{ "20261012/paired_answer": { "arm": "paired_answer", "seed": 20261012, "training": { "status": "REAL_ROLLOUT_MASK_AND_ADVANTAGE_VALIDATED", "counts": { "nonzero_advantage_tokens": 252527, "trajectories": 4352, "shared_trajectories": 600, "actual_exposed": 300, "exposed_with_nonzero_advantage": 108, "retrieved_trajectories": 3697, "shared_groups": 150, "steps": 68 }, "cost": { "rl_sequence_nonpadding_tokens": 3668903, "rl_sequence_padded_tokens": 10235968, "rl_loss_tokens": 1002320, "backend_generated_tokens": 1012764, "backend_requested_input_tokens": 6620960, "backend_calls": 11196, "aux_sequence_nonpadding_tokens_including_dp_padding": 106827, "aux_sequence_padded_tokens": 113636, "aux_loss_tokens": 253 }, "auxiliary_sources": { "groups_considered": 150, "skipped_no_verified_target": 82, "gold_checked_clean_answer": 54, "examples": 68, "answer_tokens": 253, "training_gold_with_support_title_coverage": 14 }, "cost_limitations": "Input tokens count actual backend requests, not cache-adjusted FLOPs. RL sequence tokens are one tensor pass proxy; policy/reference/rollout-logprob and backward differ in cost. Auxiliary includes zero-loss DP padding. Equal optimizer steps are not equal total computation." }, "budget": { "status": "BUDGET_AND_EXPOSURE_RECONCILED", "effective_steps": 68, "loop_seconds": 2714.451907608658, "budget_seconds": 2700.0, "overshoot_seconds": 14.451907608658075, "overshoot_fraction": 0.005352558373576954, "last_batch_seconds": 33.73658206406981, "actual_exposed": 300, "target_exposed": 300, "exposure_templates": { "train_direct": 150, "train_role": 150 }, "time_budget_reached": true, "exposure_target_reached": true, "matched_interpretation_allowed": true, "limitation": "Endpoint is first completed batch at the shared loop-time budget. Includes RL and auxiliary supervision; excludes initialization/export/evaluation/post-update sync. Actual token costs, GPU allocation time and overshoot remain separately reported. This is not exact FLOP matching." }, "timing": { "arm": "paired_answer", "gpus": 4, "training_seconds_including_setup_and_export": 3343.6007134178653, "audit_seconds": 25.512921646237373, "evaluation_seconds": 64.3174276901409, "elapsed_seconds": 3433.4450278282166, "allocated_gpu_hours_through_evaluation": 3.8149389198091295, "training_allocated_gpu_hours": 3.715111903797628 } }, "20261012/clean": { "arm": "clean", "seed": 20261012, "training": { "status": "REAL_ROLLOUT_MASK_AND_ADVANTAGE_VALIDATED", "counts": { "nonzero_advantage_tokens": 237731, "trajectories": 4992, "shared_trajectories": 0, "actual_exposed": 0, "exposed_with_nonzero_advantage": 0, "retrieved_trajectories": 4321, "shared_groups": 0, "steps": 78 }, "cost": { "rl_sequence_nonpadding_tokens": 3761824, "rl_sequence_padded_tokens": 10801024, "rl_loss_tokens": 854744, "backend_generated_tokens": 852732, "backend_requested_input_tokens": 6876330, "backend_calls": 13129 }, "auxiliary_sources": {}, "cost_limitations": "Input tokens count actual backend requests, not cache-adjusted FLOPs. RL sequence tokens are one tensor pass proxy; policy/reference/rollout-logprob and backward differ in cost. Auxiliary includes zero-loss DP padding. Equal optimizer steps are not equal total computation." }, "budget": { "status": "BUDGET_AND_EXPOSURE_RECONCILED", "effective_steps": 78, "loop_seconds": 2719.168388230726, "budget_seconds": 2700.0, "overshoot_seconds": 19.168388230726123, "overshoot_fraction": 0.007099403048417052, "last_batch_seconds": 25.682668573223054, "actual_exposed": 0, "target_exposed": 0, "exposure_templates": {}, "time_budget_reached": true, "exposure_target_reached": true, "matched_interpretation_allowed": true, "limitation": "Endpoint is first completed batch at the shared loop-time budget. Includes RL and auxiliary supervision; excludes initialization/export/evaluation/post-update sync. Actual token costs, GPU allocation time and overshoot remain separately reported. This is not exact FLOP matching." }, "timing": { "arm": "clean", "gpus": 