{ "checkpoint": "models/v000-mean", "quantized": false, "platform": "Windows-11-10.0.26200-SP0", "torch_threads": 2, "model_load_ms": 94.83249997720122, "samples": 120, "success_rate": 1.0, "abstention_rate": 0.0, "median_observation_ms": 15.581549960188568, "median_policy_ms": 2.534949977416545, "median_execution_ms": 18.842949939426035, "scope": "single-step generated browser fixtures, held-out layout templates; not arbitrary websites", "target_vps_validated": false, "results": [ { "sample_id": 0, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 39.26530003082007, "policy_ms": 4.47579997126013, "execution_ms": 85.4514999082312, "status": "ok", "code": "CLICK" }, { "sample_id": 1, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 16.616499982774258, "policy_ms": 2.3729000240564346, "execution_ms": 28.698400012217462, "status": "ok", "code": "SELECT" }, { "sample_id": 2, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 19.65469995047897, "policy_ms": 2.9569000471383333, "execution_ms": 68.5606999322772, "status": "ok", "code": "CLICK" }, { "sample_id": 3, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 17.979600001126528, "policy_ms": 2.8717999812215567, "execution_ms": 23.14290008507669, "status": "ok", "code": "TYPE" }, { "sample_id": 4, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 19.739100011065602, "policy_ms": 2.7577999280765653, "execution_ms": 19.49949993286282, "status": "ok", "code": "TYPE" }, { "sample_id": 5, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 16.746200039051473, "policy_ms": 2.7707999106496572, "execution_ms": 19.38810001593083, "status": "ok", "code": "SELECT" }, { "sample_id": 6, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 21.029299939982593, "policy_ms": 2.919499995186925, "execution_ms": 18.884699908085167, "status": "ok", "code": "SELECT" }, { "sample_id": 7, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 15.311800059862435, "policy_ms": 2.2350000217556953, "execution_ms": 26.880200020968914, "status": "ok", "code": "CLICK" }, { "sample_id": 8, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 19.12599999923259, "policy_ms": 3.3074000384658575, "execution_ms": 19.500200054608285, "status": "ok", "code": "TYPE" }, { "sample_id": 9, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 18.3537999400869, "policy_ms": 2.814000006765127, "execution_ms": 19.55550001002848, "status": "ok", "code": "SELECT" }, { "sample_id": 10, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 13.043599901720881, "policy_ms": 1.78980000782758, "execution_ms": 17.746999976225197, "status": "ok", "code": "TYPE" }, { "sample_id": 11, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 13.925999985076487, "policy_ms": 2.047999994829297, "execution_ms": 18.801199970766902, "status": "ok", "code": "SELECT" }, { "sample_id": 12, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 17.190100043080747, "policy_ms": 3.3161999890580773, "execution_ms": 19.15890001691878, "status": "ok", "code": "SELECT" }, { "sample_id": 13, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 17.293600016273558, "policy_ms": 2.692400012165308, "execution_ms": 20.335300010629, "status": "ok", "code": "SELECT" }, { "sample_id": 14, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 16.501400037668645, "policy_ms": 2.9487001011148095, "execution_ms": 16.472699935548007, "status": "ok", "code": "SELECT" }, { "sample_id": 15, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 14.902099967002869, "policy_ms": 2.5654000928625464, "execution_ms": 18.962700036354363, "status": "ok", "code": "TYPE" }, { "sample_id": 16, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 16.126300091855228, "policy_ms": 2.944599953480065, "execution_ms": 19.645099993795156, "status": "ok", "code": "TYPE" }, { "sample_id": 17, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 14.05900006648153, "policy_ms": 2.0590999629348516, "execution_ms": 19.359700032509863, "status": "ok", "code": "TYPE" }, { "sample_id": 18, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 14.621500042267144, "policy_ms": 2.1886000176891685, "execution_ms": 31.923800008371472, "status": "ok", "code": "CLICK" }, { "sample_id": 19, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 15.2591000078246, "policy_ms": 3.3369000302627683, "execution_ms": 18.572899978607893, "status": "ok", "code": "SELECT" }, { "sample_id": 20, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 16.603299998678267, "policy_ms": 3.0505999457091093, "execution_ms": 17.083900049328804, "status": "ok", "code": "SELECT" }, { "sample_id": 21, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 14.564299955964088, "policy_ms": 1.7336000455543399, "execution_ms": 16.190499998629093, "status": "ok", "code": "SELECT" }, { "sample_id": 22, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 