{"learning":{"clean_vs_damaged":{"accuracy_gain":0.06640625,"baseline_correct":152,"candidate_correct":169,"documents":256,"exact_one_sided_p":0.003816039301455021,"gained":27,"lost":10},"comparisons_to_seed":{"clean-pair":{"accuracy_gain":0.56640625,"baseline_correct":23,"candidate_correct":168,"documents":256,"exact_one_sided_p":2.2420775429197073e-44,"gained":145,"lost":0,"retention":{"documents":128,"improved":67,"mean_change":-0.0023456270093051247,"normal_99pct_upper":0.006191972733605837,"scope":"descriptive development comparison; no serving approval","standard_error":0.003314285614484069,"worsened":61}},"clean-single":{"accuracy_gain":0.5703125,"baseline_correct":23,"candidate_correct":169,"documents":256,"exact_one_sided_p":1.1210387714598537e-44,"gained":146,"lost":0,"retention":{"documents":128,"improved":66,"mean_change":-0.0021952984053393745,"normal_99pct_upper":0.006656111539709363,"scope":"descriptive development comparison; no serving approval","standard_error":0.0034361063451276157,"worsened":62}},"damaged-single":{"accuracy_gain":0.50390625,"baseline_correct":23,"candidate_correct":152,"documents":256,"exact_one_sided_p":1.4693679385278594e-39,"gained":129,"lost":0,"retention":{"documents":128,"improved":62,"mean_change":-0.0030812776106290745,"normal_99pct_upper":0.005474693857460818,"scope":"descriptive development comparison; no serving approval","standard_error":0.0033214174953765108,"worsened":66}}},"descriptive_breakdowns":{"family":{"filter":{"clean-pair":58,"clean-single":60,"damaged-single":47,"documents":64,"seed":0},"lookup":{"clean-pair":63,"clean-single":62,"damaged-single":60,"documents":64,"seed":23},"sort":{"clean-pair":47,"clean-single":47,"damaged-single":45,"documents":64,"seed":0},"total":{"clean-pair":0,"clean-single":0,"damaged-single":0,"documents":64,"seed":0}},"variant":{"0":{"clean-pair":42,"clean-single":42,"damaged-single":42,"documents":64,"seed":5},"1":{"clean-pair":45,"clean-single":44,"damaged-single":40,"documents":64,"seed":9},"2":{"clean-pair":40,"clean-single":41,"damaged-single":37,"documents":64,"seed":3},"3":{"clean-pair":41,"clean-single":42,"damaged-single":33,"documents":64,"seed":6}}},"descriptive_pair_accuracy_within_margin":true,"narrow_learning_contract_passed":true,"pair_vs_single":{"accuracy_gain":-0.00390625,"baseline_correct":169,"candidate_correct":168,"documents":256,"exact_one_sided_p":0.7734375,"gained":3,"lost":4},"prepared":"4304122faebe6a7bf8619a82f4297c81d213f2343a1c994b6e44317f0bdd8c3b","primary_arm":"clean-single","primary_conditions":{"maximum_one_sided_p":true,"minimum_accuracy_gain":true,"retention_99pct_upper_max":true},"scope":"Generated grounded tasks and previously exposed public retention probes; not a general assistant or native serving approval","selection":"2fa35a47bc63781a513f4ba111a507117bacfb69bd9f7c79a61adcfd1ec7470d","serving_approved":false,"tokens_issued":0},"serving":{"arm":"clean-pair","failure_retries":16,"identical_output_tokens":true,"pair_throughput_ratio":1.6665877725261862,"phases":{"failure":{"concurrency":2,"endpoints":["http://172.31.19.168:9090","http://172.31.24.255:9090"],"median_latency_seconds":0.6842165035195649,"model_digest":"e7f65d52f3a8f40edcc83e2e2ec6d2132d06d427844f566f90efe3b073a70b35","p95_latency_seconds":0.8832419589161873,"requests":32,"requests_per_second":2.906959299501584,"run_id":"coop-20260913-failure","scope":"Fixed