4, "training_seconds_including_setup_and_export": 3346.1614286573604, "audit_seconds": 24.076764695346355, "evaluation_seconds": 52.93807957135141, "elapsed_seconds": 3423.1988999843597, "allocated_gpu_hours_through_evaluation": 3.8035543333159554, "training_allocated_gpu_hours": 3.717957142952623 } }, "20261012/paired": { "arm": "paired", "seed": 20261012, "training": { "status": "REAL_ROLLOUT_MASK_AND_ADVANTAGE_VALIDATED", "counts": { "nonzero_advantage_tokens": 246678, "trajectories": 4992, "shared_trajectories": 600, "actual_exposed": 300, "exposed_with_nonzero_advantage": 92, "retrieved_trajectories": 4468, "shared_groups": 150, "steps": 78 }, "cost": { "rl_sequence_nonpadding_tokens": 3831624, "rl_sequence_padded_tokens": 10175296, "rl_loss_tokens": 836930, "backend_generated_tokens": 848133, "backend_requested_input_tokens": 6739735, "backend_calls": 12429 }, "auxiliary_sources": {}, "cost_limitations": "Input tokens count actual backend requests, not cache-adjusted FLOPs. RL sequence tokens are one tensor pass proxy; policy/reference/rollout-logprob and backward differ in cost. Auxiliary includes zero-loss DP padding. Equal optimizer steps are not equal total computation." }, "budget": { "status": "BUDGET_AND_EXPOSURE_RECONCILED", "effective_steps": 78, "loop_seconds": 2724.6154512800276, "budget_seconds": 2700.0, "overshoot_seconds": 24.615451280027628, "overshoot_fraction": 0.009116833807417679, "last_batch_seconds": 26.567073594778776, "actual_exposed": 300, "target_exposed": 300, "exposure_templates": { "train_role": 150, "train_direct": 150 }, "time_budget_reached": true, "exposure_target_reached": true, "matched_interpretation_allowed": true, "limitation": "Endpoint is first completed batch at the shared loop-time budget. Includes RL and auxiliary supervision; excludes initialization/export/evaluation/post-update sync. Actual token costs, GPU allocation time and overshoot remain separately reported. This is not exact FLOP matching." }, "timing": { "arm": "paired", "gpus": 4, "training_seconds_including_setup_and_export": 3414.526720385067, "audit_seconds": 22.943808436393738, "evaluation_seconds": 48.4610189339146, "elapsed_seconds": 3485.9448568820953, "allocated_gpu_hours_through_evaluation": 3.8732720632023283, "training_allocated_gpu_hours": 3.79391857820563 } }, "20261012/random_matched": { "arm": "random_matched", "seed": 20261012, "training": { "status": "REAL_ROLLOUT_MASK_AND_ADVANTAGE_VALIDATED", "counts": { "nonzero_advantage_tokens": 283649, "trajectories": 4864, "shared_trajectories": 0, "actual_exposed": 300, "exposed_with_nonzero_advantage": 114, "retrieved_trajectories": 4316, "shared_groups": 0, "steps": 76 }, "cost": { "rl_sequence_nonpadding_tokens": 3809170, "rl_sequence_padded_tokens": 10273664, "rl_loss_tokens": 937629, "backend_generated_tokens": 935697, "backend_requested_input_tokens": 7028368, "backend_calls": 13076 }, "auxiliary_sources": {}, "cost_limitations": "Input tokens count actual backend requests, not cache-adjusted FLOPs. RL sequence tokens are one tensor pass proxy; policy/reference/rollout-logprob and backward differ in cost. Auxiliary includes zero-loss DP padding. Equal optimizer steps are not equal total computation." }, "budget": { "status": "BUDGET_AND_EXPOSURE_RECONCILED", "effective_steps": 76, "loop_seconds": 2723.7498461157084, "budget_seconds": 2700.0, "overshoot_seconds": 23.74984611570835, "overshoot_fraction": 0.00879623930211415, "last_batch_seconds": 27.005504108034074, "actual_exposed": 300, "target_exposed": 300, "exposure_templates": { "train_direct": 150, "train_role": 150 }, "time_budget_reached": true, "exposure_target_reached": true, "matched_interpretation_allowed": true, "limitation": "Endpoint is first completed batch at the shared loop-time budget. Includes RL and auxiliary supervision; excludes initialization/export/evaluation/post-update sync. Actual token costs, GPU allocation time and overshoot remain separately reported. This is not exact FLOP