": 1482,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8958060204607765,
        "trm_confidence": 0.789473000549209,
        "mcts_value": 0.7953459723410123,
        "consensus_score": 0.830166878310794,
        "last_agent": "none",
        "iteration": 1,
        "query_length": 3420,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9929992234649767,
        "trm_confidence": 0.28426521696112633,
        "mcts_value": 0.41185917690810053,
        "consensus_score": 0.5479965907411591,
        "last_agent": "none",
        "iteration": 0,
        "query_length": 1850,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8060013085205403,
        "trm_confidence": 0.5559505259024953,
        "mcts_value": 0.5874778619952927,
        "consensus_score": 0.7140560561256785,
        "last_agent": "mcts",
        "iteration": 4,
        "query_length": 1979,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9246462339001165,
        "trm_confidence": 0.660862778505748,
        "mcts_value": 0.4051136030420285,
        "consensus_score": 0.7421470757232248,
        "last_agent": "hrm",
        "iteration": 1,
        "query_length": 2854,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7288285698915643,
        "trm_confidence": 0.09654123447545868,
        "mcts_value": 0.3356938520979582,
        "consensus_score": 0.3005249148591434,
        "last_agent": "trm",
        "iteration": 0,
        "query_length": 1141,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8308949979082789,
        "trm_confidence": 0.5661687200967872,
        "mcts_value": 0.02546317595496685,
        "consensus_score": 0.504627152613182,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 1717,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7440591204464212,
        "trm_confidence": 0.5544067179494655,
        "mcts_value": 0.5301409123022616,
        "consensus_score": 0.6173239368188868,
        "last_agent": "hrm",
        "iteration": 9,
        "query_length": 226,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8183444553816263,
        "trm_confidence": 0.5299927836231273,
        "mcts_value": 0.18493918750750507,
        "consensus_score": 0.5615579250574635,
        "last_agent": "none",
        "iteration": 5,
        "query_length": 2699,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8394038913571988,
        "trm_confidence": 0.2720299008072148,
        "mcts_value": 0.1620693596078362,
        "consensus_score": 0.3661769188777529,
        "last_agent": "trm",
        "iteration": 8,
        "query_length": 4810,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7257947050332835,
        "trm_confidence": 0.11032220277393885,
        "mcts_value": 0.10973091841392231,
        "consensus_score": 0.37990413330176154,
        "last_agent": "none",
        "iteration": 4,
        "query_length": 3499,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7411633773371836,
        "trm_confidence": 0.4190258018660907,
        "mcts_value": 0.2877089263480039,
        "consensus_score": 0.4612163371898409,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 4766,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9618576876322535,
        "trm_confidence": 0.16573581856333194,
        "mcts_value": 0.19029638580714042,
        "consensus_score": 0.470511406486588,
        "last_agent": "mcts",
        "iteration": 3,
        "query_length": 1267,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8699261952428934,
        "trm_confidence": 0.42415909204976393,
        "mcts_value": 0.6379199814213259,
        "consensus_score": 0.68610831058653,
        "last_agent": "hrm",
        "iteration": 0,
        "query_length": 2657,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8804675357368047,
        "trm_confidence": 0.3794571311556796,
        "mcts_value": 0.20306567679207435,
        "consensus_score": 0.4713946653578132,
        "last_agent": "hrm",
        "iteration": 8,
        "query_length": 528,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8683669203670132,
        "trm_confidence": 0.2961051612096069,
        "mcts_value": 0.20799443698331926,
        "consensus_score": 0.4618725336738305,
        "last_agent": "none",
        "iteration": 3,
        "query_length": 2258,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9029957147528913,
        "trm_confidence": 0.05306132320252652,
        "mcts_value": 0.0008571698389146299,
        "consensus_score": 0.26142683036625936,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 3870,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7515881055811068,
        "trm_confidence": 0.5800854083517826,
        "mcts_value": 0.3062177008384156,
        "consensus_score": 0.5411820895928338,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 2379,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7645534823886155,
        "trm_confidence": 0.38927637401496484,
        "mcts_value": 0.12991124070271629,
        "consensus_score": 0.4638191008772238,
        "last_agent": "none",
        "iteration": 2,
        "query_length": 4463,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7620955478922519,
        "trm_confidence": 0.27989032203406305,
        "mcts_value": 0.1166203836958097,
        "consensus_score": 0.31314048952537765,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 909,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.808227605955476,
        "trm_confidence": 0.03914402980123324,
        "mcts_value": 0.25309903289406965,
        "consensus_score": 0.3197849916812422,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 2099,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9665096359188521,
        "trm_confidence": 0.8055650852189132,
        "mcts_value": 0.08650457300765457,
        "consensus_score": 0.541007022639522,
        "last_agent": "none",
        "iteration": 1,
        "query_length": 2305,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7365906892628868,
        "trm_confidence": 0.5426186472867254,
        "mcts_value": 0.48016222358567484,
        "consensus_score": 0.6504189583593672,
        "last_agent": "trm",
        "iteration": 5,
        "query_length": 724,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7132525308832484,
        "trm_confidence": 0.24607675484191002,
        "mcts_value": 0.19865498207189433,
        "consensus_score": 0.47572385567487907,
        "last_agent": "none",
        "iteration": 6,
        "query_length": 4580,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7511023580975557,
        "trm_confidence": 0.6371909988083584,
        "mcts_value": 0.32466538481186547,
        "consensus_score": 0.569720662493389,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 4783,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8192575073664978,
        "trm_confidence": 0.42058777144530124,
        "mcts_value": 0.09210280147437415,
        "consensus_score": 0.40819981871109534,
        "last_agent": "none",
        "iteration": 2,
        "query_length": 4910,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.959843381899306,
        "trm_confidence": 0.5067325029417107,
        "mcts_value": 0.3084024260626239,
        "consensus_score": 0.5733933124412227,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 4006,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9780308174466981,
        "trm_confidence": 0.8190092519450405,
        "mcts_value": 0.34374202281113436,
        "consensus_score": 0.679446522653845,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 4847,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8385453734349244,
        "trm_confidence": 0.48651928305460523,
        "mcts_value": 0.40944883375396496,
        "consensus_score": 0.5423142727533135,
        "last_agent": "mcts",
        "iteration": 3,
        "query_length": 361,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7283919304748165,
        "trm_confidence": 0.10435241816234642,
        "mcts_value": 0.45439887808312374,
        "consensus_score": 0.4068670845445116,
        "last_agent": "mcts",
        "iteration": 2,
        "query_length": 3905,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9266618627420282,
        "trm_confidence": 0.21280124956787808,
        "mcts_value": 0.6834756638873442,
        "consensus_score": 0.6931472306574268,
        "last_agent": "none",
        "iteration": 6,
        "query_length": 532,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7157727970969165,
        "trm_confidence": 0.5822254280053311,
        "mcts_value": 0.24168142263616446,
        "consensus_score": 0.5980066293552234,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 4895,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7115625997970306,
        "trm_confidence": 0.41611869374628807,
        "mcts_value": 0.34410671336168863,
        "consensus_score": 0.395902489742884,
        "last_agent": "hrm",
        "iteration": 8,
        "query_length": 763,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8489252447317924,
        "trm_confidence": 0.3488175050052104,
        "mcts_value": 0.004098949373417123,
        "consensus_score": 0.4579438422021932,
        "last_agent": "mcts",
        "iteration": 3,
        "query_length": 4405,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8119181762674159,
        "trm_confidence": 0.40228923558521373,
        "mcts_value": 0.19246123282373728,
        "consensus_score": 0.40100721074243934,
        "last_agent": "hrm",
        "iteration": 8,
        "query_length": 757,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8609463771256752,
        "trm_confidence": 0.734629772221256,
        "mcts_value": 0.733232643276524,
        "consensus_score": 0.8473980218816325,
        "last_agent": "mcts",
        "iteration": 2,
        "query_length": 4657,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9634442273966612,
        "trm_confidence": 0.3233287242950655,
        "mcts_value": 0.0872995353312898,
        "consensus_score": 0.5202167427889899,
        "last_agent": "trm",
        "iteration": 5,
        "query_length": 1230,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9946432888953534,
        "trm_confidence": 0.5432376312320689,
        "mcts_value": 0.3802049553050477,
        "consensus_score": 0.6133588320469295,
        "last_agent": "none",
        "iteration": 4,
        "query_length": 652,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9384251675252715,
        "trm_confidence": 0.7023628971642141,
        "mcts_value": 0.23844571519841753,
        "consensus_score": 0.5871705836446232,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 339,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8561372331301683,
        "trm_confidence": 0.3506024719689236,
        "mcts_value": 0.5218305460581177,
        "consensus_score": 0.5088684237489608,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 576,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8856144334508926,
        "trm_confidence": 0.031499218465408804,
        "mcts_value": 0.36273955042497696,
        "consensus_score": 0.43106791291023394,
        "last_agent": "mcts",
        "iteration": 2,
        "query_length": 2242,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7565422909036904,
        "trm_confidence": 0.22594317178577628,
        "mcts_value": 0.06355995720741427,
        "consensus_score": 0.4381457657782145,
        "last_agent": "none",
        "iteration": 4,
        "query_length": 1131,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.992325136932654,
        "trm_confidence": 0.07204211978374034,
        "mcts_value": 0.5815010472099233,
        "consensus_score": 0.6229045853236821,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 1805,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8293229072913815,
        "trm_confidence": 0.026181879911408894,
        "mcts_value": 0.38989657417045104,
        "consensus_score": 0.5138423864817833,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 3307,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9117391010888465,
        "trm_confidence": 0.2449543416327622,
        "mcts_value": 0.732991603678876,
        "consensus_score": 0.5389179646086839,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 3812,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7352513146757804,
        "trm_confidence": 0.038003488930421665,
        "mcts_value": 0.5302491411336361,
        "consensus_score": 0.3656528491081953,
        "last_agent": "trm",
        "iteration": 5,
        "query_length": 355,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.870234434064797,
        "trm_confidence": 0.3059627671516115,
        "mcts_value": 0.34276668916021336,
        "consensus_score": 0.41894473093298157,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 1451,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7568090475018561,
        "trm_confidence": 0.18787845586226595,
        "mcts_value": 0.6334090395729941,
        "consensus_score": 0.4448893916766401,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 646,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9490057537477085,
        "trm_confidence": 0.8285426268405806,
        "mcts_value": 0.7149941480792436,
        "consensus_score": 0.9195239798114148,
        "last_agent": "trm",
        "iteration": 1,
        "query_length": 4616,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8637027082243143,
        "trm_confidence": 0.4150690578704749,
        "mcts_value": 0.6865088423172458,
        "consensus_score": 0.7494542998666873,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 3824,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8369211573705536,
        "trm_confidence": 0.2847031937971334,
        "mcts_value": 0.2664699753767556,
        "consensus_score": 0.44099947718503385,
        "last_agent": "none",
        "iteration": 0,
        "query_length": 3622,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9339609531616935,
        "trm_confidence": 0.3720073749593458,
        "mcts_value": 0.7761545231594749,
        "consensus_score": 0.7164851521456285,
        "last_agent": "none",
        "iteration": 0,
        "query_length": 1583,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7617459379039022,
        "trm_confidence": 0.6456893506063608,
        "mcts_value": 0.14472663325826077,
        "consensus_score": 0.5205384425137849,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 4357,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7709081306150634,
        "trm_confidence": 0.032953426548556077,
        "mcts_value": 0.10305271832695558,
        "consensus_score": 0.21497045423736993,
        "last_agent": "none",
        "iteration": 2,
        "query_length": 751,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9737346388886925,
        "trm_confidence": 0.46938674734886066,
        "mcts_value": 0.546948759270564,
        "consensus_score": 0.6037301654912806,
        "last_agent": "trm",
        "iteration": 8,
        "query_length": 4317,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9592410487578175,
        "trm_confidence": 0.8152849463702759,
        "mcts_value": 0.2989306274247526,
        "consensus_score": 0.6092680192247032,
        "last_agent": "hrm",
        "iteration": 2,
        "query_length": 2042,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9837080955473421,
        "trm_confidence": 0.33431545265646406,
        "mcts_value": 0.6811385287471287,
        "consensus_score": 0.6744826974177959,
        "last_agent": "hrm",
        "iteration": 2,
        "query_length": 3086,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8959575655068456,
        "trm_confidence": 0.5355627499764045,
        "mcts_value": 0.346050309650501,
        "consensus_score": 0.5382526622436733,
        "last_agent": "mcts",
        "iteration": 7,
        "query_length": 4515,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9810009075595381,
        "trm_confidence": 0.16554322366482135,
        "mcts_value": 0.09454228904780895,
        "consensus_score": 0.41322501389713184,
        "last_agent": "trm",
        "iteration": 5,
        "query_length": 167,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8310002468016519,
        "trm_confidence": 0.7272099321804025,
        "mcts_value": 0.3554712478770538,
        "consensus_score": 0.6334831251010022,
        "last_agent": "mcts",
        "iteration": 4,
        "query_length": 4867,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8778869893448848,
        "trm_confidence": 0.17733316400235186,
        "mcts_value": 0.49598773000963786,
        "consensus_score": 0.427235986841987,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 96,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9690266114772095,
        "trm_confidence": 0.41327531492474356,
        "mcts_value": 0.05242908683320221,
        "consensus_score": 0.5426863802971464,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 3741,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.828450959052662,
        "trm_confidence": 0.46470563316140123,
        "mcts_value": 0.623720838984344,
        "consensus_score": 0.66517223230718,
        "last_agent": "hrm",
        "iteration": 3,
        "query_length": 2180,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9015562565842201,
        "trm_confidence": 0.769964489790417,
        "mcts_value": 0.29730709197590294,
        "consensus_score": 0.6412922996196729,
        "last_agent": "trm",
        "iteration": 8,
        "query_length": 1853,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9209719279347199,
        "trm_confidence": 0.3774085653223031,
        "mcts_value": 0.1769154620345386,
        "consensus_score": 0.5408060869867227,
        "last_agent": "hrm",
        "iteration": 1,
        "query_length": 1940,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8880474948352626,
        "trm_confidence": 0.5888840940074078,
        "mcts_value": 0.7050565479056364,
        "consensus_score": 0.6818466783992806,
        "last_agent": "trm",
        "iteration": 1,
        "query_length": 507,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7463269264066427,
        "trm_confidence": 0.12775749309271756,
        "mcts_value": 0.18829407029016992,
        "consensus_score": 0.3600044340684342,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 3641,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9124472661527905,
        "trm_confidence": 0.1423249734377036,
        "mcts_value": 0.29059827613361366,
        "consensus_score": 0.44413782379203814,
        "last_agent": "mcts",
        "iteration": 1,
        "query_length": 1951,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7924999611858532,
        "trm_confidence": 0.08813876240274264,
        "mcts_value": 0.38104096528938425,
        "consensus_score": 0.4615655355974986,
        "last_agent": "none",
        "iteration": 0,
        "query_length": 2120,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9346633585803386,
        "trm_confidence": 0.6243199532163899,
        "mcts_value": 0.6803116479784208,
        "consensus_score": 0.7360797798925045,
        "last_agent": "hrm",
        "iteration": 9,
        "query_length": 1391,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8082431684252546,
        "trm_confidence": 0.1249026322116057,
        "mcts_value": 0.33109273139471257,
        "consensus_score": 0.3998248484936129,
        "last_agent": "trm",
        "iteration": 2,
        "query_length": 4551,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.794154031633494,
        "trm_confidence": 0.18936822561780658,
        "mcts_value": 0.09085805086171304,
        "consensus_score": 0.3458911097313942,
        "last_agent": "trm",
        "iteration": 3,
        "query_length": 4304,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7707991787861658,
        "trm_confidence": 0.5119240758573433,
        "mcts_value": 0.6238454816739297,
        "consensus_score": 0.6351092826635906,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 128,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9996488912000976,
        "trm_confidence": 0.5519494252201751,
        "mcts_value": 0.8004891104852072,
        "consensus_score": 0.865750791612955,
        "last_agent": "trm",
        "iteration": 8,
        "query_length": 1800,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7243462001271698,
        "trm_confidence": 0.009868138479946883,
        "mcts_value": 0.297938402088473,
        "consensus_score": 0.34291382027212014,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 2262,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8656488109913281,
        "trm_confidence": 0.6405517801265206,
        "mcts_value": 0.08319318191851317,
        "consensus_score": 0.47483345756656303,
        "last_agent": "none",
        "iteration": 8,
        "query_length": 352,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9984472640862603,
        "trm_confidence": 0.7842654129109966,
        "mcts_value": 0.7995273282172154,
        "consensus_score": 0.8236162569308051,
        "last_agent": "none",
        "iteration": 4,
        "query_length": 3748,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8818898194578146,
        "trm_confidence": 0.734042348906971,
        "mcts_value": 0.26947120053234996,
        "consensus_score": 0.6398681460086686,
        "last_agent": "trm",
        "iteration": 3,
        "query_length": 1085,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7585556589455966,
        "trm_confidence": 0.03816023455513779,
        "mcts_value": 0.6251626215781484,
        "consensus_score": 0.4545354267278202,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 25,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7575763484233256,
        "trm_confidence": 0.11537982339162972,
        "mcts_value": 0.5339442365121275,
        "consensus_score": 0.42190914489960163,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 3850,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9033125843363409,
        "trm_confidence": 0.654277114228846,
        "mcts_value": 0.18898441735980578,
        "consensus_score": 0.5374328446663108,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 2813,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8289871149700143,
        "trm_confidence": 0.3152982944972967,
        "mcts_value": 0.18445907233839906,
        "consensus_score": 0.49807304455598805,
        "last_agent": "trm",
        "iteration": 8,
        "query_length": 2898,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9505377373242201,
        "trm_confidence": 0.18122827258423413,
        "mcts_value": 0.003662224693813079,
        "consensus_score": 0.39270955061190405,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 4416,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7623994221125381,
        "trm_confidence": 0.3437287593091789,
        "mcts_value": 0.5365042757664858,
        "consensus_score": 0.46535255341107823,
        "last_agent": "none",
        "iteration": 4,
        "query_length": 1698,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8114629625219084,
        "trm_confidence": 0.13435011914363088,
        "mcts_value": 0.1390880405202477,
        "consensus_score": 0.34364346477826113,
        "last_agent": "hrm",
        "iteration": 5,
        "query_length": 3708,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8926757255703981,
        "trm_confidence": 0.5472648702307961,
        "mcts_value": 0.7791900694236185,
        "consensus_score": 0.722078909801799,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 1223,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7737626761015175,
        "trm_confidence": 0.49172227340870756,
        "mcts_value": 0.5084032169201707,
        "consensus_score": 0.5108536549658871,
        "last_agent": "hrm",
        "iteration": 5,
        "query_length": 663,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7556125818407456,
        "trm_confidence": 0.3040915450308808,
        "mcts_value": 0.18975375780077433,
        "consensus_score": 0.4756980510826412,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 2699,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7798092617864996,
        "trm_confidence": 0.18635909980374252,
        "mcts_value": 0.17316669164944853,
        "consensus_score": 0.30678597029114407,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 2572,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7674068423637621,
        "trm_confidence": 0.5650740523405162,
        "mcts_value": 0.2664256764332823,
        "consensus_score": 0.4399922192677869,
        "last_agent": "none",
        "iteration": 1,
        "query_length": 3936,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9348715822515172,
        "trm_confidence": 0.176103126668828,
        "mcts_value": 0.7659075071611872,
        "consensus_score": 0.6958268293709436,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 371,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7151607965950173,
        "trm_confidence": 0.5336049555333756,
        "mcts_value": 0.18811631230975562,
        "consensus_score": 0.5027439643807547,
        "last_agent": "trm",
        "iteration": 8,
        "query_length": 3017,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9342515063557144,
        "trm_confidence": 0.7829545983703684,
        "mcts_value": 0.6076479123657337,
        "consensus_score": 0.7631958887267172,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 944,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9234312783450263,
        "trm_confidence": 0.6756302582411827,
        "mcts_value": 0.6170405848403161,
        "consensus_score": 0.6962716823518058,
        "last_agent": "hrm",
        "iteration": 1,
        "query_length": 4935,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8508624810941059,
        "trm_confidence": 0.3793934835125877,
        "mcts_value": 0.4374857817217746,
        "consensus_score": 0.534934282952401,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 4806,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7740175160438414,
        "trm_confidence": 0.496429549704044,
        "mcts_value": 0.4569649040000525,
        "consensus_score": 0.5804220897718424,
        "last_agent": "mcts",
        "iteration": 2,
        "query_length": 1691,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9881878856886498,
        "trm_confidence": 0.6411262857398229,
        "mcts_value": 0.815232129761202,
        "consensus_score": 0.911685820052825,
        "last_agent": "hrm",
        "iteration": 3,
        "query_length": 1784,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8421798212914233,
        "trm_confidence": 0.7322559470279331,
        "mcts_value": 0.06962227405779058,
        "consensus_score": 0.5304943303184939,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 495,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9363378551916618,
        "trm_confidence": 0.3201360493631198,
        "mcts_value": 0.8195931501855913,
        "consensus_score": 0.7687283188648658,
        "last_agent": "mcts",
        "iteration": 3,
        "query_length": 1255,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8010569631499805,
        "trm_confidence": 0.4250948913284999,
        "mcts_value": 0.46163252778246694,
        "consensus_score": 0.6285138759041234,
        "last_agent": "mcts",
        "iteration": 0,
        "query_length": 2283,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9135753431765736,
        "trm_confidence": 0.458468610905465,
        "mcts_value": 0.4032120715775476,
        "consensus_score": 0.5027939998947915,
        "last_agent": "hrm",
        "iteration": 3,
        "query_length": 4093,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7092678511108096,
        "trm_confidence": 0.24153585322317822,
        "mcts_value": 0.575848466210233,
        "consensus_score": 0.41906020375707653,
        "last_agent": "hrm",
        "iteration": 8,
        "query_length": 2699,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9599826496461017,
