1{2 "run_id": "validation_G_debate_r2_mixed_random_mid",3 "arm": "G_debate",4 "arm_label": "G_debate_r2",5 "protocol": "structured_debate",6 "rounds": 2,7 "rounds_note": null,8 "k_repeats": null,9 "committee": "mixed_random",10 "tier": "mid",11 "models": [12 "claude-sonnet-5-5",13 "gpt-6.1-sol"14 ],15 "effort": "medium",16 "max_tokens": 4096,17 "mode": "live",18 "split": "validation",19 "n_requested": 40,20 "n_sampled": 40,21 "synthetic": false,22 "data_source": "data/processed/analytic.parquet",23 "features_source": "agentic_valuation.features",24 "geo_cache": "/Users/simon-pierreboucher/Desktop/Recherche-UQO/immobilier/agentic-valuation-quebec/data/processed/geo_features.parquet",25 "roles": [26 "appraisal_method",27 "comparables",28 "hedonic",29 "market_economist",30 "building_condition",31 "geospatial_risk",32 "underwriting",33 "buyer_wtp",34 "seller_strategy",35 "rental_usercost",36 "data_auditor",37 "skeptic"38 ],39 "prompt_hashes": {40 "appraisal_method": "4ab6a4b56e275e66",41 "comparables": "831b81649f59fa7e",42 "hedonic": "9063d328055c29d6",43 "market_economist": "623f2000f584463e",44 "building_condition": "d4c83732b1f6dc26",45 "geospatial_risk": "83d7221633a884c3",46 "underwriting": "d25a69fc8ead39ca",47 "buyer_wtp": "facb211deeccae1c",48 "seller_strategy": "3917bd2ffd8e5543",49 "rental_usercost": "e745f4c42bb39244",50 "data_auditor": "db248eb5f189de2b",51 "skeptic": "2f789b8366c5d29c"52 },53 "config_hashes": {54 "budgets.yaml": "d6d368fd34b923ce",55 "data.yaml": "c20f448c79e39726",56 "models.yaml": "80e216f864e413ce",57 "treatments.yaml": "e1f71e68e67f7701"58 },59 "schema_hash": "b0170725f58da4ed",60 "package_version": "0.1.0",61 "git_commit": null,62 "python": "3.14.4",63 "host": "MacBookPro-de-simon-pierre-2.local",64 "seed": 20261027,65 "weights_source": "none (equal weights for inverse_variance/stacking; MAD scale for robust_bayes)",66 "spend_cap_usd": 490.8608067,67 "spend_cap_hit": false,68 "start_time": "2026-10-27T23:21:34.280027+00:00",69 "end_time": "2026-10-28T00:15:48.999009+00:00",70 "wall_clock_s": 3254.7,71 "total_cost_usd_estimated": 60.0474,72 "total_tokens": 20597430,73 "n_transcripts": 40,74 "n_ok_this_session": 40,75 "n_failed_this_session": 0,76 "n_resumed_skipped": 0,77 "n_turn_errors": 0,78 "stopped_early": [],79 "errors": [],80 "prediction_columns": [81 "tx_id",82 "split",83 "arm",84 "committee",85 "aggregation",86 "point",87 "q05",88 "q25",89 "q50",90 "q75",91 "q95",92 "cost_usd",93 "latency_s",94 "tokens",95 "run_id"96 ],97 "spec": {98 "arm": "G_debate",99 "rounds": 2,100 "committee": "mixed_random",101 "split": "validation",102 "n": 40,103 "live": true,104 "tier": "mid",105 "run_id": "validation_G_debate_r2_mixed_random_mid",106 "synthetic": false,107 "seed": 20261027,108 "effort": null,109 "property_concurrency": 5,110 "out_dir": null,111 "weights_path": null,112 "max_tokens": 4096113 }114}