14.64409998152405, "policy_ms": 2.5949999690055847, "execution_ms": 16.738600097596645, "status": "ok", "code": "SELECT" }, { "sample_id": 23, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 15.111799933947623, "policy_ms": 1.7996999667957425, "execution_ms": 17.54359994083643, "status": "ok", "code": "TYPE" }, { "sample_id": 24, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 14.35959991067648, "policy_ms": 2.0147000905126333, "execution_ms": 19.103000056929886, "status": "ok", "code": "SELECT" }, { "sample_id": 25, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 15.610000002197921, "policy_ms": 2.202000003308058, "execution_ms": 18.449399969540536, "status": "ok", "code": "TYPE" }, { "sample_id": 26, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 17.05190003849566, "policy_ms": 3.314899979159236, "execution_ms": 25.99940006621182, "status": "ok", "code": "CLICK" }, { "sample_id": 27, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 17.0339000178501, "policy_ms": 2.5194999761879444, "execution_ms": 17.94970000628382, "status": "ok", "code": "TYPE" }, { "sample_id": 28, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 13.315899996086955, "policy_ms": 2.3463000543415546, "execution_ms": 15.677900053560734, "status": "ok", "code": "SELECT" }, { "sample_id": 29, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 14.399600098840892, "policy_ms": 2.669400069862604, "execution_ms": 49.62589999195188, "status": "ok", "code": "CLICK" }, { "sample_id": 30, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 15.881099971011281, "policy_ms": 2.9946999857202172, "execution_ms": 17.621599952690303, "status": "ok", "code": "SELECT" }, { "sample_id": 31, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 17.598200007341802, "policy_ms": 4.308599978685379, "execution_ms": 55.15409994404763, "status": "ok", "code": "CLICK" }, { "sample_id": 32, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 16.175299999304116, "policy_ms": 3.6958999698981643, "execution_ms": 17.239099950529635, "status": "ok", "code": "SELECT" }, { "sample_id": 33, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 15.001500025391579, "policy_ms": 3.2338000601157546, "execution_ms": 17.71430007647723, "status": "ok", "code": "TYPE" }, { "sample_id": 34, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 14.201799989677966, "policy_ms": 1.9114000024273992, "execution_ms": 47.16259997803718, "status": "ok", "code": "CLICK" }, { "sample_id": 35, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 13.829900068230927, "policy_ms": 2.2473999997600913, "execution_ms": 17.78350002132356, "status": "ok", "code": "TYPE" }, { "sample_id": 36, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 14.632800011895597, "policy_ms": 2.2116999607533216, "execution_ms": 19.842099980451167, "status": "ok", "code": "TYPE" }, { "sample_id": 37, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 17.417200026102364, "policy_ms": 2.8152000159025192, "execution_ms": 18.481899984180927, "status": "ok", "code": "SELECT" }, { "sample_id": 38, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 19.42229992710054, "policy_ms": 2.698300057090819, "execution_ms": 16.883800039067864, "status": "ok", "code": "SELECT" }, { "sample_id": 39, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 15.06789994891733, "policy_ms": 2.0670000230893493, "execution_ms": 18.08330009225756, "status": "ok", "code": "TYPE" }, { "sample_id": 40, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 15.407900093123317, "policy_ms": 2.378400065936148, "execution_ms": 17.94380007777363, "status": "ok", "code": "SELECT" }, { "sample_id": 41, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 16.018499969504774, "policy_ms": 2.269500051625073, "execution_ms": 30.739600071683526, "status": "ok", "code": "CLICK" }, { "sample_id": 42, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 14.434600016102195, "policy_ms": 2.7735999319702387, "execution_ms": 18.083500093780458, "status": "ok", "code": "TYPE" }, { "sample_id": 43, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 14.005900011397898, "policy_ms": 1.9365999614819884, "execution_ms": 32.856800011359155, "status": "ok", "code": "CLICK" }, { "sample_id": 44, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 19.68469994608313, "policy_ms": 3.2866999972611666, "execution_ms": 19.181599956937134, "status": "ok", "code": "TYPE" }, { "sample_id": 45, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 18.391699995845556, "policy_ms": 3.861800068989396, "execution_ms": 18.559099989943206, "status": "ok", "code": "SELECT" }, { "sample_id": 46, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 16.39400003477931, "policy_ms": 3.443000023253262, "execution_ms": 46.59209994133562, "status": "ok", "code": "CLICK" }, { "sample_id": 47, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 16.107599949464202, "policy_ms": 2.916299970820546, "execution_ms": 17.163899959996343, "status": "ok", "code": "SELECT" }, { "sample_id": 48, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 14.17889993172139, "policy_ms": 2.099800040014088, "execution_ms": 16.693499987013638, "status": "ok", "code": "SELECT" }, { "sample_id": 49, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 17.555300029926002, "policy_ms": 2.5318999541923404, "execution_ms": 49.185299896635115, "status": "ok", "code": "CLICK" }, { "sample_id": 50, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 14.054999919608235, "policy_ms": 2.53800000064075, "execution_ms": 57.8451999463141, "status": "ok", "code": "CLICK" }, { "sample_id": 51, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 15.612000017426908, "policy_ms": 3.152600023895502, "execution_ms": 55.50979997497052, "status": "ok", "code": "CLICK" }, { "sample_id": 52, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 15.644900035113096, "policy_ms": 2.7839000103995204, "execution_ms": 16.426200047135353, "status": "ok", "code": "TYPE" }, { "sample_id": 53, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 15.602499945089221, "policy_ms": 2.599300001747906, "execution_ms": 24.658999987877905, "status": "ok", "code": "CLICK" }, { "sample_id": 54, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 14.030199963599443, "policy_ms": 2.3668999783694744, "execution_ms": 16.39700005762279, "status": "ok", "code": "SELECT" }, { "sample_id": 55, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 0.9999998807907104, "target_confidence": 1.0, "observation_ms": 15.021800063550472, "policy_ms": 2.4136999854817986, "execution_ms": 19.459699979051948, "status": "ok", "code": "TYPE" }, { "sample_id": 56, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 16.063899965956807, "policy_ms": 3.0098999850451946, "execution_ms": 17.34220003709197, "status": "ok", "code": "SELECT" }, { "sample_id": 57, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 14.096200000494719, "policy_ms": 2.0308999810367823, "execution_ms": 16.66500000283122, "status": "ok", "code": "SELECT" }, { "sample_id": 58, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 14.93159995879978, "policy_ms": 2.5896000443026423, "execution_ms": 17.5166999688372, "status": "ok", "code": "TYPE" }, { "sample_id": 59, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 15.453699976205826, "policy_ms": 2.8801000444218516, "execution_ms": 54.95730007532984, "status": "ok", "code": "CLICK" }, { "sample_id": 60, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 13.38929997291416, "policy_ms": 2.2096000611782074, "execution_ms": 16.57179999165237, "status": "ok", "code": "SELECT" }, { "sample_id": 61, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 16.543000005185604, "policy_ms": 2.3801999632269144, "execution_ms": 19.89230001345277, "status": "ok", "code": "SELECT" }, { "sample_id": 62, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 15.148099977523088, "policy_ms": 2.1922000451013446, "execution_ms": 17.90219999384135, "status": "ok", "code": "TYPE" }, { "sample_id": 63, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 16.04280003812164, "policy_ms": 3.397200023755431, "execution_ms": 16.57950005028397, "status": "ok", "code": "SELECT" }, { "sample_id": 64, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 17.782900016754866, "policy_ms": 2.902599982917309, "execution_ms": 18.50959996227175, "status": "ok", "code": "TYPE" }, { "sample_id": 65, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 13.689000043086708, "policy_ms": 2.0917000947520137, "execution_ms": 17.4428999889642, "status": "ok", "code": "TYPE" }, { "sample_id": 66, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 14.260300085879862, "policy_ms": 2.33820010907948, "execution_ms": 23.1884999666363, "status": "ok", "code": "CLICK" }, { "sample_id": 67, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 13.83519999217242, "policy_ms": 2.8029000386595726, "execution_ms": 18.58679996803403, "status": "ok", "code": "TYPE" }, { "sample_id": 68, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 14.25320003181696, "policy_ms": 2.2649000165984035, "execution_ms": 17.223400063812733, "status": "ok", "code": "SELECT" }, { "sample_id": 69, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 13.41770007275045, "policy_ms": 2.2086999379098415, "execution_ms": 17.492099897935987, "status": "ok", "code": "SELECT" }, { "sample_id": 70, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 14.889099984429777, "policy_ms": 2.354399999603629, "execution_ms": 18.99259991478175, "status": "ok", "code": "SELECT" }, { "sample_id": 71, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 