[mixed_ct_afv_unified_scoring_v2_resenc +0.0s] Starting training (regression): device=cuda, workers=64 [mixed_ct_afv_unified_scoring_v2_resenc +1.3s] Loading manifest and fiber annotations [mixed_ct_afv_unified_scoring_v2_resenc +46.4s] Loaded 563 training fibers and 188 validation fibers [mixed_ct_afv_unified_scoring_v2_resenc +46.4s] Preparing per-crop CT z-score normalization CT per-crop z-score, epsilon 1e-06: s3://vesuvius-challenge-open-data/PHercParis4/volumes/20260411134726-2.400um-0.2m-78keV-masked.zarr::2 CT per-crop z-score, epsilon 1e-06: s3://vesuvius-challenge-open-data/PHerc0175A/volumes/20250521115057-8.640um-1.2m-116keV-masked.zarr::0 CT per-crop z-score, epsilon 1e-06: s3://vesuvius-challenge-open-data/PHerc1447/volumes/20250521151220-8.640um-1.2m-116keV-masked.zarr::0 CT per-crop z-score, epsilon 1e-06: s3://vesuvius-challenge-open-data/PHercParis4/volumes/20260411134726-2.400um-0.2m-78keV-masked.zarr::2 CT per-crop z-score, epsilon 1e-06: s3://vesuvius-challenge-open-data/PHerc0125/volumes/20250821151825-9.362um-1.2m-113keV-masked.zarr::0 [mixed_ct_afv_unified_scoring_v2_resenc +46.4s] Preparing monitor recovery fixture [mixed_ct_afv_unified_scoring_v2_resenc +50.8s] Initializing models and optimizer [mixed_ct_afv_unified_scoring_v2_resenc +50.8s] Gradient clipping at global norm 60 (0 disables clipping) [mixed_ct_afv_unified_scoring_v2_resenc +50.8s] Refinement loss: every pass for every decision; initial proposal full weight, passes x1 of their mean; selection and the attempt metric follow confidence 0.7 [mixed_ct_afv_unified_scoring_v2_resenc +52.3s] Initialized from /ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v1_ext/ckpt_077000.pt: 120 tensors left at initialization, 4 checkpoint tensors unused, 4 of another shape [mixed_ct_afv_unified_scoring_v2_resenc +52.3s] fresh: cnn.stages.0.blocks.0.conv1.all_modules.0.bias, cnn.stages.0.blocks.0.conv1.all_modules.0.weight, cnn.stages.0.blocks.0.conv1.all_modules.1.bias, cnn.stages.0.blocks.0.conv1.all_modules.1.weight, cnn.stages.0.blocks.0.conv1.conv.bias, cnn.stages.0.blocks.0.conv1.conv.weight, cnn.stages.0.blocks.0.conv1.norm.bias, cnn.stages.0.blocks.0.conv1.norm.weight, cnn.stages.0.blocks.0.conv2.all_modules.0.bias, cnn.stages.0.blocks.0.conv2.all_modules.0.weight, cnn.stages.0.blocks.0.conv2.all_modules.1.bias, cnn.stages.0.blocks.0.conv2.all_modules.1.weight, cnn.stages.0.blocks.0.conv2.conv.bias, cnn.stages.0.blocks.0.conv2.conv.weight, cnn.stages.0.blocks.0.conv2.norm.bias, cnn.stages.0.blocks.0.conv2.norm.weight, cnn.stages.1.blocks.0.conv1.all_modules.0.weight, cnn.stages.1.blocks.0.conv1.conv.weight, cnn.stages.1.blocks.0.skip.1.all_modules.0.weight, cnn.stages.1.blocks.0.skip.1.conv.weight ... [mixed_ct_afv_unified_scoring_v2_resenc +52.3s] unexpected: cnn.stages.1.blocks.0.conv1.all_modules.0.bias, cnn.stages.1.blocks.0.conv1.conv.bias, cnn.stages.1.blocks.0.conv2.all_modules.0.bias, cnn.stages.1.blocks.0.conv2.conv.bias [mixed_ct_afv_unified_scoring_v2_resenc +52.3s] mismatched: cnn.stages.1.blocks.0.conv1.all_modules.0.weight, cnn.stages.1.blocks.0.conv1.conv.weight, cnn.stages.1.blocks.0.skip.1.all_modules.0.weight, cnn.stages.1.blocks.0.skip.1.conv.weight [mixed_ct_afv_unified_scoring_v2_resenc +53.6s] Compiling training operations; first forward/backward passes may take several minutes [mixed_ct_afv_unified_scoring_v2_resenc +53.6s] Loading replay banks and preparing data loader [mixed_ct_afv_unified_scoring_v2_resenc +53.6s] 32 independent decisions per batch [mixed_ct_afv_unified_scoring_v2_resenc +53.6s] Checking mixed-source datasets and AFV checksums [mixed_ct_afv_unified_scoring_v2_resenc +78.6s] Dataset sampling: paris4=20.0%, 0175A_5mm_v1=20.0%, 1447_5mm_v1=20.0%, paris4_afv_central=20.0%, PHerc0125_3mm_v2=20.0%; AFV training draws and AFV collection order proportional to length**3 [mixed_ct_afv_unified_scoring_v2_resenc +78.6s] Task budget per source: fresh 30%, live 45%, dagger_pre_excursion 6%, dagger_recoverable 8%, dagger_terminal 8%, dagger_premature_stop 3%, dagger_ordinary 0%, synthetic_terminal 0%; replay age ceiling 3000, event cap 64, synthetic terminal fallback cap 0% [mixed_ct_afv_unified_scoring_v2_resenc +78.6s] paris4: fresh locations weighted (bend x2, compression x2) for 563/563 fibers from paris4_3cfaec19b98d_v1.npz [mixed_ct_afv_unified_scoring_v2_resenc +79.0s] 0175A_5mm_v1: fresh locations weighted (bend x2, compression x2) for 60894/60894 fibers from 0175A_5mm_v1_d4374c646e4f_v1.npz [mixed_ct_afv_unified_scoring_v2_resenc +79.6s] 1447_5mm_v1: fresh locations weighted (bend x2, compression x2) for 80356/80356 fibers from 1447_5mm_v1_e235b5235b3d_v1.npz [mixed_ct_afv_unified_scoring_v2_resenc +80.1s] paris4_afv_central: fresh locations weighted (bend x2, compression x2) for 83004/83004 fibers from paris4_afv_central_bf27920bc12f_v1.npz [mixed_ct_afv_unified_scoring_v2_resenc +80.3s] PHerc0125_3mm_v2: fresh locations weighted (bend x2, compression x2) for 41893/41893 fibers from PHerc0125_3mm_v2_432a7fb7d155_v1.npz Step 0 | event: "ct_normalization" | calibration: {"method": "crop_zscore_v1", "volumes": {"s3://vesuvius-challenge-open-data/PHercParis4/volumes/20260411134726-2.400um-0.2m-78keV-masked.zarr::2": {"method": "crop_zscore_v1", "volume": "s3://vesuvius-challenge-open-data/PHercParis4/volumes/20260411134726-2.400um-0.2m-78keV-masked.zarr::2", "epsilon": 1e-06, "shape": [18946, 8174, 8174], "chunks": [128, 128, 128], "dtype": "|u1"}, "s3://vesuvius-challenge-open-data/PHerc0175A/volumes/20250521115057-8.640um-1.2m-116keV-masked.zarr::0": {"method": "crop_zscore_v1", "volume": "s3://vesuvius-challenge-open-data/PHerc0175A/volumes/20250521115057-8.640um-1.2m-116keV-masked.zarr::0", "epsilon": 1e-06, "shape": [12748, 9363, 9363], "chunks": [128, 128, 128], "dtype": "|u1"}, "s3://vesuvius-challenge-open-data/PHerc1447/volumes/20250521151220-8.640um-1.2m-116keV-masked.zarr::0": {"method": "crop_zscore_v1", "volume": "s3://vesuvius-challenge-open-data/PHerc1447/volumes/20250521151220-8.640um-1.2m-116keV-masked.zarr::0", "epsilon": 1e-06, "shape": [24297, 8343, 8343], "chunks": [128, 128, 128], "dtype": "|u1"}, "s3://vesuvius-challenge-open-data/PHerc0125/volumes/20250821151825-9.362um-1.2m-113keV-masked.zarr::0": {"method": "crop_zscore_v1", "volume": "s3://vesuvius-challenge-open-data/PHerc0125/volumes/20250821151825-9.362um-1.2m-113keV-masked.zarr::0", "epsilon": 1e-06, "shape": [20840, 8387, 8387], "chunks": [128, 128, 128], "dtype": "|u1"}}} Step 0 | event: "dataset_configuration" | datasets: [{"name": "paris4", "kind": "paris4", "fibers": 563}, {"name": "0175A_5mm_v1", "kind": "afv", "fibers": 60894, "excluded_fibers": 406, "sha256": "d4374c646e4fb9020ea0b76b8254bc1586cdedbdaa170278ce97fceea41ffe4e", "coordinate_space": "PHerc0175A/20250521115057", "validation": {"strategy": "fiber_hash", "seed": 7349, "count": 406}, "grid_scale": 2.0}, {"name": "1447_5mm_v1", "kind": "afv", "fibers": 80356, "excluded_fibers": 406, "sha256": "e235b5235b3dfef879037a0267e7a50f5a2ba791ffdcc16c2fade3a306705fa2", "coordinate_space": "PHerc1447/20250521151220", "validation": {"strategy": "fiber_hash", "seed": 7349, "count": 406}, "grid_scale": 2.0}, {"name": "paris4_afv_central", "kind": "afv", "fibers": 83004, "excluded_fibers": 406, "sha256": "bf27920bc12f96baebfbe17c82bf4655efe67e48809c87bddcd0a1e2e01c68f6", "coordinate_space": "PHercParis4/20260411134726", "validation": {"strategy": "fiber_hash", "seed": 7349, "count": 406}, "grid_scale": 8.0}, {"name": "PHerc0125_3mm_v2", "kind": "afv", "fibers": 41893, "excluded_fibers": 406, "sha256": "432a7fb7d155046d319da87d27b05dff1cb81baae623a50af82357c1ac02649a", "coordinate_space": "PHerc0125/20250821151825", "validation": {"strategy": "fiber_hash", "seed": 7349, "count": 406}, "grid_scale": 2.0}] | names: ["paris4", "0175A_5mm_v1", "1447_5mm_v1", "paris4_afv_central", "PHerc0125_3mm_v2"] | probabilities: [0.2, 0.2, 0.2, 0.2, 0.2] | dataset_config_sha256: "754d0a401da58c93f53f8520c61439d7d5752afb1f241b3c44e839f7da921588" | input_mode: "ct" | evaluation_scope: "Source-specific held-out fibers; Paris 4 recovery fixture" Step 0 | event: "optimizer_configuration" | reset: false | data_parallel_processes: 1 | lr_restart_step: 0 | optimizer_state_entries: 0 | groups: [{"parameters": 277}] | trainable_parameters: 22466691 Step 0 | identity sampling (settings in the run configuration) [mixed_ct_afv_unified_scoring_v2_resenc +80.4s] Remote CT prefetch: 48 concurrent fetches, 512 requests per priority queue, 16 future batch plans per remote source/worker, separate async process Step 0 | event: "remote_prefetch_configuration" | enabled: true | connections: 48 | queue_size: 512 | lookahead: 16 | timeout: 120.0 | sources: 5 [mixed_ct_afv_unified_scoring_v2_resenc +80.4s] Starting data loader; batch 32 × grad steps 1 = 32 decisions per update; starting update 1 [mixed_ct_afv_unified_scoring_v2_resenc +438.5s] Update 1: data ready in 27.2s; running forward/backward and optimizer [mixed_ct_afv_unified_scoring_v2_resenc +627.1s] Update 1 complete in 188.6s; loss=4.79369 [mixed_ct_afv_unified_scoring_v2_resenc +627.1s] Update 2: waiting for data [mixed_ct_afv_unified_scoring_v2_resenc +627.1s] Update 2: data ready in 0.0s; running forward/backward and optimizer [mixed_ct_afv_unified_scoring_v2_resenc +628.0s] Update 2 complete in 1.0s; loss=5.31669 [mixed_ct_afv_unified_scoring_v2_resenc +628.0s] Update 3: waiting for data [mixed_ct_afv_unified_scoring_v2_resenc +628.0s] Update 3: data ready in 0.0s; running forward/backward and optimizer [mixed_ct_afv_unified_scoring_v2_resenc +628.9s] Update 3 complete in 0.9s; loss=4.57364 [mixed_ct_afv_unified_scoring_v2_resenc +628.9s] Update 4: waiting for data [mixed_ct_afv_unified_scoring_v2_resenc +628.9s] Update 4: data ready in 0.0s; running forward/backward and optimizer [mixed_ct_afv_unified_scoring_v2_resenc +629.8s] Update 4 complete in 0.9s; loss=5.44181 [mixed_ct_afv_unified_scoring_v2_resenc +629.8s] Update 5: waiting for data [mixed_ct_afv_unified_scoring_v2_resenc +629.8s] Update 5: data ready in 0.0s; running forward/backward and optimizer [mixed_ct_afv_unified_scoring_v2_resenc +630.7s] Update 5 complete in 0.9s; loss=4.86830 [mixed_ct_afv_unified_scoring_v2_resenc +630.7s] Startup progress complete; regular metrics every 50 updates Step 50 | last 50 updates / 1,600 crops | lr 1.50e-06 points (35/curve, tolerance 2 vox): 23,024 right / 27,080 wrong | 46.0% correct | 5,896 unknown loss 4.9802 | geometry 1.8597 | confidence 6.2411 | mean error 3.017 vox speed: 4643 ms/update | data wait 544 ms/update | 2.7 crops/s | 2.7 decisions/s | 3.70 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 936, from the previous prediction 664 confidence-labeled crops: terminal 4/1505 (0.3%) | recoverable 106/1505 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 320/1600 (20.0%), 2: 224/1600 (14.0%), 3: 352/1600 (22.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 2 chunks / 4.0 MiB | promoted 1 | preempted 0 | deferred 0 | errors 0 | lookahead 917 chunk references / 1 windows sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 55% 45% 26%>0% 26% 0 -- 89%/7%/0%/3% 62% 0/0/0/189 0175A_5mm_v1 320 53% 47% 25%>0% 25% 0 -- 86%/7%/0%/7% 61% 0/0/0/149 1447_5mm_v1 224 53% 47% 25%>0% 25% 0 -- 88%/7%/0%/5% 57% 0/0/0/106 paris4_afv_central 352 49% 51% 23%>0% 23% 0 -- 86%/7%/1%/6% 61% 0/0/0/178 PHerc0125_3mm_v2 288 52% 48% 21%>0% 21% 0 -- 86%/5%/0%/8% 62% 0/0/0/138 gradients: max 1.6e+02, clipped 34/50 updates Step 100 | last 50 updates / 1,600 crops | lr 3.00e-06 points (35/curve, tolerance 2 vox): 26,177 right / 23,170 wrong | 53.0% correct | 6,653 unknown loss 4.6073 | geometry 1.5853 | confidence 6.0440 | mean error 2.537 vox speed: 886 ms/update | data wait 1 ms/update | 35.9 crops/s | 35.9 decisions/s | 3.51 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 871, from the previous prediction 729 confidence-labeled crops: terminal 0/1493 (0.0%) | recoverable 109/1493 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 320/1600 (20.0%), 2: 416/1600 (26.0%), 3: 256/1600 (16.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 3 chunks / 6.0 MiB | promoted 1 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 58% 42% 27%>0% 27% 0 -- 89%/8%/0%/4% 58% 0/0/0/108 0175A_5mm_v1 320 55% 45% 30%>0% 30% 0 -- 84%/7%/0%/10% 56% 0/0/0/144 1447_5mm_v1 416 55% 45% 26%>0% 26% 0 -- 89%/7%/0%/4% 49% 0/0/0/186 paris4_afv_central 256 55% 45% 28%>0% 28% 0 -- 86%/5%/0%/8% 52% 0/0/0/114 PHerc0125_3mm_v2 352 56% 44% 23%>0% 23% 0 -- 84%/7%/0%/9% 62% 0/0/0/156 gradients: max 1.1e+02, clipped 24/50 updates Step 150 | last 50 updates / 1,600 crops | lr 4.50e-06 points (35/curve, tolerance 2 vox): 26,639 right / 22,395 wrong | 54.3% correct | 6,966 unknown loss 4.4207 | geometry 1.4669 | confidence 5.9075 | mean error 2.425 vox speed: 890 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 3.60 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 886, from the previous prediction 714 confidence-labeled crops: terminal 3/1490 (0.2%) | recoverable 106/1490 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 288/1600 (18.0%), 2: 192/1600 (12.0%), 3: 320/1600 (20.0%), 4: 544/1600 (34.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 9 chunks / 18.0 MiB | promoted 1 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 4 rows | mean depth 1 decisions | mean travel 8 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 57% 43% 23%>0% 23% 0 135 90%/6%/0%/4% 54% 2/0/0/107 0175A_5mm_v1 288 60% 40% 28%>0% 28% 0 -- 86%/8%/0%/5% 53% 0/0/0/114 1447_5mm_v1 192 50% 50% 22%>0% 22% 0 132 90%/9%/0%/2% 48% 1/0/0/95 paris4_afv_central 320 54% 46% 22%>0% 22% 0 -- 86%/5%/1%/8% 46% 0/0/0/148 PHerc0125_3mm_v2 544 55% 45% 26%>0% 26% 0 117 83%/6%/0%/11% 62% 1/0/0/243 gradients: max 98, clipped 20/50 updates Step 200 | last 50 updates / 1,600 crops | lr 6.00e-06 points (35/curve, tolerance 2 vox): 29,481 right / 20,982 wrong | 58.4% correct | 5,537 unknown loss 4.3793 | geometry 1.3885 | confidence 5.9816 | mean error 2.259 vox speed: 887 ms/update | data wait 1 ms/update | 35.8 crops/s | 35.8 decisions/s | 3.62 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 896, from the previous prediction 704 confidence-labeled crops: terminal 1/1516 (0.1%) | recoverable 96/1516 (6.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 288/1600 (18.0%), 2: 416/1600 (26.0%), 3: 256/1600 (16.