matching." }, "timing": { "arm": "random_matched", "gpus": 4, "training_seconds_including_setup_and_export": 3304.0648774337023, "audit_seconds": 21.311844415962696, "evaluation_seconds": 49.26294994354248, "elapsed_seconds": 3374.6536977291107, "allocated_gpu_hours_through_evaluation": 3.749615219699012, "training_allocated_gpu_hours": 3.671183197148558 } }, "20261012/random_answer": { "arm": "random_answer", "seed": 20261012, "training": { "status": "REAL_ROLLOUT_MASK_AND_ADVANTAGE_VALIDATED", "counts": { "nonzero_advantage_tokens": 342671, "trajectories": 4032, "shared_trajectories": 0, "actual_exposed": 300, "exposed_with_nonzero_advantage": 106, "retrieved_trajectories": 3219, "shared_groups": 0, "steps": 63 }, "cost": { "rl_sequence_nonpadding_tokens": 3654667, "rl_sequence_padded_tokens": 10470784, "rl_loss_tokens": 1243016, "backend_generated_tokens": 1241904, "backend_requested_input_tokens": 7080808, "backend_calls": 11683, "aux_sequence_nonpadding_tokens_including_dp_padding": 107259, "aux_sequence_padded_tokens": 113216, "aux_loss_tokens": 262 }, "auxiliary_sources": { "groups_considered": 150, "skipped_no_verified_target": 76, "gold_checked_clean_answer": 41, "examples": 74, "answer_tokens": 262, "training_gold_with_support_title_coverage": 33 }, "cost_limitations": "Input tokens count actual backend requests, not cache-adjusted FLOPs. RL sequence tokens are one tensor pass proxy; policy/reference/rollout-logprob and backward differ in cost. Auxiliary includes zero-loss DP padding. Equal optimizer steps are not equal total computation." }, "budget": { "status": "BUDGET_AND_EXPOSURE_RECONCILED", "effective_steps": 63, "loop_seconds": 2725.248705212958, "budget_seconds": 2700.0, "overshoot_seconds": 25.248705212958157, "overshoot_fraction": 0.009351372301095617, "last_batch_seconds": 45.57455317955464, "actual_exposed": 300, "target_exposed": 300, "exposure_templates": { "train_role": 150, "train_direct": 150 }, "time_budget_reached": true, "exposure_target_reached": true, "matched_interpretation_allowed": true, "limitation": "Endpoint is first completed batch at the shared loop-time budget. Includes RL and auxiliary supervision; excludes initialization/export/evaluation/post-update sync. Actual token costs, GPU allocation time and overshoot remain separately reported. This is not exact FLOP matching." }, "timing": { "arm": "random_answer", "gpus": 4, "training_seconds_including_setup_and_export": 3563.6339429002255, "audit_seconds": 21.048595015890896, "evaluation_seconds": 52.58237045072019, "elapsed_seconds": 3637.2787568569183, "allocated_gpu_hours_through_evaluation": 4.0414208409521315, "training_allocated_gpu_hours": 3.9595932698891394 } }, "20261013/random_matched": { "arm": "random_matched", "seed": 20261013, "training": { "status": "REAL_ROLLOUT_MASK_AND_ADVANTAGE_VALIDATED", "counts": { "nonzero_advantage_tokens": 271745, "trajectories": 4544, "shared_trajectories": 0, "actual_exposed": 300, "exposed_with_nonzero_advantage": 94, "retrieved_trajectories": 3948, "shared_groups": 0, "steps": 71 }, "cost": { "rl_sequence_nonpadding_tokens": 3742952, "rl_sequence_padded_tokens": 10504768, "rl_loss_tokens": 1008091, "backend_generated_tokens": 1006073, "backend_requested_input_tokens": 6749197, "backend_calls": 12150 }, "auxiliary_sources": {}, "cost_limitations": "Input tokens count actual backend requests, not cache-adjusted FLOPs. RL sequence tokens are one tensor pass proxy; policy/reference/rollout-logprob and backward differ in cost. Auxiliary includes zero-loss DP padding. Equal optimizer steps are not equal total computation." }, "budget": { "status": "BUDGET_AND_EXPOSURE_RECONCILED", "effective_steps": 71, "loop_seconds": 2713.713573261164, "budget_seconds": 2700.0, "overshoot_seconds": 13.713573261164129, "overshoot_fraction": 0.00507910120783861, "last_batch_seconds": 42.03551970701665, "actual_exposed": 300, "target_exposed": 300, "exposure_templates": { "train_direct": 150, "train_role": 150 }, "time_budget_reached": true, "exposure_target_reached": true, "matched_interpretation_allowed": true, "limitation": "Endpoint is first completed batch at the shared loop-time budget. Includes RL and auxiliary supervision; excludes initialization/export/evaluation/post-update sync. Actual token costs, GPU allocation time and overshoot remain separately reported. This is not exact FLOP matching." }, "timing": { "arm": "random_matched", "gpus": 4, "training_seconds_including_setup_and_export": 3302.784130680375, "audit_seconds": 21.651369648985565, "evaluation_seconds": 57.911904649809, "elapsed_seconds": 3382.357478618622, "allocated_gpu_hours_through_evaluation": 3.758174976242913, "training_allocated_gpu_hours": 3.6697601452004163 } }, "20261013/random_answer": { "arm": "random_answer", "seed": 20261013, "training": { "status": "REAL_ROLLOUT_MASK_AND_ADVANTAGE_VALIDATED", "counts": { "nonzero_advantage_tokens": 299568, "trajectories": 4288, "shared_trajectories": 0, "actual_exposed": 300, "exposed_with_nonzero_advantage": 128, "retrieved_trajectories": 3619, "shared_groups": 0, "steps": 67 }, "cost": { "rl_sequence_nonpadding_tokens": 3541194, "rl_sequence_padded_tokens": 10129920, "rl_loss_tokens": 1064045, "backend_generated_tokens": 1061874, "backend_requested_input_tokens": 6268705, "backend_calls": 11179, "aux_sequence_nonpadding_tokens_including_dp_padding": 116609, "aux_sequence_padded_tokens": 125184, "aux_loss_tokens": 296 }, "auxiliary_sources": { "groups_considered": 150, "skipped_no_verified_target": 75, "training_gold_with_support_title_coverage": 25, "examples": 75, "answer_tokens": 296, "gold_checked_clean_answer": 50 }, "cost_limitations": "Input tokens count actual backend requests, not cache-adjusted FLOPs. RL sequence tokens are one tensor pass proxy; policy/reference/rollout-logprob and backward differ in cost. Auxiliary includes zero-loss DP padding. Equal optimizer steps are not equal total computation." }, "budget": { "status": "BUDGET_AND_EXPOSURE_RECONCILED", "effective_steps": 67, "loop_seconds": 2710.5506856860593, "budget_seconds": 2700.0, "overshoot_seconds": 10.550685686059296, "overshoot_fraction": 0.003907661365207149, "last_batch_seconds": 37.67958049941808, "actual_exposed": 300, "target_exposed": 300, "exposure_templates": { "train_direct": 150, "train_role": 150 }, "time_budget_reached": true, "exposure_target_reached": true, "matched_interpretation_allowed": true, "limitation": "Endpoint is first completed batch at the shared loop-time budget. Includes RL and auxiliary supervision; excludes initialization/export/evaluation/post-update sync. Actual token costs, GPU allocation time and overshoot remain separately reported. This is not exact FLOP matching." }, "timing": { "arm": "random_answer", "gpus": 4, "training_seconds_including_setup_and_export": 3336.919173264876, "audit_seconds": 21.594674508087337, "evaluation_seconds": 47.31309958919883, "elapsed_seconds": 3405.8341085910797, "allocated_gpu_hours_through_evaluation": 3.7842601206567554, "training_allocated_gpu_hours": 3.7076879702943066 } }, "20261013/paired_answer": { "arm": "paired_answer", "seed": 20261013, "training": { "status": "REAL_ROLLOUT_MASK_AND_ADVANTAGE_VALIDATED", "counts": { "nonzero_advantage_tokens": 279613, "trajectories": 4224, "shared_trajectories": 600, "actual_exposed": 300, "exposed_with_nonzero_advantage": 98, "retrieved_trajectories": 3558, "shared_groups": 150, "steps": 66 }, "cost": { "rl_sequence_nonpadding_tokens": 3547595, "rl_sequence_padded_tokens": 9795456, "rl_loss_tokens": 1006341, "backend_generated_tokens": 1015457, "backend_requested_input_tokens": 6433444, "backend_calls": 11029, "aux_sequence_nonpadding_tokens_including_dp_padding": 113376, "aux_sequence_padded_tokens": 120252, "aux_loss_tokens": 314 }, "auxiliary_sources": { "groups_considered": 150, "skipped_no_verified_target": 78, "gold_checked_clean_answer": 48, "examples": 72, "answer_tokens": 314, "training_gold_with_support_title_coverage": 24 }, "cost_limitations": "Input tokens count actual