        "trm_confidence": 0.6416055173615868,
        "mcts_value": 0.7616910796400281,
        "consensus_score": 0.7309447069688255,
        "last_agent": "trm",
        "iteration": 5,
        "query_length": 2053,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7737099228984943,
        "trm_confidence": 0.15545607553068916,
        "mcts_value": 0.018312873062874588,
        "consensus_score": 0.4075293492924431,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 548,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9231984168084972,
        "trm_confidence": 0.43762291302503004,
        "mcts_value": 0.3919526077106061,
        "consensus_score": 0.574164427406082,
        "last_agent": "none",
        "iteration": 7,
        "query_length": 1559,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9664176085044609,
        "trm_confidence": 0.7033165481696211,
        "mcts_value": 0.3028680682544538,
        "consensus_score": 0.6732559343386835,
        "last_agent": "none",
        "iteration": 1,
        "query_length": 1323,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.970867806289085,
        "trm_confidence": 0.8028508591083362,
        "mcts_value": 0.2892701046838016,
        "consensus_score": 0.6214898648369789,
        "last_agent": "hrm",
        "iteration": 2,
        "query_length": 674,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7360331890304153,
        "trm_confidence": 0.04535205026991548,
        "mcts_value": 0.6295603232822491,
        "consensus_score": 0.38512672420134136,
        "last_agent": "none",
        "iteration": 1,
        "query_length": 2644,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8740720650228829,
        "trm_confidence": 0.5994547247302315,
        "mcts_value": 0.48837136070431225,
        "consensus_score": 0.7304010642182652,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 2206,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9138249081283899,
        "trm_confidence": 0.36409936512452107,
        "mcts_value": 0.28163535888200514,
        "consensus_score": 0.5133837012016812,
        "last_agent": "none",
        "iteration": 1,
        "query_length": 2109,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7782111344007834,
        "trm_confidence": 0.1041268042075554,
        "mcts_value": 0.16707883336197832,
        "consensus_score": 0.41821962101122034,
        "last_agent": "trm",
        "iteration": 3,
        "query_length": 4765,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9569228891967217,
        "trm_confidence": 0.06662118460104391,
        "mcts_value": 0.6555669235911721,
        "consensus_score": 0.5212367590721186,
        "last_agent": "mcts",
        "iteration": 2,
        "query_length": 2268,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9729389868199702,
        "trm_confidence": 0.25753376914582776,
        "mcts_value": 0.49655720574113554,
        "consensus_score": 0.5336434766722095,
        "last_agent": "hrm",
        "iteration": 6,
        "query_length": 1430,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8831006166370539,
        "trm_confidence": 0.48829954656903063,
        "mcts_value": 0.10578106858074822,
        "consensus_score": 0.5311018378462882,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 2802,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7222211658216119,
        "trm_confidence": 0.27803258546131354,
        "mcts_value": 0.2545103167234909,
        "consensus_score": 0.3355969860641119,
        "last_agent": "mcts",
        "iteration": 2,
        "query_length": 1660,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9228718929574247,
        "trm_confidence": 0.19507172003678855,
        "mcts_value": 0.682777879225739,
        "consensus_score": 0.6094001545545434,
        "last_agent": "none",
        "iteration": 8,
        "query_length": 4805,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.849758267626667,
        "trm_confidence": 0.4646397357959349,
        "mcts_value": 0.695753635586873,
        "consensus_score": 0.6684130626190737,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 3214,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9212545411893177,
        "trm_confidence": 0.05900741430404592,
        "mcts_value": 0.4235106507204567,
        "consensus_score": 0.4642425199124821,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 1681,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8404961604082992,
        "trm_confidence": 0.7125053970548564,
        "mcts_value": 0.3336501412668026,
        "consensus_score": 0.5650607856410709,
        "last_agent": "hrm",
        "iteration": 2,
        "query_length": 728,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7001704387776833,
        "trm_confidence": 0.25924101689077567,
        "mcts_value": 0.46167665754423126,
        "consensus_score": 0.5203893262975305,
        "last_agent": "trm",
        "iteration": 0,
        "query_length": 1085,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7462849376240428,
        "trm_confidence": 0.6082034279986184,
        "mcts_value": 0.10816547835548636,
        "consensus_score": 0.5128791248655455,
        "last_agent": "none",
        "iteration": 3,
        "query_length": 1982,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8236775789392226,
        "trm_confidence": 0.549736644951093,
        "mcts_value": 0.16693731092908165,
        "consensus_score": 0.5946088821893177,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 790,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9549190574203681,
        "trm_confidence": 0.8539176146019773,
        "mcts_value": 0.19145541734296373,
        "consensus_score": 0.6241425125011274,
        "last_agent": "hrm",
        "iteration": 2,
        "query_length": 3414,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8747334224618132,
        "trm_confidence": 0.769715104649675,
        "mcts_value": 0.7672015316023072,
        "consensus_score": 0.80923799917706,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 3612,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9798800375360516,
        "trm_confidence": 0.17937289642684212,
        "mcts_value": 0.46505331630125507,
        "consensus_score": 0.6080028276870187,
        "last_agent": "none",
        "iteration": 3,
        "query_length": 4694,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9956863304536023,
        "trm_confidence": 0.8822928611113594,
        "mcts_value": 0.8115874843332822,
        "consensus_score": 0.9562118942650263,
        "last_agent": "mcts",
        "iteration": 2,
        "query_length": 198,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8641848523614624,
        "trm_confidence": 0.4285375751873198,
        "mcts_value": 0.38629759737868324,
        "consensus_score": 0.5870073976257497,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 3396,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8000649779435066,
        "trm_confidence": 0.2314953845736733,
        "mcts_value": 0.08906644020564398,
        "consensus_score": 0.31128470958005844,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 4928,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7683603359658501,
        "trm_confidence": 0.507513472723641,
        "mcts_value": 0.370259011369696,
        "consensus_score": 0.4529421423043878,
        "last_agent": "mcts",
        "iteration": 3,
        "query_length": 569,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8967618582948522,
        "trm_confidence": 0.21913424302132728,
        "mcts_value": 0.05174711649358016,
        "consensus_score": 0.30020140377612037,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 4028,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9546376339408267,
        "trm_confidence": 0.1328082263232585,
        "mcts_value": 0.5183944175395805,
        "consensus_score": 0.524217698203371,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 1919,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9408256750636647,
        "trm_confidence": 0.7509014583599367,
        "mcts_value": 0.009263096790295754,
        "consensus_score": 0.5027175785440559,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 1692,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.745011873474463,
        "trm_confidence": 0.13080224976016785,
        "mcts_value": 0.4256989638627555,
        "consensus_score": 0.4788010601485034,
        "last_agent": "none",
        "iteration": 6,
        "query_length": 4799,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8482190810985331,
        "trm_confidence": 0.22466917418298077,
        "mcts_value": 0.3570747693720694,
        "consensus_score": 0.44858760028451655,
        "last_agent": "trm",
        "iteration": 2,
        "query_length": 3672,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8056637480695372,
        "trm_confidence": 0.5020157867165614,
        "mcts_value": 0.013010889083076554,
        "consensus_score": 0.42806584393091207,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 1246,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9329945307666148,
        "trm_confidence": 0.675158490223123,
        "mcts_value": 0.5928470563406139,
        "consensus_score": 0.6542262316724003,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 3108,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7736698116428354,
        "trm_confidence": 0.028665874979164216,
        "mcts_value": 0.15062604934972768,
        "consensus_score": 0.3086327923371035,
        "last_agent": "none",
        "iteration": 4,
        "query_length": 425,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9473246628050478,
        "trm_confidence": 0.7843571444504731,
        "mcts_value": 0.1706073965477051,
        "consensus_score": 0.7200109843811864,
        "last_agent": "trm",
        "iteration": 0,
        "query_length": 1525,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9855786370242724,
        "trm_confidence": 0.010672872423246276,
        "mcts_value": 0.6757135888627713,
        "consensus_score": 0.651362013950732,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 1171,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8650970517092917,
        "trm_confidence": 0.05979703219928956,
        "mcts_value": 0.6501712098150373,
        "consensus_score": 0.5477874157363418,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 2410,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8715525929563327,
        "trm_confidence": 0.11849487677955242,
        "mcts_value": 0.5707714821884884,
        "consensus_score": 0.43527644544509964,
        "last_agent": "trm",
        "iteration": 8,
        "query_length": 925,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.8417406032693464,
        "trm_confidence": 0.3781618832841875,
        "mcts_value": 0.2315820031465227,
        "consensus_score": 0.5777198799784697,
        "last_agent": "none",
        "iteration": 8,
        "query_length": 126,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7524422158699015,
        "trm_confidence": 0.14326299569614181,
        "mcts_value": 0.3026824843410715,
        "consensus_score": 0.4879762453652392,
        "last_agent": "trm",
        "iteration": 8,
        "query_length": 498,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.9578683400256693,
        "trm_confidence": 0.3244462548711551,
        "mcts_value": 0.730665495080943,
        "consensus_score": 0.6193332829221486,
        "last_agent": "trm",
        "iteration": 6,
        "query_length": 2393,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.7773682027443383,
        "trm_confidence": 0.4693664392989025,
        "mcts_value": 0.6687741749271744,
        "consensus_score": 0.5686225964387391,
        "last_agent": "hrm",
        "iteration": 0,
        "query_length": 3358,
        "has_rag_context": false
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.801920784324709,
        "trm_confidence": 0.05000797943196162,
        "mcts_value": 0.33362689209879987,
        "consensus_score": 0.4459393545472813,
        "last_agent": "none",
        "iteration": 3,
        "query_length": 2562,
        "has_rag_context": true
      },
      "label": "hrm"
    },
    {
      "features": {
        "hrm_confidence": 0.44399049870475243,
        "trm_confidence": 0.9537098201940917,
        "mcts_value": 0.7194485575615861,
        "consensus_score": 0.6949941905822187,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 1264,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5618301748473334,
        "trm_confidence": 0.7347696595957748,
        "mcts_value": 0.3105706986357273,
        "consensus_score": 0.4640513046208595,
        "last_agent": "hrm",
        "iteration": 1,
        "query_length": 943,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.1760292584653908,
        "trm_confidence": 0.7142117520188469,
        "mcts_value": 0.08585918085299075,
        "consensus_score": 0.3135822163691997,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 2713,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.021914319138401007,
        "trm_confidence": 0.783890898730759,
        "mcts_value": 0.4201730665660401,
        "consensus_score": 0.3834030287215369,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 3801,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.15000280015235107,
        "trm_confidence": 0.7041332065882151,
        "mcts_value": 0.27581539819966233,
        "consensus_score": 0.3554648631471029,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 39,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.45055688875567235,
        "trm_confidence": 0.8591323538809237,
        "mcts_value": 0.016572915728581297,
        "consensus_score": 0.44511406149096355,
        "last_agent": "mcts",
        "iteration": 3,
        "query_length": 4273,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.49324445079443197,
        "trm_confidence": 0.730086219548575,
        "mcts_value": 0.5937959316784491,
        "consensus_score": 0.5101626314664054,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 1243,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5398238930337239,
        "trm_confidence": 0.7854422765949668,
        "mcts_value": 0.43344288561357935,
        "consensus_score": 0.6720688782016373,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 3688,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.40120532900118133,
        "trm_confidence": 0.8334813283081018,
        "mcts_value": 0.10763342161337451,
        "consensus_score": 0.41469719239302494,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 958,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6199851429898666,
        "trm_confidence": 0.8357523794573152,
        "mcts_value": 0.23831112475494545,
        "consensus_score": 0.5606425829517397,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 1462,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.033064328666376076,
        "trm_confidence": 0.7392763936439497,
        "mcts_value": 0.6043824969206342,
        "consensus_score": 0.41206962239866646,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 238,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6762765013148828,
        "trm_confidence": 0.8270965361651815,
        "mcts_value": 0.3653027473934347,
        "consensus_score": 0.717118958651632,
        "last_agent": "mcts",
        "iteration": 2,
        "query_length": 4406,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.38453742310423844,
        "trm_confidence": 0.9578592375427976,
        "mcts_value": 0.23976082519376188,
        "consensus_score": 0.5984809012971757,
        "last_agent": "mcts",
        "iteration": 4,
        "query_length": 1653,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.16652177228253073,
        "trm_confidence": 0.9929581199245436,
        "mcts_value": 0.2691381459097174,
        "consensus_score": 0.3824199327195652,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 2820,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.14578435320948174,
        "trm_confidence": 0.9040104806223674,
        "mcts_value": 0.542556525391946,
        "consensus_score": 0.4834788720457953,
        "last_agent": "trm",
        "iteration": 3,
        "query_length": 2008,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3802169571295888,
        "trm_confidence": 0.9150255250490829,
        "mcts_value": 0.2944678247370635,
        "consensus_score": 0.6004279794464684,
        "last_agent": "none",
        "iteration": 6,
        "query_length": 1336,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.2981151404941536,
        "trm_confidence": 0.9700273834431048,
        "mcts_value": 0.0339131258470358,
        "consensus_score": 0.43293859472298857,
        "last_agent": "trm",
        "iteration": 1,
        "query_length": 4764,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.10775691844490401,
        "trm_confidence": 0.9241332618062232,
        "mcts_value": 0.6650295757125843,
        "consensus_score": 0.6530213762807613,
        "last_agent": "none",
        "iteration": 6,
        "query_length": 763,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.8082211137175843,
        "trm_confidence": 0.9140759829722251,
        "mcts_value": 0.2780380965244117,
        "consensus_score": 0.5993497079866142,
        "last_agent": "none",
        "iteration": 5,
        "query_length": 4129,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.4005895980957989,
        "trm_confidence": 0.7533737464888433,
        "mcts_value": 0.009228730029908863,
        "consensus_score": 0.473264372029028,
        "last_agent": "trm",
        "iteration": 2,
        "query_length": 1301,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.15195141330764292,
        "trm_confidence": 0.732474832442457,
        "mcts_value": 0.17323875809406383,
        "consensus_score": 0.2591170011809336,
        "last_agent": "hrm",
        "iteration": 3,
        "query_length": 1372,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.27485041219425316,
        "trm_confidence": 0.8135862782303556,
        "mcts_value": 0.24191127190809028,
        "consensus_score": 0.4090340974837071,
        "last_agent": "hrm",
        "iteration": 0,
        "query_length": 3025,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.4998685572957993,
        "trm_confidence": 0.7606342900469399,
        "mcts_value": 0.6545629060364406,
        "consensus_score": 0.7240604861378592,
        "last_agent": "none",
        "iteration": 5,
        "query_length": 3997,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5374037880661997,
        "trm_confidence": 0.7943744707107527,
        "mcts_value": 0.5310275204057412,
        "consensus_score": 0.6897897312440535,
        "last_agent": "hrm",
        "iteration": 9,
        "query_length": 857,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6693614822574055,
        "trm_confidence": 0.9300626704321593,
        "mcts_value": 0.2560408179709783,
        "consensus_score": 0.7043367723547445,
        "last_agent": "hrm",
        "iteration": 1,
        "query_length": 4636,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.07751066073395808,
        "trm_confidence": 0.9310272154476609,
        "mcts_value": 0.545344697158897,
        "consensus_score": 0.4247967399658241,
        "last_agent": "trm",
        "iteration": 8,
        "query_length": 3699,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6689706808174918,
        "trm_confidence": 0.8807230759274343,
        "mcts_value": 0.07641973561121224,
        "consensus_score": 0.497222599193542,
        "last_agent": "mcts",
        "iteration": 4,
        "query_length": 4261,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.19478995496854004,
        "trm_confidence": 0.8344648262713499,
        "mcts_value": 0.2884500524339905,
        "consensus_score": 0.42425547095642147,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 933,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.369795165737947,
        "trm_confidence": 0.9681229386148816,
        "mcts_value": 0.28457740178950947,
        "consensus_score": 0.5413342452478086,
        "last_agent": "none",
        "iteration": 4,
        "query_length": 2344,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.11397331765338067,
        "trm_confidence": 0.8508037974388201,
        "mcts_value": 0.26423033765407056,
        "consensus_score": 0.4228271947902419,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 2422,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6565513876998769,
        "trm_confidence": 0.7770165284605498,
        "mcts_value": 0.5312697783633291,
        "consensus_score": 0.7484645033740798,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 838,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.12883318501989313,
        "trm_confidence": 0.8579380271905042,
        "mcts_value": 0.6201656254436124,
        "consensus_score": 0.5524780423879958,
        "last_agent": "hrm",
        "iteration": 2,
        "query_length": 838,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.35780216754323463,
        "trm_confidence": 0.7345558159942782,
        "mcts_value": 0.45754047753151095,
        "consensus_score": 0.4284025278228275,
        "last_agent": "hrm",
        "iteration": 6,
        "query_length": 911,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.31772290031448275,
        "trm_confidence": 0.7202581943254159,
        "mcts_value": 0.16037220844841238,
        "consensus_score": 0.43203599133682113,
        "last_agent": "hrm",
        "iteration": 1,
        "query_length": 2005,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.1615036406046095,
        "trm_confidence": 0.9684185745144196,
        "mcts_value": 0.2807300984894874,
        "consensus_score": 0.4877444328001141,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 2460,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.06737428867692129,
        "trm_confidence": 0.7468873965882925,
        "mcts_value": 0.646735595995737,
        "consensus_score": 0.47806951804234776,
        "last_agent": "hrm",
        "iteration": 9,
        "query_length": 1293,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.07693929077895703,
        "trm_confidence": 0.976707291563801,
        "mcts_value": 0.09996145642120899,
        "consensus_score": 0.47046245046616725,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 1728,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3196784718584861,
        "trm_confidence": 0.8065494660981616,
        "mcts_value": 0.2940388128620583,
        "consensus_score": 0.4114312934648213,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 4297,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.05450232063485843,
        "trm_confidence": 0.9115817228432954,
        "mcts_value": 0.0314498724043145,
        "consensus_score": 0.36325739980954025,
        "last_agent": "none",
        "iteration": 3,
        "query_length": 2946,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5967026785359415,
        "trm_confidence": 0.728206360650649,
        "mcts_value": 0.47022154914127223,
        "consensus_score": 0.6736303093811141,
        "last_agent": "hrm",
        "iteration": 1,
        "query_length": 4255,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.38166291205082625,
        "trm_confidence": 0.8105032492619032,
        "mcts_value": 0.21092602355647808,
        "consensus_score": 0.37982354339428354,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 2722,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.14097898988819313,
        "trm_confidence": 0.8157763874389644,
        "mcts_value": 0.18745924086930824,
        "consensus_score": 0.406795340818609,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 1230,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5867465316361682,
        "trm_confidence": 0.8601716898167119,
        "mcts_value": 0.2880046945944309,
        "consensus_score": 0.595811537193605,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 1504,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6928280035829012,
        "trm_confidence": 0.9729392729552633,
        "mcts_value": 0.051257843268573244,
        "consensus_score": 0.5581614808960973,
        "last_agent": "mcts",
        "iteration": 7,
        "query_length": 2698,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.06843046817694912,
        "trm_confidence": 0.8375128873484537,
        "mcts_value": 0.41448900915880876,
        "consensus_score": 0.3654568799861959,
        "last_agent": "trm",
        "iteration": 8,
        "query_length": 158,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.7102941809372267,
        "trm_confidence": 0.917029410513186,
        "mcts_value": 0.513869027705941,
        "consensus_score": 0.7362187389027544,
        "last_agent": "hrm",
        "iteration": 8,
        "query_length": 2958,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.03135605725763654,
        "trm_confidence": 0.9607161358831008,
        "mcts_value": 0.4787219630754682,
        "consensus_score": 0.43351858069824545,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 3755,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.13657622439010203,
        "trm_confidence": 0.8260667437097314,
        "mcts_value": 0.24615378011227454,
        "consensus_score": 0.4436877883120809,
        "last_agent": "hrm",
        "iteration": 2,
        "query_length": 3462,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6621184400767522,
        "trm_confidence": 0.9664086274682653,
        "mcts_value": 0.1731040516466299,
        "consensus_score": 0.5590060150832142,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 833,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.24092127909679925,
        "trm_confidence": 0.8831566012641474,
        "mcts_value": 0.23426183125810168,
        "consensus_score": 0.3606846276110622,
        "last_agent": "none",
        "iteration": 8,
        "query_length": 3088,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.4387600739311695,
        "trm_confidence": 0.7014572013331837,
        "mcts_value": 0.5707069543934159,
        "consensus_score": 0.6216086860980519,
        "last_agent": "none",
        "iteration": 5,
        "query_length": 1208,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.23726287854895656,
        "trm_confidence": 0.8308409267947315,
        "mcts_value": 0.3812900998175645,
        "consensus_score": 0.4693614931486405,
        "last_agent": "hrm",
        "iteration": 6,
        "query_length": 3169,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.07075106842511865,
        "trm_confidence": 0.9985028251971431,
        "mcts_value": 0.3456391869928632,
        "consensus_score": 0.4077722660250702,
        "last_agent": "mcts",
        "iteration": 3,
        "query_length": 748,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.4984535248654349,
        "trm_confidence": 0.7523611394880365,
        "mcts_value": 0.3237594897871654,
        "consensus_score": 0.6087071678236574,
        "last_agent": "mcts",
        "iteration": 0,
        "query_length": 4766,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.05787850903063726,
        "trm_confidence": 0.8209820748791032,
        "mcts_value": 0.18942332955531607,
        "consensus_score": 0.45330058643701027,
        "last_agent": "hrm",
        "iteration": 2,
        "query_length": 620,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.38450059625280597,
        "trm_confidence": 0.8720922006387127,
        "mcts_value": 0.4481518822223316,
        "consensus_score": 0.5435952820129928,
        "last_agent": "hrm",
        "iteration": 9,
        "query_length": 1542,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.502503780043091,
        "trm_confidence": 0.8285177937524397,
        "mcts_value": 0.63252569593037,
        "consensus_score": 0.6284300778466518,
        "last_agent": "trm",
        "iteration": 6,
        "query_length": 3271,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.04501556211515551,