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 11 chunks / 22.0 MiB | promoted 2 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 29 rows | mean depth 1 decisions | mean travel 8 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 55% 45% 26%>0% 26% 0 132 92%/7%/0%/1% 53% 9/0/0/151 0175A_5mm_v1 288 52% 48% 21%>0% 21% 0 131 87%/7%/0%/6% 52% 4/0/0/133 1447_5mm_v1 416 52% 48% 25%>0% 25% 0 125 91%/5%/0%/4% 43% 3/0/0/198 paris4_afv_central 256 52% 48% 22%>0% 22% 0 125 89%/5%/0%/6% 44% 7/0/0/115 PHerc0125_3mm_v2 288 57% 43% 27%>0% 27% 0 124 82%/7%/0%/11% 60% 6/0/0/118 gradients: max 1e+02, clipped 18/50 updates Step 250 | last 50 updates / 1,600 crops | lr 7.50e-06 points (35/curve, tolerance 2 vox): 29,841 right / 20,106 wrong | 59.7% correct | 6,053 unknown loss 4.2118 | geometry 1.3523 | confidence 5.7189 | mean error 2.234 vox speed: 886 ms/update | data wait 1 ms/update | 35.8 crops/s | 35.8 decisions/s | 3.49 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 880, from the previous prediction 720 confidence-labeled crops: terminal 2/1500 (0.1%) | recoverable 123/1500 (8.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 320/1600 (20.0%), 2: 192/1600 (12.0%), 3: 384/1600 (24.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 12 chunks / 24.0 MiB | promoted 2 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 21 rows | mean depth 1 decisions | mean travel 8 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 54% 46% 25%>0% 25% 0 130 88%/9%/0%/3% 50% 4/0/0/174 0175A_5mm_v1 320 52% 48% 22%>0% 22% 0 124 87%/7%/0%/6% 46% 2/0/0/152 1447_5mm_v1 192 50% 50% 26%>0% 26% 0 137 87%/8%/0%/5% 40% 8/1/0/88 paris4_afv_central 384 56% 44% 23%>0% 23% 0 129 84%/9%/0%/7% 47% 7/0/0/163 PHerc0125_3mm_v2 320 55% 45% 26%>0% 26% 0 -- 84%/5%/0%/11% 58% 0/0/0/145 gradients: max 1.2e+02, clipped 21/50 updates Step 300 | last 50 updates / 1,600 crops | lr 9.00e-06 points (35/curve, tolerance 2 vox): 31,488 right / 18,218 wrong | 63.3% correct | 6,294 unknown loss 3.9525 | geometry 1.2537 | confidence 5.3975 | mean error 2.070 vox speed: 889 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 3.27 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 871, from the previous prediction 729 confidence-labeled crops: terminal 2/1499 (0.1%) | recoverable 119/1499 (7.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 384/1600 (24.0%), 2: 352/1600 (22.0%), 3: 320/1600 (20.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 15 chunks / 30.0 MiB | promoted 2 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 30 rows | mean depth 1 decisions | mean travel 9 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 54% 46% 26%>0% 26% 0 132 92%/6%/0%/2% 48% 1/0/0/118 0175A_5mm_v1 384 55% 45% 25%>0% 25% 0 129 86%/9%/0%/5% 45% 10/0/0/164 1447_5mm_v1 352 57% 43% 28%>0% 28% 0 129 84%/9%/0%/6% 40% 10/1/0/143 paris4_afv_central 320 58% 42% 27%>0% 27% 0 138 86%/7%/0%/7% 40% 6/2/0/129 PHerc0125_3mm_v2 288 49% 51% 24%>0% 24% 0 123 83%/6%/0%/12% 53% 3/1/0/145 gradients: max 1.3e+02, clipped 21/50 updates Step 350 | last 50 updates / 1,600 crops | lr 1.05e-05 points (35/curve, tolerance 2 vox): 34,510 right / 15,127 wrong | 69.5% correct | 6,363 unknown loss 3.6358 | geometry 1.0902 | confidence 5.0913 | mean error 1.847 vox speed: 887 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 2.71 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 816, from the previous prediction 784 confidence-labeled crops: terminal 2/1500 (0.1%) | recoverable 108/1500 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 320/1600 (20.0%), 2: 448/1600 (28.0%), 3: 192/1600 (12.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 15 chunks / 30.0 MiB | promoted 2 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 70 rows | mean depth 1 decisions | mean travel 10 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 54% 46% 26%>0% 26% 0 132 90%/7%/0%/3% 46% 14/0/0/119 0175A_5mm_v1 320 58% 42% 32%>0% 32% 0 118 84%/9%/0%/7% 41% 15/1/0/120 1447_5mm_v1 448 53% 47% 22%>0% 22% 0 129 87%/6%/0%/6% 29% 15/0/0/195 paris4_afv_central 192 56% 44% 24%>0% 24% 0 132 90%/5%/1%/4% 36% 18/0/0/66 PHerc0125_3mm_v2 352 57% 43% 28%>0% 28% 0 132 85%/6%/0%/9% 45% 8/0/0/144 gradients: max 1.2e+02, clipped 22/50 updates Step 400 | last 50 updates / 1,600 crops | lr 1.20e-05 points (35/curve, tolerance 2 vox): 40,457 right / 10,146 wrong | 79.9% correct | 5,397 unknown loss 2.9096 | geometry 0.8070 | confidence 4.2051 | mean error 1.446 vox speed: 932 ms/update | data wait 1 ms/update | 34.1 crops/s | 34.1 decisions/s | 1.91 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 842, from the previous prediction 758 confidence-labeled crops: terminal 1/1519 (0.1%) | recoverable 92/1519 (6.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 256/1600 (16.0%), 2: 416/1600 (26.0%), 3: 320/1600 (20.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 16 chunks / 32.0 MiB | promoted 2 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 71 rows | mean depth 1 decisions | mean travel 9 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 52% 48% 20%>0% 20% 0 131 92%/5%/0%/3% 31% 8/0/0/178 0175A_5mm_v1 256 54% 46% 22%>0% 22% 0 122 92%/4%/0%/4% 27% 10/1/0/107 1447_5mm_v1 416 56% 44% 26%>0% 26% 0 123 88%/7%/0%/5% 18% 32/0/0/150 paris4_afv_central 320 58% 42% 23%>0% 23% 0 130 88%/5%/0%/7% 26% 16/0/0/119 PHerc0125_3mm_v2 224 53% 47% 24%>0% 24% 0 116 84%/7%/0%/9% 40% 5/0/0/100 gradients: max 1.1e+02, clipped 12/50 updates Step 450 | last 50 updates / 1,600 crops | lr 1.35e-05 points (35/curve, tolerance 2 vox): 42,444 right / 7,836 wrong | 84.4% correct | 5,720 unknown loss 2.4875 | geometry 0.6704 | confidence 3.6342 | mean error 1.250 vox speed: 888 ms/update | data wait 1 ms/update | 35.8 crops/s | 35.8 decisions/s | 1.65 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 736, from the previous prediction 864 confidence-labeled crops: terminal 4/1515 (0.3%) | recoverable 109/1515 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 320/1600 (20.0%), 2: 288/1600 (18.0%), 3: 352/1600 (22.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 21 chunks / 42.0 MiB | promoted 2 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 187 rows | mean depth 1 decisions | mean travel 9 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 55% 45% 28%>0% 28% 0 133 89%/8%/0%/2% 25% 36/1/0/108 0175A_5mm_v1 320 56% 44% 28%>0% 28% 0 131 84%/11%/0%/5% 22% 46/0/0/94 1447_5mm_v1 288 52% 48% 22%>0% 22% 0 129 90%/5%/0%/5% 14% 39/0/0/99 paris4_afv_central 352 50% 50% 20%>0% 20% 0 132 88%/5%/1%/7% 21% 34/3/0/143 PHerc0125_3mm_v2 320 53% 47% 22%>0% 22% 0 126 87%/5%/0%/8% 27% 32/1/0/119 gradients: max 1.1e+02, clipped 10/50 updates Step 500 | last 50 updates / 1,600 crops | lr 1.50e-05 points (35/curve, tolerance 2 vox): 44,636 right / 6,516 wrong | 87.3% correct | 4,848 unknown loss 2.0499 | geometry 0.5614 | confidence 2.9770 | mean error 1.105 vox speed: 892 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.57 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 553, from the previous prediction 1047 confidence-labeled crops: terminal 3/1539 (0.2%) | recoverable 107/1539 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 256/1600 (16.0%), 2: 448/1600 (28.0%), 3: 352/1600 (22.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 21 chunks / 42.0 MiB | promoted 2 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 357 rows | mean depth 1 decisions | mean travel 10 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 55% 45% 22%>0% 22% 0 133 90%/8%/0%/2% 22% 52/3/0/108 0175A_5mm_v1 256 57% 43% 28%>0% 28% 0 132 90%/8%/0%/2% 17% 46/0/0/64 1447_5mm_v1 448 52% 48% 25%>0% 25% 0 125 88%/7%/0%/5% 11% 157/5/0/60 paris4_afv_central 352 53% 47% 22%>0% 22% 0 126 88%/7%/0%/5% 19% 62/1/0/103 PHerc0125_3mm_v2 192 56% 44% 19%>0% 19% 0 131 92%/2%/1%/5% 25% 40/2/0/45 gradients: max 1e+02, clipped 4/50 updates Step 550 | last 50 updates / 1,600 crops | lr 1.65e-05 points (35/curve, tolerance 2 vox): 45,273 right / 5,695 wrong | 88.8% correct | 5,032 unknown loss 2.0041 | geometry 0.5667 | confidence 2.8748 | mean error 1.062 vox speed: 887 ms/update | data wait 1 ms/update | 35.8 crops/s | 35.8 decisions/s | 1.49 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 401, from the previous prediction 1199 confidence-labeled crops: terminal 2/1540 (0.1%) | recoverable 105/1540 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 384/1600 (24.0%), 2: 256/1600 (16.0%), 3: 416/1600 (26.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 25 chunks / 50.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 477 rows | mean depth 1 decisions | mean travel 10 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 60% 40% 27%>0% 27% 0 136 92%/7%/0%/1% 21% 101/2/0/14 0175A_5mm_v1 384 51% 49% 21%>0% 21% 0 130 88%/6%/0%/6% 13% 123/2/0/66 1447_5mm_v1 256 59% 41% 26%>0% 26% 0 136 91%/7%/0%/2% 11% 88/2/0/17 paris4_afv_central 416 56% 44% 26%>0% 26% 0 132 90%/7%/0%/3% 16% 107/2/0/76 PHerc0125_3mm_v2 256 59% 41% 28%>0% 28% 0 130 87%/5%/0%/7% 28% 58/2/0/46 gradients: max 92, clipped 7/50 updates Step 600 | last 50 updates / 1,600 crops | lr 1.80e-05 points (35/curve, tolerance 2 vox): 47,179 right / 4,410 wrong | 91.5% correct | 4,411 unknown loss 1.6563 | geometry 0.4496 | confidence 2.4135 | mean error 0.948 vox speed: 942 ms/update | data wait 1 ms/update | 33.7 crops/s | 33.7 decisions/s | 1.36 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 316, from the previous prediction 1284 confidence-labeled crops: terminal 1/1543 (0.1%) | recoverable 107/1543 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 384/1600 (24.0%), 2: 288/1600 (18.0%), 3: 384/1600 (24.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 25 chunks / 50.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 550 rows | mean depth 1 decisions | mean travel 12 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 57% 43% 28%>0% 28% 0 135 88%/7%/0%/5% 16% 92/4/0/33 0175A_5mm_v1 384 56% 44% 26%>0% 26% 0 131 91%/7%/0%/3% 13% 109/2/0/59 1447_5mm_v1 288 56% 44% 28%>0% 28% 0 137 90%/6%/0%/4% 7% 128/2/0/0 paris4_afv_central 384 59% 41% 28%>0% 28% 0 134 89%/9%/0%/2% 13% 120/1/0/37 PHerc0125_3mm_v2 256 54% 46% 22%>0% 22% 0 136 90%/3%/0%/6% 22% 101/2/0/18 gradients: max 84, clipped 1/50 updates Step 650 | last 50 updates / 1,600 crops | lr 1.95e-05 points (35/curve, tolerance 2 vox): 46,847 right / 4,397 wrong | 91.4% correct | 4,756 unknown loss 1.7001 | geometry 0.4494 | confidence 2.5013 | mean error 0.951 vox speed: 888 ms/update | data wait 1 ms/update | 35.8 crops/s | 35.8 decisions/s | 1.35 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 423, from the previous prediction 1177 confidence-labeled crops: terminal 0/1541 (0.0%) | recoverable 102/1541 (6.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 256/1600 (16.0%), 2: 192/1600 (12.0%), 3: 384/1600 (24.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 28 chunks / 56.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 469 rows | mean depth 2 decisions | mean travel 14 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 57% 43% 27%>0% 27% 0 143 89%/9%/0%/2% 17% 137/3/0/29 0175A_5mm_v1 256 58% 42% 26%>0% 26% 0 131 92%/5%/0%/2% 9% 45/0/0/62 1447_5mm_v1 192 51% 49% 23%>0% 23% 0 137 90%/7%/0%/4% 4% 95/2/0/0 paris4_afv_central 384 52% 48% 21%>0% 21% 0 133 92%/6%/0%/2% 8% 125/2/0/58 PHerc0125_3mm_v2 384 57% 43% 23%>0% 23% 0 130 87%/5%/0%/8% 22% 67/1/0/99 gradients: max 64, clipped 3/50 updates Step 700 | last 50 updates / 1,600 crops | lr 2.10e-05 points (35/curve, tolerance 2 vox): 46,712 right / 4,007 wrong | 92.1% correct | 5,281 unknown loss 1.6044 | geometry 0.4150 | confidence 2.3789 | mean error 0.891 vox speed: 889 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.38 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 306, from the previous prediction 1294 confidence-labeled crops: terminal 3/1534 (0.2%) | recoverable 104/1534 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 384/1600 (24.0%), 2: 288/1600 (18.0%), 3: 256/1600 (16.0%), 4: 480/1600 (30.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 30 chunks / 60.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 628 rows | mean depth 2 decisions | mean travel 16 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 55% 45% 28%>0% 28% 0 146 90%/8%/1%/2% 15% 78/0/0/8 0175A_5mm_v1 384 53% 47% 22%>0% 22% 0 146 91%/5%/0%/4% 11% 182/1/0/0 1447_5mm_v1 288 53% 47% 26%>0% 26% 0 149 91%/6%/0%/3% 6% 129/5/56/6 paris4_afv_central 256 53% 47% 25%>0% 25% 0 139 88%/8%/0%/4% 11% 121/4/0/0 PHerc0125_3mm_v2 480 52% 48% 23%>0% 23% 0 128 86%/7%/0%/7% 21% 118/2/0/113 gradients: max 1.3e+02, clipped 2/50 updates Step 750 | last 50 updates / 1,600 crops | lr 2.25e-05 points (35/curve, tolerance 2 vox): 47,743 right / 3,546 wrong | 93.1% correct | 4,711 unknown loss 1.4569 | geometry 0.4213 | confidence 2.0712 | mean error 0.882 vox speed: 887 ms/update | data wait 1 ms/update | 35.8 crops/s | 35.8 decisions/s | 1.31 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 310, from the previous prediction 1290 confidence-labeled crops: terminal 0/1547 (0.0%) | recoverable 112/1547 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 512/1600 (32.0%), 1: 192/1600 (12.0%), 2: 352/1600 (22.0%), 3: 288/1600 (18.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 31 chunks / 62.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 36 chunk references / 1 windows live chains: 572 rows | mean depth 2 decisions | mean travel 19 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 512 55% 45% 28%>0% 28% 0 142 90%/8%/0%/2% 13% 141/1/0/90 0175A_5mm_v1 192 53% 47% 27%>0% 27% 0 144 88%/9%/0%/3% 8% 90/4/0/0 1447_5mm_v1 352 55% 45% 24%>0% 24% 0 135 90%/7%/0%/3% 6% 121/2/12/36 paris4_afv_central 288 54% 46% 23%>0% 23% 0 134 91%/5%/0%/4% 9% 133/3/5/0 PHerc0125_3mm_v2 256 60% 40% 26%>0% 26% 0 132 88%/6%/0%/6% 22% 87/0/0/15 gradients: max 75, clipped 4/50 updates Step 800 | last 50 updates / 1,600 crops | lr 2.40e-05 points (35/curve, tolerance 2 vox): 47,578 right / 3,883 wrong | 92.5% correct | 4,539 unknown loss 1.5053 | geometry 0.4070 | confidence 2.1967 | mean error 0.880 vox speed: 891 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.29 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 226, from the previous prediction 1374 confidence-labeled crops: terminal 1/1538 (0.1%) | recoverable 128/1538 (8.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 416/1600 (26.0%), 2: 192/1600 (12.0%), 3: 416/1600 (26.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 31 chunks / 62.