backend requests, not cache-adjusted FLOPs. RL sequence tokens are one tensor pass proxy; policy/reference/rollout-logprob and backward differ in cost. Auxiliary includes zero-loss DP padding. Equal optimizer steps are not equal total computation." }, "budget": { "status": "BUDGET_AND_EXPOSURE_RECONCILED", "effective_steps": 66, "loop_seconds": 2716.918354540132, "budget_seconds": 2700.0, "overshoot_seconds": 16.918354540131986, "overshoot_fraction": 0.0062660572370858425, "last_batch_seconds": 35.05141973588616, "actual_exposed": 300, "target_exposed": 300, "exposure_templates": { "train_role": 150, "train_direct": 150 }, "time_budget_reached": true, "exposure_target_reached": true, "matched_interpretation_allowed": true, "limitation": "Endpoint is first completed batch at the shared loop-time budget. Includes RL and auxiliary supervision; excludes initialization/export/evaluation/post-update sync. Actual token costs, GPU allocation time and overshoot remain separately reported. This is not exact FLOP matching." }, "timing": { "arm": "paired_answer", "gpus": 4, "training_seconds_including_setup_and_export": 3304.0175564521924, "audit_seconds": 22.049914307892323, "evaluation_seconds": 56.6587127475068, "elapsed_seconds": 3382.736637353897, "allocated_gpu_hours_through_evaluation": 3.7585962637265524, "training_allocated_gpu_hours": 3.671130618280214 } }, "20261013/clean": { "arm": "clean", "seed": 20261013, "training": { "status": "REAL_ROLLOUT_MASK_AND_ADVANTAGE_VALIDATED", "counts": { "nonzero_advantage_tokens": 269858, "trajectories": 4736, "shared_trajectories": 0, "actual_exposed": 0, "exposed_with_nonzero_advantage": 0, "retrieved_trajectories": 4127, "shared_groups": 0, "steps": 74 }, "cost": { "rl_sequence_nonpadding_tokens": 3643507, "rl_sequence_padded_tokens": 9996480, "rl_loss_tokens": 922706, "backend_generated_tokens": 920429, "backend_requested_input_tokens": 6428608, "backend_calls": 12245 }, "auxiliary_sources": {}, "cost_limitations": "Input tokens count actual backend requests, not cache-adjusted FLOPs. RL sequence tokens are one tensor pass proxy; policy/reference/rollout-logprob and backward differ in cost. Auxiliary includes zero-loss DP padding. Equal optimizer steps are not equal total computation." }, "budget": { "status": "BUDGET_AND_EXPOSURE_RECONCILED", "effective_steps": 74, "loop_seconds": 2710.1920116795227, "budget_seconds": 2700.0, "overshoot_seconds": 10.192011679522693, "overshoot_fraction": 0.0037748191405639897, "last_batch_seconds": 26.397439217194915, "actual_exposed": 0, "target_exposed": 0, "exposure_templates": {}, "time_budget_reached": true, "exposure_target_reached": true, "matched_interpretation_allowed": true, "limitation": "Endpoint is first completed batch at the shared loop-time budget. Includes RL and auxiliary supervision; excludes initialization/export/evaluation/post-update sync. Actual token costs, GPU allocation time and overshoot remain separately reported. This is not exact FLOP matching." }, "timing": { "arm": "clean", "gpus": 4, "training_seconds_including_setup_and_export": 3279.7575224544853, "audit_seconds": 21.148273780941963, "evaluation_seconds": 89.85128934960812, "elapsed_seconds": 3390.764983654022, "allocated_gpu_hours_through_evaluation": 3.7675166485044693, "training_allocated_gpu_hours": 3.644175024949428 } }, "20261013/paired": { "arm": "paired", "seed": 20261013, "training": { "status": "REAL_ROLLOUT_MASK_AND_ADVANTAGE_VALIDATED", "counts": { "nonzero_advantage_tokens": 234415, "trajectories": 4928, "shared_trajectories": 600, "actual_exposed": 300, "exposed_with_nonzero_advantage": 86, "retrieved_trajectories": 4444, "shared_groups": 150, "steps": 77 }, "cost": { "rl_sequence_nonpadding_tokens": 3809850, "rl_sequence_padded_tokens": 10649792, "rl_loss_tokens": 783056, "backend_generated_tokens": 790399, "backend_requested_input_tokens": 6788102, "backend_calls": 12405 }, "auxiliary_sources": {}, "cost_limitations": "Input tokens count actual backend requests, not cache-adjusted