        "trm_confidence": 0.7510957264335486,
        "mcts_value": 0.473920138132877,
        "consensus_score": 0.3809558026382885,
        "last_agent": "none",
        "iteration": 7,
        "query_length": 17,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.7175235392332914,
        "trm_confidence": 0.9587643401224977,
        "mcts_value": 0.07636883334349913,
        "consensus_score": 0.4889702055245333,
        "last_agent": "mcts",
        "iteration": 0,
        "query_length": 2725,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.367568488911526,
        "trm_confidence": 0.7704110467155687,
        "mcts_value": 0.28575882241038536,
        "consensus_score": 0.5336845579248805,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 149,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.1346253143879693,
        "trm_confidence": 0.9967275818499708,
        "mcts_value": 0.7449151342248853,
        "consensus_score": 0.5357726259417187,
        "last_agent": "hrm",
        "iteration": 5,
        "query_length": 500,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5478730718912188,
        "trm_confidence": 0.8995641896812363,
        "mcts_value": 0.21053906941647224,
        "consensus_score": 0.5363737352685999,
        "last_agent": "none",
        "iteration": 4,
        "query_length": 1756,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.2414694756928368,
        "trm_confidence": 0.8386683790431004,
        "mcts_value": 0.2256488687693059,
        "consensus_score": 0.3854328194994388,
        "last_agent": "mcts",
        "iteration": 4,
        "query_length": 213,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5188049086813686,
        "trm_confidence": 0.9616312906010721,
        "mcts_value": 0.11527457054933798,
        "consensus_score": 0.5706754567857266,
        "last_agent": "mcts",
        "iteration": 9,
        "query_length": 498,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.005559683494021258,
        "trm_confidence": 0.7451801783011791,
        "mcts_value": 0.030484982048218194,
        "consensus_score": 0.31285855094184184,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 1355,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.05924469930123762,
        "trm_confidence": 0.9244212877239678,
        "mcts_value": 0.4881669060107118,
        "consensus_score": 0.5418260208309533,
        "last_agent": "mcts",
        "iteration": 3,
        "query_length": 2810,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.7717402512901582,
        "trm_confidence": 0.8785568056376277,
        "mcts_value": 0.2005697978017373,
        "consensus_score": 0.5548538736806177,
        "last_agent": "trm",
        "iteration": 1,
        "query_length": 1961,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6171263197611457,
        "trm_confidence": 0.878933529808082,
        "mcts_value": 0.5865940232253513,
        "consensus_score": 0.7727262565572104,
        "last_agent": "hrm",
        "iteration": 8,
        "query_length": 1025,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.22804455185613456,
        "trm_confidence": 0.8667632214627283,
        "mcts_value": 0.2528673000162186,
        "consensus_score": 0.5356138982910263,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 4544,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3896818826228778,
        "trm_confidence": 0.8250737610156234,
        "mcts_value": 0.3531946095675046,
        "consensus_score": 0.44410749590674725,
        "last_agent": "hrm",
        "iteration": 2,
        "query_length": 4884,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5777041376588443,
        "trm_confidence": 0.7202122915399018,
        "mcts_value": 0.07680937377083678,
        "consensus_score": 0.39783808360657574,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 4191,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.631420389676864,
        "trm_confidence": 0.9700834298846228,
        "mcts_value": 0.05970587071489856,
        "consensus_score": 0.5921059756825924,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 1318,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6094912560302467,
        "trm_confidence": 0.9484593557468007,
        "mcts_value": 0.4973467618396553,
        "consensus_score": 0.6495917882552565,
        "last_agent": "mcts",
        "iteration": 7,
        "query_length": 960,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5157020450568663,
        "trm_confidence": 0.8762565341586622,
        "mcts_value": 0.33932349902157877,
        "consensus_score": 0.6283181880276507,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 2497,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3564181702938383,
        "trm_confidence": 0.7607215838511286,
        "mcts_value": 0.3180551149056891,
        "consensus_score": 0.4927122465848959,
        "last_agent": "none",
        "iteration": 0,
        "query_length": 2415,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.4534609618708469,
        "trm_confidence": 0.8748934786196683,
        "mcts_value": 0.509363800249543,
        "consensus_score": 0.6566159527138855,
        "last_agent": "none",
        "iteration": 7,
        "query_length": 1434,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.42324868604069527,
        "trm_confidence": 0.8874281904053775,
        "mcts_value": 0.477336157029331,
        "consensus_score": 0.5162430178424677,
        "last_agent": "none",
        "iteration": 3,
        "query_length": 4223,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5852719319442111,
        "trm_confidence": 0.9618961646205122,
        "mcts_value": 0.07792237608563882,
        "consensus_score": 0.5601518143056683,
        "last_agent": "none",
        "iteration": 2,
        "query_length": 84,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5033887527895373,
        "trm_confidence": 0.9482322557526242,
        "mcts_value": 0.5016128404374516,
        "consensus_score": 0.6187962383020338,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 4188,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.4246367125632823,
        "trm_confidence": 0.9169960992901469,
        "mcts_value": 0.3008444231972218,
        "consensus_score": 0.4575683593412612,
        "last_agent": "mcts",
        "iteration": 1,
        "query_length": 1885,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6722444983329169,
        "trm_confidence": 0.8120150629776024,
        "mcts_value": 0.35967682625370156,
        "consensus_score": 0.5616134951732532,
        "last_agent": "mcts",
        "iteration": 1,
        "query_length": 1137,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3238170141971061,
        "trm_confidence": 0.9447191906653171,
        "mcts_value": 0.6525504792475115,
        "consensus_score": 0.7009032232620207,
        "last_agent": "trm",
        "iteration": 1,
        "query_length": 4619,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.356989558313623,
        "trm_confidence": 0.9877288190113332,
        "mcts_value": 0.04522453596626121,
        "consensus_score": 0.5479748424819816,
        "last_agent": "mcts",
        "iteration": 9,
        "query_length": 3248,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.16212506707963395,
        "trm_confidence": 0.8656759361032029,
        "mcts_value": 0.6274337243421171,
        "consensus_score": 0.5514649168534099,
        "last_agent": "none",
        "iteration": 0,
        "query_length": 3140,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.7339216286610662,
        "trm_confidence": 0.9851488809408118,
        "mcts_value": 0.24536162360623354,
        "consensus_score": 0.5654757280560386,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 3298,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.48798433603059405,
        "trm_confidence": 0.8597245959138609,
        "mcts_value": 0.35144864419324634,
        "consensus_score": 0.5503389321323188,
        "last_agent": "mcts",
        "iteration": 1,
        "query_length": 1278,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5603984069163154,
        "trm_confidence": 0.7698196440028562,
        "mcts_value": 0.09768133976289063,
        "consensus_score": 0.42130619394362945,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 4975,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.36557083789533973,
        "trm_confidence": 0.7489620054153361,
        "mcts_value": 0.23887854732585437,
        "consensus_score": 0.47636898738027056,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 2177,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.788803394620058,
        "trm_confidence": 0.9740300979439191,
        "mcts_value": 0.7246475016671664,
        "consensus_score": 0.8554175148715983,
        "last_agent": "trm",
        "iteration": 6,
        "query_length": 3086,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.013303586213472012,
        "trm_confidence": 0.907148665365344,
        "mcts_value": 0.7733231465659587,
        "consensus_score": 0.5817125516747075,
        "last_agent": "none",
        "iteration": 8,
        "query_length": 725,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3077365828115015,
        "trm_confidence": 0.7037191641747388,
        "mcts_value": 0.3062003949110987,
        "consensus_score": 0.42661657162807,
        "last_agent": "trm",
        "iteration": 3,
        "query_length": 4154,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.28455970154793864,
        "trm_confidence": 0.8582634623491548,
        "mcts_value": 0.7354639078841835,
        "consensus_score": 0.5606811521471453,
        "last_agent": "trm",
        "iteration": 3,
        "query_length": 4690,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3809472177227391,
        "trm_confidence": 0.862992752870787,
        "mcts_value": 0.49116930984838786,
        "consensus_score": 0.6683701142621463,
        "last_agent": "hrm",
        "iteration": 3,
        "query_length": 2749,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.15940416182721573,
        "trm_confidence": 0.707512060901555,
        "mcts_value": 0.2817153276359186,
        "consensus_score": 0.28907614115112334,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 3080,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.4256674726587186,
        "trm_confidence": 0.9240195144514141,
        "mcts_value": 0.4976621895912099,
        "consensus_score": 0.594996109156145,
        "last_agent": "hrm",
        "iteration": 0,
        "query_length": 3633,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.7929585788152211,
        "trm_confidence": 0.8964987772604393,
        "mcts_value": 0.012587260670319693,
        "consensus_score": 0.5788649049472699,
        "last_agent": "trm",
        "iteration": 8,
        "query_length": 2383,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.4733439573332464,
        "trm_confidence": 0.863561470366792,
        "mcts_value": 0.5319923236550607,
        "consensus_score": 0.6947591670570525,
        "last_agent": "mcts",
        "iteration": 1,
        "query_length": 698,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5082465744857824,
        "trm_confidence": 0.707805221065471,
        "mcts_value": 0.04664323192977533,
        "consensus_score": 0.3598239848146607,
        "last_agent": "hrm",
        "iteration": 0,
        "query_length": 3014,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.11624612534982769,
        "trm_confidence": 0.8003077680152176,
        "mcts_value": 0.3159019436862092,
        "consensus_score": 0.391292788984855,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 3190,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.09900207687467694,
        "trm_confidence": 0.8957701755301245,
        "mcts_value": 0.2350631845562408,
        "consensus_score": 0.4926088050201367,
        "last_agent": "hrm",
        "iteration": 0,
        "query_length": 1857,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.839200706210048,
        "trm_confidence": 0.9827513857700444,
        "mcts_value": 0.3629901743729547,
        "consensus_score": 0.7617934664041085,
        "last_agent": "trm",
        "iteration": 5,
        "query_length": 3531,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.08786009173036866,
        "trm_confidence": 0.7094743992078996,
        "mcts_value": 0.3743384467839459,
        "consensus_score": 0.3370680983960174,
        "last_agent": "mcts",
        "iteration": 7,
        "query_length": 267,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.7094734398843789,
        "trm_confidence": 0.9659936674383465,
        "mcts_value": 0.4267944139338113,
        "consensus_score": 0.755332857837817,
        "last_agent": "none",
        "iteration": 2,
        "query_length": 684,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5675345712066688,
        "trm_confidence": 0.7592446155589834,
        "mcts_value": 0.5426028944636934,
        "consensus_score": 0.7035572940915822,
        "last_agent": "hrm",
        "iteration": 9,
        "query_length": 4548,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6229253486081775,
        "trm_confidence": 0.7685008694895777,
        "mcts_value": 0.381856515082727,
        "consensus_score": 0.5755932611771244,
        "last_agent": "trm",
        "iteration": 6,
        "query_length": 1734,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.16567819614037949,
        "trm_confidence": 0.7865659020882856,
        "mcts_value": 0.05659224693895597,
        "consensus_score": 0.3355917611519778,
        "last_agent": "mcts",
        "iteration": 9,
        "query_length": 2722,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3241416909846743,
        "trm_confidence": 0.856292689826732,
        "mcts_value": 0.351223555127338,
        "consensus_score": 0.452408114827305,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 1895,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.619377773100635,
        "trm_confidence": 0.9149868270181939,
        "mcts_value": 0.4685106965131719,
        "consensus_score": 0.7402103738682992,
        "last_agent": "hrm",
        "iteration": 3,
        "query_length": 4145,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6540348403488278,
        "trm_confidence": 0.9188231378797913,
        "mcts_value": 0.5441732431263956,
        "consensus_score": 0.7749456771829988,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 895,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6285553989241485,
        "trm_confidence": 0.944174705112871,
        "mcts_value": 0.38230153859435984,
        "consensus_score": 0.7370670141569622,
        "last_agent": "none",
        "iteration": 4,
        "query_length": 2051,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.14264086330260134,
        "trm_confidence": 0.7317827070972884,
        "mcts_value": 0.531137302466143,
        "consensus_score": 0.4478308355211137,
        "last_agent": "mcts",
        "iteration": 4,
        "query_length": 2832,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6023345605360149,
        "trm_confidence": 0.8474831002892362,
        "mcts_value": 0.6691200500211939,
        "consensus_score": 0.6203354566909851,
        "last_agent": "trm",
        "iteration": 6,
        "query_length": 1953,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3507914131324489,
        "trm_confidence": 0.8842845551414735,
        "mcts_value": 0.3126509759824731,
        "consensus_score": 0.5102851239046965,
        "last_agent": "hrm",
        "iteration": 6,
        "query_length": 1140,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6065421070084859,
        "trm_confidence": 0.8416433353061938,
        "mcts_value": 0.5530048631860084,
        "consensus_score": 0.6353391340031346,
        "last_agent": "trm",
        "iteration": 8,
        "query_length": 2918,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3764406418636302,
        "trm_confidence": 0.7460476616122571,
        "mcts_value": 0.18432514629470334,
        "consensus_score": 0.483438947642208,
        "last_agent": "hrm",
        "iteration": 2,
        "query_length": 3714,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.34580787111028655,
        "trm_confidence": 0.7765373968172902,
        "mcts_value": 0.6417948119039376,
        "consensus_score": 0.5643082009684584,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 1183,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.1430520289014284,
        "trm_confidence": 0.962455254641549,
        "mcts_value": 0.7069185384482377,
        "consensus_score": 0.5105718196838507,
        "last_agent": "mcts",
        "iteration": 4,
        "query_length": 4681,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.4297221008494155,
        "trm_confidence": 0.9066256638279759,
        "mcts_value": 0.3128610767072595,
        "consensus_score": 0.6121956697556076,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 118,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.19953631197384059,
        "trm_confidence": 0.9677453011628979,
        "mcts_value": 0.14785643143045163,
        "consensus_score": 0.46694010105577965,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 377,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.09414792784677961,
        "trm_confidence": 0.7749011136801479,
        "mcts_value": 0.5807953018467219,
        "consensus_score": 0.42856095587348325,
        "last_agent": "hrm",
        "iteration": 5,
        "query_length": 2635,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.33591852705329256,
        "trm_confidence": 0.7036422702766131,
        "mcts_value": 0.3092393207762463,
        "consensus_score": 0.5110895146507483,
        "last_agent": "hrm",
        "iteration": 3,
        "query_length": 3341,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.48181244318081007,
        "trm_confidence": 0.7503813941796591,
        "mcts_value": 0.5103600743829874,
        "consensus_score": 0.5081839680702516,
        "last_agent": "none",
        "iteration": 4,
        "query_length": 2594,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.17949808914376916,
        "trm_confidence": 0.7008121821027361,
        "mcts_value": 0.3334761664417101,
        "consensus_score": 0.37141425116139576,
        "last_agent": "trm",
        "iteration": 2,
        "query_length": 2971,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3466169105373836,
        "trm_confidence": 0.7544569324329081,
        "mcts_value": 0.46250204482550167,
        "consensus_score": 0.5362509531242983,
        "last_agent": "mcts",
        "iteration": 1,
        "query_length": 1516,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.49903474535210157,
        "trm_confidence": 0.7524933512462464,
        "mcts_value": 0.2352663374726087,
        "consensus_score": 0.5521504143015393,
        "last_agent": "none",
        "iteration": 2,
        "query_length": 3875,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.024672623991961514,
        "trm_confidence": 0.8491077538652446,
        "mcts_value": 0.16877198340705082,
        "consensus_score": 0.39231990419162194,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 550,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6895519670085203,
        "trm_confidence": 0.9646233128314313,
        "mcts_value": 0.7519493168716266,
        "consensus_score": 0.8417272431529607,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 2840,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.7314607128206364,
        "trm_confidence": 0.8958361673473945,
        "mcts_value": 0.07999904434156259,
        "consensus_score": 0.532251537072416,
        "last_agent": "none",
        "iteration": 0,
        "query_length": 2961,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.49541895353943594,
        "trm_confidence": 0.8335796397998163,
        "mcts_value": 0.6170285539474784,
        "consensus_score": 0.7468757653775413,
        "last_agent": "hrm",
        "iteration": 1,
        "query_length": 1504,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.042337164682284135,
        "trm_confidence": 0.966226869441021,
        "mcts_value": 0.6532135052245751,
        "consensus_score": 0.5248709106886035,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 2778,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.09166962753153295,
        "trm_confidence": 0.777922675625991,
        "mcts_value": 0.07928118116289253,
        "consensus_score": 0.2369557329818372,
        "last_agent": "none",
        "iteration": 7,
        "query_length": 1110,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.4600966895664224,
        "trm_confidence": 0.7649900314617617,
        "mcts_value": 0.4567481143485402,
        "consensus_score": 0.5684452702636409,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 2275,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.06392551939140412,
        "trm_confidence": 0.9507588913807878,
        "mcts_value": 0.6294178813732334,
        "consensus_score": 0.6457235801481944,
        "last_agent": "mcts",
        "iteration": 9,
        "query_length": 417,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.27840269980272137,
        "trm_confidence": 0.8880298192981482,
        "mcts_value": 0.5354546137596964,
        "consensus_score": 0.6213747875897874,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 58,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3684733249395634,
        "trm_confidence": 0.7183912197320327,
        "mcts_value": 0.07360348928504164,
        "consensus_score": 0.33577568669676794,
        "last_agent": "trm",
        "iteration": 3,
        "query_length": 2862,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.1085917895613392,
        "trm_confidence": 0.7551424371635501,
        "mcts_value": 0.029457983415894704,
        "consensus_score": 0.3723787239843709,
        "last_agent": "none",
        "iteration": 7,
        "query_length": 2098,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3399464330694429,
        "trm_confidence": 0.7084876916914343,
        "mcts_value": 0.17191870566237222,
        "consensus_score": 0.42612954689413557,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 551,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.4661264689151281,
        "trm_confidence": 0.7863427044884104,
        "mcts_value": 0.037099782425568675,
        "consensus_score": 0.5210284254474875,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 4645,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.1626991710875187,
        "trm_confidence": 0.839563741469072,
        "mcts_value": 0.6009076245985423,
        "consensus_score": 0.5816609984811155,
        "last_agent": "none",
        "iteration": 2,
        "query_length": 4304,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.610584907402012,
        "trm_confidence": 0.8388396813450355,
        "mcts_value": 0.5052515886628856,
        "consensus_score": 0.680192234511322,
        "last_agent": "mcts",
        "iteration": 7,
        "query_length": 2476,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.03248932695593583,
        "trm_confidence": 0.8115133779488601,
        "mcts_value": 0.5599191457885969,
        "consensus_score": 0.45457359326509383,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 3796,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3896249423489808,
        "trm_confidence": 0.7628281844378286,
        "mcts_value": 0.15414269285299517,
        "consensus_score": 0.34837371019302776,
        "last_agent": "none",
        "iteration": 8,
        "query_length": 3598,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.30799173649505523,
        "trm_confidence": 0.8140706751697913,
        "mcts_value": 0.5819996073446608,
        "consensus_score": 0.5321880521911325,
        "last_agent": "trm",
        "iteration": 8,
        "query_length": 1375,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3445939379059129,
        "trm_confidence": 0.8982604587774398,
        "mcts_value": 0.5969309674681113,
        "consensus_score": 0.6048412620007652,
        "last_agent": "mcts",
        "iteration": 4,
        "query_length": 3424,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.07248691640425747,
        "trm_confidence": 0.7953167004427881,
        "mcts_value": 0.2884535276545265,
        "consensus_score": 0.31019042439305106,
        "last_agent": "mcts",
        "iteration": 4,
        "query_length": 3260,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.0318471571811411,
        "trm_confidence": 0.9792133543616941,
        "mcts_value": 0.5463945301768476,
        "consensus_score": 0.5106987179032431,
        "last_agent": "none",
        "iteration": 0,
        "query_length": 2172,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.137782560158677,
        "trm_confidence": 0.7355455757064404,
        "mcts_value": 0.17450190345222677,
        "consensus_score": 0.32953794684888726,
        "last_agent": "hrm",
        "iteration": 5,
        "query_length": 2690,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.06068349066973206,
        "trm_confidence": 0.7485608081886223,
        "mcts_value": 0.3564966240656323,
        "consensus_score": 0.3543952817215828,
        "last_agent": "hrm",
        "iteration": 6,
        "query_length": 1661,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5716627543553109,
        "trm_confidence": 0.7072688062262754,
        "mcts_value": 0.06577214345905116,
        "consensus_score": 0.3736063176058892,
        "last_agent": "none",
        "iteration": 1,
        "query_length": 4420,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.4792595224651311,
        "trm_confidence": 0.8852980720509989,
        "mcts_value": 0.7748521953486893,
        "consensus_score": 0.6308243504765374,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 2024,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3305727687781483,
        "trm_confidence": 0.8869214510686106,
        "mcts_value": 0.3021399182018856,
        "consensus_score": 0.592793378378774,
        "last_agent": "hrm",
        "iteration": 1,
        "query_length": 2143,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.4630549898545562,
        "trm_confidence": 0.8137375183274645,
        "mcts_value": 0.6908313485013811,
        "consensus_score": 0.5673948289702623,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 1973,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.12103597238851165,
        "trm_confidence": 0.9113190636932357,
        "mcts_value": 0.4327473909675736,
        "consensus_score": 0.5230702186281557,
        "last_agent": "none",
        "iteration": 5,
        "query_length": 1658,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5752077479652343,
        "trm_confidence": 0.773235722525673,
        "mcts_value": 0.023207103765942348,
        "consensus_score": 0.49898927377377894,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 2778,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.43705698931283243,
        "trm_confidence": 0.7483137048459819,
        "mcts_value": 0.0388928802268296,
        "consensus_score": 0.4260458548997197,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 2065,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.49476042604650194,
        "trm_confidence": 0.8095554109063201,
        "mcts_value": 0.4216848270686829,
        "consensus_score": 0.5168609890926189,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 3593,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.37745678874719507,
        "trm_confidence": 0.900049985707135,
        "mcts_value": 0.2572833569043462,
        "consensus_score": 0.4287739523404067,
        "last_agent": "mcts",
        "iteration": 2,
        "query_length": 3286,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.19934137811910552,
        "trm_confidence": 0.9959877752574893,
        "mcts_value": 0.7301188829975164,
        "consensus_score": 0.5755449884856995,
        "last_agent": "trm",
        "iteration": 3,
