{ "checkpoint": "models/v000-mean", "quantized": true, "platform": "Windows-11-10.0.26200-SP0", "torch_threads": 2, "model_load_ms": 25.02419997472316, "samples": 120, "success_rate": 1.0, "abstention_rate": 0.0, "median_observation_ms": 14.41470003919676, "median_policy_ms": 3.57734999852255, "median_execution_ms": 17.8553499863483, "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": 37.62129996903241, "policy_ms": 12.632400030270219, "execution_ms": 71.28469995222986, "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": 15.9093999536708, "policy_ms": 3.0161000322550535, "execution_ms": 25.469899992458522, "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": 18.087800010107458, "policy_ms": 5.51899999845773, "execution_ms": 26.201299973763525, "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": 15.898199984803796, "policy_ms": 5.092400009743869, "execution_ms": 21.35419996920973, "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": 15.813399921171367, "policy_ms": 5.034400033764541, "execution_ms": 17.902899999171495, "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": 16.981099965050817, "policy_ms": 5.512900068424642, "execution_ms": 18.0193999549374, "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": 14.82200005557388, "policy_ms": 5.799200036562979, "execution_ms": 17.239099950529635, "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": 13.865399989299476, "policy_ms": 6.007699994370341, "execution_ms": 33.872000058181584, "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": 17.255400074645877, "policy_ms": 5.548899993300438, "execution_ms": 19.663000013679266, "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": 19.132500048726797, "policy_ms": 5.646299920044839, "execution_ms": 19.738900009542704, "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": 12.606699950993061, "policy_ms": 3.804299980401993, "execution_ms": 16.540999989956617, "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": 14.412199961952865, "policy_ms": 4.534500068984926, "execution_ms": 16.125399968586862, "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": 15.338700031861663, "policy_ms": 5.266900057904422, "execution_ms": 17.172600026242435, "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": 15.84109989926219, "policy_ms": 5.727100069634616, "execution_ms": 16.551899956539273, "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": 16.812600078992546, "policy_ms": 3.6643000785261393, "execution_ms": 16.701900050975382, "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": 15.536199905909598, "policy_ms": 6.461300072260201, "execution_ms": 21.07290003914386, "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": 18.391499994322658, "policy_ms": 4.103200044482946, "execution_ms": 18.96160002797842, "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": 13.639400014653802, "policy_ms": 2.788799931295216, "execution_ms": 17.299399944022298, "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": 13.328199973329902, "policy_ms": 3.21670004632324, "execution_ms": 22.127500036731362, "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": 13.830000068992376, "policy_ms": 6.633899989537895, "execution_ms": 19.3853999953717, "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": 18.035999964922667, "policy_ms": 4.082100000232458, "execution_ms": 16.994900070130825, "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": 15.832200064323843, "policy_ms": 5.283599952235818, "execution_ms": 17.181200091727078, "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": 15.869900002144277, "policy_ms": 3.853800008073449, "execution_ms": 19.205500022508204, "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": 13.350500026717782, "policy_ms": 2.565900096669793, "execution_ms": 16.938899992965162, "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": 14.414899982511997, "policy_ms": 4.378500045277178, "execution_ms": 17.37589994445443, "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": 17.135400092229247, "policy_ms": 4.474699962884188, "execution_ms": 17.42829999420792, "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": 16.44999999552965, "policy_ms": 3.6712000146508217, "execution_ms": 28.61389995086938, "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": 14.34950006660074, "policy_ms": 4.304699948988855, "execution_ms": 17.689900007098913, "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": 13.385800062678754, "policy_ms": 2.7074000099673867, "execution_ms": 15.752900042571127, "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": 13.923300080932677, "policy_ms": 3.7194000324234366, "execution_ms": 53.11719991732389, "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": 15.35100000910461, "policy_ms": 4.512999905273318, "execution_ms": 18.279499956406653, "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": 16.698100022040308, "policy_ms": 4.132899921387434, "execution_ms": 53.41589997988194, "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": 14.364800066687167, "policy_ms": 4.093099967576563, "execution_ms": 16.438699909485877, "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": 14.601000002585351, "policy_ms": 2.8792000375688076, "execution_ms": 19.969500019215047, "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": 13.527199975214899, "policy_ms": 2.602200023829937, "execution_ms": 35.12849996332079, "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": 13.413999928161502, "policy_ms": 5.127100041136146, "execution_ms": 17.4235999584198, "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": 14.238900039345026, "policy_ms": 4.187999991700053, "execution_ms": 16.33759995456785, "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": 16.04399993084371, "policy_ms": 3.3961000153794885, "execution_ms": 15.964599908329546, "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": 26.540399994701147, "policy_ms": 4.631899995729327, "execution_ms": 