{ "checkpoint": "lexical-role-baseline", "quantized": false, "platform": "Windows-11-10.0.26200-SP0", "torch_threads": 2, "model_load_ms": 0.001400010660290718, "samples": 120, "success_rate": 1.0, "abstention_rate": 0.0, "median_observation_ms": 14.419150014873594, "median_policy_ms": 0.6543999770656228, "median_execution_ms": 18.23655003681779, "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": null, "target_confidence": null, "observation_ms": 44.16080005466938, "policy_ms": 0.8844000985845923, "execution_ms": 67.56510003469884, "status": "ok", "code": "CLICK" }, { "sample_id": 1, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 19.639500067569315, "policy_ms": 0.615399912931025, "execution_ms": 33.42659992631525, "status": "ok", "code": "TYPE" }, { "sample_id": 2, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 16.99519995599985, "policy_ms": 0.650599948130548, "execution_ms": 18.854000023566186, "status": "ok", "code": "TYPE" }, { "sample_id": 3, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 16.24879997689277, "policy_ms": 0.4392999690026045, "execution_ms": 35.071799997240305, "status": "ok", "code": "CLICK" }, { "sample_id": 4, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 18.25359999202192, "policy_ms": 0.8140000281855464, "execution_ms": 54.35280001256615, "status": "ok", "code": "CLICK" }, { "sample_id": 5, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 16.733899945393205, "policy_ms": 1.1953000212088227, "execution_ms": 18.512400100007653, "status": "ok", "code": "TYPE" }, { "sample_id": 6, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 19.15399997960776, "policy_ms": 1.2266000267118216, "execution_ms": 21.094099967740476, "status": "ok", "code": "SELECT" }, { "sample_id": 7, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 15.534100006334484, "policy_ms": 0.9629999985918403, "execution_ms": 18.566299928352237, "status": "ok", "code": "SELECT" }, { "sample_id": 8, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 18.995100050233305, "policy_ms": 1.178800011985004, "execution_ms": 18.797900062054396, "status": "ok", "code": "SELECT" }, { "sample_id": 9, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 15.997200040146708, "policy_ms": 0.7844000356271863, "execution_ms": 21.156599977985024, "status": "ok", "code": "TYPE" }, { "sample_id": 10, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 16.03860000614077, "policy_ms": 0.4400999750941992, "execution_ms": 18.02049996331334, "status": "ok", "code": "SELECT" }, { "sample_id": 11, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 14.232499990612268, "policy_ms": 0.5302000790834427, "execution_ms": 55.32719998154789, "status": "ok", "code": "CLICK" }, { "sample_id": 12, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 16.24310004990548, "policy_ms": 0.800000037997961, "execution_ms": 17.804200062528253, "status": "ok", "code": "TYPE" }, { "sample_id": 13, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 16.25879993662238, "policy_ms": 0.6303000263869762, "execution_ms": 28.08949991595, "status": "ok", "code": "CLICK" }, { "sample_id": 14, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 16.071700025349855, "policy_ms": 0.9278999641537666, "execution_ms": 16.698800027370453, "status": "ok", "code": "SELECT" }, { "sample_id": 15, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 13.84399994276464, "policy_ms": 0.354900024831295, "execution_ms": 18.176699988543987, "status": "ok", "code": "TYPE" }, { "sample_id": 16, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 15.37020003888756, "policy_ms": 1.0884000221267343, "execution_ms": 16.861099982634187, "status": "ok", "code": "TYPE" }, { "sample_id": 17, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 17.149100080132484, "policy_ms": 0.9626999963074923, "execution_ms": 20.279100048355758, "status": "ok", "code": "SELECT" }, { "sample_id": 18, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 18.968299962580204, "policy_ms": 0.5751000717282295, "execution_ms": 26.933499961160123, "status": "ok", "code": "CLICK" }, { "sample_id": 19, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 14.034099993295968, "policy_ms": 0.672800000756979, "execution_ms": 17.166699981316924, "status": "ok", "code": "SELECT" }, { "sample_id": 20, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 15.259800013154745, "policy_ms": 0.9447999764233828, "execution_ms": 49.03200000990182, "status": "ok", "code": "CLICK" }, { "sample_id": 21, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 13.620699988678098, "policy_ms": 0.6331000477075577, "execution_ms": 27.49260002747178, "status": "ok", "code": "CLICK" }, { "sample_id": 22, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 14.103499939665198, "policy_ms": 0.5707000382244587, "execution_ms": 24.60339991375804, "status": "ok", "code": "CLICK" }, { "sample_id": 23, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 13.339199940674007, "policy_ms": 0.5837000207975507, "execution_ms": 19.531299942173064, "status": "ok", "code": "SELECT" }, { "sample_id": 24, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 15.748800011351705, "policy_ms": 0.8015000494197011, "execution_ms": 30.768100055865943, "status": "ok", "code": "CLICK" }, { "sample_id": 25, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 14.36999998986721, "policy_ms": 0.37589995190501213, "execution_ms": 23.56660005170852, "status": "ok", "code": "CLICK" }, { "sample_id": 26, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 14.280900009907782, "policy_ms": 0.7805000059306622, "execution_ms": 16.527900006622076, "status": "ok", "code": "TYPE" }, { "sample_id": 27, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 12.191000045277178, "policy_ms": 0.345399952493608, "execution_ms": 15.357499942183495, "status": "ok", "code": "SELECT" }, { "sample_id": 28, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 12.898499961011112, "policy_ms": 0.5040999967604876, "execution_ms": 16.09469996765256, "status": "ok", "code": "TYPE" }, { "sample_id": 29, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 13.182599912397563, "policy_ms": 0.3445000620558858, "execution_ms": 29.846600024029613, "status": "ok", "code": "CLICK" }, { "sample_id": 30, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 15.733600012026727, "policy_ms": 1.0106000117957592, "execution_ms": 16.95099996868521, "status": "ok", "code": "TYPE" }, { "sample_id": 31, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 15.682500088587403, "policy_ms": 0.9608999826014042, "execution_ms": 15.943999984301627, "status": "ok", "code": "SELECT" }, { "sample_id": 32, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 13.335500028915703, "policy_ms": 0.5293000722303987, "execution_ms": 16.483500017784536, "status": "ok", "code": "TYPE" }, { "sample_id": 33, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 12.888799887150526, "policy_ms": 0.44920004438608885, "execution_ms": 36.82060004211962, "status": "ok", "code": "CLICK" }, { "sample_id": 34, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 14.28410003427416, "policy_ms": 1.007299986667931, "execution_ms": 16.155699966475368, "status": "ok", "code": "TYPE" }, { "sample_id": 35, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 12.47499999590218, "policy_ms": 0.40060002356767654, "execution_ms": 36.80640005040914, "status": "ok", "code": "CLICK" }, { "sample_id": 36, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 13.016300043091178, "policy_ms": 0.48829999286681414, "execution_ms": 15.87659993674606, "status": "ok", "code": "TYPE" }, { "sample_id": 37, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 14.751999988220632, "policy_ms": 0.7793999975547194, "execution_ms": 14.470600057393312, "status": "ok", "code": "SELECT" }, { "sample_id": 38, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 14.468300039879978, "policy_ms": 0.6741000106558204, "execution_ms": 19.78059997782111, "status": "ok", "code": "TYPE" }, { "sample_id": 39, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 14.707599999383092, "policy_ms": 0.4615000216290355, "execution_ms": 23.1457999907434, "status": "ok", "code": "TYPE" }, { "sample_id": 40, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 15.119099989533424, "policy_ms": 0.8006000425666571, "execution_ms": 35.60920001473278, "status": "ok", "code": "CLICK" }, { "sample_id": 41, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 12.995700002647936, "policy_ms": 0.5558000411838293, "execution_ms": 15.2473000343889, "status": "ok", "code": "SELECT" }, { "sample_id": 42, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 14.171299990266562, "policy_ms": 0.924500054679811, "execution_ms": 15.520500019192696, "status": "ok", "code": "SELECT" }, { "sample_id": 43, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 12.335099978372455, "policy_ms": 0.39770000148564577, "execution_ms": 15.11550007853657, "status": "ok", "code": "SELECT" }, { "sample_id": 44, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 14.5817999728024, "policy_ms": 0.8651999523863196, "execution_ms": 30.093299923464656, "status": "ok", "code": "CLICK" }, { "sample_id": 45, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 11.988200014457107, "policy_ms": 0.45759999193251133, "execution_ms": 19.70379997510463, "status": "ok", "code": "SELECT" }, { "sample_id": 46, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 13.02190008573234, "policy_ms": 0.5883999401703477, "execution_ms": 14.98619990888983, "status": "ok", "code": "SELECT" }, { "sample_id": 47, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 14.309999998658895, "policy_ms": 0.6582000060006976, "execution_ms": 