        "query_length": 3725,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5567073453453225,
        "trm_confidence": 0.7973216992424815,
        "mcts_value": 0.667820394630638,
        "consensus_score": 0.7490087098029837,
        "last_agent": "mcts",
        "iteration": 1,
        "query_length": 477,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5885682385532721,
        "trm_confidence": 0.7443840739914134,
        "mcts_value": 0.33371270642218903,
        "consensus_score": 0.5228312239336328,
        "last_agent": "trm",
        "iteration": 1,
        "query_length": 2627,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.616400595471275,
        "trm_confidence": 0.7743840974982654,
        "mcts_value": 0.5376609440049948,
        "consensus_score": 0.736773953874989,
        "last_agent": "mcts",
        "iteration": 1,
        "query_length": 3719,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3676503385665803,
        "trm_confidence": 0.7771451422092329,
        "mcts_value": 0.2169944690565093,
        "consensus_score": 0.41062309332611235,
        "last_agent": "none",
        "iteration": 6,
        "query_length": 4963,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.48208269744353577,
        "trm_confidence": 0.8407400310916351,
        "mcts_value": 0.09494959287482052,
        "consensus_score": 0.5587096279787113,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 2008,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.565282606096974,
        "trm_confidence": 0.896171629186319,
        "mcts_value": 0.5701631380009995,
        "consensus_score": 0.7522257705070587,
        "last_agent": "none",
        "iteration": 6,
        "query_length": 4150,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.38184870645861857,
        "trm_confidence": 0.7298528928596371,
        "mcts_value": 0.041792505404248316,
        "consensus_score": 0.33275672524540734,
        "last_agent": "hrm",
        "iteration": 0,
        "query_length": 4818,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3956058611748936,
        "trm_confidence": 0.7481014374828604,
        "mcts_value": 0.5125608281675542,
        "consensus_score": 0.6164655197628086,
        "last_agent": "none",
        "iteration": 3,
        "query_length": 4570,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3636044808504141,
        "trm_confidence": 0.8689032280259077,
        "mcts_value": 0.5037191858411157,
        "consensus_score": 0.5207597230038525,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 3767,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5042486778558699,
        "trm_confidence": 0.9487663423609357,
        "mcts_value": 0.20782446818532185,
        "consensus_score": 0.6027481629327691,
        "last_agent": "none",
        "iteration": 0,
        "query_length": 1989,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.0860525064016166,
        "trm_confidence": 0.876995566672904,
        "mcts_value": 0.06177547004860453,
        "consensus_score": 0.3917751143950009,
        "last_agent": "hrm",
        "iteration": 3,
        "query_length": 1594,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.578535731544378,
        "trm_confidence": 0.9286219063090528,
        "mcts_value": 0.23085907844237044,
        "consensus_score": 0.6211278092362573,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 1251,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.1637657571093312,
        "trm_confidence": 0.7978487926010698,
        "mcts_value": 0.27442913148171366,
        "consensus_score": 0.46048109136816084,
        "last_agent": "trm",
        "iteration": 5,
        "query_length": 572,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.09808957441878606,
        "trm_confidence": 0.7942464664235027,
        "mcts_value": 0.18531774254454983,
        "consensus_score": 0.4311539880440103,
        "last_agent": "none",
        "iteration": 6,
        "query_length": 3206,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.053949052911980266,
        "trm_confidence": 0.9369717945206762,
        "mcts_value": 0.21249967947120593,
        "consensus_score": 0.3471608064460887,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 653,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.003914871612376854,
        "trm_confidence": 0.8853357809165663,
        "mcts_value": 0.1447452394219721,
        "consensus_score": 0.37177355591158734,
        "last_agent": "mcts",
        "iteration": 4,
        "query_length": 1187,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.371068074401949,
        "trm_confidence": 0.9294172657425274,
        "mcts_value": 0.6216606633028641,
        "consensus_score": 0.7357961992476911,
        "last_agent": "none",
        "iteration": 5,
        "query_length": 1580,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.2177393810241916,
        "trm_confidence": 0.9755993845183886,
        "mcts_value": 0.7762695635461,
        "consensus_score": 0.5760450510508233,
        "last_agent": "mcts",
        "iteration": 7,
        "query_length": 137,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3886885142379816,
        "trm_confidence": 0.8843818657757279,
        "mcts_value": 0.4842358700709443,
        "consensus_score": 0.5161014917841068,
        "last_agent": "mcts",
        "iteration": 2,
        "query_length": 2527,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3443884439931525,
        "trm_confidence": 0.8793338021890114,
        "mcts_value": 0.19052828786785117,
        "consensus_score": 0.5479787111122544,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 364,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.1315284405602689,
        "trm_confidence": 0.8229103229646983,
        "mcts_value": 0.5289007220143683,
        "consensus_score": 0.4841381476335501,
        "last_agent": "hrm",
        "iteration": 0,
        "query_length": 2661,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6112116300243542,
        "trm_confidence": 0.944729083478624,
        "mcts_value": 0.73414717924184,
        "consensus_score": 0.8447240816360495,
        "last_agent": "trm",
        "iteration": 3,
        "query_length": 45,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.20920003901734632,
        "trm_confidence": 0.8826423427616783,
        "mcts_value": 0.1737092300537618,
        "consensus_score": 0.34463198824139984,
        "last_agent": "none",
        "iteration": 7,
        "query_length": 1519,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.2619664334960157,
        "trm_confidence": 0.9282858938551011,
        "mcts_value": 0.17001383871513442,
        "consensus_score": 0.4441448910463415,
        "last_agent": "trm",
        "iteration": 6,
        "query_length": 4000,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.7900517104476441,
        "trm_confidence": 0.9624469462822163,
        "mcts_value": 0.7333061975991593,
        "consensus_score": 0.9215178589969355,
        "last_agent": "none",
        "iteration": 5,
        "query_length": 4511,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.24394936531507527,
        "trm_confidence": 0.7250686096445434,
        "mcts_value": 0.16107211186535897,
        "consensus_score": 0.31068481388053937,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 1115,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6754832080856815,
        "trm_confidence": 0.9878137816945693,
        "mcts_value": 0.5690404010302622,
        "consensus_score": 0.77931625562138,
        "last_agent": "trm",
        "iteration": 3,
        "query_length": 3555,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6151419235787853,
        "trm_confidence": 0.7560076354843795,
        "mcts_value": 0.1979787353524825,
        "consensus_score": 0.47408006726392415,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 2981,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.46509443130234623,
        "trm_confidence": 0.8353896755029916,
        "mcts_value": 0.11728955828719136,
        "consensus_score": 0.5072568831779007,
        "last_agent": "mcts",
        "iteration": 2,
        "query_length": 4989,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.21559775166278877,
        "trm_confidence": 0.7486539804831354,
        "mcts_value": 0.21486731881020213,
        "consensus_score": 0.3838754859924959,
        "last_agent": "none",
        "iteration": 2,
        "query_length": 3421,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.09827180114003921,
        "trm_confidence": 0.8625215714989685,
        "mcts_value": 0.525298829132584,
        "consensus_score": 0.5841534924297165,
        "last_agent": "none",
        "iteration": 4,
        "query_length": 1177,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.027041610671315285,
        "trm_confidence": 0.7294949773377593,
        "mcts_value": 0.45409808520273953,
        "consensus_score": 0.455749205666904,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 2817,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3136839190701999,
        "trm_confidence": 0.9330721164870491,
        "mcts_value": 0.2527021236687303,
        "consensus_score": 0.4344241706773192,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 361,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5423567683286818,
        "trm_confidence": 0.9481024043823998,
        "mcts_value": 0.6234234921929721,
        "consensus_score": 0.7469709789523773,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 1487,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3592651604263909,
        "trm_confidence": 0.7002473850504056,
        "mcts_value": 0.21240272293132584,
        "consensus_score": 0.3487292223594683,
        "last_agent": "hrm",
        "iteration": 2,
        "query_length": 2419,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.03534708308710173,
        "trm_confidence": 0.7210700112660109,
        "mcts_value": 0.5689164364087479,
        "consensus_score": 0.40419839345802605,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 611,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.4595328841269329,
        "trm_confidence": 0.7427713846981341,
        "mcts_value": 0.3861801905176152,
        "consensus_score": 0.5676242344019248,
        "last_agent": "none",
        "iteration": 1,
        "query_length": 342,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.37637053417760125,
        "trm_confidence": 0.7229915984352644,
        "mcts_value": 0.6146379513451925,
        "consensus_score": 0.6057302688509411,
        "last_agent": "none",
        "iteration": 7,
        "query_length": 2800,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.03228201472915102,
        "trm_confidence": 0.9034034662515389,
        "mcts_value": 0.6965553049876887,
        "consensus_score": 0.5170159200652188,
        "last_agent": "trm",
        "iteration": 0,
        "query_length": 195,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.22673522654323894,
        "trm_confidence": 0.9811045002628791,
        "mcts_value": 0.43307446871534455,
        "consensus_score": 0.4817267760636323,
        "last_agent": "none",
        "iteration": 0,
        "query_length": 4955,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.706902151582988,
        "trm_confidence": 0.9181032717884066,
        "mcts_value": 0.1172332492359439,
        "consensus_score": 0.5384099377645399,
        "last_agent": "hrm",
        "iteration": 2,
        "query_length": 3042,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5361609677126532,
        "trm_confidence": 0.8660294109660951,
        "mcts_value": 0.49629320506058805,
        "consensus_score": 0.610438570706895,
        "last_agent": "hrm",
        "iteration": 6,
        "query_length": 2899,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6022105841966343,
        "trm_confidence": 0.7462911156168053,
        "mcts_value": 0.492155691889362,
        "consensus_score": 0.6981002830659674,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 3711,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3660627178418953,
        "trm_confidence": 0.9784715944336495,
        "mcts_value": 0.0031767683220894345,
        "consensus_score": 0.36175125003510994,
        "last_agent": "none",
        "iteration": 1,
        "query_length": 4906,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.4947540701137711,
        "trm_confidence": 0.8055107172550791,
        "mcts_value": 0.5578112066415875,
        "consensus_score": 0.6374013195390709,
        "last_agent": "trm",
        "iteration": 1,
        "query_length": 4264,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.009179028716510931,
        "trm_confidence": 0.8064405757412797,
        "mcts_value": 0.19821918663560692,
        "consensus_score": 0.3040663745259947,
        "last_agent": "mcts",
        "iteration": 3,
        "query_length": 2701,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5759017228908051,
        "trm_confidence": 0.9279779192782271,
        "mcts_value": 0.11509313567569202,
        "consensus_score": 0.6012243745561644,
        "last_agent": "mcts",
        "iteration": 7,
        "query_length": 2335,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.38411527252869104,
        "trm_confidence": 0.8944281790844549,
        "mcts_value": 0.3490320696210835,
        "consensus_score": 0.46494980643922096,
        "last_agent": "none",
        "iteration": 5,
        "query_length": 1838,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.08605914289116712,
        "trm_confidence": 0.7492760161762799,
        "mcts_value": 0.6194796224353809,
        "consensus_score": 0.4419111985897207,
        "last_agent": "none",
        "iteration": 3,
        "query_length": 3590,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.48948486094393795,
        "trm_confidence": 0.9694751335534431,
        "mcts_value": 0.7991866758820104,
        "consensus_score": 0.7127005379587096,
        "last_agent": "none",
        "iteration": 1,
        "query_length": 4412,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5642706425765245,
        "trm_confidence": 0.8036610975370052,
        "mcts_value": 0.6188160411145123,
        "consensus_score": 0.6567639625907616,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 2564,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.33312734089856844,
        "trm_confidence": 0.7812761180158937,
        "mcts_value": 0.025418611170304407,
        "consensus_score": 0.440740922080869,
        "last_agent": "none",
        "iteration": 1,
        "query_length": 4051,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.03307344027948186,
        "trm_confidence": 0.869594964101483,
        "mcts_value": 0.06104188617818761,
        "consensus_score": 0.36046003504232005,
        "last_agent": "hrm",
        "iteration": 0,
        "query_length": 81,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.16002249463825832,
        "trm_confidence": 0.801608690026225,
        "mcts_value": 0.13244891523472083,
        "consensus_score": 0.41737562537958395,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 1632,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5280922234212754,
        "trm_confidence": 0.9879043657522382,
        "mcts_value": 0.5697330752523868,
        "consensus_score": 0.7372284971038164,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 2780,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.39737379960595426,
        "trm_confidence": 0.8450405954043092,
        "mcts_value": 0.2412642627233361,
        "consensus_score": 0.41561107133448993,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 2378,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.06432983471771407,
        "trm_confidence": 0.7826905153250104,
        "mcts_value": 0.13020762131912073,
        "consensus_score": 0.35348883970009126,
        "last_agent": "none",
        "iteration": 6,
        "query_length": 692,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.27174780465748594,
        "trm_confidence": 0.7270515929468473,
        "mcts_value": 0.2211892338025744,
        "consensus_score": 0.4507547077068259,
        "last_agent": "trm",
        "iteration": 6,
        "query_length": 2732,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.7209489134411255,
        "trm_confidence": 0.8460946246914985,
        "mcts_value": 0.564267053498185,
        "consensus_score": 0.709832775968962,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 670,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.2817517105607826,
        "trm_confidence": 0.982752778995942,
        "mcts_value": 0.22593093578389653,
        "consensus_score": 0.596693593524408,
        "last_agent": "none",
        "iteration": 7,
        "query_length": 1684,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.01577439131129125,
        "trm_confidence": 0.7185656431076729,
        "mcts_value": 0.37090365019669297,
        "consensus_score": 0.2827274615509427,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 4032,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.47099881769875523,
        "trm_confidence": 0.8638162200244239,
        "mcts_value": 0.22295367841930777,
        "consensus_score": 0.43205074719581693,
        "last_agent": "none",
        "iteration": 8,
        "query_length": 4342,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.1947221200387112,
        "trm_confidence": 0.7256672923055179,
        "mcts_value": 0.007799355637569309,
        "consensus_score": 0.296203379145196,
        "last_agent": "mcts",
        "iteration": 7,
        "query_length": 717,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6616562017705362,
        "trm_confidence": 0.8812921911396764,
        "mcts_value": 0.31759699748431286,
        "consensus_score": 0.6899362138102705,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 2900,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.2428662903497912,
        "trm_confidence": 0.9832436939981397,
        "mcts_value": 0.023142295682260235,
        "consensus_score": 0.36495223526773574,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 945,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6688805237376906,
        "trm_confidence": 0.8574002528330715,
        "mcts_value": 0.22141704712565802,
        "consensus_score": 0.5867200628239739,
        "last_agent": "none",
        "iteration": 3,
        "query_length": 3712,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.0670001734640036,
        "trm_confidence": 0.7472804539865435,
        "mcts_value": 0.3255789958687113,
        "consensus_score": 0.4143307249261845,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 694,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3267956106620368,
        "trm_confidence": 0.9321549283957002,
        "mcts_value": 0.027350993797126976,
        "consensus_score": 0.3519870391985993,
        "last_agent": "mcts",
        "iteration": 0,
        "query_length": 126,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.26675615856256124,
        "trm_confidence": 0.8486474023806047,
        "mcts_value": 0.4196913407843851,
        "consensus_score": 0.5078201302744079,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 879,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.24076419485499137,
        "trm_confidence": 0.7954192441326038,
        "mcts_value": 0.4069230558316348,
        "consensus_score": 0.500321991851245,
        "last_agent": "none",
        "iteration": 3,
        "query_length": 3450,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5232461696556046,
        "trm_confidence": 0.9082135039558441,
        "mcts_value": 0.10914988406229703,
        "consensus_score": 0.5017448951596606,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 2644,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.13977011727098437,
        "trm_confidence": 0.9565491839452641,
        "mcts_value": 0.434483402192434,
        "consensus_score": 0.4951348535874965,
        "last_agent": "trm",
        "iteration": 0,
        "query_length": 3296,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.7027971798793406,
        "trm_confidence": 0.8364775682002916,
        "mcts_value": 0.2111804508894396,
        "consensus_score": 0.6567402210062185,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 3329,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.4528137893677984,
        "trm_confidence": 0.7608033363603153,
        "mcts_value": 0.36276653795130237,
        "consensus_score": 0.6143671543969972,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 722,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.4481637236190501,
        "trm_confidence": 0.741486100448232,
        "mcts_value": 0.6248777808889668,
        "consensus_score": 0.5105112890996961,
        "last_agent": "none",
        "iteration": 8,
        "query_length": 4318,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.06857363256335977,
        "trm_confidence": 0.7224344602682048,
        "mcts_value": 0.4306979431022975,
        "consensus_score": 0.436828020490445,
        "last_agent": "mcts",
        "iteration": 0,
        "query_length": 2642,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6867730459739563,
        "trm_confidence": 0.8570271543534677,
        "mcts_value": 0.6646629381862011,
        "consensus_score": 0.7986559333145749,
        "last_agent": "hrm",
        "iteration": 5,
        "query_length": 1851,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.07578925298603742,
        "trm_confidence": 0.8517926511733398,
        "mcts_value": 0.4953156926761809,
        "consensus_score": 0.4010946427909876,
        "last_agent": "trm",
        "iteration": 5,
        "query_length": 2697,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.16010157541766123,
        "trm_confidence": 0.7898537470823938,
        "mcts_value": 0.42773108304892316,
        "consensus_score": 0.516586144254161,
        "last_agent": "trm",
        "iteration": 0,
        "query_length": 333,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.29374569321155797,
        "trm_confidence": 0.9582303084863814,
        "mcts_value": 0.5655772628314892,
        "consensus_score": 0.586373694517166,
        "last_agent": "hrm",
        "iteration": 1,
        "query_length": 1770,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5498426640572068,
        "trm_confidence": 0.725907902350479,
        "mcts_value": 0.578520264476097,
        "consensus_score": 0.6824129043635416,
        "last_agent": "trm",
        "iteration": 6,
        "query_length": 905,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.22196676364259524,
        "trm_confidence": 0.995382212980279,
        "mcts_value": 0.06800281593644064,
        "consensus_score": 0.4408387822400072,
        "last_agent": "mcts",
        "iteration": 3,
        "query_length": 2222,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.35901101594477813,
        "trm_confidence": 0.7021256062509844,
        "mcts_value": 0.15898145305339606,
        "consensus_score": 0.3442777347215891,
        "last_agent": "none",
        "iteration": 0,
        "query_length": 4410,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.2907794325942135,
        "trm_confidence": 0.9636906617159372,
        "mcts_value": 0.5865100538924193,
        "consensus_score": 0.6649694420812382,
        "last_agent": "hrm",
        "iteration": 3,
        "query_length": 4692,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.45162383408007106,
        "trm_confidence": 0.9075638979000847,
        "mcts_value": 0.0008014781372515946,
        "consensus_score": 0.3603478898903641,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 3176,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.7951723642940054,
        "trm_confidence": 0.981269577631884,
        "mcts_value": 0.018015779119747756,
        "consensus_score": 0.6565220316400938,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 2362,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.17350022601120896,
        "trm_confidence": 0.7020147352490155,
        "mcts_value": 0.45230798396929145,
        "consensus_score": 0.5012118061301566,
        "last_agent": "none",
        "iteration": 6,
        "query_length": 1952,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6700725888611195,
        "trm_confidence": 0.7931087773863589,
        "mcts_value": 0.30442181656221784,
        "consensus_score": 0.5786261709853329,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 2346,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.33108310202008395,
        "trm_confidence": 0.8658014693792242,
        "mcts_value": 0.6330995001890818,
        "consensus_score": 0.6918048944891201,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 1477,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5166902976733988,
        "trm_confidence": 0.8326445672393485,
        "mcts_value": 0.614193224751526,
        "consensus_score": 0.6671058116347,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 1003,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5990841615461594,
        "trm_confidence": 0.902918133015186,
        "mcts_value": 0.16752512087819482,
        "consensus_score": 0.5561840478842063,
        "last_agent": "hrm",
        "iteration": 6,
        "query_length": 4964,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5226621376728493,
        "trm_confidence": 0.9567887526441046,
        "mcts_value": 0.11064255655673892,
        "consensus_score": 0.5933025634730631,
        "last_agent": "none",
        "iteration": 7,
        "query_length": 3743,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6129866001129773,
        "trm_confidence": 0.8176127385585569,
        "mcts_value": 0.2971114670527823,
        "consensus_score": 0.5900054324972238,
        "last_agent": "hrm",
        "iteration": 0,
        "query_length": 3059,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.4624508846282009,
        "trm_confidence": 0.915452737638381,
        "mcts_value": 0.40492024675669264,
        "consensus_score": 0.6562619177500998,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 2326,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6278566491726219,
        "trm_confidence": 0.9877220014239958,
        "mcts_value": 0.8770151232588164,
        "consensus_score": 0.9251168545389185,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 4461,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.134780155862351,
        "trm_confidence": 0.8828288105490902,
        "mcts_value": 0.42535114381200245,
        "consensus_score": 0.53406200788668,
        "last_agent": "trm",
        "iteration": 5,
        "query_length": 3053,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.24109591575028186,
        "trm_confidence": 0.7328967697575094,
        "mcts_value": 0.320223246140162,
        "consensus_score": 0.4178703697231394,
        "last_agent": "trm",
        "iteration": 1,
        "query_length": 4663,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.41234914508910964,
        "trm_confidence": 0.9512851878566462,
        "mcts_value": 0.041476680211027434,
        "consensus_score": 0.3716685895144577,
        "last_agent": "none",
        "iteration": 5,
        "query_length": 28,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.645435040183728,
        "trm_confidence": 0.9115377837997666,
        "mcts_value": 0.1852714048300609,
        "consensus_score": 0.5991834652475413,
        "last_agent": "mcts",
        "iteration": 1,
        "query_length": 1162,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.47153643452004507,
        "trm_confidence": 0.96663864825443,
        "mcts_value": 0.7677453812636504,
        "consensus_score": 0.6803563821819433,
        "last_agent": "none",
        "iteration": 3,
        "query_length": 3270,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.7422740536411904,
        "trm_confidence": 0.8440240225583628,
        "mcts_value": 0.08515006658387513,
        "consensus_score": 0.6275423880481276,