FLOPs. RL sequence tokens are one tensor pass proxy; policy/reference/rollout-logprob and backward differ in cost. Auxiliary includes zero-loss DP padding. Equal optimizer steps are not equal total computation." }, "budget": { "status": "BUDGET_AND_EXPOSURE_RECONCILED", "effective_steps": 77, "loop_seconds": 2702.411171090789, "budget_seconds": 2700.0, "overshoot_seconds": 2.4111710907891393, "overshoot_fraction": 0.0008930263299218311, "last_batch_seconds": 27.053393561393023, "actual_exposed": 300, "target_exposed": 300, "exposure_templates": { "train_direct": 150, "train_role": 150 }, "time_budget_reached": true, "exposure_target_reached": true, "matched_interpretation_allowed": true, "limitation": "Endpoint is first completed batch at the shared loop-time budget. Includes RL and auxiliary supervision; excludes initialization/export/evaluation/post-update sync. Actual token costs, GPU allocation time and overshoot remain separately reported. This is not exact FLOP matching." }, "timing": { "arm": "paired", "gpus": 4, "training_seconds_including_setup_and_export": 3334.201490016654, "audit_seconds": 21.028885998763144, "evaluation_seconds": 45.426083605736494, "elapsed_seconds": 3400.6660578250885, "allocated_gpu_hours_through_evaluation": 3.778517842027876, "training_allocated_gpu_hours": 3.704668322240727 } }, "20261014/paired_answer": { "arm": "paired_answer", "seed": 20261014, "training": { "status": "REAL_ROLLOUT_MASK_AND_ADVANTAGE_VALIDATED", "counts": { "nonzero_advantage_tokens": 264318, "trajectories": 4352, "shared_trajectories": 600, "actual_exposed": 300, "exposed_with_nonzero_advantage": 108, "retrieved_trajectories": 3777, "shared_groups": 150, "steps": 68 }, "cost": { "rl_sequence_nonpadding_tokens": 3533022, "rl_sequence_padded_tokens": 9975232, "rl_loss_tokens": 900063, "backend_generated_tokens": 908865, "backend_requested_input_tokens": 6112117, "backend_calls": 10738, "aux_sequence_nonpadding_tokens_including_dp_padding": 91389, "aux_sequence_padded_tokens": 101220, "aux_loss_tokens": 238 }, "auxiliary_sources": { "groups_considered": 150, "skipped_no_verified_target": 90, "examples": 60, "answer_tokens": 238, "training_gold_with_support_title_coverage": 18, "gold_checked_clean_answer": 42 }, "cost_limitations": "Input tokens count actual backend requests, not cache-adjusted FLOPs. RL sequence tokens are one tensor pass proxy; policy/reference/rollout-logprob and backward differ in cost. Auxiliary includes zero-loss DP padding. Equal optimizer steps are not equal total computation." }, "budget": { "status": "BUDGET_AND_EXPOSURE_RECONCILED", "effective_steps": 68, "loop_seconds": 2724.7475920431316, "budget_seconds": 2700.0, "overshoot_seconds": 24.74759204313159, "overshoot_fraction": 0.009165774830789397, "last_batch_seconds": 26.105257497169077, "actual_exposed": 300, "target_exposed": 300, "exposure_templates": { "train_direct": 150, "train_role": 150 }, "time_budget_reached": true, "exposure_target_reached": true, "matched_interpretation_allowed": true, "limitation": "Endpoint is first completed batch at the shared loop-time budget. Includes RL and auxiliary supervision; excludes initialization/export/evaluation/post-update sync. Actual token costs, GPU allocation time and overshoot remain separately reported. This is not exact FLOP matching." }, "timing": { "arm": "paired_answer", "gpus": 4, "training_seconds_including_setup_and_export": 3326.9173680711538, "audit_seconds": 23.79169140011072, "evaluation_seconds": 68.29739535413682, "elapsed_seconds": 3419.0189802646637, "allocated_gpu_hours_through_evaluation": 3.7989099780718485, "training_allocated_gpu_hours": 3.6965748534123932 } }, "20261014/paired": { "arm": "paired", "seed": 20261014, "training": { "status": "REAL_ROLLOUT_MASK_AND_ADVANTAGE_VALIDATED", "counts": { "nonzero_advantage_tokens": 270688, "trajectories": 4480, "shared_trajectories": 600, "actual_exposed": 300, "exposed_with_nonzero_advantage": 96, "retrieved_trajectories": 3594, "shared_groups": 