15.935900039039552, "status": "ok", "code": "TYPE" }, { "sample_id": 48, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 12.75830005761236, "policy_ms": 0.470499973744154, "execution_ms": 16.083700000308454, "status": "ok", "code": "TYPE" }, { "sample_id": 49, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 14.898000052198768, "policy_ms": 0.4702999722212553, "execution_ms": 15.781099908053875, "status": "ok", "code": "TYPE" }, { "sample_id": 50, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 16.115900012664497, "policy_ms": 0.3816999960690737, "execution_ms": 31.260699965059757, "status": "ok", "code": "CLICK" }, { "sample_id": 51, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 12.905900017358363, "policy_ms": 0.5972000071778893, "execution_ms": 14.75949992891401, "status": "ok", "code": "SELECT" }, { "sample_id": 52, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 13.056700001470745, "policy_ms": 0.5821000086143613, "execution_ms": 14.593300060369074, "status": "ok", "code": "SELECT" }, { "sample_id": 53, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 13.520499924197793, "policy_ms": 0.5411000456660986, "execution_ms": 51.81269999593496, "status": "ok", "code": "CLICK" }, { "sample_id": 54, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 14.77579993661493, "policy_ms": 0.979200005531311, "execution_ms": 16.562399920076132, "status": "ok", "code": "TYPE" }, { "sample_id": 55, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 14.839299954473972, "policy_ms": 0.8650999516248703, "execution_ms": 14.928199932910502, "status": "ok", "code": "SELECT" }, { "sample_id": 56, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 13.910999987274408, "policy_ms": 0.5905999569222331, "execution_ms": 29.894700041040778, "status": "ok", "code": "CLICK" }, { "sample_id": 57, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 12.491600005887449, "policy_ms": 0.6194000598043203, "execution_ms": 27.068500057794154, "status": "ok", "code": "CLICK" }, { "sample_id": 58, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 16.75559999421239, "policy_ms": 1.045699929818511, "execution_ms": 17.84610003232956, "status": "ok", "code": "SELECT" }, { "sample_id": 59, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 15.737999929115176, "policy_ms": 0.6213000742718577, "execution_ms": 17.932799994014204, "status": "ok", "code": "SELECT" }, { "sample_id": 60, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 14.018699992448092, "policy_ms": 0.5952999927103519, "execution_ms": 15.688900020904839, "status": "ok", "code": "SELECT" }, { "sample_id": 61, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 18.1864999467507, "policy_ms": 0.9354000212624669, "execution_ms": 15.699899988248944, "status": "ok", "code": "SELECT" }, { "sample_id": 62, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 15.992000000551343, "policy_ms": 0.98690006416291, "execution_ms": 15.52999997511506, "status": "ok", "code": "TYPE" }, { "sample_id": 63, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 14.69190011266619, "policy_ms": 0.7239000406116247, "execution_ms": 30.035499949008226, "status": "ok", "code": "CLICK" }, { "sample_id": 64, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 13.679000083357096, "policy_ms": 0.8476999355480075, "execution_ms": 16.526299994438887, "status": "ok", "code": "TYPE" }, { "sample_id": 65, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 12.900499976240098, "policy_ms": 0.47889992129057646, "execution_ms": 37.6344999531284, "status": "ok", "code": "CLICK" }, { "sample_id": 66, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 13.2619000505656, "policy_ms": 0.7680000271648169, "execution_ms": 24.721599998883903, "status": "ok", "code": "CLICK" }, { "sample_id": 67, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 12.547199963591993, "policy_ms": 0.49700005911290646, "execution_ms": 18.470700015313923, "status": "ok", "code": "TYPE" }, { "sample_id": 68, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 11.782900081016123, "policy_ms": 0.3533000126481056, "execution_ms": 39.68659997917712, "status": "ok", "code": "CLICK" }, { "sample_id": 69, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 12.033800012432039, "policy_ms": 0.4251999780535698, "execution_ms": 27.57269993890077, "status": "ok", "code": "CLICK" }, { "sample_id": 70, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 13.44229991082102, "policy_ms": 0.6742000114172697, "execution_ms": 25.500599993392825, "status": "ok", "code": "CLICK" }, { "sample_id": 71, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 11.803000001236796, "policy_ms": 0.38810004480183125, "execution_ms": 18.064299947582185, "status": "ok", "code": "SELECT" }, { "sample_id": 72, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 14.015699969604611, "policy_ms": 0.8902999106794596, "execution_ms": 