        "last_agent": "hrm",
        "iteration": 9,
        "query_length": 3237,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.36512558615465096,
        "trm_confidence": 0.8032111504777462,
        "mcts_value": 0.2056665511746096,
        "consensus_score": 0.5178225856236568,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 2685,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.4046702589584559,
        "trm_confidence": 0.8868074033068132,
        "mcts_value": 0.5594746319742098,
        "consensus_score": 0.6177706909658732,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 4443,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.7244777124892807,
        "trm_confidence": 0.9920046051237716,
        "mcts_value": 0.7677781225671295,
        "consensus_score": 0.8246683798676082,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 1130,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.1163748177424833,
        "trm_confidence": 0.9285844535692526,
        "mcts_value": 0.2544184440051873,
        "consensus_score": 0.42227243777758766,
        "last_agent": "hrm",
        "iteration": 3,
        "query_length": 2995,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.21715587039051465,
        "trm_confidence": 0.9460314234351094,
        "mcts_value": 0.6305312909373195,
        "consensus_score": 0.5126486087433115,
        "last_agent": "none",
        "iteration": 8,
        "query_length": 3133,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5205522309687037,
        "trm_confidence": 0.8205096326507446,
        "mcts_value": 0.6593660351069702,
        "consensus_score": 0.7269518942863782,
        "last_agent": "mcts",
        "iteration": 4,
        "query_length": 579,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.24396749425820094,
        "trm_confidence": 0.9338072634357416,
        "mcts_value": 0.6702832678201668,
        "consensus_score": 0.5816335703502248,
        "last_agent": "none",
        "iteration": 8,
        "query_length": 2354,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6118934963390785,
        "trm_confidence": 0.713319189073177,
        "mcts_value": 0.5332547377561968,
        "consensus_score": 0.6960586417493916,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 1201,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.08920373448121506,
        "trm_confidence": 0.7860506664040456,
        "mcts_value": 0.3438548600718579,
        "consensus_score": 0.35037941774769443,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 3864,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.15847387327825596,
        "trm_confidence": 0.7429590213229365,
        "mcts_value": 0.12792101257159041,
        "consensus_score": 0.28827365637353636,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 2006,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6984138608003901,
        "trm_confidence": 0.884507314571012,
        "mcts_value": 0.4727101899991634,
        "consensus_score": 0.724117013380798,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 1770,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6713096838839204,
        "trm_confidence": 0.7961752187349718,
        "mcts_value": 0.43288967658355837,
        "consensus_score": 0.5889604169542866,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 747,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.4488211404725031,
        "trm_confidence": 0.9005330712117321,
        "mcts_value": 0.7086021561543003,
        "consensus_score": 0.7204733703895366,
        "last_agent": "mcts",
        "iteration": 7,
        "query_length": 724,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6098893298605026,
        "trm_confidence": 0.7319166049037223,
        "mcts_value": 0.3784285480519917,
        "consensus_score": 0.5565481761013581,
        "last_agent": "hrm",
        "iteration": 6,
        "query_length": 1530,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.18342038240570238,
        "trm_confidence": 0.946638182171232,
        "mcts_value": 0.6811593682621512,
        "consensus_score": 0.610614999599084,
        "last_agent": "trm",
        "iteration": 3,
        "query_length": 3571,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.14906470567047758,
        "trm_confidence": 0.9447500065979817,
        "mcts_value": 0.043762371778592296,
        "consensus_score": 0.43036790751419335,
        "last_agent": "hrm",
        "iteration": 3,
        "query_length": 1191,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5791896333506865,
        "trm_confidence": 0.7741202761842156,
        "mcts_value": 0.20043586225665735,
        "consensus_score": 0.4599537712140579,
        "last_agent": "mcts",
        "iteration": 3,
        "query_length": 2706,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.05766527097728564,
        "trm_confidence": 0.7952965662415622,
        "mcts_value": 0.11022538489293426,
        "consensus_score": 0.3044729840498236,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 2683,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.6029620503797231,
        "trm_confidence": 0.7638024120280091,
        "mcts_value": 0.5907938874476786,
        "consensus_score": 0.6884096166109919,
        "last_agent": "hrm",
        "iteration": 2,
        "query_length": 1257,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.08530128330633818,
        "trm_confidence": 0.7633662983313895,
        "mcts_value": 0.2520781283405512,
        "consensus_score": 0.3820795576734962,
        "last_agent": "mcts",
        "iteration": 3,
        "query_length": 4228,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.23824827363124398,
        "trm_confidence": 0.7888764206137916,
        "mcts_value": 0.5446763936859894,
        "consensus_score": 0.5995966421557799,
        "last_agent": "hrm",
        "iteration": 3,
        "query_length": 4627,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.03645149074883067,
        "trm_confidence": 0.7834486851676804,
        "mcts_value": 0.5042875204423015,
        "consensus_score": 0.375596595433501,
        "last_agent": "mcts",
        "iteration": 4,
        "query_length": 3641,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.06196604299688403,
        "trm_confidence": 0.7553414752537317,
        "mcts_value": 0.228067362442038,
        "consensus_score": 0.43891547507428164,
        "last_agent": "trm",
        "iteration": 6,
        "query_length": 1758,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3671711066549766,
        "trm_confidence": 0.7116732462322645,
        "mcts_value": 0.05643153072067,
        "consensus_score": 0.3869341087454756,
        "last_agent": "none",
        "iteration": 8,
        "query_length": 4955,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.03268675557397803,
        "trm_confidence": 0.7898234484449573,
        "mcts_value": 0.017092359289111358,
        "consensus_score": 0.2658728350317865,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 447,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.52843234883557,
        "trm_confidence": 0.8722166517345464,
        "mcts_value": 0.046973836607357426,
        "consensus_score": 0.4131610208865,
        "last_agent": "trm",
        "iteration": 3,
        "query_length": 1212,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.537218265182701,
        "trm_confidence": 0.8317105336684291,
        "mcts_value": 0.6348304972281507,
        "consensus_score": 0.7195942206794689,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 2895,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.7122174030713968,
        "trm_confidence": 0.9556537585022351,
        "mcts_value": 0.6248546442004216,
        "consensus_score": 0.824602298589196,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 2946,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.4075676356567753,
        "trm_confidence": 0.8914861071737631,
        "mcts_value": 0.6735839520535416,
        "consensus_score": 0.6438178188689948,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 3894,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.22118242784204983,
        "trm_confidence": 0.9885658304105771,
        "mcts_value": 0.6305436490773224,
        "consensus_score": 0.6648145353061471,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 1067,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.44713971051353746,
        "trm_confidence": 0.8720577585125344,
        "mcts_value": 0.1291584908205778,
        "consensus_score": 0.5777778604336374,
        "last_agent": "none",
        "iteration": 2,
        "query_length": 2380,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3700644311505463,
        "trm_confidence": 0.9387385014415938,
        "mcts_value": 0.17713630425104762,
        "consensus_score": 0.43310515579166653,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 4026,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.44960067505898793,
        "trm_confidence": 0.7270410946114799,
        "mcts_value": 0.15686753079097213,
        "consensus_score": 0.5197820503039892,
        "last_agent": "mcts",
        "iteration": 3,
        "query_length": 2904,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3199410028080757,
        "trm_confidence": 0.7254575330664803,
        "mcts_value": 0.28191721973716205,
        "consensus_score": 0.4927010141245345,
        "last_agent": "trm",
        "iteration": 5,
        "query_length": 1870,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.1887215862478396,
        "trm_confidence": 0.7784964746898831,
        "mcts_value": 0.6295415304385765,
        "consensus_score": 0.46526891754299554,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 2389,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.15580876886389092,
        "trm_confidence": 0.7265170706252492,
        "mcts_value": 0.03659760596095116,
        "consensus_score": 0.3643137689828136,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 1120,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5036437005204956,
        "trm_confidence": 0.8986148813894688,
        "mcts_value": 0.7505958370597642,
        "consensus_score": 0.8098702430062522,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 3312,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.41157849497836363,
        "trm_confidence": 0.9377275301005535,
        "mcts_value": 0.5901441016930987,
        "consensus_score": 0.5621351771273859,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 531,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5229840904691805,
        "trm_confidence": 0.7254924259377215,
        "mcts_value": 0.5835719402202927,
        "consensus_score": 0.5559205968361439,
        "last_agent": "hrm",
        "iteration": 5,
        "query_length": 3444,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.1135682279040384,
        "trm_confidence": 0.9880678013568904,
        "mcts_value": 0.3099394152532933,
        "consensus_score": 0.5101540764015787,
        "last_agent": "none",
        "iteration": 2,
        "query_length": 3436,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.0931857133634963,
        "trm_confidence": 0.7069873180014079,
        "mcts_value": 0.34132687164971404,
        "consensus_score": 0.41111798093237595,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 2463,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.19816711313021182,
        "trm_confidence": 0.8245844241216375,
        "mcts_value": 0.01128953256772494,
        "consensus_score": 0.28600632181700203,
        "last_agent": "none",
        "iteration": 3,
        "query_length": 2538,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.09542007485860418,
        "trm_confidence": 0.7060433414489806,
        "mcts_value": 0.4036851991817658,
        "consensus_score": 0.42763328845763604,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 1285,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.10596334580580775,
        "trm_confidence": 0.8703268350523573,
        "mcts_value": 0.4345196590148425,
        "consensus_score": 0.38258895880717203,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 1557,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.2038095686689596,
        "trm_confidence": 0.8283441428088825,
        "mcts_value": 0.49952058948680605,
        "consensus_score": 0.5787469937813052,
        "last_agent": "trm",
        "iteration": 8,
        "query_length": 2092,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.023670709999474536,
        "trm_confidence": 0.861503439799904,
        "mcts_value": 0.5505167267521973,
        "consensus_score": 0.48756021343289146,
        "last_agent": "hrm",
        "iteration": 2,
        "query_length": 656,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5498646131944848,
        "trm_confidence": 0.9806820359409224,
        "mcts_value": 0.41912753188379326,
        "consensus_score": 0.6740757416147484,
        "last_agent": "mcts",
        "iteration": 4,
        "query_length": 2834,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.30222002578153945,
        "trm_confidence": 0.7870927770123011,
        "mcts_value": 0.4710654708838269,
        "consensus_score": 0.428114476243183,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 2188,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.517244726214594,
        "trm_confidence": 0.9047679380297499,
        "mcts_value": 0.5252141294078266,
        "consensus_score": 0.6707013811352175,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 4448,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5782538078279272,
        "trm_confidence": 0.9196957346546054,
        "mcts_value": 0.19366752533222478,
        "consensus_score": 0.5205335237756176,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 2537,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.27759100049047153,
        "trm_confidence": 0.8100229296789879,
        "mcts_value": 0.42875768468675873,
        "consensus_score": 0.449435213499085,
        "last_agent": "hrm",
        "iteration": 8,
        "query_length": 731,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.06612097058965101,
        "trm_confidence": 0.9118732335049741,
        "mcts_value": 0.3815523241631893,
        "consensus_score": 0.35882608252140374,
        "last_agent": "mcts",
        "iteration": 7,
        "query_length": 50,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.2867772756099233,
        "trm_confidence": 0.9847758442317486,
        "mcts_value": 0.4988418394014734,
        "consensus_score": 0.5416091682199714,
        "last_agent": "none",
        "iteration": 6,
        "query_length": 3350,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.14486151824914825,
        "trm_confidence": 0.8111884784516898,
        "mcts_value": 0.07931269493394683,
        "consensus_score": 0.2690967240029296,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 3213,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5723089366326094,
        "trm_confidence": 0.7005789123620018,
        "mcts_value": 0.011999521163016088,
        "consensus_score": 0.4120130151634491,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 2620,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.37732918644693036,
        "trm_confidence": 0.9138658874296303,
        "mcts_value": 0.7930651348065357,
        "consensus_score": 0.6026172795010194,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 441,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.22386318014259662,
        "trm_confidence": 0.9786790051379077,
        "mcts_value": 0.1251749172286063,
        "consensus_score": 0.47523542337757685,
        "last_agent": "hrm",
        "iteration": 5,
        "query_length": 3564,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5440501535428177,
        "trm_confidence": 0.9773857535449653,
        "mcts_value": 0.2729229773638274,
        "consensus_score": 0.578132939434112,
        "last_agent": "mcts",
        "iteration": 1,
        "query_length": 2733,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.5210214009226612,
        "trm_confidence": 0.7430050210628615,
        "mcts_value": 0.4532371791200645,
        "consensus_score": 0.5784902038906994,
        "last_agent": "none",
        "iteration": 4,
        "query_length": 2534,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.7436476874790247,
        "trm_confidence": 0.9438716043087283,
        "mcts_value": 0.46161146207938814,
        "consensus_score": 0.7803539423196538,
        "last_agent": "none",
        "iteration": 5,
        "query_length": 159,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3321384555464614,
        "trm_confidence": 0.9596295566866301,
        "mcts_value": 0.34033824077037494,
        "consensus_score": 0.5050176867248054,
        "last_agent": "hrm",
        "iteration": 5,
        "query_length": 4605,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.34360040070387166,
        "trm_confidence": 0.898138343593137,
        "mcts_value": 0.1738642953167907,
        "consensus_score": 0.5452360923540416,
        "last_agent": "none",
        "iteration": 3,
        "query_length": 1367,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3708734210569285,
        "trm_confidence": 0.7622059999079367,
        "mcts_value": 0.35534973457636193,
        "consensus_score": 0.5618726214324747,
        "last_agent": "none",
        "iteration": 1,
        "query_length": 485,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.4451081419245728,
        "trm_confidence": 0.8594284516516211,
        "mcts_value": 0.38743479820397003,
        "consensus_score": 0.47119544529397467,
        "last_agent": "trm",
        "iteration": 5,
        "query_length": 356,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.4561927932840184,
        "trm_confidence": 0.8028064541133494,
        "mcts_value": 0.34978475478808896,
        "consensus_score": 0.45490497659531093,
        "last_agent": "hrm",
        "iteration": 0,
        "query_length": 4638,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.399815127929161,
        "trm_confidence": 0.7572655439716671,
        "mcts_value": 0.11687132043692414,
        "consensus_score": 0.5009945487675382,
        "last_agent": "hrm",
        "iteration": 2,
        "query_length": 3720,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.2547861301271446,
        "trm_confidence": 0.9627238049078941,
        "mcts_value": 0.3670222299040831,
        "consensus_score": 0.6199826651155291,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 3757,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.3682753491268558,
        "trm_confidence": 0.7206816855197664,
        "mcts_value": 0.6139025598324387,
        "consensus_score": 0.6570031105111498,
        "last_agent": "trm",
        "iteration": 1,
        "query_length": 301,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.09011705955441619,
        "trm_confidence": 0.7946610237708518,
        "mcts_value": 0.4826493062063232,
        "consensus_score": 0.4105788027159111,
        "last_agent": "mcts",
        "iteration": 3,
        "query_length": 1686,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.11852945767324931,
        "trm_confidence": 0.8493902676335996,
        "mcts_value": 0.3238043581681488,
        "consensus_score": 0.3592431389123262,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 2869,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.14902826731190888,
        "trm_confidence": 0.8565661461551899,
        "mcts_value": 0.057772755366731966,
        "consensus_score": 0.29491744671302983,
        "last_agent": "mcts",
        "iteration": 0,
        "query_length": 3267,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.48758750266237133,
        "trm_confidence": 0.848045247328965,
        "mcts_value": 0.06533413356703832,
        "consensus_score": 0.4551926859669786,
        "last_agent": "none",
        "iteration": 6,
        "query_length": 1058,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.0511193831067477,
        "trm_confidence": 0.8116235234548076,
        "mcts_value": 0.15665045493738797,
        "consensus_score": 0.3331483723501024,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 2360,
        "has_rag_context": false
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.37819121088269464,
        "trm_confidence": 0.9938865043372864,
        "mcts_value": 0.07211910142969438,
        "consensus_score": 0.4461136263330919,
        "last_agent": "none",
        "iteration": 2,
        "query_length": 1219,
        "has_rag_context": true
      },
      "label": "trm"
    },
    {
      "features": {
        "hrm_confidence": 0.31723435453434623,
        "trm_confidence": 0.23776372466300788,
        "mcts_value": 0.6380934441153502,
        "consensus_score": 0.33376791876446466,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 1345,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.1686840121841821,
        "trm_confidence": 0.3768754430606373,
        "mcts_value": 0.9407471422986178,
        "consensus_score": 0.4322633759443156,
        "last_agent": "none",
        "iteration": 7,
        "query_length": 3990,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.42855636392941093,
        "trm_confidence": 0.1846502295230367,
        "mcts_value": 0.6646314497486088,
        "consensus_score": 0.36041945971320166,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 2756,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.3837846892832857,
        "trm_confidence": 0.17542621208214185,
        "mcts_value": 0.8260592450783548,
        "consensus_score": 0.4523110645038243,
        "last_agent": "mcts",
        "iteration": 4,
        "query_length": 1628,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.13536975187858913,
        "trm_confidence": 0.22041493479425212,
        "mcts_value": 0.6886788593507293,
        "consensus_score": 0.4286284993938186,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 1277,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.4640450577224613,
        "trm_confidence": 0.3360216995983519,
        "mcts_value": 0.8607849384132342,
        "consensus_score": 0.6346544965847443,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 908,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.3822976453806744,
        "trm_confidence": 0.08681387071692682,
        "mcts_value": 0.856150003967066,
        "consensus_score": 0.42029892280404163,
        "last_agent": "trm",
        "iteration": 6,
        "query_length": 4107,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5750886628472248,
        "trm_confidence": 0.30967271573066185,
        "mcts_value": 0.7515370227892435,
        "consensus_score": 0.5961329131448321,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 3365,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.18185877193841446,
        "trm_confidence": 0.1690427244553648,
        "mcts_value": 0.7922239563195673,
        "consensus_score": 0.33405727528610984,
        "last_agent": "hrm",
        "iteration": 6,
        "query_length": 424,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.4683262010865425,
        "trm_confidence": 0.3830163896608705,
        "mcts_value": 0.9798412171603175,
        "consensus_score": 0.6001141642393371,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 1590,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6673941740574011,
        "trm_confidence": 0.27658075424446743,
        "mcts_value": 0.7172150697731274,
        "consensus_score": 0.6137456131966376,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 3704,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.41029586862175155,
        "trm_confidence": 0.08314398616690488,
        "mcts_value": 0.6646202905804774,
        "consensus_score": 0.44384845643016324,
        "last_agent": "trm",
        "iteration": 5,
        "query_length": 3689,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.12670689546912067,
        "trm_confidence": 0.2750004870208167,
        "mcts_value": 0.9640358696520795,
        "consensus_score": 0.48467150832871886,
        "last_agent": "trm",
        "iteration": 6,
        "query_length": 4998,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.284618700591374,
        "trm_confidence": 0.45444631860734425,
        "mcts_value": 0.6116286456675187,
        "consensus_score": 0.4650280632525767,
        "last_agent": "none",
        "iteration": 5,
        "query_length": 4661,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.1919416816053799,
        "trm_confidence": 0.5104898707022725,
        "mcts_value": 0.7477884184725011,
        "consensus_score": 0.4453370629321546,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 4404,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.4319160371123925,
        "trm_confidence": 0.05127219031728598,
        "mcts_value": 0.815710392901433,
        "consensus_score": 0.36859303552538447,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 3731,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.13051870523620435,
        "trm_confidence": 0.1534902256461637,
        "mcts_value": 0.934368432366737,
        "consensus_score": 0.37076724564825825,
        "last_agent": "none",
        "iteration": 8,
        "query_length": 750,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6336349293494238,
        "trm_confidence": 0.20652809834757668,
        "mcts_value": 0.874701583416583,
        "consensus_score": 0.6033733348792163,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 2360,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.06683838837145224,
        "trm_confidence": 0.49840071878243597,
        "mcts_value": 0.9629660746374693,
        "consensus_score": 0.5549764698683062,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 2452,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5794184689020143,
        "trm_confidence": 0.5395671972826748,
        "mcts_value": 0.7286489352876984,
        "consensus_score": 0.6159413233746899,
        "last_agent": "hrm",
        "iteration": 9,
        "query_length": 1597,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6056284183515128,
        "trm_confidence": 0.22258915323155873,
        "mcts_value": 0.8358025321613871,
        "consensus_score": 0.5513833644009682,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 2747,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.26967294262331243,
        "trm_confidence": 0.6975536817441912,
        "mcts_value": 0.9548156081322211,
        "consensus_score": 0.5711938651033375,
        "last_agent": "none",
        "iteration": 7,
        "query_length": 4031,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6360324407605084,
        "trm_confidence": 0.5117975638356418,
        "mcts_value": 0.9193728763200538,
        "consensus_score": 0.7144308741981286,
        "last_agent": "none",
        "iteration": 6,
        "query_length": 716,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.014257931579064364,
        "trm_confidence": 0.35213206337937436,
        "mcts_value": 0.8037201252899288,
        "consensus_score": 0.4612308667842808,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 869,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.3702812658390706,
        "trm_confidence": 0.02343849829797033,
        "mcts_value": 0.7561939959074084,
        "consensus_score": 0.3599041751153518,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 2137,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.4832856761452664,
        "trm_confidence": 0.30456142559419025,
        "mcts_value": 0.6124963915962599,
        "consensus_score": 0.4499168030676857,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 2461,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.3317891311203637,
        "trm_confidence": 0.4066099439054674,
        "mcts_value": 0.7752107125082925,
        "consensus_score": 0.40580152529265495,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 194,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.14516101266756698,
