1{2 "run_id": "validation_I_placebo_r2_mixed_random_mid",3 "arm": "I_placebo",4 "arm_label": "I_placebo",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 "placebo_1",27 "placebo_2",28 "placebo_3",29 "placebo_4",30 "placebo_5",31 "placebo_6",32 "placebo_7",33 "placebo_8",34 "placebo_9",35 "placebo_10",36 "placebo_11",37 "placebo_12"38 ],39 "prompt_hashes": {40 "placebo_1": "2e639018993a7f90",41 "placebo_2": "3b8dc64ee9d7510c",42 "placebo_3": "e53e5b360a75a3d7",43 "placebo_4": "c7ad48395054051d",44 "placebo_5": "60f3ff16d47dee84",45 "placebo_6": "25278ebbadf591ea",46 "placebo_7": "6ca049c56da54247",47 "placebo_8": "79f63b31af873c44",48 "placebo_9": "f615020da100bf33",49 "placebo_10": "924b26c724c0e358",50 "placebo_11": "fcb29d21820098bd",51 "placebo_12": "0866555d436b3cc8"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": 219.6186177000003,67 "spend_cap_hit": false,68 "start_time": "2026-10-28T01:04:08.788224+00:00",69 "end_time": "2026-10-28T01:47:24.975263+00:00",70 "wall_clock_s": 2596.2,71 "total_cost_usd_estimated": 54.076,72 "total_tokens": 19058150,73 "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 "2drbGwU359hwygoXYa5zVk",80 "3iiuPNCNsVjLp9ph1fxGyA",81 "3onE5QdWUVcAhfm7SMx7H1"82 ],83 "errors": [],84 "prediction_columns": [85 "tx_id",86 "split",87 "arm",88 "committee",89 "aggregation",90 "point",91 "q05",92 "q25",93 "q50",94 "q75",95 "q95",96 "cost_usd",97 "latency_s",98 "tokens",99 "run_id"100 ],101 "spec": {102 "arm": "I_placebo",103 "rounds": 2,104 "committee": "mixed_random",105 "split": "validation",106 "n": 40,107 "live": true,108 "tier": "mid",109 "run_id": "validation_I_placebo_r2_mixed_random_mid",110 "synthetic": false,111 "seed": 20261027,112 "effort": null,113 "property_concurrency": 5,114 "out_dir": null,115 "weights_path": null,116 "max_tokens": 4096117 }118}