15.118100098334253, "policy_ms": 2.049700007773936, "execution_ms": 16.765999956987798, "status": "ok", "code": "SELECT" }, { "sample_id": 72, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 13.800599961541593, "policy_ms": 2.3164000594988465, "execution_ms": 16.07240003068, "status": "ok", "code": "SELECT" }, { "sample_id": 73, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 16.42070000525564, "policy_ms": 2.664300031028688, "execution_ms": 48.22770005557686, "status": "ok", "code": "CLICK" }, { "sample_id": 74, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 14.748099958524108, "policy_ms": 2.309100003913045, "execution_ms": 17.563499975949526, "status": "ok", "code": "SELECT" }, { "sample_id": 75, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 15.617599943652749, "policy_ms": 2.6485000271350145, "execution_ms": 19.47709999512881, "status": "ok", "code": "TYPE" }, { "sample_id": 76, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 16.095299972221255, "policy_ms": 2.246699994429946, "execution_ms": 36.16459993645549, "status": "ok", "code": "CLICK" }, { "sample_id": 77, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 0.9999996423721313, "target_confidence": 1.0, "observation_ms": 13.93000001553446, "policy_ms": 1.8730999436229467, "execution_ms": 17.385800019837916, "status": "ok", "code": "TYPE" }, { "sample_id": 78, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 14.462099992670119, "policy_ms": 1.9548999844118953, "execution_ms": 17.557399929501116, "status": "ok", "code": "TYPE" }, { "sample_id": 79, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 12.694700038991868, "policy_ms": 1.8824998987838626, "execution_ms": 16.00139995571226, "status": "ok", "code": "SELECT" }, { "sample_id": 80, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 15.65910002682358, "policy_ms": 2.6213000528514385, "execution_ms": 16.734899953007698, "status": "ok", "code": "SELECT" }, { "sample_id": 81, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 13.176799984648824, "policy_ms": 1.9902000203728676, "execution_ms": 17.17170001938939, "status": "ok", "code": "TYPE" }, { "sample_id": 82, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 16.945500043220818, "policy_ms": 2.790499944239855, "execution_ms": 28.858299949206412, "status": "ok", "code": "CLICK" }, { "sample_id": 83, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 17.11849996354431, "policy_ms": 1.7516999505460262, "execution_ms": 17.47860002797097, "status": "ok", "code": "TYPE" }, { "sample_id": 84, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 16.13450003787875, "policy_ms": 3.758199978619814, "execution_ms": 18.592400010675192, "status": "ok", "code": "TYPE" }, { "sample_id": 85, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 15.310899936594069, "policy_ms": 1.8527000211179256, "execution_ms": 16.89169998280704, "status": "ok", "code": "SELECT" }, { "sample_id": 86, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 13.286700006574392, "policy_ms": 2.0032000029459596, "execution_ms": 15.58500004466623, "status": "ok", "code": "SELECT" }, { "sample_id": 87, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 14.731299947015941, "policy_ms": 2.5716000236570835, "execution_ms": 15.785299940034747, "status": "ok", "code": "SELECT" }, { "sample_id": 88, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 16.0413000266999, "policy_ms": 1.6233999049291015, "execution_ms": 63.821999938227236, "status": "ok", "code": "CLICK" }, { "sample_id": 89, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 23.811299935914576, "policy_ms": 3.3231000415980816, "execution_ms": 33.75049994792789, "status": "ok", "code": "CLICK" }, { "sample_id": 90, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 20.766600035130978, "policy_ms": 2.5644999695941806, "execution_ms": 16.564899939112365, "status": "ok", "code": "SELECT" }, { "sample_id": 91, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 15.304500004276633, "policy_ms": 2.670199959538877, "execution_ms": 37.092900020070374, "status": "ok", "code": "CLICK" }, { "sample_id": 92, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 14.011699939146638, "policy_ms": 2.076699980534613, "execution_ms": 23.877799976617098, "status": "ok", "code": "CLICK" }, { "sample_id": 93, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 13.011000002734363, "policy_ms": 1.8856000388041139, "execution_ms": 24.29840003605932, "status": "ok", "code": "CLICK" }, { "sample_id": 94, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 16.379600041545928, "policy_ms": 1.985599985346198, "execution_ms": 54.59110008087009, "status": "ok", "code": "CLICK" }, { "sample_id": 95, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 16.296499990858138, "policy_ms": 3.0744000105187297, "execution_ms": 