19.121500081382692, "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": 15.889800037257373, "policy_ms": 3.1840000301599503, "execution_ms": 17.70540000870824, "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": 15.134099987335503, "policy_ms": 3.0033999355509877, "execution_ms": 16.632999991998076, "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": 14.889399986714125, "policy_ms": 3.1727999448776245, "execution_ms": 25.77980002388358, "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": 14.59840009920299, "policy_ms": 4.682299913838506, "execution_ms": 16.406500013545156, "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": 14.68430005479604, "policy_ms": 4.088599933311343, "execution_ms": 28.929500025697052, "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": 14.376199920661747, "policy_ms": 4.986200015991926, "execution_ms": 16.860099975019693, "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": 16.623200033791363, "policy_ms": 4.373600007966161, "execution_ms": 18.432200071401894, "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": 14.561800053343177, "policy_ms": 3.499399987049401, "execution_ms": 27.99790003336966, "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": 14.240399934351444, "policy_ms": 5.295600043609738, "execution_ms": 15.72590006981045, "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": 13.309900066815317, "policy_ms": 2.8980999486520886, "execution_ms": 15.681900084018707, "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": 14.854899956844747, "policy_ms": 3.838400007225573, "execution_ms": 55.447100079618394, "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": 13.326299958862364, "policy_ms": 2.8696999652311206, "execution_ms": 23.991799913346767, "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": 13.349700020626187, "policy_ms": 3.5822999197989702, "execution_ms": 56.39919999521226, "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": 14.78339999448508, "policy_ms": 3.729499992914498, "execution_ms": 16.672699945047498, "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": 14.769100002013147, "policy_ms": 3.4248000010848045, "execution_ms": 29.20390001963824, "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": 13.078800053335726, "policy_ms": 3.207899979315698, "execution_ms": 15.42360009625554, "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": 14.118699938990176, "policy_ms": 6.170599954202771, "execution_ms": 20.675099920481443, "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": 14.673000085167587, "policy_ms": 3.6236000014469028, "execution_ms": 15.989199979230762, "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": 14.943099929951131, "policy_ms": 2.6851000729948282, "execution_ms": 15.859800041653216, "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": 13.816399965435266, "policy_ms": 3.6018999526277184, "execution_ms": 17.008199938572943, "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": 15.36870002746582, "policy_ms": 3.6346999695524573, "execution_ms": 53.78079996444285, "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": 13.50350002758205, "policy_ms": 3.5181998973712325, "execution_ms": 15.322600025683641, "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": 17.149899969808757, "policy_ms": 2.9458000790327787, "execution_ms": 16.737199970521033, "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": 14.436300029046834, "policy_ms": 3.234799951314926, "execution_ms": 17.42679998278618, "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": 17.384600010700524, "policy_ms": 5.293700029142201, "execution_ms": 18.279399955645204, "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": 16.040600021369755, "policy_ms": 3.9652999257668853, "execution_ms": 17.908999929204583, "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": 14.787400024943054, "policy_ms": 3.342699958011508, "execution_ms": 17.94659998267889, "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": 14.00720002129674, "policy_ms": 2.9574999352917075, "execution_ms": 32.74679998867214, "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": 13.619599980302155, "policy_ms": 3.4682999830693007, "execution_ms": 17.2487999079749, "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": 14.932500082068145, "policy_ms": 3.6065001040697098, "execution_ms": 18.082199967466295, "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": 13.849299983121455, "policy_ms": 3.2056999625638127, "execution_ms": 15.345799969509244, "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": 13.045099913142622, "policy_ms": 3.0004000291228294, "execution_ms": 15.859699924476445, "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": 12.964699999429286, "policy_ms": 2.8626000275835395, "execution_ms": 15.342899947427213, "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": 13.942399993538857, "policy_ms": 2.713400055654347, "execution_ms": 15.963400015607476, "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": 16.47229993250221, "policy_ms": 4.1096999775618315, "execution_ms": 54.6372999669984, "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": 12.743699946440756, "policy_ms": 2.6798000326380134, "execution_ms": 15.190300066024065, "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": 12.755700037814677, "policy_ms": 3.222699975594878, "execution_ms": 15.89470007456839, "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": 16.391800018027425, "policy_ms": 3.490199916996062, "execution_ms": 25.451899971812963, "status": "ok", "code": "CLICK" }, { "sample_id": 77, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": 0.9999997615814209, "target_confidence": 1.0, "observation_ms": 13.050799956545234, "policy_ms": 3.091600025072694, "execution_ms": 16.81780000217259, "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": 14.967800001613796, "policy_ms": 3.2140000257641077, "execution_ms": 17.405399936251342, "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": 