14.916999964043498, "status": "ok", "code": "SELECT" }, { "sample_id": 73, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 13.399000046774745, "policy_ms": 0.36049995105713606, "execution_ms": 15.539199928753078, "status": "ok", "code": "SELECT" }, { "sample_id": 74, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 15.09320002514869, "policy_ms": 0.8718000026419759, "execution_ms": 16.901400056667626, "status": "ok", "code": "TYPE" }, { "sample_id": 75, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 12.592799961566925, "policy_ms": 0.542100053280592, "execution_ms": 15.988199971616268, "status": "ok", "code": "SELECT" }, { "sample_id": 76, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 11.824400047771633, "policy_ms": 0.38129999302327633, "execution_ms": 15.170400030910969, "status": "ok", "code": "SELECT" }, { "sample_id": 77, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 13.546600006520748, "policy_ms": 0.7453999714925885, "execution_ms": 15.744400094263256, "status": "ok", "code": "TYPE" }, { "sample_id": 78, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 13.20859999395907, "policy_ms": 0.5144999595358968, "execution_ms": 15.13299997895956, "status": "ok", "code": "SELECT" }, { "sample_id": 79, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 14.504700084216893, "policy_ms": 1.0025999508798122, "execution_ms": 15.456799999810755, "status": "ok", "code": "SELECT" }, { "sample_id": 80, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 13.464399962686002, "policy_ms": 0.6858999840915203, "execution_ms": 53.98220010101795, "status": "ok", "code": "CLICK" }, { "sample_id": 81, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 13.151800027117133, "policy_ms": 0.5308999679982662, "execution_ms": 25.827899924479425, "status": "ok", "code": "CLICK" }, { "sample_id": 82, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 13.75400007236749, "policy_ms": 0.9355999063700438, "execution_ms": 24.42569995764643, "status": "ok", "code": "CLICK" }, { "sample_id": 83, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 12.96979992184788, "policy_ms": 0.6698000943288207, "execution_ms": 15.828400035388768, "status": "ok", "code": "TYPE" }, { "sample_id": 84, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 14.359999913722277, "policy_ms": 0.9399000555276871, "execution_ms": 16.569300089031458, "status": "ok", "code": "TYPE" }, { "sample_id": 85, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 14.517500065267086, "policy_ms": 0.6384999724105, "execution_ms": 15.053500072099268, "status": "ok", "code": "SELECT" }, { "sample_id": 86, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 12.57750007789582, "policy_ms": 0.4347999347373843, "execution_ms": 31.86029999051243, "status": "ok", "code": "CLICK" }, { "sample_id": 87, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 17.500899964943528, "policy_ms": 0.6910000229254365, "execution_ms": 34.30270007811487, "status": "ok", "code": "CLICK" }, { "sample_id": 88, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 15.543899964541197, "policy_ms": 0.8764000376686454, "execution_ms": 23.99320004042238, "status": "ok", "code": "CLICK" }, { "sample_id": 89, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 18.68330000434071, "policy_ms": 1.850700005888939, "execution_ms": 18.987199990078807, "status": "ok", "code": "SELECT" }, { "sample_id": 90, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 18.260999931953847, "policy_ms": 0.8619000436738133, "execution_ms": 52.262900047935545, "status": "ok", "code": "CLICK" }, { "sample_id": 91, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 15.592999989166856, "policy_ms": 0.7587999571114779, "execution_ms": 16.66369999293238, "status": "ok", "code": "TYPE" }, { "sample_id": 92, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 16.388299991376698, "policy_ms": 0.9327000007033348, "execution_ms": 16.459499951452017, "status": "ok", "code": "TYPE" }, { "sample_id": 93, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 12.450300040654838, "policy_ms": 0.48509996850043535, "execution_ms": 17.858700011856854, "status": "ok", "code": "TYPE" }, { "sample_id": 94, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 14.805999933741987, "policy_ms": 0.8106000022962689, "execution_ms": 34.32700003031641, "status": "ok", "code": "CLICK" }, { "sample_id": 95, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 17.680800054222345, "policy_ms": 0.6501999450847507, "execution_ms": 39.23270001541823, "status": "ok", "code": "CLICK" }, { "sample_id": 96, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 14.850700041279197, "policy_ms": 1.3064000522717834, "execution_ms": 17.951800022274256, "status": "ok", "code": "TYPE" }, { "sample_id": 97, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 16.61170006264001, "policy_ms": 