150, "steps": 70 }, "cost": { "rl_sequence_nonpadding_tokens": 3648797, "rl_sequence_padded_tokens": 10065088, "rl_loss_tokens": 981068, "backend_generated_tokens": 993165, "backend_requested_input_tokens": 6854791, "backend_calls": 11939 }, "auxiliary_sources": {}, "cost_limitations": "Input tokens count actual backend requests, not cache-adjusted FLOPs. RL sequence tokens are one tensor pass proxy; policy/reference/rollout-logprob and backward differ in cost. Auxiliary includes zero-loss DP padding. Equal optimizer steps are not equal total computation." }, "budget": { "status": "BUDGET_AND_EXPOSURE_RECONCILED", "effective_steps": 70, "loop_seconds": 2710.549219547771, "budget_seconds": 2700.0, "overshoot_seconds": 10.549219547770917, "overshoot_fraction": 0.003907118351026195, "last_batch_seconds": 27.81757804378867, "actual_exposed": 300, "target_exposed": 300, "exposure_templates": { "train_direct": 150, "train_role": 150 }, "time_budget_reached": true, "exposure_target_reached": true, "matched_interpretation_allowed": true, "limitation": "Endpoint is first completed batch at the shared loop-time budget. Includes RL and auxiliary supervision; excludes initialization/export/evaluation/post-update sync. Actual token costs, GPU allocation time and overshoot remain separately reported. This is not exact FLOP matching." }, "timing": { "arm": "paired", "gpus": 4, "training_seconds_including_setup_and_export": 3328.482014004141, "audit_seconds": 21.141326579265296, "evaluation_seconds": 52.04592780675739, "elapsed_seconds": 3401.675806045532, "allocated_gpu_hours_through_evaluation": 3.779639784495036, "training_allocated_gpu_hours": 3.6983133488934903 } }, "20261014/random_matched": { "arm": "random_matched", "seed": 20261014, "training": { "status": "REAL_ROLLOUT_MASK_AND_ADVANTAGE_VALIDATED", "counts": { "nonzero_advantage_tokens": 250140, "trajectories": 4672, "shared_trajectories": 0, "actual_exposed": 300, "exposed_with_nonzero_advantage": 106, "retrieved_trajectories": 4079, "shared_groups": 0, "steps": 73 }, "cost": { "rl_sequence_nonpadding_tokens": 3528793, "rl_sequence_padded_tokens": 10054528, "rl_loss_tokens": 837992, "backend_generated_tokens": 835615, "backend_requested_input_tokens": 6259461, "backend_calls": 11973 }, "auxiliary_sources": {}, "cost_limitations": "Input tokens count actual backend requests, not cache-adjusted FLOPs. RL sequence tokens are one tensor pass proxy; policy/reference/rollout-logprob and backward differ in cost. Auxiliary includes zero-loss DP padding. Equal optimizer steps are not equal total computation." }, "budget": { "status": "BUDGET_AND_EXPOSURE_RECONCILED", "effective_steps": 73, "loop_seconds": 2715.6321745608, "budget_seconds": 2700.0, "overshoot_seconds": 15.632174560800195, "overshoot_fraction": 0.005789694281777891, "last_batch_seconds": 33.50570999272168, "actual_exposed": 300, "target_exposed": 300, "exposure_templates": { "train_direct": 150, "train_role": 150 }, "time_budget_reached": true, "exposure_target_reached": true, "matched_interpretation_allowed": true, "limitation": "Endpoint is first completed batch at the shared loop-time budget. Includes RL and auxiliary supervision; excludes initialization/export/evaluation/post-update sync. Actual token costs, GPU allocation time and overshoot remain separately reported. This is not exact FLOP matching." }, "timing": { "arm": "random_matched", "gpus": 4, "training_seconds_including_setup_and_export": 3328.8213895261288, "audit_seconds": 24.528389466926455, "evaluation_seconds": 44.02744354493916, "elapsed_seconds": 3397.390631914139, "allocated_gpu_hours_through_evaluation": 3.7748784799045985, "training_allocated_gpu_hours": 3.6986904328068095 } }, "20261014/clean": { "arm": "clean", "seed": 20261014, "training": { "status": "REAL_ROLLOUT_MASK_AND_ADVANTAGE_VALIDATED", "counts": { "nonzero_advantage_tokens": 246540, "trajectories": 4672, "shared_trajectories": 0, "actual_exposed": 0, "exposed_with_nonzero_advantage": 0, "retrieved_trajectories": 