14.972600038163364, "policy_ms": 2.8268001042306423, "execution_ms": 15.729099977761507, "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": 14.751699985936284, "policy_ms": 3.3696999307721853, "execution_ms": 15.580100007355213, "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": 12.43179989978671, "policy_ms": 2.8056000592187047, "execution_ms": 17.807799973525107, "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": 14.738199999555945, "policy_ms": 3.5724000772461295, "execution_ms": 44.675700017251074, "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": 12.236100039444864, "policy_ms": 2.6415999745950103, "execution_ms": 16.449499991722405, "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": 13.924599974416196, "policy_ms": 3.60769999679178, "execution_ms": 17.42260006722063, "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": 13.466799980960786, "policy_ms": 2.74529994931072, "execution_ms": 16.5047999471426, "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": 13.812400051392615, "policy_ms": 2.6603000005707145, "execution_ms": 16.211100039072335, "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": 15.411000000312924, "policy_ms": 4.1949000442400575, "execution_ms": 16.126899980008602, "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": 13.425800018012524, "policy_ms": 2.6032000314444304, "execution_ms": 31.431799987331033, "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": 13.225400005467236, "policy_ms": 3.0793999321758747, "execution_ms": 23.149600019678473, "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": 13.442200026474893, "policy_ms": 3.201100043952465, "execution_ms": 15.241599990986288, "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": 13.873999938368797, "policy_ms": 3.424399998039007, "execution_ms": 48.66520001087338, "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": 13.487900025211275, "policy_ms": 3.0587000073865056, "execution_ms": 22.639399976469576, "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": 14.161800034344196, "policy_ms": 2.711100038141012, "execution_ms": 22.43150002323091, "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": 12.239300063811243, "policy_ms": 2.923400024883449, "execution_ms": 25.97399998921901, "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": 19.015799975022674, "policy_ms": 8.061100030317903, "execution_ms": 62.51029996201396, "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": 20.582800032570958, "policy_ms": 4.975900053977966, "execution_ms": 59.358500060625374, "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": 14.282400021329522, "policy_ms": 3.287100000306964, "execution_ms": 15.02649998292327, "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": 15.70639992132783, "policy_ms": 4.549300065264106, "execution_ms": 29.984500026330352, "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": 13.901899917982519, "policy_ms": 2.962200087495148, "execution_ms": 18.34650000091642, "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": 15.403400058858097, "policy_ms": 3.374099964275956, "execution_ms": 24.916700087487698, "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": 14.324600109830499, "policy_ms": 3.590900101698935, "execution_ms": 22.811400005593896, "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": 12.611599988304079, "policy_ms": 2.905000001192093, "execution_ms": 15.321100014261901, "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": 14.414500095881522, "policy_ms": 3.623099997639656, "execution_ms": 16.4383000228554, "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": 15.251999953761697, "policy_ms": 3.1798999989405274, "execution_ms": 16.58080006018281, "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": 15.724699944257736, "policy_ms": 4.033600096590817, "execution_ms": 15.671700006350875, "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": 12.63809995725751, "policy_ms": 2.73499998729676, "execution_ms": 15.813999925740063, "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": 14.218099997378886, "policy_ms": 3.1020999886095524, "execution_ms": 29.38760002143681, "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": 14.07080003991723, "policy_ms": 4.406499909237027, "execution_ms": 53.650799905881286, "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": 18.183300038799644, "policy_ms": 3.753499942831695, "execution_ms": 52.17579996678978, "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": 13.622700003907084, "policy_ms": 3.5475000040605664, "execution_ms": 56.21059995610267, "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": 13.845799956470728, "policy_ms": 3.569000051356852, "execution_ms": 21.023000008426607, "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": 12.917899992316961, "policy_ms": 2.942000050097704, "execution_ms": 16.059600049629807, "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": 14.517100062221289, "policy_ms": 3.9793000323697925, "execution_ms": 47.290600021369755, "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": 13.212600024417043, "policy_ms": 3.5225999308750033, "execution_ms": 16.48279989603907, "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": 13.99080001283437, "policy_ms": 3.449199954047799, "execution_ms": 16.65330003015697, "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": 12.152399984188378, "policy_ms": 2.692400012165308, "execution_ms": 26.97710006032139, "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": 13.995500048622489, "policy_ms": 4.225799930281937, "execution_ms": 22.803800064139068, "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": 14.969900017604232, "policy_ms": 3.659200039692223, "execution_ms": 54.38180000055581, "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": 12.350099976174533, "policy_ms": 2.7945999754592776, "execution_ms": 15.37889998871833, "status": "ok", "code": "SELECT" } ] }