committed tasks, trusted operated replicas; no payments or permissionless admission","seconds":11.008066059090197,"successful":32,"unavailable":[{"endpoint":"http://172.31.24.255:9090","error":"URLError"}]},"pair":{"concurrency":2,"endpoints":["http://172.31.19.168:9090","http://172.31.24.255:9090"],"median_latency_seconds":0.3348826081492007,"model_digest":"e7f65d52f3a8f40edcc83e2e2ec6d2132d06d427844f566f90efe3b073a70b35","p95_latency_seconds":0.6630818145349622,"requests":32,"requests_per_second":4.841100021743467,"run_id":"coop-20260913-pair","scope":"Fixed committed tasks, trusted operated replicas; no payments or permissionless admission","seconds":6.610067930072546,"successful":32,"unavailable":[]},"single":{"concurrency":2,"endpoints":["http://172.31.19.168:9090"],"median_latency_seconds":0.6831884048879147,"model_digest":"e7f65d52f3a8f40edcc83e2e2ec6d2132d06d427844f566f90efe3b073a70b35","p95_latency_seconds":0.8729198360815644,"requests":32,"requests_per_second":2.904797515948055,"run_id":"coop-20260913-single","scope":"Fixed committed tasks, trusted operated replicas; no payments or permissionless admission","seconds":11.016258387826383,"successful":32,"unavailable":[]}},"scope":"Fixed task workload, equal client concurrency; operated replicas and volatile retry cache"},"training":{"clean-pair":{"allocated_gpu_seconds_including_checkpoints":2587.949524672,"host_interface_byte_deltas":[{"docker0":{"rx":0,"tx":0},"enp39s0":{"rx":883196259917,"tx":890332783732}},{"docker0":{"rx":0,"tx":0},"enp39s0":{"rx":888863729720,"tx":884939143853}}],"median_update_seconds":7.120141245000241,"network_scope":"Whole host interface during measured training; not isolated NCCL bytes","parameter_digest":"e7f65d52f3a8f40edcc83e2e2ec6d2132d06d427844f566f90efe3b073a70b35","peak_cuda_allocated_bytes":[38443578368,38443545600],"training_seconds_including_checkpoints":1294.041107028,"update_seconds":917.0302770500002,"updates":128,"world":2},"clean-single":{"allocated_gpu_seconds_including_checkpoints":986.840472131,"host_interface_byte_deltas":[{"docker0":{"rx":0,"tx":0},"enp39s0":{"rx":152708,"tx":19637212}}],"median_update_seconds":1.6131280920000108,"network_scope":"Whole host interface during measured training; not isolated NCCL bytes","parameter_digest":"98d64446325b9359e4b2934d313140c2de9fe11c03133b7045811b6fa21f6f8f","peak_cuda_allocated_bytes":[31598899712],"training_seconds_including_checkpoints":986.840472131,"update_seconds":208.6959044090014,"updates":128,"world":1},"damaged-single":{"allocated_gpu_seconds_including_checkpoints":665.013934828,"host_interface_byte_deltas":[{"docker0":{"rx":0,"tx":0},"enp39s0":{"rx":100937,"tx":14661614}}],"median_update_seconds":1.6226522289999252,"network_scope":"Whole host interface during measured training; not isolated NCCL bytes","parameter_digest":"eec8055c909aafe424392d9ff434c063d9d6b885afad1deb2ebfe4241284741c","peak_cuda_allocated_bytes":[31598899712],"training_seconds_including_checkpoints":665.013934828,"update_seconds":209.547250749,"updates":128,"world":1},"pair_training_loop_speedup":0.7626036505111171,"pair_update_speedup":0.22757798693447331,"timer_scope":"Training loop through final digest, including checkpoints; excludes seed verification, model loading and process-group initialization. Retain process resource logs and whole-instance lifetime separately."}}