0.552099896594882, "execution_ms": 18.892599968239665, "status": "ok", "code": "TYPE" }, { "sample_id": 98, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 13.599199941381812, "policy_ms": 0.5657000001519918, "execution_ms": 18.29640008509159, "status": "ok", "code": "SELECT" }, { "sample_id": 99, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 17.335299984551966, "policy_ms": 0.9585000807419419, "execution_ms": 23.56340002734214, "status": "ok", "code": "CLICK" }, { "sample_id": 100, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 14.68399993609637, "policy_ms": 0.8733000140637159, "execution_ms": 28.93310005310923, "status": "ok", "code": "CLICK" }, { "sample_id": 101, "template": "table", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 18.984500085934997, "policy_ms": 0.4178999224677682, "execution_ms": 33.350200043059886, "status": "ok", "code": "CLICK" }, { "sample_id": 102, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 16.587800113484263, "policy_ms": 0.8635000558570027, "execution_ms": 18.836400005966425, "status": "ok", "code": "TYPE" }, { "sample_id": 103, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 13.54469999205321, "policy_ms": 0.37420005537569523, "execution_ms": 16.46219997201115, "status": "ok", "code": "TYPE" }, { "sample_id": 104, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 15.141700045205653, "policy_ms": 0.9451999794691801, "execution_ms": 18.88480002526194, "status": "ok", "code": "TYPE" }, { "sample_id": 105, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 11.78149995394051, "policy_ms": 0.3524000057950616, "execution_ms": 15.675799921154976, "status": "ok", "code": "TYPE" }, { "sample_id": 106, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 12.68280006479472, "policy_ms": 0.4248999757692218, "execution_ms": 16.675100079737604, "status": "ok", "code": "TYPE" }, { "sample_id": 107, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 15.878700069151819, "policy_ms": 0.9968999074772, "execution_ms": 16.802300000563264, "status": "ok", "code": "SELECT" }, { "sample_id": 108, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 17.157299909740686, "policy_ms": 0.8647000649943948, "execution_ms": 31.02210001088679, "status": "ok", "code": "CLICK" }, { "sample_id": 109, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 17.74039992596954, "policy_ms": 1.3216000515967607, "execution_ms": 23.970499983988702, "status": "ok", "code": "TYPE" }, { "sample_id": 110, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 15.85009996779263, "policy_ms": 0.5281000630930066, "execution_ms": 16.252800007350743, "status": "ok", "code": "SELECT" }, { "sample_id": 111, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 14.949799980968237, "policy_ms": 0.8281000191345811, "execution_ms": 17.168399994261563, "status": "ok", "code": "SELECT" }, { "sample_id": 112, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 13.862899970263243, "policy_ms": 0.5413000471889973, "execution_ms": 17.31850008945912, "status": "ok", "code": "TYPE" }, { "sample_id": 113, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 13.006699969992042, "policy_ms": 0.7114999461919069, "execution_ms": 49.89020002540201, "status": "ok", "code": "CLICK" }, { "sample_id": 114, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 14.234700007364154, "policy_ms": 0.9977999143302441, "execution_ms": 16.336999949999154, "status": "ok", "code": "SELECT" }, { "sample_id": 115, "template": "nested", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 13.109100051224232, "policy_ms": 0.6169999251142144, "execution_ms": 15.93759993556887, "status": "ok", "code": "TYPE" }, { "sample_id": 116, "template": "table", "success": true, "abstained": false, "expected_action": "T", "predicted_action": "T", "action_confidence": null, "target_confidence": null, "observation_ms": 14.565299963578582, "policy_ms": 0.5788999842479825, "execution_ms": 16.50369993876666, "status": "ok", "code": "TYPE" }, { "sample_id": 117, "template": "nested", "success": true, "abstained": false, "expected_action": "C", "predicted_action": "C", "action_confidence": null, "target_confidence": null, "observation_ms": 13.3552000625059, "policy_ms": 0.42269995901733637, "execution_ms": 28.218899969942868, "status": "ok", "code": "CLICK" }, { "sample_id": 118, "template": "table", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 11.856800061650574, "policy_ms": 0.6411999929696321, "execution_ms": 14.822600060142577, "status": "ok", "code": "SELECT" }, { "sample_id": 119, "template": "nested", "success": true, "abstained": false, "expected_action": "O", "predicted_action": "O", "action_confidence": null, "target_confidence": null, "observation_ms": 14.885099953971803, "policy_ms": 1.077700057066977, "execution_ms": 14.775300049223006, "status": "ok", "code": "SELECT" } ] }