4060, "shared_groups": 0, "steps": 73 }, "cost": { "rl_sequence_nonpadding_tokens": 3806705, "rl_sequence_padded_tokens": 10908544, "rl_loss_tokens": 982897, "backend_generated_tokens": 980405, "backend_requested_input_tokens": 6863404, "backend_calls": 12371 }, "auxiliary_sources": {}, "cost_limitations": "Input tokens count actual backend requests, not cache-adjusted FLOPs. RL sequence tokens are one tensor pass proxy; policy/reference/rollout-logprob and backward differ in cost. Auxiliary includes zero-loss DP padding. Equal optimizer steps are not equal total computation." }, "budget": { "status": "BUDGET_AND_EXPOSURE_RECONCILED", "effective_steps": 73, "loop_seconds": 2716.0083371615037, "budget_seconds": 2700.0, "overshoot_seconds": 16.008337161503732, "overshoot_fraction": 0.0059290137635199525, "last_batch_seconds": 27.549102265387774, "actual_exposed": 0, "target_exposed": 0, "exposure_templates": {}, "time_budget_reached": true, "exposure_target_reached": true, "matched_interpretation_allowed": true, "limitation": "Endpoint is first completed batch at the shared loop-time budget. Includes RL and auxiliary supervision; excludes initialization/export/evaluation/post-update sync. Actual token costs, GPU allocation time and overshoot remain separately reported. This is not exact FLOP matching." }, "timing": { "arm": "clean", "gpus": 4, "training_seconds_including_setup_and_export": 3302.788107490167, "audit_seconds": 23.864731727167964, "evaluation_seconds": 62.77103084884584, "elapsed_seconds": 3389.4356598854065, "allocated_gpu_hours_through_evaluation": 3.766039622094896, "training_allocated_gpu_hours": 3.6697645638779632 } }, "20261014/random_answer": { "arm": "random_answer", "seed": 20261014, "training": { "status": "REAL_ROLLOUT_MASK_AND_ADVANTAGE_VALIDATED", "counts": { "nonzero_advantage_tokens": 255445, "trajectories": 4480, "shared_trajectories": 0, "actual_exposed": 300, "exposed_with_nonzero_advantage": 116, "retrieved_trajectories": 3845, "shared_groups": 0, "steps": 70 }, "cost": { "rl_sequence_nonpadding_tokens": 3421336, "rl_sequence_padded_tokens": 9502592, "rl_loss_tokens": 865800, "backend_generated_tokens": 862893, "backend_requested_input_tokens": 5652356, "backend_calls": 10894, "aux_sequence_nonpadding_tokens_including_dp_padding": 103864, "aux_sequence_padded_tokens": 110276, "aux_loss_tokens": 267 }, "auxiliary_sources": { "groups_considered": 150, "skipped_no_verified_target": 80, "examples": 70, "answer_tokens": 267, "gold_checked_clean_answer": 45, "training_gold_with_support_title_coverage": 25 }, "cost_limitations": "Input tokens count actual backend requests, not cache-adjusted FLOPs. RL sequence tokens are one tensor pass proxy; policy/reference/rollout-logprob and backward differ in cost. Auxiliary includes zero-loss DP padding. Equal optimizer steps are not equal total computation." }, "budget": { "status": "BUDGET_AND_EXPOSURE_RECONCILED", "effective_steps": 70, "loop_seconds": 2714.7741301264614, "budget_seconds": 2700.0, "overshoot_seconds": 14.774130126461387, "overshoot_fraction": 0.005471900046837508, "last_batch_seconds": 36.14861673209816, "actual_exposed": 300, "target_exposed": 300, "exposure_templates": { "train_direct": 150, "train_role": 150 }, "time_budget_reached": true, "exposure_target_reached": true, "matched_interpretation_allowed": true, "limitation": "Endpoint is first completed batch at the shared loop-time budget. Includes RL and auxiliary supervision; excludes initialization/export/evaluation/post-update sync. Actual token costs, GPU allocation time and overshoot remain separately reported. This is not exact FLOP matching." }, "timing": { "arm": "random_answer", "gpus": 4, "training_seconds_including_setup_and_export": 3403.9832347370684, "audit_seconds": 26.96418576594442, "evaluation_seconds": 212.24144984222949, "elapsed_seconds": 3643.2243885993958, "allocated_gpu_hours_through_evaluation": 4.048027098443773, "training_allocated_gpu_hours": 3.7822035941522985 } } }