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 656 rows | mean depth 2 decisions | mean travel 18 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 56% 44% 30%>0% 30% 0 143 89%/10%/0%/2% 15% 113/2/0/0 0175A_5mm_v1 416 57% 43% 27%>0% 27% 0 148 88%/8%/0%/4% 10% 176/6/19/1 1447_5mm_v1 192 52% 48% 20%>0% 20% 0 140 93%/5%/0%/3% 5% 68/1/0/25 paris4_afv_central 416 55% 45% 29%>0% 29% 0 138 86%/10%/0%/5% 10% 171/2/45/16 PHerc0125_3mm_v2 320 57% 43% 27%>0% 27% 0 131 88%/7%/0%/6% 19% 128/1/0/9 gradients: max 66, clipped 2/50 updates Step 850 | last 50 updates / 1,600 crops | lr 2.54e-05 points (35/curve, tolerance 2 vox): 48,227 right / 3,146 wrong | 93.9% correct | 4,627 unknown loss 1.3202 | geometry 0.3401 | confidence 1.9602 | mean error 0.793 vox speed: 886 ms/update | data wait 1 ms/update | 35.9 crops/s | 35.9 decisions/s | 1.25 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 290, from the previous prediction 1310 confidence-labeled crops: terminal 1/1541 (0.1%) | recoverable 109/1541 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 320/1600 (20.0%), 2: 480/1600 (30.0%), 3: 160/1600 (10.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 31 chunks / 62.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 1331 chunk references / 1 windows live chains: 638 rows | mean depth 3 decisions | mean travel 24 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 50% 50% 23%>0% 23% 0 136 91%/7%/0%/1% 14% 127/0/0/79 0175A_5mm_v1 320 54% 46% 25%>0% 25% 0 133 90%/7%/0%/3% 8% 132/5/12/14 1447_5mm_v1 480 53% 47% 25%>0% 25% 0 142 90%/6%/0%/4% 4% 226/6/0/0 paris4_afv_central 160 51% 49% 23%>0% 23% 0 138 86%/6%/0%/8% 10% 58/2/0/20 PHerc0125_3mm_v2 224 58% 42% 27%>0% 27% 0 141 86%/8%/0%/5% 20% 95/2/0/0 gradients: max 71, clipped 4/50 updates Step 900 | last 50 updates / 1,600 crops | lr 2.69e-05 points (35/curve, tolerance 2 vox): 48,503 right / 3,149 wrong | 93.9% correct | 4,348 unknown loss 1.3034 | geometry 0.3532 | confidence 1.9004 | mean error 0.795 vox speed: 886 ms/update | data wait 1 ms/update | 35.8 crops/s | 35.8 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 253, from the previous prediction 1347 confidence-labeled crops: terminal 2/1549 (0.1%) | recoverable 106/1549 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 384/1600 (24.0%), 2: 288/1600 (18.0%), 3: 352/1600 (22.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 31 chunks / 62.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 668 rows | mean depth 3 decisions | mean travel 24 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 51% 49% 28%>0% 28% 0 141 89%/9%/0%/2% 13% 126/1/0/0 0175A_5mm_v1 384 52% 48% 23%>0% 23% 0 134 91%/5%/0%/4% 7% 120/1/0/63 1447_5mm_v1 288 56% 44% 25%>0% 25% 0 133 91%/7%/0%/2% 4% 105/6/0/23 paris4_afv_central 352 52% 48% 25%>0% 25% 0 144 90%/7%/0%/3% 8% 169/7/0/0 PHerc0125_3mm_v2 320 53% 47% 24%>0% 24% 0 147 87%/8%/1%/5% 19% 148/3/22/1 gradients: max 78, clipped 1/50 updates Step 950 | last 50 updates / 1,600 crops | lr 2.84e-05 points (35/curve, tolerance 2 vox): 48,179 right / 3,432 wrong | 93.4% correct | 4,389 unknown loss 1.3921 | geometry 0.3608 | confidence 2.0625 | mean error 0.831 vox speed: 888 ms/update | data wait 1 ms/update | 35.8 crops/s | 35.8 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 251, from the previous prediction 1349 confidence-labeled crops: terminal 3/1554 (0.2%) | recoverable 99/1554 (6.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 320/1600 (20.0%), 2: 224/1600 (14.0%), 3: 288/1600 (18.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 31 chunks / 62.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 645 rows | mean depth 3 decisions | mean travel 25 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 53% 47% 24%>0% 24% 0 146 94%/5%/0%/1% 14% 168/2/0/26 0175A_5mm_v1 320 54% 46% 23%>0% 23% 0 141 93%/5%/0%/2% 8% 147/1/0/0 1447_5mm_v1 224 58% 42% 27%>0% 27% 0 139 90%/7%/0%/3% 4% 80/2/0/14 paris4_afv_central 288 55% 45% 27%>0% 27% 0 138 90%/6%/0%/4% 9% 129/5/0/0 PHerc0125_3mm_v2 352 54% 46% 27%>0% 27% 0 136 86%/8%/1%/6% 17% 121/2/28/42 gradients: max 68, clipped 3/50 updates Step 1,000 | last 50 updates / 1,600 crops | lr 2.99e-05 points (35/curve, tolerance 2 vox): 48,793 right / 3,129 wrong | 94.0% correct | 4,078 unknown loss 1.2490 | geometry 0.3495 | confidence 1.7989 | mean error 0.799 vox speed: 887 ms/update | data wait 1 ms/update | 35.8 crops/s | 35.8 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 267, from the previous prediction 1333 confidence-labeled crops: terminal 1/1551 (0.1%) | recoverable 115/1551 (7.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 352/1600 (22.0%), 2: 192/1600 (12.0%), 3: 416/1600 (26.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 31 chunks / 62.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 640 rows | mean depth 4 decisions | mean travel 30 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 54% 46% 22%>0% 22% 0 141 90%/7%/0%/3% 11% 193/3/0/0 0175A_5mm_v1 352 55% 45% 25%>0% 25% 0 134 91%/6%/0%/3% 6% 153/3/0/6 1447_5mm_v1 192 55% 45% 24%>0% 24% 0 151 91%/8%/0%/1% 5% 86/3/0/0 paris4_afv_central 416 52% 48% 25%>0% 25% 0 142 89%/7%/0%/3% 7% 137/4/1/63 PHerc0125_3mm_v2 224 55% 45% 24%>0% 24% 0 132 87%/8%/0%/5% 19% 71/0/0/30 gradients: max 81, clipped 3/50 updates Step 1,000 | dagger_launched: true Step 1,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 251, "stopped": 0, "mean_example_error": 1.1993961750018982, "geometry_examples": 31, "examples": 32, "inference_and_measurement_seconds": 1.9064888395369053, "total_diagnostic_seconds": 4.034865327179432} Step 1,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 3/8 (37.5%) | false stops 3 1.5-3: recovered 5/8 (62.5%) | false stops 0 3-4.5: recovered 0/8 (0.0%) | false stops 0 4.5-6: recovered 1/8 (12.5%) | false stops 1 entire proposed 8-point prefix correct (distance): 11/28 (39.3%) gate @ 0.50: false stops 1/14 (7.1%) | accepted wrong 16/27 (59.3%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 2.482 2.400 following 16 464 1.555 1.495 recoverable 12 348 3.719 3.607 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 1.555 1.495 332 32 812 2.482 2.400 improved/worsened/compared states: 5/1/28 Step 1,000 | batch_diagnostics_seconds: 4.037600364536047 | recovery_seconds: 13.120556827634573 Step 1,050 | last 50 updates / 1,600 crops | lr 2.99e-05 points (35/curve, tolerance 2 vox): 47,077 right / 3,322 wrong | 93.4% correct | 5,601 unknown loss 1.3075 | geometry 0.3727 | confidence 1.8698 | mean error 0.829 vox speed: 908 ms/update | data wait 1 ms/update | 24.9 crops/s | 24.9 decisions/s | 1.29 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 230, from the previous prediction 1370 confidence-labeled crops: terminal 3/1529 (0.2%) | recoverable 102/1529 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 256/1600 (16.0%), 2: 256/1600 (16.0%), 3: 384/1600 (24.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 31 chunks / 62.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 649 rows | mean depth 4 decisions | mean travel 30 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 54% 46% 30%>0% 30% 0 141 89%/9%/0%/2% 11% 107/1/0/26 0175A_5mm_v1 256 58% 42% 26%>0% 26% 0 134 91%/5%/0%/3% 8% 107/2/0/0 1447_5mm_v1 256 52% 48% 21%>0% 21% 0 177 89%/6%/0%/5% 4% 124/1/1/0 paris4_afv_central 384 56% 44% 24%>0% 24% 0 133 91%/5%/0%/4% 9% 140/4/0/28 PHerc0125_3mm_v2 416 56% 44% 26%>0% 26% 0 142 86%/6%/0%/7% 18% 171/3/0/13 gradients: max 65, clipped 2/50 updates Step 1,073 | dagger_states: 306 | dagger_source_step: 1000 | dagger_caches: 1 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/decisions_001000.npz" | dagger_fibers: 128 | dagger_supply: {"states": 306, "traces": 128, "fibers": 128, "rows": {"terminal": 2, "premature_stop": 123, "recoverable": 4, "pre_excursion": 6, "ordinary": 170}, "events": {"terminal": 1, "premature_stop": 123, "recoverable": 3, "pre_excursion": 3, "ordinary": 65}, "supervision": {"following": 296, "recoverable": 4, "terminal": 2, "unknown": 4}, "departed_traces": 4, "terminal_traces": 1, "stop_reasons": {"confidence": 127, "unannotated": 1}, "traced_voxels": 2537.4866199842095, "seconds": 86.88353578001261, "voxels_per_second": 29.205609523179113} | dagger_coverage: {"requested": 128, "seeds": 128, "skipped": [], "epoch": 0} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 0 | dagger_launch_interval: null | dagger_publication_age: 73 | dataset: "paris4" Step 1,100 | last 50 updates / 1,600 crops | lr 2.99e-05 points (35/curve, tolerance 2 vox): 47,405 right / 3,716 wrong | 92.7% correct | 4,879 unknown loss 1.3768 | geometry 0.3813 | confidence 1.9909 | mean error 0.833 vox speed: 1075 ms/update | data wait 1 ms/update | 29.5 crops/s | 29.5 decisions/s | 1.29 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 220, from the previous prediction 1380 confidence-labeled crops: terminal 3/1541 (0.2%) | recoverable 101/1541 (6.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 224/1600 (14.0%), 2: 256/1600 (16.0%), 3: 320/1600 (20.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 31 chunks / 62.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 659 rows | mean depth 4 decisions | mean travel 33 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 56% 44% 27%>0% 27% 0 145 89%/9%/0%/2% 15% 197/2/0/0 0175A_5mm_v1 224 50% 50% 22%>0% 22% 0 152 92%/4%/0%/4% 7% 111/0/0/0 1447_5mm_v1 256 53% 47% 25%>0% 25% 0 164 91%/5%/0%/4% 5% 121/2/7/0 paris4_afv_central 320 57% 43% 23%>0% 23% 0 138 90%/5%/0%/5% 9% 117/1/0/22 PHerc0125_3mm_v2 352 54% 46% 26%>0% 26% 0 132 88%/6%/1%/5% 17% 113/3/0/49 gradients: max 1e+02, clipped 2/50 updates Step 1,150 | last 50 updates / 1,600 crops | lr 2.99e-05 points (35/curve, tolerance 2 vox): 49,354 right / 2,425 wrong | 95.3% correct | 4,221 unknown loss 1.1170 | geometry 0.3191 | confidence 1.5958 | mean error 0.751 vox speed: 921 ms/update | data wait 1 ms/update | 34.5 crops/s | 34.5 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 240, from the previous prediction 1360 confidence-labeled crops: terminal 2/1553 (0.1%) | recoverable 95/1553 (6.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 256/1600 (16.0%), 2: 384/1600 (24.0%), 3: 384/1600 (24.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 32 chunks / 64.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 665 rows | mean depth 4 decisions | mean travel 35 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 56% 44% 23%>0% 23% 0 142 92%/6%/0%/2% 11% 169/1/0/0 0175A_5mm_v1 256 57% 43% 32%>0% 32% 0 158 91%/7%/0%/3% 7% 109/2/0/0 1447_5mm_v1 384 55% 45% 23%>0% 23% 0 147 92%/6%/0%/2% 3% 144/3/35/28 paris4_afv_central 384 54% 46% 26%>0% 26% 0 137 91%/6%/0%/3% 7% 161/2/0/14 PHerc0125_3mm_v2 192 54% 46% 23%>0% 23% 0 142 88%/5%/1%/7% 15% 82/3/0/6 gradients: max 72, clipped 2/50 updates Step 1,200 | last 50 updates / 1,600 crops | lr 2.99e-05 points (35/curve, tolerance 2 vox): 48,041 right / 2,797 wrong | 94.5% correct | 5,162 unknown loss 1.1700 | geometry 0.3186 | confidence 1.7027 | mean error 0.752 vox speed: 888 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.24 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 211, from the previous prediction 1389 confidence-labeled crops: terminal 4/1533 (0.3%) | recoverable 111/1533 (7.2%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 384/1600 (24.0%), 2: 288/1600 (18.0%), 3: 224/1600 (14.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 32 chunks / 64.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 657 rows | mean depth 5 decisions | mean travel 40 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 55% 45% 22%>0% 22% 0 142 91%/7%/0%/2% 12% 159/2/0/0 0175A_5mm_v1 384 59% 41% 26%>0% 26% 0 157 88%/7%/0%/5% 6% 159/1/0/0 1447_5mm_v1 288 57% 43% 31%>0% 31% 0 134 88%/7%/0%/6% 4% 84/1/0/39 paris4_afv_central 224 55% 45% 25%>0% 25% 0 140 88%/7%/0%/5% 5% 100/1/0/0 PHerc0125_3mm_v2 352 56% 44% 28%>0% 28% 0 147 88%/7%/1%/4% 16% 155/2/0/0 gradients: max 68, clipped 1/50 updates Step 1,250 | last 50 updates / 1,600 crops | lr 2.99e-05 points (35/curve, tolerance 2 vox): 49,040 right / 3,039 wrong | 94.2% correct | 3,921 unknown loss 1.2583 | geometry 0.3428 | confidence 1.8311 | mean error 0.780 vox speed: 888 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 314, from the previous prediction 1286 confidence-labeled crops: terminal 5/1558 (0.3%) | recoverable 120/1558 (7.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 608/1600 (38.0%), 1: 320/1600 (20.0%), 2: 288/1600 (18.0%), 3: 160/1600 (10.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 32 chunks / 64.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 568 rows | mean depth 5 decisions | mean travel 41 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 608 55% 45% 25%>0% 25% 0 132 91%/7%/0%/1% 12% 166/1/45/106 0175A_5mm_v1 320 55% 45% 29%>0% 29% 0 141 87%/10%/0%/3% 5% 108/4/45/37 1447_5mm_v1 288 53% 47% 26%>0% 26% 0 132 89%/8%/0%/3% 5% 118/2/0/17 paris4_afv_central 160 49% 51% 22%>0% 22% 0 182 91%/5%/1%/3% 11% 82/0/0/0 PHerc0125_3mm_v2 224 58% 42% 27%>0% 27% 0 146 89%/5%/1%/5% 17% 94/4/0/0 gradients: max 78, clipped 2/50 updates Step 1,300 | last 50 updates / 1,600 crops | lr 2.99e-05 points (35/curve, tolerance 2 vox): 49,172 right / 2,685 wrong | 94.8% correct | 4,143 unknown loss 1.0439 | geometry 0.2898 | confidence 1.5082 | mean error 0.717 vox speed: 888 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1590 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 256, from the previous prediction 1334 confidence-labeled crops: terminal 4/1543 (0.3%) | recoverable 118/1543 (7.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 288/1600 (18.0%), 2: 480/1600 (30.0%), 3: 288/1600 (18.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 32 chunks / 64.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 653 rows | mean depth 5 decisions | mean travel 44 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 52% 45% 22%>3% 19% 6 152 89%/7%/0%/4% 12% 113/2/0/3 0175A_5mm_v1 288 57% 43% 26%>0% 26% 0 147 90%/6%/0%/4% 4% 125/3/0/0 1447_5mm_v1 480 54% 46% 23%>0% 23% 0 134 89%/8%/0%/3% 4% 173/1/0/50 paris4_afv_central 288 55% 45% 24%>0% 24% 0 171 90%/8%/0%/2% 6% 129/1/0/0 PHerc0125_3mm_v2 288 46% 54% 20%>0% 20% 0 140 86%/7%/1%/6% 16% 113/8/42/42 gradients: max 57, clipped 0/50 updates Step 1,350 | last 50 updates / 1,600 crops | lr 2.99e-05 points (35/curve, tolerance 2 vox): 48,577 right / 2,751 wrong | 94.6% correct | 4,672 unknown loss 1.1495 | geometry 0.3160 | confidence 1.6671 | mean error 0.747 vox speed: 887 ms/update | data wait 1 ms/update | 35.8 crops/s | 35.8 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 234, from the previous prediction 1366 confidence-labeled crops: terminal 2/1542 (0.1%) | recoverable 110/1542 (7.1%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 288/1600 (18.0%), 2: 224/1600 (14.0%), 3: 416/1600 (26.