        "trm_confidence": 0.3666476585589127,
        "mcts_value": 0.7803698989263979,
        "consensus_score": 0.38677221181038673,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 1590,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.17421580970351525,
        "trm_confidence": 0.013584305345812252,
        "mcts_value": 0.8555022223073749,
        "consensus_score": 0.3285983529082227,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 3200,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.4947874216922198,
        "trm_confidence": 0.6969688604377766,
        "mcts_value": 0.984805745527544,
        "consensus_score": 0.6845326939486411,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 651,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6900972403096294,
        "trm_confidence": 0.45779103982578184,
        "mcts_value": 0.8447041417793998,
        "consensus_score": 0.7439608524389874,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 3863,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.33902091207983737,
        "trm_confidence": 0.6431585950876396,
        "mcts_value": 0.879769484128925,
        "consensus_score": 0.6820165532035585,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 4575,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.34527141232123215,
        "trm_confidence": 0.05654023144514711,
        "mcts_value": 0.6278444007194423,
        "consensus_score": 0.3879543488102126,
        "last_agent": "mcts",
        "iteration": 7,
        "query_length": 3040,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.3633985489127985,
        "trm_confidence": 0.08223578013738105,
        "mcts_value": 0.7119183154685265,
        "consensus_score": 0.4627728820362914,
        "last_agent": "trm",
        "iteration": 5,
        "query_length": 4476,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.564362042106087,
        "trm_confidence": 0.07407016937927796,
        "mcts_value": 0.6820202394092734,
        "consensus_score": 0.4630363159325126,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 4580,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.043136798864328404,
        "trm_confidence": 0.3141644686405523,
        "mcts_value": 0.6760243546432525,
        "consensus_score": 0.284938943098287,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 1201,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.07943945317119827,
        "trm_confidence": 0.3972852005837342,
        "mcts_value": 0.7797415764132528,
        "consensus_score": 0.35592691464910986,
        "last_agent": "none",
        "iteration": 8,
        "query_length": 1697,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.47916888501006816,
        "trm_confidence": 0.12721069746834832,
        "mcts_value": 0.9903578369125293,
        "consensus_score": 0.43506473555601977,
        "last_agent": "trm",
        "iteration": 6,
        "query_length": 2501,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6109251809065226,
        "trm_confidence": 0.42224326203559465,
        "mcts_value": 0.6873645332478184,
        "consensus_score": 0.5088025226690918,
        "last_agent": "mcts",
        "iteration": 7,
        "query_length": 2055,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.02293006118458668,
        "trm_confidence": 0.5126228955174705,
        "mcts_value": 0.9904803457432356,
        "consensus_score": 0.5632771727200497,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 2794,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6077083785703474,
        "trm_confidence": 0.3758999568929495,
        "mcts_value": 0.753722663641583,
        "consensus_score": 0.610171703036418,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 4911,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.3561551971414369,
        "trm_confidence": 0.5026040999614456,
        "mcts_value": 0.7905138172334725,
        "consensus_score": 0.5024261765259104,
        "last_agent": "hrm",
        "iteration": 9,
        "query_length": 1564,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5601399601255522,
        "trm_confidence": 0.23818074832487054,
        "mcts_value": 0.6783086408838899,
        "consensus_score": 0.4221763496498987,
        "last_agent": "none",
        "iteration": 5,
        "query_length": 2296,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.049455501039777416,
        "trm_confidence": 0.02133663638725458,
        "mcts_value": 0.9384726117379951,
        "consensus_score": 0.32945521022237767,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 4636,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.621376823613001,
        "trm_confidence": 0.19819455347125778,
        "mcts_value": 0.9836682156268193,
        "consensus_score": 0.6644694770882553,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 3536,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5708519291478197,
        "trm_confidence": 0.5551556028865148,
        "mcts_value": 0.7691664572171499,
        "consensus_score": 0.68311193027234,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 2364,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6037348568884349,
        "trm_confidence": 0.14812180701757874,
        "mcts_value": 0.9227691911279663,
        "consensus_score": 0.46662190675527215,
        "last_agent": "trm",
        "iteration": 6,
        "query_length": 4698,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.4900071078471968,
        "trm_confidence": 0.005143697344983877,
        "mcts_value": 0.715641312477804,
        "consensus_score": 0.3746702590207091,
        "last_agent": "none",
        "iteration": 8,
        "query_length": 3910,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.10939178398195547,
        "trm_confidence": 0.6368554815027963,
        "mcts_value": 0.9586427648589736,
        "consensus_score": 0.5115745385178366,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 4515,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.04569406413506189,
        "trm_confidence": 0.05244518738162813,
        "mcts_value": 0.6337379577753245,
        "consensus_score": 0.2619376192221952,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 2060,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.4125013925652396,
        "trm_confidence": 0.373128423350449,
        "mcts_value": 0.7779343037127447,
        "consensus_score": 0.5588003864963985,
        "last_agent": "mcts",
        "iteration": 7,
        "query_length": 1259,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.24520848957413544,
        "trm_confidence": 0.008550824040745563,
        "mcts_value": 0.9989242970186536,
        "consensus_score": 0.3426862126548425,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 906,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.43921226307516803,
        "trm_confidence": 0.36424863017582754,
        "mcts_value": 0.975993037481997,
        "consensus_score": 0.6527592104987346,
        "last_agent": "none",
        "iteration": 5,
        "query_length": 3920,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.43377415349500237,
        "trm_confidence": 0.6543300057320453,
        "mcts_value": 0.6012244419456372,
        "consensus_score": 0.6206815241310022,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 3779,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.38558216552118657,
        "trm_confidence": 0.4329781655522392,
        "mcts_value": 0.9296955784400505,
        "consensus_score": 0.4945350377049861,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 4454,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5190127646113992,
        "trm_confidence": 0.5880399059653311,
        "mcts_value": 0.623520423129399,
        "consensus_score": 0.6609301701610041,
        "last_agent": "mcts",
        "iteration": 4,
        "query_length": 3238,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.3318393511494707,
        "trm_confidence": 0.5881757144394681,
        "mcts_value": 0.6270480030517179,
        "consensus_score": 0.4297499172597767,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 4775,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.568117721217664,
        "trm_confidence": 0.08919866975081975,
        "mcts_value": 0.8045080957744134,
        "consensus_score": 0.41219687973020713,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 428,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6647447717237248,
        "trm_confidence": 0.5362233846202891,
        "mcts_value": 0.8920854940579495,
        "consensus_score": 0.773119623409771,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 2224,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.047843523276939165,
        "trm_confidence": 0.4591671770872801,
        "mcts_value": 0.6731612735946729,
        "consensus_score": 0.30378115873790856,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 1205,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.21972065941926494,
        "trm_confidence": 0.3403108164411346,
        "mcts_value": 0.9808802901635648,
        "consensus_score": 0.5619753283503452,
        "last_agent": "trm",
        "iteration": 5,
        "query_length": 3782,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6851660581129289,
        "trm_confidence": 0.5212923059696694,
        "mcts_value": 0.8637869100021912,
        "consensus_score": 0.6137079280840779,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 3905,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.3023762398867548,
        "trm_confidence": 0.19880407143507609,
        "mcts_value": 0.6586492571457411,
        "consensus_score": 0.3103753646082625,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 427,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.1187878525693058,
        "trm_confidence": 0.5065164465219747,
        "mcts_value": 0.8362149634342894,
        "consensus_score": 0.4262214121144573,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 1139,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.2240733361205828,
        "trm_confidence": 0.15755356804221354,
        "mcts_value": 0.8087989548182943,
        "consensus_score": 0.3911291754933931,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 2561,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.10797053546654987,
        "trm_confidence": 0.5040983132964151,
        "mcts_value": 0.7212994472591383,
        "consensus_score": 0.46085584376523214,
        "last_agent": "trm",
        "iteration": 5,
        "query_length": 619,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.45737663965518527,
        "trm_confidence": 0.2515373237743721,
        "mcts_value": 0.8075721740968564,
        "consensus_score": 0.4933148914359465,
        "last_agent": "trm",
        "iteration": 5,
        "query_length": 2215,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5448118163257429,
        "trm_confidence": 0.5586054984192101,
        "mcts_value": 0.6628045296607167,
        "consensus_score": 0.6380608076129142,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 2990,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.41688106510175704,
        "trm_confidence": 0.6434554695583351,
        "mcts_value": 0.7873350971167254,
        "consensus_score": 0.5910117383513462,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 3784,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.0365974075749261,
        "trm_confidence": 0.5197425037389355,
        "mcts_value": 0.7485121950034508,
        "consensus_score": 0.3650812535923534,
        "last_agent": "trm",
        "iteration": 5,
        "query_length": 4709,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.009780903592923794,
        "trm_confidence": 0.0015646433231779942,
        "mcts_value": 0.813886686969084,
        "consensus_score": 0.27720953764304135,
        "last_agent": "hrm",
        "iteration": 6,
        "query_length": 3222,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.15990553020219894,
        "trm_confidence": 0.41454500477946354,
        "mcts_value": 0.8454375750864443,
        "consensus_score": 0.44914922680662356,
        "last_agent": "trm",
        "iteration": 5,
        "query_length": 2849,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.020882475609089554,
        "trm_confidence": 0.2156888734253879,
        "mcts_value": 0.6947442412509153,
        "consensus_score": 0.3908294681314898,
        "last_agent": "mcts",
        "iteration": 9,
        "query_length": 509,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.3971523913098409,
        "trm_confidence": 0.5801162774355986,
        "mcts_value": 0.8303334474327215,
        "consensus_score": 0.5497920590294468,
        "last_agent": "hrm",
        "iteration": 6,
        "query_length": 1739,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.2202815043024345,
        "trm_confidence": 0.6145396436116048,
        "mcts_value": 0.6163479959086697,
        "consensus_score": 0.5661205774551992,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 4524,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.3425514376363715,
        "trm_confidence": 0.38830366684107404,
        "mcts_value": 0.7525395606350557,
        "consensus_score": 0.39851664037438295,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 2853,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.4623165669453584,
        "trm_confidence": 0.36958590048149587,
        "mcts_value": 0.8613619248014569,
        "consensus_score": 0.53127614641898,
        "last_agent": "trm",
        "iteration": 5,
        "query_length": 3363,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5771517973843117,
        "trm_confidence": 0.2946381896294308,
        "mcts_value": 0.6179765614036828,
        "consensus_score": 0.47953969430487003,
        "last_agent": "hrm",
        "iteration": 6,
        "query_length": 1062,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.10591547503264781,
        "trm_confidence": 0.1158720510815211,
        "mcts_value": 0.6699667811082876,
        "consensus_score": 0.3289569323109717,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 4479,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.1950315914457207,
        "trm_confidence": 0.4535299440877101,
        "mcts_value": 0.9622469588269951,
        "consensus_score": 0.44503595517348216,
        "last_agent": "mcts",
        "iteration": 7,
        "query_length": 569,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.16648476941725376,
        "trm_confidence": 0.5799478637875408,
        "mcts_value": 0.8145279108890514,
        "consensus_score": 0.6143474836122634,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 4423,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.1690423424352558,
        "trm_confidence": 0.6051152673501498,
        "mcts_value": 0.916165307338229,
        "consensus_score": 0.5194811534365443,
        "last_agent": "hrm",
        "iteration": 9,
        "query_length": 899,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.3735339537579074,
        "trm_confidence": 0.36744546699302055,
        "mcts_value": 0.7153546813656679,
        "consensus_score": 0.5555309376028053,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 3732,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.25031874715615304,
        "trm_confidence": 0.33524288546700354,
        "mcts_value": 0.7085552531300148,
        "consensus_score": 0.4301542438819369,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 462,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.3185152324082315,
        "trm_confidence": 0.43265168677086213,
        "mcts_value": 0.6075406316151813,
        "consensus_score": 0.4449816343500297,
        "last_agent": "none",
        "iteration": 4,
        "query_length": 2300,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.13140114571483114,
        "trm_confidence": 0.5294149801296674,
        "mcts_value": 0.7646622780518945,
        "consensus_score": 0.43943960399395965,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 3893,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.08541045230398672,
        "trm_confidence": 0.0799841057774221,
        "mcts_value": 0.9509708316865555,
        "consensus_score": 0.33218573041392463,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 4325,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.17168609017650976,
        "trm_confidence": 0.059342319940571306,
        "mcts_value": 0.9052748329580975,
        "consensus_score": 0.3615277096685557,
        "last_agent": "mcts",
        "iteration": 9,
        "query_length": 4637,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.3905206749498208,
        "trm_confidence": 0.5069341912244236,
        "mcts_value": 0.8933379393453402,
        "consensus_score": 0.6332088422339671,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 4970,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.22198446331259672,
        "trm_confidence": 0.45778120270185196,
        "mcts_value": 0.9082437330553728,
        "consensus_score": 0.5256325518603668,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 581,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.605980924748164,
        "trm_confidence": 0.09900190161810396,
        "mcts_value": 0.8664868661221519,
        "consensus_score": 0.49704436327276325,
        "last_agent": "mcts",
        "iteration": 4,
        "query_length": 945,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.34543005581446923,
        "trm_confidence": 0.24886331304467244,
        "mcts_value": 0.9958633986686976,
        "consensus_score": 0.48502767116761686,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 1037,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.19863066706415367,
        "trm_confidence": 0.049342653749843375,
        "mcts_value": 0.8902833582950973,
        "consensus_score": 0.4160742159399531,
        "last_agent": "none",
        "iteration": 4,
        "query_length": 2372,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5222729624500508,
        "trm_confidence": 0.5388251171454546,
        "mcts_value": 0.7001506395959293,
        "consensus_score": 0.6319043541606525,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 1844,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.16983775796793601,
        "trm_confidence": 0.19196700055773697,
        "mcts_value": 0.9960582456223053,
        "consensus_score": 0.5153263112987674,
        "last_agent": "none",
        "iteration": 5,
        "query_length": 2822,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.026012463273550022,
        "trm_confidence": 0.3876334659856583,
        "mcts_value": 0.9020105510002314,
        "consensus_score": 0.4537276117936565,
        "last_agent": "trm",
        "iteration": 6,
        "query_length": 440,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.25493629665410233,
        "trm_confidence": 0.5940680810031305,
        "mcts_value": 0.6650998877199426,
        "consensus_score": 0.44282997311359035,
        "last_agent": "hrm",
        "iteration": 9,
        "query_length": 4586,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.05273146371199239,
        "trm_confidence": 0.47453250954306,
        "mcts_value": 0.795691626875996,
        "consensus_score": 0.4991054978548028,
        "last_agent": "mcts",
        "iteration": 9,
        "query_length": 3727,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5383433396880142,
        "trm_confidence": 0.35403079482493055,
        "mcts_value": 0.8688062889198347,
        "consensus_score": 0.6696075213804851,
        "last_agent": "none",
        "iteration": 8,
        "query_length": 929,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.347716446502645,
        "trm_confidence": 0.5108732447446628,
        "mcts_value": 0.6941808068055524,
        "consensus_score": 0.4581345169301566,
        "last_agent": "hrm",
        "iteration": 5,
        "query_length": 670,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.12727081704664014,
        "trm_confidence": 0.23202471738705419,
        "mcts_value": 0.7738441442948825,
        "consensus_score": 0.35718845589604403,
        "last_agent": "hrm",
        "iteration": 8,
        "query_length": 1700,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.4561891312377787,
        "trm_confidence": 0.4899456525584073,
        "mcts_value": 0.9946360303688364,
        "consensus_score": 0.611289438660992,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 4474,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6812791930683556,
        "trm_confidence": 0.6048970737743725,
        "mcts_value": 0.642552259724387,
        "consensus_score": 0.6036385357633317,
        "last_agent": "none",
        "iteration": 8,
        "query_length": 3610,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.2007419475322064,
        "trm_confidence": 0.09521921165463454,
        "mcts_value": 0.7292512824581884,
        "consensus_score": 0.35160109700351094,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 3222,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6570274695412733,
        "trm_confidence": 0.447013837431379,
        "mcts_value": 0.6861428669917617,
        "consensus_score": 0.6766481894648053,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 3175,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.23164454587012664,
        "trm_confidence": 0.13660021839519046,
        "mcts_value": 0.9397847274576645,
        "consensus_score": 0.3890140917354511,
        "last_agent": "hrm",
        "iteration": 5,
        "query_length": 766,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.1898918508955939,
        "trm_confidence": 0.4539586088835965,
        "mcts_value": 0.7303854411117326,
        "consensus_score": 0.36868510796468457,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 339,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.15464950354903867,
        "trm_confidence": 0.14046663778904223,
        "mcts_value": 0.6703398997124024,
        "consensus_score": 0.22200390169644207,
        "last_agent": "none",
        "iteration": 7,
        "query_length": 2176,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6963623972143801,
        "trm_confidence": 0.6428773479732098,
        "mcts_value": 0.70917592470996,
        "consensus_score": 0.6258203849304502,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 4246,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5618521589751637,
        "trm_confidence": 0.08836900559639752,
        "mcts_value": 0.9377654186636974,
        "consensus_score": 0.45421863074809,
        "last_agent": "trm",
        "iteration": 6,
        "query_length": 3960,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.07515609478057111,
        "trm_confidence": 0.40200810157588607,
        "mcts_value": 0.9528339996382009,
        "consensus_score": 0.4078011047565753,
        "last_agent": "hrm",
        "iteration": 8,
        "query_length": 4111,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.05815794228443604,
        "trm_confidence": 0.40547793712618185,
        "mcts_value": 0.7819077169902374,
        "consensus_score": 0.39688453282745606,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 4693,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.1915902853443494,
        "trm_confidence": 0.11026234512208015,
        "mcts_value": 0.9499588801890357,
        "consensus_score": 0.37856657901553775,
        "last_agent": "mcts",
        "iteration": 4,
        "query_length": 4286,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.09959808498961463,
        "trm_confidence": 0.07795548095061157,
        "mcts_value": 0.8640470535528122,
        "consensus_score": 0.2736772647485496,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 3947,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.03259280621015039,
        "trm_confidence": 0.5320502470804781,
        "mcts_value": 0.9776460461986113,
        "consensus_score": 0.5615267158398527,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 1656,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.31715748456687903,
        "trm_confidence": 0.5996026240044333,
        "mcts_value": 0.7142035966761665,
        "consensus_score": 0.5893732969234061,
        "last_agent": "none",
        "iteration": 7,
        "query_length": 3651,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5981688076295373,
        "trm_confidence": 0.2799658617472436,
        "mcts_value": 0.8908062965073484,
        "consensus_score": 0.62336407779224,
        "last_agent": "hrm",
        "iteration": 6,
        "query_length": 849,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5430080698155945,
        "trm_confidence": 0.6842541888902517,
        "mcts_value": 0.8099641519819655,
        "consensus_score": 0.6679487027569171,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 4395,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.2078571164724059,
        "trm_confidence": 0.060513189288164355,
        "mcts_value": 0.6535556277500726,
        "consensus_score": 0.29288749713939705,
        "last_agent": "hrm",
        "iteration": 5,
        "query_length": 378,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.1721493554123952,
        "trm_confidence": 0.42832623493944827,
        "mcts_value": 0.9174310997566425,
        "consensus_score": 0.5862051805525512,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 328,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.41325844502884623,
        "trm_confidence": 0.3546277978318974,
        "mcts_value": 0.7379064775863817,
        "consensus_score": 0.43165487407108916,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 939,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.4287118607378963,
        "trm_confidence": 0.04911206963964689,
        "mcts_value": 0.7069330857830289,
        "consensus_score": 0.3622404120999607,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 1259,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.2870207995972183,
        "trm_confidence": 0.22152209969067274,
        "mcts_value": 0.8420707400012313,
        "consensus_score": 0.5334308840117128,
        "last_agent": "none",
        "iteration": 6,
        "query_length": 3194,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5221610491550703,
        "trm_confidence": 0.4772597995887381,
        "mcts_value": 0.9637205262561082,
        "consensus_score": 0.5642874467466807,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 1468,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.15794658994125135,
        "trm_confidence": 0.5453969607405275,
        "mcts_value": 0.683773134610504,
        "consensus_score": 0.41955721819814945,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 1135,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.43199000437925217,
        "trm_confidence": 0.608219663374262,
        "mcts_value": 0.7772711948981151,
        "consensus_score": 0.5279878332237377,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 1249,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.3377762158718462,
        "trm_confidence": 0.6468872542889508,
        "mcts_value": 0.7456433448967901,
        "consensus_score": 0.5857472691254665,
        "last_agent": "hrm",
        "iteration": 6,
        "query_length": 3167,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6524794138598203,
        "trm_confidence": 0.3511632894458381,
        "mcts_value": 0.6260862679598145,
        "consensus_score": 0.5439942579752518,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 4146,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.19074201448203928,
        "trm_confidence": 0.4611717367721946,
        "mcts_value": 0.7753628362120595,
        "consensus_score": 0.4896878806847168,
        "last_agent": "trm",
        "iteration": 5,
        "query_length": 2747,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.27813371894457856,
        "trm_confidence": 0.0030806453324853985,
        "mcts_value": 0.8576347429346562,
        "consensus_score": 0.36893259910581094,
        "last_agent": "mcts",
        "iteration": 4,