16.227299929596484, "status": "ok", "code": "SELECT" }, { "sample_id": 96, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 17.221099929884076, "policy_ms": 2.3589000338688493, "execution_ms": 26.841900078579783, "status": "ok", "code": "CLICK" }, { "sample_id": 97, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 15.73899993672967, "policy_ms": 2.517700078897178, "execution_ms": 17.348799970932305, "status": "ok", "code": "SELECT" }, { "sample_id": 98, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 18.34059995599091, "policy_ms": 3.5897999769076705, "execution_ms": 57.149400003254414, "status": "ok", "code": "CLICK" }, { "sample_id": 99, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 13.47820006776601, "policy_ms": 2.0010999869555235, "execution_ms": 19.055900047533214, "status": "ok", "code": "TYPE" }, { "sample_id": 100, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 15.886600012890995, "policy_ms": 2.6661999290809035, "execution_ms": 45.8870999282226, "status": "ok", "code": "CLICK" }, { "sample_id": 101, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 15.560599975287914, "policy_ms": 2.4163000052794814, "execution_ms": 54.38099999446422, "status": "ok", "code": "CLICK" }, { "sample_id": 102, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 16.889899969100952, "policy_ms": 2.011100063100457, "execution_ms": 16.504700062796474, "status": "ok", "code": "SELECT" }, { "sample_id": 103, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 16.29449997562915, "policy_ms": 2.780499984510243, "execution_ms": 19.519400084391236, "status": "ok", "code": "TYPE" }, { "sample_id": 104, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 13.983900076709688, "policy_ms": 2.109499997459352, "execution_ms": 17.442299984395504, "status": "ok", "code": "TYPE" }, { "sample_id": 105, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 16.304599936120212, "policy_ms": 2.9320999747142196, "execution_ms": 17.409599968232214, "status": "ok", "code": "SELECT" }, { "sample_id": 106, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 14.55680001527071, "policy_ms": 1.893899985589087, "execution_ms": 18.181099905632436, "status": "ok", "code": "SELECT" }, { "sample_id": 107, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 17.123700003139675, "policy_ms": 2.4634000146761537, "execution_ms": 48.8061000360176, "status": "ok", "code": "CLICK" }, { "sample_id": 108, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 17.840699991211295, "policy_ms": 2.7535000117495656, "execution_ms": 35.18350003287196, "status": "ok", "code": "CLICK" }, { "sample_id": 109, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 17.224400071427226, "policy_ms": 3.4622999373823404, "execution_ms": 54.714999976567924, "status": "ok", "code": "CLICK" }, { "sample_id": 110, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 15.835199970752, "policy_ms": 2.4546999484300613, "execution_ms": 37.75389993097633, "status": "ok", "code": "CLICK" }, { "sample_id": 111, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 15.625800006091595, "policy_ms": 2.4775000056251884, "execution_ms": 55.337300058454275, "status": "ok", "code": "CLICK" }, { "sample_id": 112, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 13.766699936240911, "policy_ms": 2.698600059375167, "execution_ms": 17.538800020702183, "status": "ok", "code": "SELECT" }, { "sample_id": 113, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 15.187599929049611, "policy_ms": 2.510800026357174, "execution_ms": 30.082399956882, "status": "ok", "code": "CLICK" }, { "sample_id": 114, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 13.821200001984835, "policy_ms": 3.1963998917490244, "execution_ms": 16.73319994006306, "status": "ok", "code": "SELECT" }, { "sample_id": 115, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 17.30060006957501, "policy_ms": 2.5875000283122063, "execution_ms": 17.220100038684905, "status": "ok", "code": "TYPE" }, { "sample_id": 116, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 13.151900027878582, "policy_ms": 1.94390001706779, "execution_ms": 35.48570000566542, "status": "ok", "code": "CLICK" }, { "sample_id": 117, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 14.91240004543215, "policy_ms": 2.590400050394237, "execution_ms": 40.888600051403046, "status": "ok", "code": "CLICK" }, { "sample_id": 118, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 14.83240001834929, "policy_ms": 2.840799978002906, "execution_ms": 54.049600032158196, "status": "ok", "code": "CLICK" }, { "sample_id": 119, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": 1.0, "target_confidence": 1.0, "observation_ms": 14.717500074766576, "policy_ms": 1.794799929484725, "execution_ms": 17.267199931666255, "status": "ok", "code": "SELECT" } ] }