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 32 chunks / 64.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 641 rows | mean depth 6 decisions | mean travel 47 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 54% 46% 23%>0% 23% 0 142 88%/11%/0%/2% 12% 133/1/0/15 0175A_5mm_v1 288 58% 42% 26%>0% 26% 0 141 93%/5%/0%/2% 6% 121/0/0/0 1447_5mm_v1 224 55% 45% 26%>0% 26% 0 137 91%/6%/0%/3% 3% 101/0/0/0 paris4_afv_central 416 57% 43% 25%>0% 25% 0 159 88%/6%/0%/6% 7% 131/2/70/48 PHerc0125_3mm_v2 352 54% 46% 25%>0% 25% 0 134 89%/7%/0%/4% 14% 155/3/0/7 gradients: max 70, clipped 2/50 updates Step 1,400 | last 50 updates / 1,600 crops | lr 2.98e-05 points (35/curve, tolerance 2 vox): 49,107 right / 2,287 wrong | 95.6% correct | 4,606 unknown loss 0.9950 | geometry 0.2802 | confidence 1.4295 | mean error 0.711 vox speed: 896 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 324, from the previous prediction 1276 confidence-labeled crops: terminal 4/1534 (0.3%) | recoverable 109/1534 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 288/1600 (18.0%), 2: 416/1600 (26.0%), 3: 256/1600 (16.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 32 chunks / 64.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 578 rows | mean depth 6 decisions | mean travel 49 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 57% 43% 23%>0% 23% 0 162 91%/7%/1%/2% 10% 138/0/0/0 0175A_5mm_v1 288 50% 50% 21%>0% 21% 0 146 93%/5%/0%/2% 5% 143/1/0/0 1447_5mm_v1 416 56% 44% 23%>0% 23% 0 136 87%/8%/0%/5% 3% 140/3/59/43 paris4_afv_central 256 59% 41% 24%>0% 24% 0 136 89%/8%/0%/3% 5% 94/5/0/12 PHerc0125_3mm_v2 320 57% 43% 24%>0% 24% 0 134 85%/6%/1%/9% 15% 63/2/0/74 gradients: max 57, clipped 0/50 updates Step 1,450 | last 50 updates / 1,600 crops | lr 2.98e-05 points (35/curve, tolerance 2 vox): 49,270 right / 2,248 wrong | 95.6% correct | 4,482 unknown loss 0.9501 | geometry 0.2872 | confidence 1.3258 | mean error 0.704 vox speed: 888 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 235, from the previous prediction 1365 confidence-labeled crops: terminal 3/1532 (0.2%) | recoverable 117/1532 (7.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 192/1600 (12.0%), 2: 352/1600 (22.0%), 3: 416/1600 (26.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 32 chunks / 64.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 656 rows | mean depth 6 decisions | mean travel 49 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 55% 45% 26%>0% 26% 0 153 89%/8%/0%/3% 9% 169/0/49/5 0175A_5mm_v1 192 53% 47% 25%>0% 25% 0 150 88%/10%/0%/3% 5% 91/1/0/0 1447_5mm_v1 352 55% 45% 29%>0% 29% 0 140 87%/8%/0%/5% 3% 141/0/0/17 paris4_afv_central 416 55% 45% 25%>0% 25% 0 134 89%/6%/0%/5% 6% 148/3/0/38 PHerc0125_3mm_v2 256 58% 42% 26%>0% 26% 0 146 88%/6%/0%/6% 13% 107/0/0/0 gradients: max 69, clipped 1/50 updates Step 1,500 | last 50 updates / 1,600 crops | lr 2.98e-05 points (35/curve, tolerance 2 vox): 49,430 right / 2,884 wrong | 94.5% correct | 3,686 unknown loss 1.1255 | geometry 0.3423 | confidence 1.5665 | mean error 0.754 vox speed: 888 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 233, from the previous prediction 1367 confidence-labeled crops: terminal 2/1554 (0.1%) | recoverable 135/1554 (8.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 416/1600 (26.0%), 2: 352/1600 (22.0%), 3: 256/1600 (16.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 32 chunks / 64.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 636 rows | mean depth 7 decisions | mean travel 55 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 54% 46% 24%>0% 24% 0 148 89%/9%/0%/3% 15% 115/3/50/3 0175A_5mm_v1 416 56% 44% 24%>0% 24% 0 149 91%/8%/0%/1% 5% 160/4/31/22 1447_5mm_v1 352 60% 40% 29%>0% 29% 0 136 86%/11%/0%/3% 4% 118/2/0/22 paris4_afv_central 256 57% 43% 23%>0% 23% 0 142 90%/8%/0%/2% 6% 109/2/0/0 PHerc0125_3mm_v2 320 58% 42% 27%>0% 27% 0 148 87%/7%/0%/6% 14% 134/3/0/0 gradients: max 70, clipped 1/50 updates Step 1,550 | last 50 updates / 1,600 crops | lr 2.98e-05 points (35/curve, tolerance 2 vox): 49,959 right / 2,109 wrong | 95.9% correct | 3,932 unknown loss 0.9660 | geometry 0.2700 | confidence 1.3919 | mean error 0.688 vox speed: 886 ms/update | data wait 1 ms/update | 35.8 crops/s | 35.8 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 221, from the previous prediction 1379 confidence-labeled crops: terminal 4/1551 (0.3%) | recoverable 103/1551 (6.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 224/1600 (14.0%), 2: 384/1600 (24.0%), 3: 320/1600 (20.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 684 rows | mean depth 6 decisions | mean travel 48 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 55% 45% 28%>0% 28% 0 146 89%/9%/0%/2% 5% 145/2/0/0 0175A_5mm_v1 224 54% 46% 25%>0% 25% 0 144 92%/7%/0%/1% 6% 104/1/0/0 1447_5mm_v1 384 61% 39% 27%>0% 27% 0 138 88%/9%/0%/3% 5% 148/2/0/2 paris4_afv_central 320 55% 45% 25%>0% 25% 0 144 94%/4%/0%/2% 5% 144/2/0/0 PHerc0125_3mm_v2 352 49% 51% 22%>0% 22% 0 147 89%/4%/1%/6% 12% 143/4/40/38 gradients: max 58, clipped 0/50 updates Step 1,600 | last 50 updates / 1,600 crops | lr 2.98e-05 points (35/curve, tolerance 2 vox): 49,346 right / 2,284 wrong | 95.6% correct | 4,370 unknown loss 1.0219 | geometry 0.2857 | confidence 1.4725 | mean error 0.702 vox speed: 889 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 252, from the previous prediction 1348 confidence-labeled crops: terminal 4/1536 (0.3%) | recoverable 97/1536 (6.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 256/1600 (16.0%), 2: 416/1600 (26.0%), 3: 320/1600 (20.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 638 rows | mean depth 7 decisions | mean travel 57 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 51% 49% 22%>0% 22% 0 152 95%/4%/0%/1% 10% 99/0/8/11 0175A_5mm_v1 256 55% 45% 26%>0% 26% 0 142 90%/6%/0%/4% 4% 115/1/0/0 1447_5mm_v1 416 54% 46% 23%>0% 23% 0 151 90%/6%/0%/4% 3% 193/6/0/0 paris4_afv_central 320 58% 42% 26%>0% 26% 0 143 88%/8%/0%/4% 8% 133/2/19/1 PHerc0125_3mm_v2 384 57% 43% 26%>0% 26% 0 134 87%/6%/1%/6% 15% 98/3/0/68 gradients: max 63, clipped 2/50 updates Step 1,650 | last 50 updates / 1,600 crops | lr 2.98e-05 points (35/curve, tolerance 2 vox): 49,010 right / 2,919 wrong | 94.4% correct | 4,071 unknown loss 1.1564 | geometry 0.3272 | confidence 1.6584 | mean error 0.765 vox speed: 888 ms/update | data wait 1 ms/update | 35.8 crops/s | 35.8 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 289, from the previous prediction 1311 confidence-labeled crops: terminal 3/1545 (0.2%) | recoverable 129/1545 (8.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 128/1600 (8.0%), 2: 320/1600 (20.0%), 3: 352/1600 (22.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 603 rows | mean depth 7 decisions | mean travel 60 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 55% 45% 25%>0% 25% 0 158 90%/7%/0%/3% 11% 172/2/0/0 0175A_5mm_v1 128 49% 51% 24%>0% 24% 0 153 91%/6%/0%/3% 6% 65/2/0/0 1447_5mm_v1 320 55% 45% 28%>0% 28% 0 139 87%/10%/0%/3% 4% 144/1/1/0 paris4_afv_central 352 57% 43% 25%>0% 25% 0 140 88%/10%/0%/2% 7% 114/2/11/39 PHerc0125_3mm_v2 416 54% 46% 21%>0% 21% 0 131 88%/6%/0%/6% 13% 108/3/0/83 gradients: max 67, clipped 3/50 updates Step 1,700 | last 50 updates / 1,600 crops | lr 2.98e-05 points (35/curve, tolerance 2 vox): 49,156 right / 2,513 wrong | 95.1% correct | 4,331 unknown loss 1.0891 | geometry 0.3145 | confidence 1.5492 | mean error 0.741 vox speed: 891 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1597 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 210, from the previous prediction 1387 confidence-labeled crops: terminal 2/1546 (0.1%) | recoverable 108/1546 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 320/1600 (20.0%), 2: 224/1600 (14.0%), 3: 192/1600 (12.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 670 rows | mean depth 7 decisions | mean travel 53 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 53% 46% 23%>1% 22% 3 150 92%/6%/0%/2% 10% 184/1/8/22 0175A_5mm_v1 320 55% 45% 23%>0% 23% 0 168 88%/7%/0%/5% 4% 145/1/0/0 1447_5mm_v1 224 58% 42% 20%>0% 20% 0 141 91%/7%/0%/2% 2% 93/0/0/0 paris4_afv_central 192 62% 38% 28%>0% 28% 0 158 90%/8%/0%/2% 7% 72/0/0/0 PHerc0125_3mm_v2 416 58% 42% 26%>0% 26% 0 140 88%/6%/0%/5% 13% 176/8/0/0 gradients: max 58, clipped 0/50 updates Step 1,750 | last 50 updates / 1,600 crops | lr 2.97e-05 points (35/curve, tolerance 2 vox): 49,328 right / 2,453 wrong | 95.3% correct | 4,219 unknown loss 1.0700 | geometry 0.3352 | confidence 1.4696 | mean error 0.750 vox speed: 885 ms/update | data wait 1 ms/update | 35.9 crops/s | 35.9 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 282, from the previous prediction 1318 confidence-labeled crops: terminal 2/1544 (0.1%) | recoverable 80/1544 (5.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 192/1600 (12.0%), 2: 192/1600 (12.0%), 3: 448/1600 (28.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 653 rows | mean depth 7 decisions | mean travel 60 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 55% 45% 23%>0% 23% 0 141 93%/5%/0%/2% 10% 104/0/0/68 0175A_5mm_v1 192 51% 49% 21%>0% 21% 0 164 92%/5%/0%/3% 4% 95/1/0/0 1447_5mm_v1 192 50% 50% 22%>0% 22% 0 158 91%/6%/0%/3% 3% 96/1/0/0 paris4_afv_central 448 55% 45% 24%>0% 24% 0 155 91%/5%/0%/4% 6% 202/3/0/0 PHerc0125_3mm_v2 384 53% 47% 23%>0% 23% 0 137 90%/5%/0%/5% 12% 156/2/0/24 gradients: max 74, clipped 2/50 updates Step 1,800 | last 50 updates / 1,600 crops | lr 2.97e-05 points (35/curve, tolerance 2 vox): 48,821 right / 3,004 wrong | 94.2% correct | 4,175 unknown loss 1.1453 | geometry 0.3303 | confidence 1.6300 | mean error 0.783 vox speed: 888 ms/update | data wait 1 ms/update | 35.8 crops/s | 35.8 decisions/s | 1.25 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 247, from the previous prediction 1353 confidence-labeled crops: terminal 2/1540 (0.1%) | recoverable 113/1540 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 256/1600 (16.0%), 2: 224/1600 (14.0%), 3: 288/1600 (18.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 665 rows | mean depth 8 decisions | mean travel 61 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 59% 41% 27%>0% 27% 0 142 89%/9%/0%/2% 11% 145/0/0/13 0175A_5mm_v1 256 55% 45% 22%>0% 22% 0 153 90%/7%/0%/3% 7% 114/3/0/0 1447_5mm_v1 224 57% 43% 25%>0% 25% 0 178 91%/5%/0%/4% 3% 97/1/0/0 paris4_afv_central 288 53% 47% 23%>0% 23% 0 152 90%/7%/0%/3% 8% 131/2/1/3 PHerc0125_3mm_v2 448 48% 52% 23%>0% 23% 0 137 87%/7%/0%/6% 12% 178/3/0/56 gradients: max 68, clipped 2/50 updates Step 1,850 | last 50 updates / 1,600 crops | lr 2.97e-05 points (35/curve, tolerance 2 vox): 48,292 right / 2,545 wrong | 95.0% correct | 5,163 unknown loss 1.0980 | geometry 0.3081 | confidence 1.5798 | mean error 0.740 vox speed: 985 ms/update | data wait 1 ms/update | 32.3 crops/s | 32.3 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 223, from the previous prediction 1377 confidence-labeled crops: terminal 0/1533 (0.0%) | recoverable 112/1533 (7.3%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 256/1600 (16.0%), 2: 320/1600 (20.0%), 3: 288/1600 (18.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 640 rows | mean depth 7 decisions | mean travel 60 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 60% 40% 34%>0% 34% 0 149 88%/10%/0%/2% 12% 115/0/0/0 0175A_5mm_v1 256 54% 46% 26%>0% 26% 0 156 87%/8%/0%/5% 4% 118/0/0/0 1447_5mm_v1 320 59% 41% 24%>0% 24% 0 177 88%/8%/0%/3% 3% 131/2/0/0 paris4_afv_central 288 54% 46% 25%>0% 25% 0 138 90%/6%/0%/4% 6% 113/3/0/19 PHerc0125_3mm_v2 448 54% 46% 26%>0% 26% 0 139 91%/4%/0%/6% 14% 163/2/0/43 gradients: max 62, clipped 1/50 updates Step 1,900 | last 50 updates / 1,600 crops | lr 2.97e-05 points (35/curve, tolerance 2 vox): 49,717 right / 2,348 wrong | 95.5% correct | 3,935 unknown loss 0.9851 | geometry 0.3149 | confidence 1.3404 | mean error 0.727 vox speed: 888 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 187, from the previous prediction 1413 confidence-labeled crops: terminal 3/1547 (0.2%) | recoverable 123/1547 (8.0%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 160/1600 (10.0%), 2: 352/1600 (22.0%), 3: 320/1600 (20.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 1144 chunk references / 1 windows live chains: 679 rows | mean depth 8 decisions | mean travel 68 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 56% 44% 26%>0% 26% 0 148 92%/7%/0%/1% 10% 184/1/0/0 0175A_5mm_v1 160 57% 43% 29%>0% 29% 0 158 89%/7%/0%/4% 4% 69/0/0/0 1447_5mm_v1 352 58% 42% 27%>0% 27% 0 152 90%/7%/0%/3% 2% 145/4/8/4 paris4_afv_central 320 55% 45% 28%>0% 28% 0 142 88%/8%/0%/4% 5% 145/2/0/0 PHerc0125_3mm_v2 352 55% 45% 25%>0% 25% 0 142 85%/9%/1%/5% 13% 136/2/0/23 gradients: max 65, clipped 1/50 updates Step 1,950 | last 50 updates / 1,600 crops | lr 2.97e-05 points (35/curve, tolerance 2 vox): 49,099 right / 2,594 wrong | 95.0% correct | 4,307 unknown loss 1.0629 | geometry 0.3261 | confidence 1.4736 | mean error 0.719 vox speed: 887 ms/update | data wait 1 ms/update | 35.8 crops/s | 35.8 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 245, from the previous prediction 1355 confidence-labeled crops: terminal 5/1549 (0.3%) | recoverable 106/1549 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 160/1600 (10.0%), 2: 320/1600 (20.0%), 3: 352/1600 (22.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 645 rows | mean depth 8 decisions | mean travel 66 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 54% 46% 28%>0% 28% 0 139 88%/9%/0%/3% 11% 174/1/70/4 0175A_5mm_v1 160 63% 37% 27%>0% 28% 0 179 92%/6%/0%/2% 3% 57/2/54/2 1447_5mm_v1 320 56% 44% 22%>0% 22% 0 133 91%/6%/0%/3% 5% 110/3/0/31 paris4_afv_central 352 52% 48% 23%>0% 23% 0 134 91%/6%/1%/2% 8% 131/1/0/38 PHerc0125_3mm_v2 384 55% 45% 21%>0% 21% 0 152 89%/5%/0%/5% 14% 173/2/0/0 gradients: max 58, clipped 0/50 updates Step 2,000 | last 50 updates / 1,600 crops | lr 2.97e-05 points (35/curve, tolerance 2 vox): 49,285 right / 2,085 wrong | 95.9% correct | 4,630 unknown loss 0.9472 | geometry 0.2592 | confidence 1.3759 | mean error 0.686 vox speed: 890 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 316, from the previous prediction 1284 confidence-labeled crops: terminal 2/1531 (0.1%) | recoverable 119/1531 (7.