        "query_length": 3946,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.057343918929729384,
        "trm_confidence": 0.19296580365101634,
        "mcts_value": 0.7722439609923336,
        "consensus_score": 0.3719495015694868,
        "last_agent": "trm",
        "iteration": 6,
        "query_length": 3265,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.4097169332098666,
        "trm_confidence": 0.6768629970444513,
        "mcts_value": 0.9665208135659924,
        "consensus_score": 0.7579703734932746,
        "last_agent": "none",
        "iteration": 6,
        "query_length": 3559,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.4791859970651187,
        "trm_confidence": 0.3437572785462063,
        "mcts_value": 0.6547327136015789,
        "consensus_score": 0.3959238509215329,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 4657,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.12701811871100846,
        "trm_confidence": 0.13127201260054583,
        "mcts_value": 0.7263325940353923,
        "consensus_score": 0.32405022804403866,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 2794,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.14610881840587236,
        "trm_confidence": 0.03224543007153378,
        "mcts_value": 0.7747408030151097,
        "consensus_score": 0.31022441439405146,
        "last_agent": "hrm",
        "iteration": 9,
        "query_length": 4839,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.4162918703423202,
        "trm_confidence": 0.3053725646623523,
        "mcts_value": 0.7634403070509002,
        "consensus_score": 0.5108381359765598,
        "last_agent": "none",
        "iteration": 7,
        "query_length": 1037,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6584276987172909,
        "trm_confidence": 0.2996197271071803,
        "mcts_value": 0.8030835374132496,
        "consensus_score": 0.49022691502051574,
        "last_agent": "none",
        "iteration": 5,
        "query_length": 3610,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.3866693525184374,
        "trm_confidence": 0.5001864687890575,
        "mcts_value": 0.8162287407554648,
        "consensus_score": 0.5226244825444922,
        "last_agent": "hrm",
        "iteration": 8,
        "query_length": 1884,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.2815183694024254,
        "trm_confidence": 0.12524734995951484,
        "mcts_value": 0.7645123010483306,
        "consensus_score": 0.37322348205449835,
        "last_agent": "mcts",
        "iteration": 9,
        "query_length": 3920,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.20955924034800916,
        "trm_confidence": 0.47760280306181196,
        "mcts_value": 0.8374452765271861,
        "consensus_score": 0.5562316817182031,
        "last_agent": "none",
        "iteration": 6,
        "query_length": 3385,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.007527602320017867,
        "trm_confidence": 0.0860795408424368,
        "mcts_value": 0.629478332855336,
        "consensus_score": 0.17223016157143217,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 2200,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.3003525811907735,
        "trm_confidence": 0.2763886248904426,
        "mcts_value": 0.7099726767881527,
        "consensus_score": 0.38919696122072633,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 601,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.542797116870872,
        "trm_confidence": 0.1411816092963175,
        "mcts_value": 0.6546417611118621,
        "consensus_score": 0.4924810905196795,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 1604,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6647009394575618,
        "trm_confidence": 0.32413276472680647,
        "mcts_value": 0.6346983073152016,
        "consensus_score": 0.567224436161121,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 2763,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6203528193725827,
        "trm_confidence": 0.1900075922950634,
        "mcts_value": 0.8021590618237545,
        "consensus_score": 0.5925447551355442,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 4554,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.16650482550703383,
        "trm_confidence": 0.3097693113012998,
        "mcts_value": 0.7173500691167987,
        "consensus_score": 0.42494158391789427,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 2654,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.4091297662980429,
        "trm_confidence": 0.36351973519098435,
        "mcts_value": 0.8783515072913102,
        "consensus_score": 0.5612175890826414,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 2878,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.2544274738110137,
        "trm_confidence": 0.10481394578673936,
        "mcts_value": 0.7550002142155073,
        "consensus_score": 0.3907813448478427,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 1784,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5918205291007997,
        "trm_confidence": 0.0763380145377019,
        "mcts_value": 0.9439400227462136,
        "consensus_score": 0.5763738663568887,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 3207,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.19636694379014258,
        "trm_confidence": 0.30947440618900923,
        "mcts_value": 0.690009451873861,
        "consensus_score": 0.35235996368713696,
        "last_agent": "trm",
        "iteration": 8,
        "query_length": 1010,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.3544618486405938,
        "trm_confidence": 0.6221514364555268,
        "mcts_value": 0.6883515339398161,
        "consensus_score": 0.6134674023819895,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 3167,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6148680140308268,
        "trm_confidence": 0.4971168583490874,
        "mcts_value": 0.6719545858810111,
        "consensus_score": 0.6090791467164668,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 2568,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.15169861868926238,
        "trm_confidence": 0.08440420871119089,
        "mcts_value": 0.8729957144008273,
        "consensus_score": 0.3628212159351468,
        "last_agent": "none",
        "iteration": 6,
        "query_length": 4551,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5704315813364017,
        "trm_confidence": 0.5241789003278413,
        "mcts_value": 0.7295918187007265,
        "consensus_score": 0.641652207466697,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 518,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.11266319650705656,
        "trm_confidence": 0.19303345284654838,
        "mcts_value": 0.7558684743931463,
        "consensus_score": 0.4298103324610913,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 958,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.33947217517053047,
        "trm_confidence": 0.5016657593251701,
        "mcts_value": 0.6864448220328685,
        "consensus_score": 0.46653265726760484,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 4322,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.1756367402124386,
        "trm_confidence": 0.4417197013131317,
        "mcts_value": 0.9150250250158817,
        "consensus_score": 0.4415408780782525,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 4388,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.13150250773626537,
        "trm_confidence": 0.6008398704647704,
        "mcts_value": 0.8526673542319743,
        "consensus_score": 0.4390067739467064,
        "last_agent": "none",
        "iteration": 8,
        "query_length": 3149,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5604702739715944,
        "trm_confidence": 0.41662668123237145,
        "mcts_value": 0.8126018927370995,
        "consensus_score": 0.6531839456316204,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 3290,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5192746395010424,
        "trm_confidence": 0.16653692066503153,
        "mcts_value": 0.6911244604345537,
        "consensus_score": 0.5551067285620601,
        "last_agent": "trm",
        "iteration": 5,
        "query_length": 4692,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.17009010337397337,
        "trm_confidence": 0.25263195544389805,
        "mcts_value": 0.9434692481935456,
        "consensus_score": 0.49832513457294436,
        "last_agent": "none",
        "iteration": 5,
        "query_length": 3510,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.17058772796290878,
        "trm_confidence": 0.1649605688983235,
        "mcts_value": 0.7945433693066476,
        "consensus_score": 0.2785450063396625,
        "last_agent": "none",
        "iteration": 6,
        "query_length": 109,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.43486322474956446,
        "trm_confidence": 0.4386586274023478,
        "mcts_value": 0.6124623243306682,
        "consensus_score": 0.40948815318225856,
        "last_agent": "none",
        "iteration": 6,
        "query_length": 3444,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.2883823068083915,
        "trm_confidence": 0.32557134992486714,
        "mcts_value": 0.8328668272978313,
        "consensus_score": 0.423929493564191,
        "last_agent": "mcts",
        "iteration": 9,
        "query_length": 1517,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.15675407418460977,
        "trm_confidence": 0.21751818744351983,
        "mcts_value": 0.9642656169000174,
        "consensus_score": 0.378303847804623,
        "last_agent": "none",
        "iteration": 6,
        "query_length": 1006,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.13275246708368724,
        "trm_confidence": 0.6435778470574205,
        "mcts_value": 0.8532422948258542,
        "consensus_score": 0.463810763672716,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 1720,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.2487519092491989,
        "trm_confidence": 0.4168497941370826,
        "mcts_value": 0.8815079231064198,
        "consensus_score": 0.46041202899874717,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 4798,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.43066103537470446,
        "trm_confidence": 0.15270442919272204,
        "mcts_value": 0.7412751318883843,
        "consensus_score": 0.474132405597394,
        "last_agent": "hrm",
        "iteration": 9,
        "query_length": 3198,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.30771116271093707,
        "trm_confidence": 0.541816860194706,
        "mcts_value": 0.6175433133671664,
        "consensus_score": 0.5191797458724317,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 4071,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.3722451902992266,
        "trm_confidence": 0.1430559856244378,
        "mcts_value": 0.693978218101228,
        "consensus_score": 0.4483388707386439,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 330,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.34069695390044646,
        "trm_confidence": 0.4205945634849206,
        "mcts_value": 0.8026179113292021,
        "consensus_score": 0.5087920630048397,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 1868,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.4406615999824062,
        "trm_confidence": 0.44450554110256446,
        "mcts_value": 0.9516453804946554,
        "consensus_score": 0.6341105772697317,
        "last_agent": "mcts",
        "iteration": 7,
        "query_length": 2152,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.017872285996915215,
        "trm_confidence": 0.005219501996662123,
        "mcts_value": 0.8368011549795481,
        "consensus_score": 0.35695934625039155,
        "last_agent": "none",
        "iteration": 4,
        "query_length": 3910,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.13871986073570491,
        "trm_confidence": 0.6270115867856038,
        "mcts_value": 0.845995105179858,
        "consensus_score": 0.47526133858339714,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 4759,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.10552091639326548,
        "trm_confidence": 0.22136289713036456,
        "mcts_value": 0.6962971814096776,
        "consensus_score": 0.3206504448984908,
        "last_agent": "hrm",
        "iteration": 5,
        "query_length": 763,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5814443788879998,
        "trm_confidence": 0.11138242296333094,
        "mcts_value": 0.6918609344958324,
        "consensus_score": 0.5576115038375773,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 4467,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.08420000060926532,
        "trm_confidence": 0.5717740472353708,
        "mcts_value": 0.8234256267071626,
        "consensus_score": 0.44807033138676555,
        "last_agent": "none",
        "iteration": 7,
        "query_length": 3520,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.03884019056702028,
        "trm_confidence": 0.11803313328209751,
        "mcts_value": 0.841056260930314,
        "consensus_score": 0.26488164351115007,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 2176,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.3590681153579105,
        "trm_confidence": 0.4365026033918791,
        "mcts_value": 0.9687685283097158,
        "consensus_score": 0.5596107174327927,
        "last_agent": "hrm",
        "iteration": 5,
        "query_length": 1213,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.24055214801083497,
        "trm_confidence": 0.31692561198959995,
        "mcts_value": 0.7177291899339334,
        "consensus_score": 0.4262157207483272,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 3072,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.10465821764253724,
        "trm_confidence": 0.2449994628955808,
        "mcts_value": 0.7951904659530589,
        "consensus_score": 0.42507819257167323,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 2077,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.036105681083131086,
        "trm_confidence": 0.6531526690349112,
        "mcts_value": 0.7666804094272028,
        "consensus_score": 0.46803354790699225,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 4934,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.2033292341002282,
        "trm_confidence": 0.2885614547813878,
        "mcts_value": 0.9749441286315138,
        "consensus_score": 0.5407177872229005,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 2260,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.38893016798324637,
        "trm_confidence": 0.4825663175691036,
        "mcts_value": 0.9979292693656894,
        "consensus_score": 0.6518343651367351,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 4872,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.20618925247647954,
        "trm_confidence": 0.01596017330247862,
        "mcts_value": 0.8253024529881239,
        "consensus_score": 0.37203127646855805,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 1101,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.20165745105494864,
        "trm_confidence": 0.3569922506583052,
        "mcts_value": 0.6658824561963457,
        "consensus_score": 0.47551362976539985,
        "last_agent": "hrm",
        "iteration": 5,
        "query_length": 2120,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6552332642343762,
        "trm_confidence": 0.08169258417520159,
        "mcts_value": 0.9013474455802182,
        "consensus_score": 0.6313694248295101,
        "last_agent": "none",
        "iteration": 8,
        "query_length": 4067,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.35430370894841656,
        "trm_confidence": 0.43584077185655556,
        "mcts_value": 0.6215098838274221,
        "consensus_score": 0.4319182228520328,
        "last_agent": "none",
        "iteration": 7,
        "query_length": 1305,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.02930306778478665,
        "trm_confidence": 0.30763723660522463,
        "mcts_value": 0.8210191759091919,
        "consensus_score": 0.3287501243915139,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 2544,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.12262977807574112,
        "trm_confidence": 0.15365290566384016,
        "mcts_value": 0.7198587585408276,
        "consensus_score": 0.3275975984973937,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 4865,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.027919921899200804,
        "trm_confidence": 0.15917366193599397,
        "mcts_value": 0.622156824686119,
        "consensus_score": 0.338304391695657,
        "last_agent": "trm",
        "iteration": 8,
        "query_length": 202,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6505977835323311,
        "trm_confidence": 0.2318247714734927,
        "mcts_value": 0.7275878370330569,
        "consensus_score": 0.6036590132102162,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 3878,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.19376230117373294,
        "trm_confidence": 0.460903398328293,
        "mcts_value": 0.9577157077922971,
        "consensus_score": 0.4550532188029849,
        "last_agent": "mcts",
        "iteration": 7,
        "query_length": 3258,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.002607855654334512,
        "trm_confidence": 0.22732648150856585,
        "mcts_value": 0.6251053111722411,
        "consensus_score": 0.30291933501906637,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 399,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6576675279857744,
        "trm_confidence": 0.0879573147862478,
        "mcts_value": 0.9964800123105004,
        "consensus_score": 0.5637231036945398,
        "last_agent": "trm",
        "iteration": 8,
        "query_length": 1547,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5223972689147434,
        "trm_confidence": 0.08690500903747479,
        "mcts_value": 0.7465488619637961,
        "consensus_score": 0.4183129050235457,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 4194,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.032799915867008554,
        "trm_confidence": 0.12207420247613396,
        "mcts_value": 0.934884085944662,
        "consensus_score": 0.2729516512169948,
        "last_agent": "none",
        "iteration": 5,
        "query_length": 3021,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.3698200768977788,
        "trm_confidence": 0.4605242469841963,
        "mcts_value": 0.8016841206715465,
        "consensus_score": 0.5209919709580493,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 1230,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6063193015724208,
        "trm_confidence": 0.015857791395556153,
        "mcts_value": 0.621716435099165,
        "consensus_score": 0.42072620067370714,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 1575,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.47712759383618397,
        "trm_confidence": 0.0856861934113269,
        "mcts_value": 0.843366698617572,
        "consensus_score": 0.5037122487838062,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 556,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.24115339617428277,
        "trm_confidence": 0.5273167309174219,
        "mcts_value": 0.7772207997433388,
        "consensus_score": 0.5238051665640774,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 3470,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.32142392562430644,
        "trm_confidence": 0.11908708393095507,
        "mcts_value": 0.6680921202370974,
        "consensus_score": 0.2734781653997501,
        "last_agent": "mcts",
        "iteration": 4,
        "query_length": 1676,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.051226483406671686,
        "trm_confidence": 0.3158319118734727,
        "mcts_value": 0.6291845364640543,
        "consensus_score": 0.2811010337150727,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 3814,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.31411206359138144,
        "trm_confidence": 0.2625795571528991,
        "mcts_value": 0.7618972994685068,
        "consensus_score": 0.35376953006622147,
        "last_agent": "hrm",
        "iteration": 6,
        "query_length": 360,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.39260310105547147,
        "trm_confidence": 0.06913540039949356,
        "mcts_value": 0.640948132006049,
        "consensus_score": 0.2881484759127241,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 3999,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.0031143173662041,
        "trm_confidence": 0.13508210923749067,
        "mcts_value": 0.6550539140529478,
        "consensus_score": 0.21159523105528902,
        "last_agent": "mcts",
        "iteration": 9,
        "query_length": 102,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.28351870855068445,
        "trm_confidence": 0.4987368986319067,
        "mcts_value": 0.7623777589951879,
        "consensus_score": 0.5402621918668641,
        "last_agent": "trm",
        "iteration": 5,
        "query_length": 1984,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.39537985421393407,
        "trm_confidence": 0.09709532221116647,
        "mcts_value": 0.7750447562620203,
        "consensus_score": 0.38116134291072923,
        "last_agent": "none",
        "iteration": 6,
        "query_length": 4513,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.4571445090238169,
        "trm_confidence": 0.3565729185311946,
        "mcts_value": 0.7843045065167165,
        "consensus_score": 0.4729652904733782,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 3495,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.04349742889371404,
        "trm_confidence": 0.18068504975469057,
        "mcts_value": 0.7917218730871143,
        "consensus_score": 0.39944153958867057,
        "last_agent": "hrm",
        "iteration": 9,
        "query_length": 2940,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.2960341540506385,
        "trm_confidence": 0.29864449828811546,
        "mcts_value": 0.8408351304006768,
        "consensus_score": 0.4173046031450657,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 1658,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6199640440317623,
        "trm_confidence": 0.25792774231741794,
        "mcts_value": 0.6619457645715343,
        "consensus_score": 0.4232269367444519,
        "last_agent": "trm",
        "iteration": 6,
        "query_length": 4463,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.34327889208174345,
        "trm_confidence": 0.6694379580513925,
        "mcts_value": 0.9667826343447138,
        "consensus_score": 0.7348815233350119,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 4262,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5295282585288805,
        "trm_confidence": 0.40413730215330446,
        "mcts_value": 0.7527501648675294,
        "consensus_score": 0.6060753755312994,
        "last_agent": "trm",
        "iteration": 5,
        "query_length": 2839,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.2118643256670007,
        "trm_confidence": 0.6903129267689389,
        "mcts_value": 0.9830369736776687,
        "consensus_score": 0.6040799449989437,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 3918,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.3883777802635978,
        "trm_confidence": 0.21499190556342018,
        "mcts_value": 0.9139847653152711,
        "consensus_score": 0.5274748910061812,
        "last_agent": "trm",
        "iteration": 6,
        "query_length": 3172,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.4677165338014153,
        "trm_confidence": 0.5336812877933015,
        "mcts_value": 0.9450656503012365,
        "consensus_score": 0.6655198675925663,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 3227,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.064791257116833,
        "trm_confidence": 0.3878136514494973,
        "mcts_value": 0.7049164642875259,
        "consensus_score": 0.40851230489962564,
        "last_agent": "mcts",
        "iteration": 9,
        "query_length": 2647,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.44256067596182397,
        "trm_confidence": 0.6990868447642679,
        "mcts_value": 0.6049027207460108,
        "consensus_score": 0.5410932845820328,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 4789,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.17563905026679527,
        "trm_confidence": 0.2428720177572709,
        "mcts_value": 0.7701661899108189,
        "consensus_score": 0.4238671003608342,
        "last_agent": "trm",
        "iteration": 6,
        "query_length": 679,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6425597511360087,
        "trm_confidence": 0.28482017194039755,
        "mcts_value": 0.6160125051197055,
        "consensus_score": 0.4800928459389961,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 3975,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.028807240036680212,
        "trm_confidence": 0.10169072715622485,
        "mcts_value": 0.9455299801055019,
        "consensus_score": 0.28795394821004383,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 1272,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.24066451373569545,
        "trm_confidence": 0.5360884799176838,
        "mcts_value": 0.7751073468262039,
        "consensus_score": 0.5658005925109914,
        "last_agent": "hrm",
        "iteration": 5,
        "query_length": 2241,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5429223745586418,
        "trm_confidence": 0.48654437294048475,
        "mcts_value": 0.7431240129661425,
        "consensus_score": 0.6240343280929974,
        "last_agent": "hrm",
        "iteration": 9,
        "query_length": 1190,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.02963761542642319,
        "trm_confidence": 0.6081220430624733,
        "mcts_value": 0.735873490187804,
        "consensus_score": 0.4743636632100025,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 1890,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5241162636843427,
        "trm_confidence": 0.46632505648051575,
        "mcts_value": 0.8893638012649838,
        "consensus_score": 0.5673579013326266,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 2080,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.4556607463952512,
        "trm_confidence": 0.6330503900641488,
        "mcts_value": 0.8631025294335881,
        "consensus_score": 0.7362659819170875,
        "last_agent": "trm",
        "iteration": 8,
        "query_length": 1645,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5733551887565057,
        "trm_confidence": 0.45122373788403347,
        "mcts_value": 0.7454935918237665,
        "consensus_score": 0.5054773788721658,
        "last_agent": "mcts",
        "iteration": 4,
        "query_length": 3981,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6662827140882281,
        "trm_confidence": 0.2994956771298266,
        "mcts_value": 0.9537992592418003,
        "consensus_score": 0.6966713983480932,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 126,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.697918453147421,
        "trm_confidence": 0.02115344262594493,
        "mcts_value": 0.6581350844908569,
        "consensus_score": 0.47853617716369623,
        "last_agent": "none",
        "iteration": 8,
        "query_length": 2610,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5051054710091956,
        "trm_confidence": 0.15147969429688585,
        "mcts_value": 0.7022510975876359,
        "consensus_score": 0.518753072114909,
        "last_agent": "none",
        "iteration": 5,
        "query_length": 4041,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.13226202532353212,
        "trm_confidence": 0.16659054134282192,
        "mcts_value": 0.9133653943335903,
        "consensus_score": 0.45989340141352153,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 998,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6511273501213394,
        "trm_confidence": 0.3933144844251073,
        "mcts_value": 0.782212362475966,
        "consensus_score": 0.5864203019824493,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 39,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.28815102182140645,
        "trm_confidence": 0.21852605082475737,
        "mcts_value": 0.7734939894749928,
        "consensus_score": 0.3918626493212357,
        "last_agent": "hrm",
        "iteration": 8,
        "query_length": 489,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.14084314868811096,
        "trm_confidence": 0.268945602243145,
        "mcts_value": 0.8727266902585276,
        "consensus_score": 0.35681294157532756,
        "last_agent": "none",
        "iteration": 7,
        "query_length": 1968,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.11439160427818405,
        "trm_confidence": 0.15855390735182287,
        "mcts_value": 0.7418622304225458,
        "consensus_score": 0.4018918310559654,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 557,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.13211278804007376,
        "trm_confidence": 0.5091159917392016,
        "mcts_value": 0.7324797627224844,
        "consensus_score": 0.4516118533956279,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 645,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.24738488351716395,
        "trm_confidence": 0.2693562444969951,
        "mcts_value": 0.6591013780288429,
        "consensus_score": 0.38983348186601663,
        "last_agent": "mcts",
        "iteration": 9,
        "query_length": 150,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.4521926221421078,
        "trm_confidence": 0.6654542154042549,
        "mcts_value": 0.9729888205344299,
        "consensus_score": 0.7603564601768946,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 742,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.3100398441795226,
        "trm_confidence": 0.460622617268879,
        "mcts_value": 0.74292359315087,
        "consensus_score": 0.48991505681969105,
        "last_agent": "hrm",
        "iteration": 9,
        "query_length": 1338,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.1429320477494029,
        "trm_confidence": 0.6607633266673058,
        "mcts_value": 0.9319946595939554,
        "consensus_score": 0.5096725046576749,
        "last_agent": "none",
        "iteration": 4,
        "query_length": 1889,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.517131593916873,
        "trm_confidence": 0.6723713247332932,
        "mcts_value": 0.9274807310812058,
        "consensus_score": 0.6597691423142072,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 2021,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.31192002079093006,
        "trm_confidence": 0.004095015871525065,
        "mcts_value": 0.8149594502277363,
        "consensus_score": 0.39188393433789537,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 4475,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.37127967759295266,
        "trm_confidence": 0.12477220220576545,
        "mcts_value": 0.9726623447572149,
        "consensus_score": 0.5187609223686881,
        "last_agent": "hrm",
        "iteration": 5,
        "query_length": 678,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.07748641831384613,
        "trm_confidence": 0.07435221386350471,
        "mcts_value": 0.8741260969676969,
        "consensus_score": 0.3899769556847119,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 1605,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.09897539965443619,
        "trm_confidence": 0.36286645309387816,
        "mcts_value": 0.9184595034591311,
        "consensus_score": 0.379425550262937,
        "last_agent": "hrm",
        "iteration": 8,
        "query_length": 3230,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.14541151475870714,
        "trm_confidence": 0.4780262591640515,
        "mcts_value": 0.9777923425290989,
        "consensus_score": 0.4594016570260412,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 492,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.699204268274178,
        "trm_confidence": 0.018148359663660893,
        "mcts_value": 0.8292384565366064,
        "consensus_score": 0.4155932454202992,
        "last_agent": "hrm",
        "iteration": 6,
        "query_length": 4669,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6388077833881113,
        "trm_confidence": 0.6163674834515231,
        "mcts_value": 0.9535348944394526,
        "consensus_score": 0.6648724705235719,
        "last_agent": "trm",
        "iteration": 6,
        "query_length": 3148,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.46426661995738583,
        "trm_confidence": 0.1660000042337644,
        "mcts_value": 0.7250444271136965,
        "consensus_score": 0.4164542020271105,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 1780,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.21438169265579282,
        "trm_confidence": 0.016469676876294147,
        "mcts_value": 0.9857296347539451,
        "consensus_score": 0.3565071821891915,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 3180,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.43823421407178675,
        "trm_confidence": 0.6124293493895462,
        "mcts_value": 0.645086426531144,
        "consensus_score": 0.5377537730249061,
        "last_agent": "none",
        "iteration": 5,
        "query_length": 2388,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.4140413718557169,
        "trm_confidence": 0.07080024932932218,
        "mcts_value": 0.9143867285802508,
        "consensus_score": 0.47384688078716997,
        "last_agent": "none",
        "iteration": 7,
        "query_length": 3661,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6867783549341924,
        "trm_confidence": 0.12323032392527893,
        "mcts_value": 0.6440642384213293,
        "consensus_score": 0.4232395505850826,
        "last_agent": "hrm",
        "iteration": 5,
        "query_length": 3098,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.17469002741858916,
        "trm_confidence": 0.3005963681150697,
        "mcts_value": 0.7774678297930802,
        "consensus_score": 0.4522891119879232,
        "last_agent": "none",
        "iteration": 5,
        "query_length": 4265,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.13872532157599726,
        "trm_confidence": 0.6681974928109701,
        "mcts_value": 0.7225384948710823,
        "consensus_score": 0.5719325584429859,
        "last_agent": "trm",
        "iteration": 6,
        "query_length": 2384,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.24272674878274395,
        "trm_confidence": 0.58772811267745,
        "mcts_value": 0.80752569966152,
        "consensus_score": 0.5964265337467927,
        "last_agent": "hrm",
        "iteration": 6,
        "query_length": 1102,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.0007363807922716958,
        "trm_confidence": 0.6732710189801955,
        "mcts_value": 0.7477560894079189,
        "consensus_score": 0.483846978274696,
        "last_agent": "none",
        "iteration": 8,
        "query_length": 2180,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.2030334913415154,
        "trm_confidence": 0.6801032874356393,
        "mcts_value": 0.8202565843277612,
        "consensus_score": 0.5921314702046956,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 680,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.04504428752245054,
        "trm_confidence": 0.24516113759752034,
        "mcts_value": 0.715144560196313,
        "consensus_score": 0.27033614762577307,
        "last_agent": "none",
        "iteration": 8,
        "query_length": 1430,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.23301783778983257,
        "trm_confidence": 0.46925515268030815,
        "mcts_value": 0.7211019432982609,
        "consensus_score": 0.40431259810438114,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 1304,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.12227880344444905,
        "trm_confidence": 0.6529222532052206,
        "mcts_value": 0.6404797783707098,
        "consensus_score": 0.42131746934726383,
        "last_agent": "trm",
        "iteration": 8,
        "query_length": 4349,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6154667532299842,
        "trm_confidence": 0.10362325407179326,
        "mcts_value": 0.6776406769395111,
        "consensus_score": 0.36932701209435204,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 1739,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.08971358945199735,
        "trm_confidence": 0.6875615806959464,
        "mcts_value": 0.7303964181662499,
        "consensus_score": 0.5379049220026138,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 3616,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.2691948469916442,
        "trm_confidence": 0.22703799524254756,
        "mcts_value": 0.879296909595676,
        "consensus_score": 0.5019016364555173,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 2247,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.20623477405137147,
        "trm_confidence": 0.3426701595313342,
        "mcts_value": 0.8274639271393289,
        "consensus_score": 0.48153238358586203,
        "last_agent": "hrm",
        "iteration": 6,
        "query_length": 3062,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.0066050093393553826,
        "trm_confidence": 0.3988879217624059,
        "mcts_value": 0.8338452848863351,
        "consensus_score": 0.43983349585107095,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 180,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.33249821550006176,
        "trm_confidence": 0.37215949992821973,
        "mcts_value": 0.8297870637085434,
        "consensus_score": 0.4249424071794077,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 2264,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.38958886100466183,
        "trm_confidence": 0.04028922516076817,
        "mcts_value": 0.819223843937053,
        "consensus_score": 0.4359574126581124,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 1477,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.4976213779106428,
        "trm_confidence": 0.2008256766826197,
        "mcts_value": 0.6188499666382289,
        "consensus_score": 0.4694400731677854,
        "last_agent": "hrm",
        "iteration": 8,
        "query_length": 4359,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5849432813373271,
        "trm_confidence": 0.42391415051634174,
        "mcts_value": 0.7761809380865067,
        "consensus_score": 0.5087604503757408,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 1394,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.4301947589066826,
        "trm_confidence": 0.007278671057098373,
        "mcts_value": 0.8566582715646531,
        "consensus_score": 0.452916335726504,
        "last_agent": "hrm",
        "iteration": 9,
        "query_length": 3093,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.21256089197690825,
        "trm_confidence": 0.2948914923860667,
        "mcts_value": 0.6235903081559845,
        "consensus_score": 0.3129685723986355,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 761,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.46808633399039357,
        "trm_confidence": 0.18380072281210846,
        "mcts_value": 0.6452869132028206,
        "consensus_score": 0.47467249283102203,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 914,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5578799743408257,
        "trm_confidence": 0.5487661061438933,
        "mcts_value": 0.7565794654347361,
        "consensus_score": 0.5961871104646532,
        "last_agent": "hrm",
        "iteration": 8,
        "query_length": 4110,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.05688528576326571,
        "trm_confidence": 0.304222998883027,
        "mcts_value": 0.832162734472168,
        "consensus_score": 0.33430254268044207,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 664,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.2522673022310073,
        "trm_confidence": 0.20211872466786202,
        "mcts_value": 0.836309499455014,
        "consensus_score": 0.3708157987302678,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 3976,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5227074134042587,
        "trm_confidence": 0.2929715439256572,
        "mcts_value": 0.6533044577724123,
        "consensus_score": 0.3927404693595567,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 637,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.48722867299686506,
        "trm_confidence": 0.17999942599883725,
        "mcts_value": 0.6839737252374378,
        "consensus_score": 0.5287782484460767,
        "last_agent": "hrm",
        "iteration": 8,
        "query_length": 3467,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6081947137154111,
        "trm_confidence": 0.2531792725155051,
        "mcts_value": 0.8460758153320737,
        "consensus_score": 0.5789780492314458,
        "last_agent": "none",
        "iteration": 7,
        "query_length": 3122,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.4046425342306847,
        "trm_confidence": 0.5281018359588862,
        "mcts_value": 0.6659661634434454,
        "consensus_score": 0.5069134633177373,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 2624,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.4088338487835222,
        "trm_confidence": 0.2943976070261187,
        "mcts_value": 0.6693235039583664,
        "consensus_score": 0.36055634085907,
        "last_agent": "mcts",
        "iteration": 4,
        "query_length": 3216,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6244276612236238,
        "trm_confidence": 0.3819486118099557,
        "mcts_value": 0.907731983807261,
        "consensus_score": 0.5591491781164382,
        "last_agent": "none",
        "iteration": 7,
        "query_length": 2719,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.20376871458816245,
        "trm_confidence": 0.08672342650185047,
        "mcts_value": 0.8997910516490358,
        "consensus_score": 0.4768974041987981,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 2961,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.06804884644945763,
        "trm_confidence": 0.044298543381425914,
        "mcts_value": 0.9294697959298257,
        "consensus_score": 0.3989925746828216,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 455,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5875125004474275,
        "trm_confidence": 0.04033643008293962,
        "mcts_value": 0.9615448783904088,
        "consensus_score": 0.5530153846951598,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 4449,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.48023325325328375,
        "trm_confidence": 0.0729001088879177,
        "mcts_value": 0.9932021996304753,
        "consensus_score": 0.45510478965721624,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 1130,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.07023052249153387,
        "trm_confidence": 0.5343709379002983,
        "mcts_value": 0.6393357067489559,
        "consensus_score": 0.3275165787235706,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 3365,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6170942399050675,
        "trm_confidence": 0.45535938149292754,
        "mcts_value": 0.9201799249273803,
        "consensus_score": 0.6287184664962302,
        "last_agent": "hrm",
        "iteration": 10,
        "query_length": 1562,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6204802934349398,
        "trm_confidence": 0.47657581474524546,
        "mcts_value": 0.7102043357819398,
        "consensus_score": 0.6815545649292236,
        "last_agent": "mcts",
        "iteration": 5,
        "query_length": 2450,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.4676245308653993,
        "trm_confidence": 0.6398424062285161,
        "mcts_value": 0.8450968626801308,
        "consensus_score": 0.6270051948636078,
        "last_agent": "none",
        "iteration": 9,
        "query_length": 159,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6278599542833151,
        "trm_confidence": 0.4760425988010662,
        "mcts_value": 0.9151010875463841,
        "consensus_score": 0.5865291244442012,
        "last_agent": "none",
        "iteration": 5,
        "query_length": 4674,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6771893195888536,
        "trm_confidence": 0.025759580280082583,
        "mcts_value": 0.9874601893508905,
        "consensus_score": 0.5399593299964925,
        "last_agent": "hrm",
        "iteration": 9,
        "query_length": 1672,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.42664222096266324,
        "trm_confidence": 0.1000087329786002,
        "mcts_value": 0.6343635619799833,
        "consensus_score": 0.38564670521397426,
        "last_agent": "hrm",
        "iteration": 5,
        "query_length": 98,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6376250648109357,
        "trm_confidence": 0.5212521749410738,
        "mcts_value": 0.6320077702106438,
        "consensus_score": 0.6833264663421412,
        "last_agent": "mcts",
        "iteration": 7,
        "query_length": 3425,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.431414967236463,
        "trm_confidence": 0.2460915708799335,
        "mcts_value": 0.7433629170008382,
        "consensus_score": 0.5228777985021462,
        "last_agent": "mcts",
        "iteration": 9,
        "query_length": 796,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.36891543450466274,
        "trm_confidence": 0.15641123171088717,
        "mcts_value": 0.697120372098111,
        "consensus_score": 0.31499351078862187,
        "last_agent": "trm",
        "iteration": 8,
        "query_length": 3342,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.06133603798457597,
        "trm_confidence": 0.6968683253433593,
        "mcts_value": 0.8895709118212334,
        "consensus_score": 0.6481281384391293,
        "last_agent": "none",
        "iteration": 7,
        "query_length": 660,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5162963128563078,
        "trm_confidence": 0.6154580997694961,
        "mcts_value": 0.6974274223222228,
        "consensus_score": 0.6726605315321647,
        "last_agent": "mcts",
        "iteration": 9,
        "query_length": 4994,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.4834933691149103,
        "trm_confidence": 0.09271597870094306,
        "mcts_value": 0.9279959791115091,
        "consensus_score": 0.4242386564553582,
        "last_agent": "mcts",
        "iteration": 9,
        "query_length": 519,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6350832535941797,
        "trm_confidence": 0.46171324980985295,
        "mcts_value": 0.807750446876551,
        "consensus_score": 0.6140783324445211,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 4002,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.3950473148643869,
        "trm_confidence": 0.3561326175534258,
        "mcts_value": 0.9041101138745393,
        "consensus_score": 0.5971434892949538,
        "last_agent": "mcts",
        "iteration": 8,
        "query_length": 3863,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5284664924573852,
        "trm_confidence": 0.359632315666183,
        "mcts_value": 0.6250207145109178,
        "consensus_score": 0.4193117982181257,
        "last_agent": "trm",
        "iteration": 8,
        "query_length": 2544,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5556592843548339,
        "trm_confidence": 0.21017367604051104,
        "mcts_value": 0.8094508678178738,
        "consensus_score": 0.46547605311069346,
        "last_agent": "hrm",
        "iteration": 6,
        "query_length": 1578,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5918639565147593,
        "trm_confidence": 0.5046427330550736,
        "mcts_value": 0.8783734409608674,
        "consensus_score": 0.593090888657487,
        "last_agent": "hrm",
        "iteration": 6,
        "query_length": 2486,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.17372031475182587,
        "trm_confidence": 0.36777085299629275,
        "mcts_value": 0.8746896798232204,
        "consensus_score": 0.5304262511627166,
        "last_agent": "hrm",
        "iteration": 8,
        "query_length": 4420,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.45735500062212225,
        "trm_confidence": 0.3846274245379609,
        "mcts_value": 0.7331467002465364,
        "consensus_score": 0.4443217392962154,
        "last_agent": "mcts",
        "iteration": 10,
        "query_length": 4435,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.13890753564258362,
        "trm_confidence": 0.2638956732793761,
        "mcts_value": 0.7893444052789862,
        "consensus_score": 0.3639475105300018,
        "last_agent": "trm",
        "iteration": 5,
        "query_length": 3220,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.622438379720267,
        "trm_confidence": 0.6589407839330126,
        "mcts_value": 0.7849101441121128,
        "consensus_score": 0.6326577884340824,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 4804,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6045200488979079,
        "trm_confidence": 0.016588181483114306,
        "mcts_value": 0.9979445614554436,
        "consensus_score": 0.48993697408678183,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 3119,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.35960297881693765,
        "trm_confidence": 0.38811491739756926,
        "mcts_value": 0.8325627905169265,
        "consensus_score": 0.477250480649796,
        "last_agent": "hrm",
        "iteration": 8,
        "query_length": 2855,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5141036811251816,
        "trm_confidence": 0.471118384598279,
        "mcts_value": 0.8584046895883358,
        "consensus_score": 0.5809657554330427,
        "last_agent": "hrm",
        "iteration": 9,
        "query_length": 2349,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.13660095437682052,
        "trm_confidence": 0.445975889224767,
        "mcts_value": 0.6513189291697111,
        "consensus_score": 0.3691818176349714,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 2369,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.33445363477286755,
        "trm_confidence": 0.42516981740064813,
        "mcts_value": 0.6662468722161602,
        "consensus_score": 0.4105705631007749,
        "last_agent": "none",
        "iteration": 8,
        "query_length": 4914,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.15644903475187527,
        "trm_confidence": 0.5496442951602336,
        "mcts_value": 0.7246232553575843,
        "consensus_score": 0.5688682950106064,
        "last_agent": "hrm",
        "iteration": 7,
        "query_length": 1908,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.11063520378319223,
        "trm_confidence": 0.5899335993964124,
        "mcts_value": 0.7042277265701904,
        "consensus_score": 0.3852363074032394,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 4243,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5931454356934307,
        "trm_confidence": 0.01928333490401931,
        "mcts_value": 0.8070959148296653,
        "consensus_score": 0.5286054450237737,
        "last_agent": "hrm",
        "iteration": 4,
        "query_length": 719,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.04279058165206425,
        "trm_confidence": 0.583271527160927,
        "mcts_value": 0.8116704876951447,
        "consensus_score": 0.3847070222502957,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 3913,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.6169720930483836,
        "trm_confidence": 0.44480049095375845,
        "mcts_value": 0.7859484495431563,
        "consensus_score": 0.6406127620558477,
        "last_agent": "trm",
        "iteration": 6,
        "query_length": 3837,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.0733583824003766,
        "trm_confidence": 0.3703074200514973,
        "mcts_value": 0.984468956358673,
        "consensus_score": 0.43055863777047165,
        "last_agent": "trm",
        "iteration": 5,
        "query_length": 2794,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.5440256379818951,
        "trm_confidence": 0.6081611824354874,
        "mcts_value": 0.9191736468715476,
        "consensus_score": 0.6490340366787835,
        "last_agent": "trm",
        "iteration": 7,
        "query_length": 3514,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.21164017849959813,
        "trm_confidence": 0.39304547328913053,
        "mcts_value": 0.8987430492545462,
        "consensus_score": 0.43500295675164846,
        "last_agent": "trm",
        "iteration": 9,
        "query_length": 3651,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.03215589797470598,
        "trm_confidence": 0.6947210968425096,
        "mcts_value": 0.9509811659011822,
        "consensus_score": 0.5088419364607382,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 2991,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.3868629492725526,
        "trm_confidence": 0.1678673810126457,
        "mcts_value": 0.707013079346613,
        "consensus_score": 0.39933539882401803,
        "last_agent": "trm",
        "iteration": 8,
        "query_length": 4828,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.4160259913511262,
        "trm_confidence": 0.10340999377874809,
        "mcts_value": 0.8770966001067517,
        "consensus_score": 0.5287339820537985,
        "last_agent": "mcts",
        "iteration": 4,
        "query_length": 2363,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.46720028290912374,
        "trm_confidence": 0.2578597578154601,
        "mcts_value": 0.6228235667727845,
        "consensus_score": 0.4797042324734294,
        "last_agent": "mcts",
        "iteration": 6,
        "query_length": 1202,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.03190645756728122,
        "trm_confidence": 0.30115856844220246,
        "mcts_value": 0.7831844040664724,
        "consensus_score": 0.4002960159816925,
        "last_agent": "mcts",
        "iteration": 7,
        "query_length": 433,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.21620392844068265,
        "trm_confidence": 0.010104361945315464,
        "mcts_value": 0.8672223942746995,
        "consensus_score": 0.2956293396370816,
        "last_agent": "mcts",
        "iteration": 7,
        "query_length": 800,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.1684574328960351,
        "trm_confidence": 0.540957317932618,
        "mcts_value": 0.8777532055014203,
        "consensus_score": 0.48635142172215373,
        "last_agent": "hrm",
        "iteration": 8,
        "query_length": 3036,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.03276345927082932,
        "trm_confidence": 0.3748118647412166,
        "mcts_value": 0.7542224514574996,
        "consensus_score": 0.4863477374654152,
        "last_agent": "none",
        "iteration": 10,
        "query_length": 3405,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.24146628227404615,
        "trm_confidence": 0.11219073612816403,
        "mcts_value": 0.9022786037697009,
        "consensus_score": 0.4518281475111357,
        "last_agent": "mcts",
        "iteration": 7,
        "query_length": 683,
        "has_rag_context": false
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.23558057312119612,
        "trm_confidence": 0.46696633168896406,
        "mcts_value": 0.9309570648942034,
        "consensus_score": 0.5177396166273129,
        "last_agent": "trm",
        "iteration": 4,
        "query_length": 2024,
        "has_rag_context": true
      },
      "label": "mcts"
    },
    {
      "features": {
        "hrm_confidence": 0.3382837374618919,
        "trm_confidence": 0.07576130578580771,
        "mcts_value": 0.9800137852816538,
        "consensus_score": 0.5022516420516963,
        "last_agent": "trm",
        "iteration": 10,
        "query_length": 2337,
        "has_rag_context": false
      },
      "label": "mcts"
    }
  ]
}