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 480/1600 (30.0%), 2: 288/1600 (18.0%), 3: 288/1600 (18.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 589 rows | mean depth 9 decisions | mean travel 75 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 59% 41% 26%>0% 26% 0 134 88%/9%/0%/3% 9% 77/0/1/55 0175A_5mm_v1 480 52% 48% 27%>0% 27% 0 150 88%/8%/0%/4% 5% 178/3/0/51 1447_5mm_v1 288 60% 40% 23%>0% 23% 0 139 89%/7%/0%/4% 4% 103/4/0/12 paris4_afv_central 288 53% 47% 24%>0% 24% 0 138 90%/7%/0%/3% 6% 123/0/0/12 PHerc0125_3mm_v2 224 52% 48% 21%>0% 21% 0 151 84%/7%/0%/8% 11% 108/6/0/0 gradients: max 78, clipped 2/50 updates Step 2,000 | dagger_launched: true Step 2,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 249, "stopped": 0, "mean_example_error": 0.5914091049663482, "geometry_examples": 31, "examples": 32, "inference_and_measurement_seconds": 2.019128628075123, "total_diagnostic_seconds": 4.194197792559862} Step 2,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 5/8 (62.5%) | false stops 1 1.5-3: recovered 4/8 (50.0%) | false stops 0 3-4.5: recovered 6/8 (75.0%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 0 entire proposed 8-point prefix correct (distance): 17/28 (60.7%) gate @ 0.50: false stops 1/19 (5.3%) | accepted wrong 9/26 (34.6%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.528 1.550 following 16 464 0.921 0.921 recoverable 12 348 2.338 2.390 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 0.921 0.921 332 32 812 1.528 1.550 improved/worsened/compared states: 0/2/28 Step 2,000 | batch_diagnostics_seconds: 4.197066497057676 | recovery_seconds: 15.623627558350563 Step 2,050 | last 50 updates / 1,600 crops | lr 2.97e-05 points (35/curve, tolerance 2 vox): 49,729 right / 2,057 wrong | 96.0% correct | 4,214 unknown loss 0.9715 | geometry 0.2614 | confidence 1.4200 | mean error 0.666 vox speed: 1288 ms/update | data wait 1 ms/update | 18.6 crops/s | 18.6 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 290, from the previous prediction 1310 confidence-labeled crops: terminal 3/1544 (0.2%) | recoverable 106/1544 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 352/1600 (22.0%), 2: 352/1600 (22.0%), 3: 288/1600 (18.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 610 rows | mean depth 9 decisions | mean travel 74 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 50% 50% 23%>0% 23% 0 146 90%/9%/0%/2% 9% 128/0/0/0 0175A_5mm_v1 352 51% 49% 26%>0% 26% 0 134 91%/5%/0%/3% 5% 53/0/0/120 1447_5mm_v1 352 59% 41% 27%>0% 27% 0 142 88%/7%/0%/5% 3% 141/3/0/3 paris4_afv_central 288 56% 44% 26%>0% 26% 0 142 90%/6%/0%/3% 5% 128/1/0/0 PHerc0125_3mm_v2 352 55% 45% 20%>0% 20% 0 160 89%/6%/1%/4% 13% 160/3/0/0 gradients: max 46, clipped 0/50 updates Step 2,100 | last 50 updates / 1,600 crops | lr 2.96e-05 points (35/curve, tolerance 2 vox): 49,290 right / 2,332 wrong | 95.5% correct | 4,378 unknown loss 0.9753 | geometry 0.2826 | confidence 1.3854 | mean error 0.691 vox speed: 1329 ms/update | data wait 1 ms/update | 24.0 crops/s | 24.0 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 254, from the previous prediction 1346 confidence-labeled crops: terminal 4/1547 (0.3%) | recoverable 118/1547 (7.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 160/1600 (10.0%), 2: 448/1600 (28.0%), 3: 352/1600 (22.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 614 rows | mean depth 8 decisions | mean travel 68 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 57% 43% 24%>0% 24% 0 165 90%/8%/0%/2% 9% 138/2/0/0 0175A_5mm_v1 160 56% 44% 25%>0% 25% 0 138 90%/7%/1%/2% 4% 18/0/0/52 1447_5mm_v1 448 55% 45% 26%>0% 26% 0 136 88%/8%/0%/4% 4% 169/3/0/34 paris4_afv_central 352 58% 42% 28%>0% 28% 0 150 88%/7%/1%/5% 8% 147/2/17/1 PHerc0125_3mm_v2 320 56% 44% 23%>0% 23% 0 159 89%/7%/0%/4% 12% 142/1/0/0 gradients: max 60, clipped 1/50 updates Step 2,150 | last 50 updates / 1,600 crops | lr 2.96e-05 points (35/curve, tolerance 2 vox): 49,294 right / 2,209 wrong | 95.7% correct | 4,497 unknown loss 0.9778 | geometry 0.2959 | confidence 1.3638 | mean error 0.699 vox speed: 1348 ms/update | data wait 1 ms/update | 23.6 crops/s | 23.6 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 244, from the previous prediction 1356 confidence-labeled crops: terminal 4/1540 (0.3%) | recoverable 109/1540 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 160/1600 (10.0%), 1: 384/1600 (24.0%), 2: 256/1600 (16.0%), 3: 384/1600 (24.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 643 rows | mean depth 9 decisions | mean travel 76 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 160 59% 41% 26%>0% 26% 0 170 91%/7%/0%/2% 6% 66/0/0/0 0175A_5mm_v1 384 56% 44% 29%>0% 29% 0 160 91%/8%/0%/2% 6% 168/2/0/0 1447_5mm_v1 256 54% 46% 23%>0% 23% 0 136 88%/6%/0%/6% 3% 85/0/0/33 paris4_afv_central 384 55% 45% 24%>0% 24% 0 137 90%/7%/0%/3% 4% 172/4/8/0 PHerc0125_3mm_v2 416 57% 43% 26%>0% 26% 0 144 87%/6%/1%/6% 14% 152/1/96/27 gradients: max 51, clipped 0/50 updates Step 2,200 | last 50 updates / 1,600 crops | lr 2.96e-05 points (35/curve, tolerance 2 vox): 49,849 right / 2,109 wrong | 95.9% correct | 4,042 unknown loss 0.9137 | geometry 0.2538 | confidence 1.3197 | mean error 0.667 vox speed: 1338 ms/update | data wait 1 ms/update | 23.8 crops/s | 23.8 decisions/s | 1.14 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 237, from the previous prediction 1363 confidence-labeled crops: terminal 2/1549 (0.1%) | recoverable 118/1549 (7.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 416/1600 (26.0%), 2: 352/1600 (22.0%), 3: 256/1600 (16.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 642 rows | mean depth 9 decisions | mean travel 71 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 55% 45% 22%>0% 22% 0 175 90%/8%/0%/2% 8% 144/0/0/0 0175A_5mm_v1 416 60% 40% 26%>0% 26% 0 157 89%/7%/0%/3% 7% 165/4/0/0 1447_5mm_v1 352 53% 47% 25%>0% 25% 0 138 87%/9%/0%/4% 2% 133/2/0/32 paris4_afv_central 256 55% 45% 19%>0% 19% 0 136 90%/7%/0%/4% 7% 80/1/5/35 PHerc0125_3mm_v2 256 53% 47% 21%>0% 21% 0 147 90%/5%/0%/4% 10% 120/1/0/0 gradients: max 69, clipped 2/50 updates Step 2,250 | last 50 updates / 1,600 crops | lr 2.96e-05 points (35/curve, tolerance 2 vox): 48,943 right / 2,498 wrong | 95.1% correct | 4,559 unknown loss 0.9830 | geometry 0.2980 | confidence 1.3700 | mean error 0.737 vox speed: 1342 ms/update | data wait 1 ms/update | 23.7 crops/s | 23.7 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1599 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 208, from the previous prediction 1391 confidence-labeled crops: terminal 3/1534 (0.2%) | recoverable 113/1534 (7.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 480/1600 (30.0%), 1: 224/1600 (14.0%), 2: 288/1600 (18.0%), 3: 288/1600 (18.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 707 rows | mean depth 9 decisions | mean travel 71 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 480 52% 47% 26%>0% 26% 1 149 87%/9%/0%/4% 10% 182/0/54/45 0175A_5mm_v1 224 53% 47% 21%>0% 21% 0 151 90%/8%/1%/2% 5% 106/2/0/0 1447_5mm_v1 288 54% 46% 25%>0% 25% 0 156 93%/3%/0%/4% 3% 132/3/0/0 paris4_afv_central 288 55% 45% 23%>0% 23% 0 146 87%/8%/0%/6% 4% 129/2/0/0 PHerc0125_3mm_v2 320 51% 49% 25%>0% 25% 0 141 88%/6%/0%/6% 12% 158/2/0/0 gradients: max 67, clipped 1/50 updates Step 2,300 | last 50 updates / 1,600 crops | lr 2.96e-05 points (35/curve, tolerance 2 vox): 48,894 right / 2,256 wrong | 95.6% correct | 4,850 unknown loss 0.9525 | geometry 0.2998 | confidence 1.3054 | mean error 0.709 vox speed: 1331 ms/update | data wait 1 ms/update | 23.9 crops/s | 23.9 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 308, from the previous prediction 1292 confidence-labeled crops: terminal 2/1538 (0.1%) | recoverable 100/1538 (6.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 256/1600 (16.0%), 2: 320/1600 (20.0%), 3: 384/1600 (24.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 631 rows | mean depth 10 decisions | mean travel 82 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 50% 50% 26%>0% 26% 0 134 92%/6%/0%/2% 10% 73/0/0/134 0175A_5mm_v1 256 52% 48% 21%>0% 21% 0 158 90%/5%/0%/4% 5% 123/0/0/1 1447_5mm_v1 320 53% 47% 23%>0% 23% 0 147 87%/8%/1%/4% 3% 150/4/9/0 paris4_afv_central 384 52% 48% 22%>0% 22% 0 149 89%/6%/0%/4% 7% 186/1/0/0 PHerc0125_3mm_v2 224 56% 44% 25%>0% 25% 0 156 90%/5%/0%/5% 10% 99/2/0/0 gradients: max 68, clipped 3/50 updates Step 2,350 | last 50 updates / 1,600 crops | lr 2.95e-05 points (35/curve, tolerance 2 vox): 49,637 right / 2,373 wrong | 95.4% correct | 3,990 unknown loss 0.9563 | geometry 0.2844 | confidence 1.3438 | mean error 0.699 vox speed: 1329 ms/update | data wait 1 ms/update | 24.0 crops/s | 24.0 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 206, from the previous prediction 1394 confidence-labeled crops: terminal 1/1553 (0.1%) | recoverable 131/1553 (8.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 384/1600 (24.0%), 2: 288/1600 (18.0%), 3: 288/1600 (18.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 709 rows | mean depth 9 decisions | mean travel 75 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 55% 45% 25%>0% 25% 0 141 90%/9%/0%/1% 8% 125/0/0/4 0175A_5mm_v1 384 54% 46% 23%>0% 23% 0 159 89%/8%/0%/3% 7% 177/4/0/0 1447_5mm_v1 288 55% 45% 29%>0% 29% 0 141 87%/10%/0%/3% 3% 129/0/35/2 paris4_afv_central 288 51% 49% 20%>0% 20% 0 145 91%/7%/0%/2% 5% 140/3/1/0 PHerc0125_3mm_v2 352 55% 45% 26%>0% 26% 0 149 87%/7%/0%/6% 13% 138/3/1/21 gradients: max 74, clipped 3/50 updates Step 2,400 | last 50 updates / 1,600 crops | lr 2.95e-05 points (35/curve, tolerance 2 vox): 48,919 right / 2,622 wrong | 94.9% correct | 4,459 unknown loss 1.0622 | geometry 0.3163 | confidence 1.4918 | mean error 0.723 vox speed: 1335 ms/update | data wait 1 ms/update | 23.8 crops/s | 23.8 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 234, from the previous prediction 1366 confidence-labeled crops: terminal 4/1539 (0.3%) | recoverable 119/1539 (7.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 320/1600 (20.0%), 2: 224/1600 (14.0%), 3: 320/1600 (20.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 673 rows | mean depth 10 decisions | mean travel 77 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 52% 48% 25%>0% 25% 0 154 89%/9%/0%/2% 10% 169/1/0/0 0175A_5mm_v1 320 57% 43% 27%>0% 27% 0 150 87%/9%/0%/4% 6% 137/3/0/2 1447_5mm_v1 224 61% 39% 28%>0% 28% 0 149 91%/6%/0%/3% 3% 88/0/0/0 paris4_afv_central 320 55% 45% 27%>0% 27% 0 136 88%/8%/1%/4% 5% 115/0/1/28 PHerc0125_3mm_v2 384 51% 49% 24%>0% 24% 0 144 88%/6%/1%/6% 17% 164/4/0/23 gradients: max 65, clipped 1/50 updates Step 2,450 | last 50 updates / 1,600 crops | lr 2.95e-05 points (35/curve, tolerance 2 vox): 49,682 right / 1,824 wrong | 96.5% correct | 4,494 unknown loss 0.8529 | geometry 0.2636 | confidence 1.1786 | mean error 0.668 vox speed: 1344 ms/update | data wait 1 ms/update | 23.7 crops/s | 23.7 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 216, from the previous prediction 1384 confidence-labeled crops: terminal 0/1539 (0.0%) | recoverable 127/1539 (8.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 384/1600 (24.0%), 2: 352/1600 (22.0%), 3: 352/1600 (22.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 692 rows | mean depth 11 decisions | mean travel 89 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 55% 45% 27%>0% 27% 0 150 89%/9%/0%/2% 9% 129/1/0/0 0175A_5mm_v1 384 53% 47% 26%>0% 26% 0 138 88%/8%/0%/4% 5% 136/0/0/43 1447_5mm_v1 352 56% 44% 25%>0% 25% 0 153 90%/7%/0%/3% 3% 156/5/0/0 paris4_afv_central 352 50% 50% 27%>0% 27% 0 141 89%/7%/0%/4% 4% 175/5/0/0 PHerc0125_3mm_v2 224 57% 43% 27%>0% 27% 0 147 83%/8%/0%/8% 10% 96/3/0/0 gradients: max 51, clipped 0/50 updates Step 2,500 | last 50 updates / 1,600 crops | lr 2.95e-05 points (35/curve, tolerance 2 vox): 49,606 right / 2,245 wrong | 95.7% correct | 4,149 unknown loss 0.9708 | geometry 0.3070 | confidence 1.3276 | mean error 0.705 vox speed: 1331 ms/update | data wait 1 ms/update | 23.9 crops/s | 23.9 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 250, from the previous prediction 1350 confidence-labeled crops: terminal 2/1548 (0.1%) | recoverable 116/1548 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 480/1600 (30.0%), 1: 96/1600 (6.0%), 2: 352/1600 (22.0%), 3: 352/1600 (22.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 666 rows | mean depth 11 decisions | mean travel 86 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 480 52% 48% 21%>0% 21% 0 137 91%/6%/0%/3% 8% 174/0/75/56 0175A_5mm_v1 96 59% 41% 31%>0% 31% 0 145 82%/11%/0%/6% 7% 39/2/0/0 1447_5mm_v1 352 50% 50% 22%>0% 22% 0 141 89%/7%/0%/4% 2% 160/4/24/17 paris4_afv_central 352 55% 45% 25%>0% 25% 0 139 87%/9%/1%/4% 6% 150/5/0/9 PHerc0125_3mm_v2 320 55% 45% 26%>0% 26% 0 158 91%/7%/0%/2% 12% 143/5/0/0 gradients: max 59, clipped 0/50 updates Step 2,550 | last 50 updates / 1,600 crops | lr 2.95e-05 points (35/curve, tolerance 2 vox): 48,303 right / 2,797 wrong | 94.5% correct | 4,900 unknown loss 1.0664 | geometry 0.3172 | confidence 1.4985 | mean error 0.737 vox speed: 1327 ms/update | data wait 1 ms/update | 24.0 crops/s | 24.0 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 266, from the previous prediction 1334 confidence-labeled crops: terminal 6/1540 (0.4%) | recoverable 110/1540 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 288/1600 (18.0%), 2: 224/1600 (14.0%), 3: 352/1600 (22.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 613 rows | mean depth 9 decisions | mean travel 76 vox | 6 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 59% 41% 28%>0% 28% 0 145 88%/9%/0%/4% 11% 70/1/0/61 0175A_5mm_v1 288 56% 44% 24%>0% 24% 0 181 90%/7%/0%/3% 6% 127/3/5/0 1447_5mm_v1 224 55% 45% 24%>0% 24% 0 138 92%/4%/0%/4% 4% 100/2/0/0 paris4_afv_central 352 56% 44% 28%>0% 28% 0 134 88%/7%/1%/5% 7% 122/2/0/32 PHerc0125_3mm_v2 416 52% 48% 24%>0% 24% 0 145 89%/6%/0%/4% 15% 194/5/16/4 gradients: max 70, clipped 1/50 updates Step 2,600 | last 50 updates / 1,600 crops | lr 2.94e-05 points (35/curve, tolerance 2 vox): 48,004 right / 3,037 wrong | 94.0% correct | 4,959 unknown loss 1.2104 | geometry 0.3602 | confidence 1.7003 | mean error 0.780 vox speed: 1345 ms/update | data wait 1 ms/update | 23.7 crops/s | 23.7 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1598 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 328, from the previous prediction 1270 confidence-labeled crops: terminal 4/1533 (0.3%) | recoverable 110/1533 (7.2%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 512/1600 (32.0%), 1: 192/1600 (12.0%), 2: 160/1600 (10.0%), 3: 256/1600 (16.0%), 4: 480/1600 (30.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 575 rows | mean depth 11 decisions | mean travel 90 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 512 56% 43% 24%>0% 24% 2 166 90%/8%/0%/2% 11% 222/2/0/0 0175A_5mm_v1 192 57% 43% 22%>0% 22% 0 157 92%/4%/0%/4% 5% 83/1/0/0 1447_5mm_v1 160 56% 44% 25%>0% 25% 0 145 86%/7%/0%/7% 3% 71/1/0/0 paris4_afv_central 256 55% 45% 24%>0% 24% 0 154 91%/7%/0%/2% 7% 116/1/0/0 PHerc0125_3mm_v2 480 55% 45% 24%>0% 24% 0 131 86%/6%/0%/8% 14% 83/1/0/135 gradients: max 70, clipped 2/50 updates Step 2,619 | dagger_states: 7528 | dagger_source_step: 2000 | dagger_caches: 1 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/0175A_5mm_v1/decisions_002000.npz" | dagger_fibers: 384 | dagger_supply: {"states": 7528, "traces": 384, "fibers": 384, "rows": {"terminal": 193, "premature_stop": 103, "recoverable": 31, "pre_excursion": 211, "ordinary": 6978}, "events": {"terminal": 25, "premature_stop": 103, "recoverable": 12, "pre_excursion": 38, "ordinary": 376}, "supervision": {"following": 7289, "recoverable": 31, "terminal": 193, "unknown": 15}, "departed_traces": 33, "terminal_traces": 25, "stop_reasons": {"after_failure": 23, "confidence": 105, "unannotated": 256}, "traced_voxels": 114767.63069187712, "seconds": 839.8896092437208, "voxels_per_second": 136.64608947266262} | dagger_coverage: {"requested": 384, "seeds": 384, "skipped": [], "epoch": 0} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 0 | dagger_launch_interval: null | dagger_publication_age: 619 | dataset: "0175A_5mm_v1" Step 2,650 | last 50 updates / 1,600 crops | lr 2.94e-05 points (35/curve, tolerance 2 vox): 49,415 right / 2,326 wrong | 95.5% correct | 4,259 unknown loss 0.9617 | geometry 0.2924 | confidence 1.3386 | mean error 0.714 vox speed: 1004 ms/update | data wait 1 ms/update | 31.7 crops/s | 31.7 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 269, from the previous prediction 1331 confidence-labeled crops: terminal 4/1540 (0.3%) | recoverable 94/1540 (6.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 512/1600 (32.0%), 2: 192/1600 (12.0%), 3: 352/1600 (22.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 639 rows | mean depth 12 decisions | mean travel 98 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 53% 47% 20%>0% 20% 0 143 92%/6%/0%/2% 9% 136/0/0/0 0175A_5mm_v1 512 52% 48% 23%>0% 23% 0 163 90%/6%/0%/4% 5% 150/5/37/95 1447_5mm_v1 192 53% 47% 25%>0% 25% 0 167 90%/7%/0%/3% 2% 90/0/0/0 paris4_afv_central 352 55% 45% 26%>0% 26% 0 155 87%/8%/1%/4% 6% 157/5/0/0 PHerc0125_3mm_v2 256 55% 45% 23%>0% 23% 0 145 92%/2%/1%/5% 14% 106/2/0/8 gradients: max 55, clipped 0/50 updates Step 2,700 | last 50 updates / 1,600 crops | lr 2.94e-05 points (35/curve, tolerance 2 vox): 49,328 right / 2,158 wrong | 95.8% correct | 4,514 unknown loss 0.9186 | geometry 0.2634 | confidence 1.3104 | mean error 0.658 vox speed: 889 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 236, from the previous prediction 1364 confidence-labeled crops: terminal 4/1540 (0.3%) | recoverable 102/1540 (6.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 448/1600 (28.0%), 2: 288/1600 (18.0%), 3: 352/1600 (22.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 676 rows | mean depth 10 decisions | mean travel 82 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 46% 54% 18%>0% 18% 0 146 93%/5%/0%/3% 6% 104/0/1/0 0175A_5mm_v1 448 55% 45% 25%>0% 25% 0 133 89%/6%/0%/5% 4% 135/0/0/66 1447_5mm_v1 288 55% 45% 26%>0% 26% 0 168 91%/4%/0%/4% 4% 131/1/40/0 paris4_afv_central 352 56% 44% 27%>0% 27% 0 155 90%/8%/1%/2% 7% 156/1/7/0 PHerc0125_3mm_v2 320 53% 47% 24%>0% 24% 0 157 86%/8%/0%/5% 13% 150/0/0/0 gradients: max 43, clipped 0/50 updates Step 2,750 | last 50 updates / 1,600 crops | lr 2.94e-05 points (35/curve, tolerance 2 vox): 49,799 right / 2,145 wrong | 95.9% correct | 4,056 unknown loss 0.9639 | geometry 0.2670 | confidence 1.3939 | mean error 0.664 vox speed: 888 ms/update | data wait 1 ms/update | 35.8 crops/s | 35.8 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 331, from the previous prediction 1269 confidence-labeled crops: terminal 2/1556 (0.1%) | recoverable 101/1556 (6.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 224/1600 (14.0%), 2: 448/1600 (28.0%), 3: 352/1600 (22.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 581 rows | mean depth 12 decisions | mean travel 99 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 58% 42% 24%>0% 24% 0 150 91%/8%/0%/1% 10% 121/1/0/0 0175A_5mm_v1 224 56% 44% 23%>0% 23% 0 141 91%/4%/0%/5% 4% 75/0/0/23 1447_5mm_v1 448 52% 48% 24%>0% 24% 0 137 92%/5%/0%/3% 5% 118/3/9/95 paris4_afv_central 352 54% 46% 24%>0% 24% 0 140 89%/8%/0%/3% 7% 152/4/0/10 PHerc0125_3mm_v2 288 60% 40% 28%>0% 28% 0 164 91%/7%/0%/2% 13% 115/4/0/0 gradients: max 63, clipped 1/50 updates Step 2,800 | last 50 updates / 1,600 crops | lr 2.94e-05 points (35/curve, tolerance 2 vox): 49,960 right / 2,137 wrong | 95.9% correct | 3,903 unknown loss 0.9130 | geometry 0.2578 | confidence 1.3104 | mean error 0.674 vox speed: 887 ms/update | data wait 1 ms/update | 35.8 crops/s | 35.8 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 244, from the previous prediction 1356 confidence-labeled crops: terminal 2/1550 (0.1%) | recoverable 119/1550 (7.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 320/1600 (20.0%), 2: 320/1600 (20.0%), 3: 256/1600 (16.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 651 rows | mean depth 12 decisions | mean travel 99 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 53% 47% 25%>0% 25% 0 159 90%/8%/0%/2% 9% 209/3/0/0 0175A_5mm_v1 320 56% 44% 27%>0% 27% 0 142 89%/6%/0%/5% 5% 142/2/0/0 1447_5mm_v1 320 52% 48% 23%>0% 23% 0 133 92%/5%/0%/3% 3% 71/1/0/81 paris4_afv_central 256 57% 43% 30%>0% 30% 0 144 87%/11%/0%/2% 4% 111/2/0/0 PHerc0125_3mm_v2 256 54% 46% 23%>0% 23% 0 154 87%/9%/0%/4% 13% 118/0/0/0 gradients: max 70, clipped 1/50 updates Step 2,850 | last 50 updates / 1,600 crops | lr 2.93e-05 points (35/curve, tolerance 2 vox): 49,579 right / 2,134 wrong | 95.9% correct | 4,287 unknown loss 0.9658 | geometry 0.2841 | confidence 1.3634 | mean error 0.669 vox speed: 891 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 258, from the previous prediction 1342 confidence-labeled crops: terminal 5/1548 (0.3%) | recoverable 119/1548 (7.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 224/1600 (14.0%), 2: 320/1600 (20.0%), 3: 416/1600 (26.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 652 rows | mean depth 11 decisions | mean travel 93 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 57% 43% 25%>0% 25% 0 138 89%/7%/0%/4% 9% 106/0/82/46 0175A_5mm_v1 224 51% 49% 24%>0% 24% 0 152 90%/7%/0%/3% 5% 110/2/0/0 1447_5mm_v1 320 57% 43% 24%>0% 24% 0 136 88%/8%/1%/3% 3% 129/4/0/8 paris4_afv_central 416 56% 44% 24%>0% 24% 0 141 88%/7%/0%/4% 6% 184/2/0/0 PHerc0125_3mm_v2 288 57% 43% 23%>0% 23% 0 150 89%/8%/0%/2% 13% 123/1/0/0 gradients: max 58, clipped 0/50 updates Step 2,900 | last 50 updates / 1,600 crops | lr 2.93e-05 points (35/curve, tolerance 2 vox): 48,644 right / 2,715 wrong | 94.7% correct | 4,641 unknown loss 1.0518 | geometry 0.3057 | confidence 1.4922 | mean error 0.740 vox speed: 892 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 321, from the previous prediction 1279 confidence-labeled crops: terminal 4/1534 (0.3%) | recoverable 120/1534 (7.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 352/1600 (22.0%), 2: 160/1600 (10.0%), 3: 288/1600 (18.0%), 4: 480/1600 (30.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 584 rows | mean depth 13 decisions | mean travel 105 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 54% 46% 27%>0% 27% 0 138 89%/10%/0%/2% 11% 145/3/0/2 0175A_5mm_v1 352 52% 48% 22%>0% 22% 0 143 88%/7%/0%/5% 5% 104/1/128/64 1447_5mm_v1 160 56% 44% 24%>0% 24% 0 145 92%/6%/0%/2% 3% 70/1/0/0 paris4_afv_central 288 56% 44% 25%>0% 25% 0 145 91%/7%/0%/1% 4% 126/1/0/0 PHerc0125_3mm_v2 480 57% 43% 25%>0% 25% 0 138 85%/7%/1%/8% 13% 139/1/116/68 gradients: max 72, clipped 3/50 updates Step 2,950 | last 50 updates / 1,600 crops | lr 2.93e-05 points (35/curve, tolerance 2 vox): 49,539 right / 1,665 wrong | 96.7% correct | 4,796 unknown loss 0.8688 | geometry 0.2427 | confidence 1.2522 | mean error 0.646 vox speed: 890 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 268, from the previous prediction 1332 confidence-labeled crops: terminal 1/1543 (0.1%) | recoverable 110/1543 (7.1%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 480/1600 (30.0%), 2: 352/1600 (22.0%), 3: 256/1600 (16.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 601 rows | mean depth 12 decisions | mean travel 99 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 52% 48% 21%>0% 21% 0 151 91%/4%/0%/5% 8% 92/0/0/0 0175A_5mm_v1 480 61% 39% 29%>0% 29% 0 136 89%/7%/0%/4% 4% 147/0/0/40 1447_5mm_v1 352 55% 45% 26%>0% 26% 0 151 90%/8%/0%/2% 3% 155/1/64/2 paris4_afv_central 256 55% 45% 26%>0% 26% 0 154 91%/7%/0%/2% 5% 114/1/0/0 PHerc0125_3mm_v2 320 57% 43% 24%>0% 24% 0 134 88%/7%/0%/4% 14% 93/3/0/45 gradients: max 66, clipped 1/50 updates Step 3,000 | last 50 updates / 1,600 crops | lr 2.93e-05 points (35/curve, tolerance 2 vox): 48,011 right / 2,674 wrong | 94.7% correct | 5,315 unknown loss 1.0230 | geometry 0.2774 | confidence 1.4913 | mean error 0.711 vox speed: 889 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1593 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 315, from the previous prediction 1278 confidence-labeled crops: terminal 8/1517 (0.5%) | recoverable 93/1517 (6.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 288/1600 (18.0%), 2: 288/1600 (18.0%), 3: 96/1600 (6.0%), 4: 544/1600 (34.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 581 rows | mean depth 13 decisions | mean travel 102 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 52% 47% 24%>2% 23% 4 228 91%/5%/1%/3% 9% 179/0/0/0 0175A_5mm_v1 288 55% 45% 25%>0% 25% 0 140 89%/6%/0%/5% 4% 93/1/0/36 1447_5mm_v1 288 59% 41% 30%>0% 30% 0 146 90%/7%/0%/3% 4% 106/0/36/12 paris4_afv_central 96 61% 39% 28%>0% 28% 0 157 90%/7%/0%/3% 7% 37/0/0/0 PHerc0125_3mm_v2 544 53% 47% 21%>0% 21% 0 136 85%/6%/1%/9% 12% 166/7/0/91 gradients: max 55, clipped 0/50 updates Step 3,000 | dagger_launched: true Step 3,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 248, "stopped": 1, "mean_example_error": 0.5324990124471726, "geometry_examples": 31, "examples": 32, "inference_and_measurement_seconds": 2.027054790407419, "total_diagnostic_seconds": 3.9414797984063625} Step 3,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 5/8 (62.5%) | false stops 1 1.5-3: recovered 5/8 (62.5%) | false stops 1 3-4.5: recovered 5/8 (62.5%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 0 entire proposed 8-point prefix correct (distance): 21/28 (75.0%) gate @ 0.50: false stops 2/23 (8.7%) | accepted wrong 4/25 (16.0%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.470 1.393 following 16 464 0.969 0.894 recoverable 12 348 2.139 2.058 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 0.969 0.894 332 32 812 1.470 1.393 improved/worsened/compared states: 3/1/28 Step 3,000 | batch_diagnostics_seconds: 3.9447672329843044 | recovery_seconds: 15.499918017536402 Step 3,050 | last 50 updates / 1,600 crops | lr 2.92e-05 points (35/curve, tolerance 2 vox): 48,799 right / 2,575 wrong | 95.0% correct | 4,626 unknown loss 1.0388 | geometry 0.3010 | confidence 1.4756 | mean error 0.723 vox speed: 1303 ms/update | data wait 1 ms/update | 18.5 crops/s | 18.5 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 230, from the previous prediction 1370 confidence-labeled crops: terminal 2/1540 (0.1%) | recoverable 106/1540 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 256/1600 (16.0%), 2: 224/1600 (14.0%), 3: 480/1600 (30.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 651 rows | mean depth 14 decisions | mean travel 110 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 53% 47% 26%>0% 26% 0 139 89%/8%/0%/3% 12% 160/0/42/5 0175A_5mm_v1 256 58% 42% 29%>0% 29% 0 139 89%/6%/0%/4% 6% 108/0/0/0 1447_5mm_v1 224 54% 46% 25%>0% 25% 0 157 90%/6%/0%/4% 4% 103/3/0/0 paris4_afv_central 480 55% 45% 24%>0% 24% 0 167 90%/6%/0%/4% 6% 165/2/23/53 PHerc0125_3mm_v2 288 60% 40% 27%>0% 27% 0 134 89%/6%/0%/4% 13% 115/2/0/0 gradients: max 68, clipped 2/50 updates Step 3,100 | last 50 updates / 1,600 crops | lr 2.92e-05 points (35/curve, tolerance 2 vox): 49,020 right / 2,135 wrong | 95.8% correct | 4,845 unknown loss 0.9396 | geometry 0.2633 | confidence 1.3527 | mean error 0.684 vox speed: 1365 ms/update | data wait 1 ms/update | 23.3 crops/s | 23.3 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1591 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 318, from the previous prediction 1273 confidence-labeled crops: terminal 2/1532 (0.1%) | recoverable 115/1532 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 480/1600 (30.0%), 1: 224/1600 (14.0%), 2: 384/1600 (24.0%), 3: 224/1600 (14.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 563 rows | mean depth 13 decisions | mean travel 103 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 480 52% 47% 24%>1% 23% 6 299 91%/6%/0%/2% 9% 96/0/0/128 0175A_5mm_v1 224 57% 43% 24%>0% 24% 0 168 90%/5%/0%/4% 4% 97/1/0/0 1447_5mm_v1 384 51% 49% 22%>0% 22% 0 137 88%/8%/0%/4% 3% 152/3/0/38 paris4_afv_central 224 55% 45% 25%>0% 25% 0 138 88%/8%/0%/4% 4% 98/4/0/2 PHerc0125_3mm_v2 288 58% 42% 26%>0% 26% 0 147 84%/9%/1%/7% 11% 120/1/0/0 gradients: max 50, clipped 0/50 updates Step 3,150 | last 50 updates / 1,600 crops | lr 2.92e-05 points (35/curve, tolerance 2 vox): 49,558 right / 1,878 wrong | 96.3% correct | 4,564 unknown loss 0.8846 | geometry 0.2539 | confidence 1.2615 | mean error 0.665 vox speed: 1338 ms/update | data wait 1 ms/update | 23.8 crops/s | 23.8 decisions/s | 1.14 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 326, from the previous prediction 1274 confidence-labeled crops: terminal 0/1539 (0.0%) | recoverable 91/1539 (5.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 480/1600 (30.0%), 1: 288/1600 (18.0%), 2: 352/1600 (22.0%), 3: 256/1600 (16.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 33 chunks / 66.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 574 rows | mean depth 13 decisions | mean travel 105 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 480 58% 42% 22%>0% 22% 0 134 93%/5%/0%/3% 9% 176/1/0/27 0175A_5mm_v1 288 55% 45% 26%>0% 26% 0 155 89%/8%/0%/3% 4% 131/4/0/0 1447_5mm_v1 352 49% 51% 20%>0% 20% 0 143 92%/5%/0%/3% 2% 96/2/0/84 paris4_afv_central 256 58% 42% 28%>0% 28% 0 146 88%/6%/0%/7% 6% 71/3/0/36 PHerc0125_3mm_v2 224 55% 45% 21%>0% 21% 0 155 88%/6%/0%/6% 11% 100/0/0/0 gradients: max 50, clipped 0/50 updates Step 3,200 | last 50 updates / 1,600 crops | lr 2.92e-05 points (35/curve, tolerance 2 vox): 49,783 right / 1,707 wrong | 96.7% correct | 4,510 unknown loss 0.8151 | geometry 0.2320 | confidence 1.1662 | mean error 0.622 vox speed: 1356 ms/update | data wait 1 ms/update | 23.4 crops/s | 23.4 decisions/s | 1.12 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 291, from the previous prediction 1309 confidence-labeled crops: terminal 0/1549 (0.0%) | recoverable 103/1549 (6.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 416/1600 (26.0%), 2: 480/1600 (30.0%), 3: 128/1600 (8.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 36 chunks / 72.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 578 rows | mean depth 14 decisions | mean travel 117 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 57% 43% 26%>0% 26% 0 141 92%/7%/0%/1% 8% 152/1/0/0 0175A_5mm_v1 416 58% 42% 27%>0% 27% 0 140 90%/6%/0%/4% 5% 117/1/123/59 1447_5mm_v1 480 57% 43% 25%>0% 25% 0 133 90%/7%/0%/3% 2% 154/3/0/53 paris4_afv_central 128 54% 46% 23%>0% 23% 0 158 92%/5%/0%/3% 6% 59/2/0/0 PHerc0125_3mm_v2 224 57% 43% 30%>0% 30% 0 176 88%/6%/0%/6% 11% 96/2/0/0 gradients: max 73, clipped 2/50 updates Step 3,250 | last 50 updates / 1,600 crops | lr 2.91e-05 points (35/curve, tolerance 2 vox): 49,384 right / 1,960 wrong | 96.2% correct | 4,656 unknown loss 0.8337 | geometry 0.2285 | confidence 1.2103 | mean error 0.635 vox speed: 1361 ms/update | data wait 1 ms/update | 23.4 crops/s | 23.4 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 293, from the previous prediction 1307 confidence-labeled crops: terminal 5/1540 (0.3%) | recoverable 111/1540 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 160/1600 (10.0%), 1: 192/1600 (12.0%), 2: 576/1600 (36.0%), 3: 352/1600 (22.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 36 chunks / 72.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 611 rows | mean depth 13 decisions | mean travel 104 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 160 59% 41% 30%>0% 30% 0 151 89%/11%/0%/0% 9% 66/0/0/0 0175A_5mm_v1 192 57% 43% 26%>0% 26% 0 136 93%/5%/0%/2% 6% 81/0/0/2 1447_5mm_v1 576 53% 47% 24%>0% 24% 0 136 90%/6%/0%/4% 4% 173/3/0/100 paris4_afv_central 352 55% 45% 25%>0% 25% 0 169 89%/7%/0%/4% 5% 158/6/0/0 PHerc0125_3mm_v2 320 58% 42% 23%>0% 23% 0 165 86%/7%/1%/7% 12% 133/1/0/0 gradients: max 48, clipped 0/50 updates Step 3,300 | last 50 updates / 1,600 crops | lr 2.91e-05 points (35/curve, tolerance 2 vox): 50,185 right / 1,998 wrong | 96.2% correct | 3,817 unknown loss 0.8849 | geometry 0.2698 | confidence 1.2302 | mean error 0.677 vox speed: 1362 ms/update | data wait 1 ms/update | 23.4 crops/s | 23.4 decisions/s | 1.13 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 230, from the previous prediction 1370 confidence-labeled crops: terminal 3/1547 (0.2%) | recoverable 103/1547 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 256/1600 (16.0%), 2: 256/1600 (16.0%), 3: 352/1600 (22.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 36 chunks / 72.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 670 rows | mean depth 12 decisions | mean travel 98 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 55% 45% 25%>0% 25% 0 155 90%/9%/0%/1% 9% 189/0/0/0 0175A_5mm_v1 256 57% 43% 27%>0% 27% 0 139 90%/8%/0%/2% 4% 93/1/0/17 1447_5mm_v1 256 56% 44% 24%>0% 24% 0 143 91%/5%/0%/4% 2% 113/0/0/0 paris4_afv_central 352 52% 48% 24%>0% 24% 0 144 91%/5%/0%/4% 3% 120/5/49/49 PHerc0125_3mm_v2 320 51% 49% 28%>0% 28% 0 149 88%/4%/1%/6% 11% 155/3/16/1 gradients: max 70, clipped 1/50 updates Step 3,350 | last 50 updates / 1,600 crops | lr 2.91e-05 points (35/curve, tolerance 2 vox): 49,850 right / 1,687 wrong | 96.7% correct | 4,463 unknown loss 0.8013 | geometry 0.2328 | confidence 1.1369 | mean error 0.639 vox speed: 1364 ms/update | data wait 1 ms/update | 23.3 crops/s | 23.3 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1590 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 299, from the previous prediction 1291 confidence-labeled crops: terminal 5/1542 (0.3%) | recoverable 124/1542 (8.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 480/1600 (30.0%), 1: 160/1600 (10.0%), 2: 448/1600 (28.0%), 3: 288/1600 (18.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 36 chunks / 72.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 556 rows | mean depth 14 decisions | mean travel 115 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 480 57% 40% 25%>2% 23% 7 309 90%/7%/1%/2% 9% 116/0/128/78 0175A_5mm_v1 160 62% 38% 26%>0% 26% 0 180 89%/6%/1%/4% 3% 61/4/0/0 1447_5mm_v1 448 53% 47% 25%>0% 25% 0 138 88%/9%/0%/3% 3% 209/7/9/0 paris4_afv_central 288 57% 43% 30%>0% 30% 0 136 87%/8%/0%/5% 4% 105/3/0/20 PHerc0125_3mm_v2 224 59% 41% 29%>0% 29% 0 142 85%/7%/0%/8% 7% 65/1/77/27 gradients: max 72, clipped 2/50 updates Step 3,400 | last 50 updates / 1,600 crops | lr 2.91e-05 points (35/curve, tolerance 2 vox): 48,809 right / 1,954 wrong | 96.2% correct | 5,237 unknown loss 0.9589 | geometry 0.2722 | confidence 1.3734 | mean error 0.665 vox speed: 1354 ms/update | data wait 1 ms/update | 23.5 crops/s | 23.5 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1596 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 269, from the previous prediction 1327 confidence-labeled crops: terminal 1/1539 (0.1%) | recoverable 115/1539 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 320/1600 (20.0%), 2: 320/1600 (20.0%), 3: 224/1600 (14.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 644 rows | mean depth 14 decisions | mean travel 115 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 55% 44% 29%>1% 28% 4 235 88%/9%/0%/2% 8% 89/0/0/37 0175A_5mm_v1 320 54% 46% 24%>0% 24% 0 168 89%/8%/0%/3% 5% 147/4/3/0 1447_5mm_v1 320 52% 48% 24%>0% 24% 0 145 91%/7%/0%/3% 3% 151/0/43/2 paris4_afv_central 224 54% 46% 22%>0% 22% 0 144 92%/4%/0%/4% 6% 104/1/0/0 PHerc0125_3mm_v2 448 52% 48% 26%>0% 26% 0 138 86%/7%/0%/7% 13% 153/2/0/61 gradients: max 64, clipped 1/50 updates Step 3,450 | last 50 updates / 1,600 crops | lr 2.90e-05 points (35/curve, tolerance 2 vox): 49,049 right / 2,161 wrong | 95.8% correct | 4,790 unknown loss 0.9183 | geometry 0.2708 | confidence 1.2950 | mean error 0.674 vox speed: 1344 ms/update | data wait 1 ms/update | 23.7 crops/s | 23.7 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1594 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 284, from the previous prediction 1310 confidence-labeled crops: terminal 3/1539 (0.2%) | recoverable 122/1539 (7.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 544/1600 (34.0%), 2: 288/1600 (18.0%), 3: 160/1600 (10.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 578 rows | mean depth 12 decisions | mean travel 96 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 52% 46% 25%>1% 23% 4 210 91%/7%/1%/1% 10% 176/1/0/17 0175A_5mm_v1 544 58% 42% 26%>0% 26% 0 145 87%/8%/0%/5% 4% 121/1/58/110 1447_5mm_v1 288 56% 44% 23%>0% 23% 0 155 90%/6%/0%/5% 2% 126/1/0/0 paris4_afv_central 160 57% 42% 29%>0% 29% 0 170 86%/9%/0%/4% 9% 68/0/0/0 PHerc0125_3mm_v2 192 55% 45% 23%>0% 23% 0 151 87%/8%/0%/5% 14% 87/3/0/0 gradients: max 44, clipped 0/50 updates Step 3,500 | last 50 updates / 1,600 crops | lr 2.90e-05 points (35/curve, tolerance 2 vox): 49,265 right / 1,892 wrong | 96.3% correct | 4,843 unknown loss 0.9126 | geometry 0.2870 | confidence 1.2511 | mean error 0.686 vox speed: 1361 ms/update | data wait 1 ms/update | 23.4 crops/s | 23.4 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1597 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 223, from the previous prediction 1374 confidence-labeled crops: terminal 0/1531 (0.0%) | recoverable 119/1531 (7.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 288/1600 (18.0%), 2: 224/1600 (14.0%), 3: 384/1600 (24.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 655 rows | mean depth 15 decisions | mean travel 123 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 54% 46% 26%>1% 25% 3 185 90%/9%/0%/2% 7% 175/2/0/0 0175A_5mm_v1 288 60% 40% 28%>0% 28% 0 136 86%/8%/0%/6% 4% 74/0/0/42 1447_5mm_v1 224 59% 41% 29%>0% 29% 0 164 87%/10%/0%/3% 2% 92/2/0/0 paris4_afv_central 384 55% 45% 22%>0% 22% 0 151 90%/5%/0%/5% 5% 172/0/0/0 PHerc0125_3mm_v2 320 56% 44% 25%>0% 25% 0 146 87%/7%/0%/6% 13% 142/5/0/0 gradients: max 65, clipped 2/50 updates Step 3,550 | last 50 updates / 1,600 crops | lr 2.90e-05 points (35/curve, tolerance 2 vox): 48,498 right / 2,531 wrong | 95.0% correct | 4,971 unknown loss 1.0002 | geometry 0.2986 | confidence 1.4032 | mean error 0.711 vox speed: 1342 ms/update | data wait 1 ms/update | 23.7 crops/s | 23.7 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1587 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 221, from the previous prediction 1366 confidence-labeled crops: terminal 7/1540 (0.5%) | recoverable 115/1540 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 288/1600 (18.0%), 2: 320/1600 (20.0%), 3: 256/1600 (16.0%), 4: 480/1600 (30.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 681 rows | mean depth 13 decisions | mean travel 109 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 46% 49% 23%>5% 18% 7 362 91%/5%/2%/2% 11% 125/0/0/0 0175A_5mm_v1 288 55% 45% 27%>0% 27% 0 141 91%/8%/0%/1% 4% 129/2/0/0 1447_5mm_v1 320 53% 47% 25%>0% 25% 0 177 90%/8%/0%/2% 3% 149/4/0/0 paris4_afv_central 256 52% 48% 25%>0% 25% 0 147 89%/7%/0%/3% 7% 106/2/48/18 PHerc0125_3mm_v2 480 56% 44% 27%>0% 27% 0 140 84%/7%/1%/8% 14% 172/8/23/39 gradients: max 71, clipped 2/50 updates Step 3,600 | last 50 updates / 1,600 crops | lr 2.89e-05 points (35/curve, tolerance 2 vox): 50,254 right / 1,837 wrong | 96.5% correct | 3,909 unknown loss 0.8081 | geometry 0.2548 | confidence 1.1065 | mean error 0.656 vox speed: 1349 ms/update | data wait 1 ms/update | 23.6 crops/s | 23.6 decisions/s | 1.12 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1596 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 248, from the previous prediction 1348 confidence-labeled crops: terminal 2/1544 (0.1%) | recoverable 95/1544 (6.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 352/1600 (22.0%), 2: 384/1600 (24.0%), 3: 352/1600 (22.0%), 4: 160/1600 (10.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 662 rows | mean depth 12 decisions | mean travel 96 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 51% 48% 23%>1% 22% 4 203 91%/7%/0%/2% 7% 169/1/30/1 0175A_5mm_v1 352 52% 48% 22%>0% 22% 0 142 91%/5%/0%/4% 5% 168/5/0/0 1447_5mm_v1 384 51% 49% 22%>0% 22% 0 154 93%/5%/0%/2% 2% 176/3/25/14 paris4_afv_central 352 57% 43% 26%>0% 26% 0 135 88%/7%/1%/5% 6% 84/2/0/66 PHerc0125_3mm_v2 160 52% 48% 27%>0% 27% 0 136 87%/8%/0%/6% 12% 65/0/0/12 gradients: max 71, clipped 3/50 updates Step 3,649 | dagger_states: 8446 | dagger_source_step: 3000 | dagger_caches: 1 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/1447_5mm_v1/decisions_003000.npz" | dagger_fibers: 384 | dagger_supply: {"states": 8446, "traces": 384, "fibers": 384, "rows": {"terminal": 308, "premature_stop": 47, "recoverable": 30, "pre_excursion": 287, "ordinary": 7759}, "events": {"terminal": 40, "premature_stop": 47, "recoverable": 13, "pre_excursion": 50, "ordinary": 381}, "supervision": {"following": 8092, "recoverable": 25, "terminal": 308, "unknown": 21}, "departed_traces": 48, "terminal_traces": 40, "stop_reasons": {"after_failure": 39, "confidence": 48, "unannotated": 297}, "traced_voxels": 128315.30427106367, "seconds": 893.1941162534058, "voxels_per_second": 143.6589224403933} | dagger_coverage: {"requested": 384, "seeds": 384, "skipped": [], "epoch": 0} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 0 | dagger_launch_interval: null | dagger_publication_age: 649 | dataset: "1447_5mm_v1" Step 3,650 | last 50 updates / 1,600 crops | lr 2.89e-05 points (35/curve, tolerance 2 vox): 49,502 right / 2,136 wrong | 95.9% correct | 4,362 unknown loss 0.9009 | geometry 0.2689 | confidence 1.2640 | mean error 0.673 vox speed: 1310 ms/update | data wait 1 ms/update | 24.3 crops/s | 24.3 decisions/s | 1.14 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 266, from the previous prediction 1334 confidence-labeled crops: terminal 2/1553 (0.1%) | recoverable 112/1553 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 256/1600 (16.0%), 2: 352/1600 (22.0%), 3: 256/1600 (16.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 611 rows | mean depth 15 decisions | mean travel 123 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 58% 42% 26%>0% 26% 0 154 90%/8%/0%/2% 8% 148/0/0/0 0175A_5mm_v1 256 55% 45% 21%>0% 21% 0 148 88%/7%/0%/5% 4% 114/3/0/0 1447_5mm_v1 352 56% 44% 27%>0% 27% 0 137 91%/7%/0%/2% 2% 105/3/0/51 paris4_afv_central 256 54% 46% 27%>0% 27% 0 144 90%/8%/0%/2% 6% 102/6/0/15 PHerc0125_3mm_v2 384 57% 43% 26%>0% 26% 0 139 90%/5%/1%/5% 14% 142/2/0/25 gradients: max 65, clipped 1/50 updates Step 3,700 | last 50 updates / 1,600 crops | lr 2.89e-05 points (35/curve, tolerance 2 vox): 49,332 right / 2,284 wrong | 95.6% correct | 4,384 unknown loss 0.8735 | geometry 0.2621 | confidence 1.2228 | mean error 0.672 vox speed: 887 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 221, from the previous prediction 1379 confidence-labeled crops: terminal 4/1548 (0.3%) | recoverable 108/1548 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 320/1600 (20.0%), 2: 320/1600 (20.0%), 3: 320/1600 (20.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 703 rows | mean depth 13 decisions | mean travel 109 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 54% 46% 24%>0% 24% 0 152 89%/9%/0%/2% 9% 117/1/0/0 0175A_5mm_v1 320 59% 41% 24%>0% 24% 0 144 91%/6%/0%/3% 4% 132/5/0/0 1447_5mm_v1 320 53% 47% 21%>0% 21% 0 141 92%/5%/0%/3% 4% 127/1/0/22 paris4_afv_central 320 53% 47% 24%>0% 24% 0 150 87%/9%/0%/4% 5% 145/2/0/4 PHerc0125_3mm_v2 384 53% 47% 26%>0% 26% 0 147 90%/5%/1%/4% 13% 182/4/0/0 gradients: max 62, clipped 2/50 updates Step 3,750 | last 50 updates / 1,600 crops | lr 2.89e-05 points (35/curve, tolerance 2 vox): 49,160 right / 2,868 wrong | 94.5% correct | 3,972 unknown loss 1.0071 | geometry 0.3311 | confidence 1.3520 | mean error 0.756 vox speed: 891 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1581 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 284, from the previous prediction 1297 confidence-labeled crops: terminal 4/1551 (0.3%) | recoverable 103/1551 (6.6%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 480/1600 (30.0%), 1: 352/1600 (22.0%), 2: 224/1600 (14.0%), 3: 192/1600 (12.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 622 rows | mean depth 12 decisions | mean travel 100 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 480 48% 49% 27%>4% 23% 9 397 89%/8%/0%/2% 11% 179/1/3/54 0175A_5mm_v1 352 56% 44% 26%>0% 26% 0 151 92%/5%/0%/3% 5% 154/4/0/0 1447_5mm_v1 224 55% 45% 25%>0% 25% 0 150 92%/4%/0%/4% 3% 100/2/0/0 paris4_afv_central 192 56% 44% 26%>0% 26% 0 143 91%/6%/0%/3% 8% 78/0/0/7 PHerc0125_3mm_v2 352 53% 47% 24%>0% 24% 0 135 88%/7%/0%/5% 12% 111/4/0/53 gradients: max 74, clipped 1/50 updates Step 3,800 | last 50 updates / 1,600 crops | lr 2.88e-05 points (35/curve, tolerance 2 vox): 49,544 right / 1,653 wrong | 96.8% correct | 4,803 unknown loss 0.8046 | geometry 0.2205 | confidence 1.1684 | mean error 0.610 vox speed: 890 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 195, from the previous prediction 1405 confidence-labeled crops: terminal 3/1542 (0.2%) | recoverable 101/1542 (6.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 256/1600 (16.0%), 2: 480/1600 (30.0%), 3: 288/1600 (18.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 719 rows | mean depth 14 decisions | mean travel 117 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 54% 46% 25%>0% 25% 0 153 89%/7%/0%/4% 8% 117/2/0/0 0175A_5mm_v1 256 55% 45% 23%>0% 23% 0 138 91%/5%/0%/4% 4% 90/3/54/24 1447_5mm_v1 480 54% 46% 27%>0% 27% 0 152 91%/7%/0%/2% 2% 221/1/0/0 paris4_afv_central 288 49% 51% 24%>0% 24% 0 148 90%/6%/0%/4% 3% 146/3/0/0 PHerc0125_3mm_v2 320 55% 45% 23%>0% 23% 0 140 87%/6%/1%/7% 11% 145/5/0/0 gradients: max 47, clipped 0/50 updates Step 3,850 | last 50 updates / 1,600 crops | lr 2.88e-05 points (35/curve, tolerance 2 vox): 48,865 right / 2,170 wrong | 95.7% correct | 4,965 unknown loss 0.8944 | geometry 0.2641 | confidence 1.2606 | mean error 0.670 vox speed: 888 ms/update | data wait 1 ms/update | 35.8 crops/s | 35.8 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 195, from the previous prediction 1405 confidence-labeled crops: terminal 3/1534 (0.2%) | recoverable 107/1534 (7.0%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 384/1600 (24.0%), 2: 224/1600 (14.0%), 3: 256/1600 (16.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 665 rows | mean depth 15 decisions | mean travel 122 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 61% 39% 29%>0% 29% 0 144 91%/7%/0%/2% 9% 155/0/0/7 0175A_5mm_v1 384 51% 49% 22%>0% 22% 0 142 90%/6%/0%/4% 5% 162/2/0/25 1447_5mm_v1 224 60% 40% 24%>0% 24% 0 143 88%/7%/0%/5% 2% 90/0/0/0 paris4_afv_central 256 55% 45% 25%>0% 25% 0 156 88%/7%/0%/4% 5% 116/2/0/0 PHerc0125_3mm_v2 320 56% 44% 30%>0% 30% 0 137 86%/7%/1%/7% 11% 142/3/15/0 gradients: max 82, clipped 2/50 updates Step 3,900 | last 50 updates / 1,600 crops | lr 2.88e-05 points (35/curve, tolerance 2 vox): 50,189 right / 2,227 wrong | 95.8% correct | 3,584 unknown loss 0.9007 | geometry 0.2583 | confidence 1.2847 | mean error 0.670 vox speed: 886 ms/update | data wait 1 ms/update | 35.9 crops/s | 35.9 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1594 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 231, from the previous prediction 1363 confidence-labeled crops: terminal 4/1557 (0.3%) | recoverable 127/1557 (8.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 608/1600 (38.0%), 1: 288/1600 (18.0%), 2: 288/1600 (18.0%), 3: 224/1600 (14.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 635 rows | mean depth 12 decisions | mean travel 99 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 608 56% 43% 26%>1% 25% 3 214 90%/8%/1%/2% 8% 224/1/0/35 0175A_5mm_v1 288 52% 48% 29%>0% 29% 0 141 86%/10%/0%/4% 4% 123/2/0/16 1447_5mm_v1 288 60% 40% 26%>0% 26% 0 147 88%/10%/0%/2% 2% 115/3/0/0 paris4_afv_central 224 60% 40% 26%>0% 26% 0 160 93%/4%/0%/3% 7% 89/1/0/0 PHerc0125_3mm_v2 192 56% 44% 29%>0% 29% 0 150 89%/7%/0%/4% 12% 84/1/0/0 gradients: max 57, clipped 0/50 updates Step 3,950 | last 50 updates / 1,600 crops | lr 2.87e-05 points (35/curve, tolerance 2 vox): 49,061 right / 2,309 wrong | 95.5% correct | 4,630 unknown loss 0.9617 | geometry 0.2770 | confidence 1.3695 | mean error 0.688 vox speed: 893 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1589 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 326, from the previous prediction 1263 confidence-labeled crops: terminal 3/1541 (0.2%) | recoverable 92/1541 (6.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 288/1600 (18.0%), 2: 256/1600 (16.0%), 3: 224/1600 (14.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 597 rows | mean depth 16 decisions | mean travel 128 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 47% 51% 21%>2% 19% 8 444 92%/5%/0%/3% 11% 89/0/0/123 0175A_5mm_v1 288 50% 50% 26%>0% 26% 0 147 89%/8%/0%/3% 4% 120/3/0/23 1447_5mm_v1 256 62% 38% 27%>0% 27% 0 156 89%/8%/0%/3% 2% 97/5/0/0 paris4_afv_central 224 54% 46% 28%>0% 28% 0 163 91%/7%/0%/3% 3% 104/3/0/0 PHerc0125_3mm_v2 416 51% 49% 20%>0% 20% 0 144 90%/4%/1%/6% 12% 187/5/42/18 gradients: max 59, clipped 0/50 updates Step 4,000 | last 50 updates / 1,600 crops | lr 2.87e-05 points (35/curve, tolerance 2 vox): 49,351 right / 2,149 wrong | 95.8% correct | 4,500 unknown loss 0.8455 | geometry 0.2367 | confidence 1.2177 | mean error 0.632 vox speed: 889 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.14 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1593 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 218, from the previous prediction 1375 confidence-labeled crops: terminal 9/1542 (0.6%) | recoverable 118/1542 (7.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 384/1600 (24.0%), 2: 416/1600 (26.0%), 3: 256/1600 (16.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 655 rows | mean depth 15 decisions | mean travel 120 vox | 8 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 55% 43% 28%>2% 25% 7 297 91%/8%/0%/1% 8% 138/0/0/0 0175A_5mm_v1 384 58% 42% 28%>0% 28% 0 141 89%/7%/0%/3% 4% 161/1/0/0 1447_5mm_v1 416 57% 43% 25%>0% 25% 0 161 88%/7%/0%/5% 3% 179/4/0/0 paris4_afv_central 256 59% 41% 27%>0% 27% 0 171 87%/8%/2%/4% 7% 106/4/0/0 PHerc0125_3mm_v2 224 59% 41% 29%>0% 29% 0 142 85%/6%/1%/8% 13% 71/3/0/21 gradients: max 51, clipped 0/50 updates Step 4,000 | dagger_launched: true Step 4,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 256, "stopped": 0, "mean_example_error": 0.4349463111720979, "geometry_examples": 32, "examples": 32, "inference_and_measurement_seconds": 2.4206608571112156, "total_diagnostic_seconds": 4.363641291856766} Step 4,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 6/8 (75.0%) | false stops 1 1.5-3: recovered 5/8 (62.5%) | false stops 0 3-4.5: recovered 5/8 (62.5%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 0 entire proposed 8-point prefix correct (distance): 21/28 (75.0%) gate @ 0.50: false stops 0/22 (0.0%) | accepted wrong 4/25 (16.0%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.494 1.449 following 16 464 0.986 0.986 recoverable 12 348 2.172 2.066 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 0.986 0.986 332 32 812 1.494 1.449 improved/worsened/compared states: 2/1/28 Step 4,000 | batch_diagnostics_seconds: 4.367306906729937 | recovery_seconds: 16.672586116939783 Step 4,050 | last 50 updates / 1,600 crops | lr 2.87e-05 points (35/curve, tolerance 2 vox): 48,640 right / 3,057 wrong | 94.1% correct | 4,303 unknown loss 1.1283 | geometry 0.3357 | confidence 1.5853 | mean error 0.754 vox speed: 1346 ms/update | data wait 1 ms/update | 17.7 crops/s | 17.7 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1592 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 289, from the previous prediction 1303 confidence-labeled crops: terminal 1/1544 (0.1%) | recoverable 121/1544 (7.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 352/1600 (22.0%), 2: 192/1600 (12.0%), 3: 224/1600 (14.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 597 rows | mean depth 15 decisions | mean travel 118 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 53% 45% 26%>2% 24% 4 266 90%/8%/0%/2% 11% 201/2/0/0 0175A_5mm_v1 352 55% 45% 25%>0% 25% 0 140 90%/8%/0%/3% 4% 122/3/41/35 1447_5mm_v1 192 50% 50% 27%>0% 27% 0 163 90%/9%/0%/1% 1% 96/2/0/0 paris4_afv_central 224 54% 46% 26%>0% 26% 0 167 87%/6%/0%/7% 6% 97/3/47/5 PHerc0125_3mm_v2 384 57% 43% 28%>0% 28% 0 133 87%/7%/0%/6% 17% 81/1/0/85 gradients: max 74, clipped 2/50 updates Step 4,100 | last 50 updates / 1,600 crops | lr 2.86e-05 points (35/curve, tolerance 2 vox): 49,284 right / 1,876 wrong | 96.3% correct | 4,840 unknown loss 0.8598 | geometry 0.2217 | confidence 1.2762 | mean error 0.609 vox speed: 1364 ms/update | data wait 1 ms/update | 23.3 crops/s | 23.3 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 223, from the previous prediction 1377 confidence-labeled crops: terminal 5/1539 (0.3%) | recoverable 108/1539 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 448/1600 (28.0%), 2: 352/1600 (22.0%), 3: 224/1600 (14.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 691 rows | mean depth 15 decisions | mean travel 123 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 52% 48% 24%>0% 24% 0 165 92%/7%/0%/2% 8% 92/0/0/0 0175A_5mm_v1 448 52% 48% 24%>0% 24% 0 133 90%/6%/0%/4% 4% 164/6/0/50 1447_5mm_v1 352 53% 47% 28%>0% 28% 0 153 89%/9%/0%/3% 2% 164/1/0/0 paris4_afv_central 224 51% 49% 21%>0% 21% 0 157 92%/5%/0%/3% 8% 110/2/0/0 PHerc0125_3mm_v2 384 58% 42% 27%>0% 27% 0 142 85%/7%/1%/7% 14% 161/5/0/0 gradients: max 44, clipped 0/50 updates Step 4,150 | last 50 updates / 1,600 crops | lr 2.86e-05 points (35/curve, tolerance 2 vox): 49,825 right / 1,735 wrong | 96.6% correct | 4,440 unknown loss 0.7992 | geometry 0.2329 | confidence 1.1325 | mean error 0.629 vox speed: 1380 ms/update | data wait 1 ms/update | 23.0 crops/s | 23.0 decisions/s | 1.14 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 290, from the previous prediction 1310 confidence-labeled crops: terminal 2/1548 (0.1%) | recoverable 105/1548 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 320/1600 (20.0%), 2: 448/1600 (28.0%), 3: 288/1600 (18.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 625 rows | mean depth 15 decisions | mean travel 121 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 58% 42% 27%>0% 27% 0 165 89%/9%/0%/2% 9% 108/0/0/0 0175A_5mm_v1 320 50% 50% 21%>0% 21% 0 135 93%/5%/0%/2% 3% 111/3/0/49 1447_5mm_v1 448 55% 45% 25%>0% 25% 0 138 90%/6%/0%/4% 2% 148/3/16/52 paris4_afv_central 288 58% 42% 25%>0% 25% 0 149 90%/6%/1%/3% 6% 121/1/0/0 PHerc0125_3mm_v2 288 52% 48% 21%>0% 21% 0 147 88%/7%/0%/5% 10% 137/4/0/0 gradients: max 67, clipped 1/50 updates Step 4,200 | last 50 updates / 1,600 crops | lr 2.86e-05 points (35/curve, tolerance 2 vox): 48,983 right / 2,227 wrong | 95.7% correct | 4,790 unknown loss 0.9767 | geometry 0.2849 | confidence 1.3837 | mean error 0.690 vox speed: 1382 ms/update | data wait 1 ms/update | 23.0 crops/s | 23.0 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1599 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 254, from the previous prediction 1345 confidence-labeled crops: terminal 0/1539 (0.0%) | recoverable 133/1539 (8.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 224/1600 (14.0%), 2: 384/1600 (24.0%), 3: 352/1600 (22.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 608 rows | mean depth 15 decisions | mean travel 122 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 52% 48% 25%>0% 24% 1 180 88%/9%/0%/3% 10% 169/1/0/0 0175A_5mm_v1 224 53% 47% 26%>0% 26% 0 147 91%/8%/0%/1% 4% 87/1/0/19 1447_5mm_v1 384 59% 41% 31%>0% 31% 0 133 87%/7%/0%/6% 3% 77/2/0/81 paris4_afv_central 352 59% 41% 31%>0% 31% 0 142 86%/9%/0%/5% 6% 143/2/5/3 PHerc0125_3mm_v2 288 54% 46% 23%>0% 23% 0 141 88%/8%/0%/4% 15% 132/4/0/0 gradients: max 59, clipped 0/50 updates Step 4,250 | last 50 updates / 1,600 crops | lr 2.85e-05 points (35/curve, tolerance 2 vox): 49,278 right / 2,012 wrong | 96.1% correct | 4,710 unknown loss 0.8464 | geometry 0.2471 | confidence 1.1986 | mean error 0.655 vox speed: 1379 ms/update | data wait 1 ms/update | 23.1 crops/s | 23.1 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1591 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 239, from the previous prediction 1352 confidence-labeled crops: terminal 7/1539 (0.5%) | recoverable 96/1539 (6.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 256/1600 (16.0%), 2: 384/1600 (24.0%), 3: 256/1600 (16.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 653 rows | mean depth 16 decisions | mean travel 131 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 52% 45% 22%>2% 20% 7 343 93%/6%/1%/1% 9% 137/0/128/23 0175A_5mm_v1 256 56% 44% 25%>0% 25% 0 161 90%/7%/0%/3% 5% 112/1/0/0 1447_5mm_v1 384 54% 46% 22%>0% 22% 0 140 91%/5%/0%/4% 3% 176/4/0/0 paris4_afv_central 256 49% 51% 24%>0% 24% 0 138 89%/6%/0%/5% 3% 84/2/0/46 PHerc0125_3mm_v2 352 59% 41% 26%>0% 26% 0 142 85%/7%/1%/7% 11% 144/3/0/0 gradients: max 50, clipped 0/50 updates Step 4,300 | last 50 updates / 1,600 crops | lr 2.85e-05 points (35/curve, tolerance 2 vox): 49,487 right / 1,967 wrong | 96.2% correct | 4,546 unknown loss 0.8630 | geometry 0.2823 | confidence 1.1612 | mean error 0.674 vox speed: 1362 ms/update | data wait 1 ms/update | 23.4 crops/s | 23.4 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 298, from the previous prediction 1302 confidence-labeled crops: terminal 1/1546 (0.1%) | recoverable 113/1546 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 480/1600 (30.0%), 1: 256/1600 (16.0%), 2: 352/1600 (22.0%), 3: 288/1600 (18.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 607 rows | mean depth 16 decisions | mean travel 132 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 480 53% 47% 23%>0% 23% 0 136 89%/9%/0%/2% 9% 126/1/0/99 0175A_5mm_v1 256 53% 47% 23%>0% 23% 0 173 89%/7%/0%/3% 3% 120/2/0/0 1447_5mm_v1 352 54% 46% 24%>0% 24% 0 149 92%/6%/0%/3% 3% 161/1/0/0 paris4_afv_central 288 57% 43% 28%>0% 28% 0 146 88%/9%/0%/3% 4% 106/3/0/17 PHerc0125_3mm_v2 224 58% 42% 25%>0% 25% 0 157 89%/4%/0%/7% 12% 94/2/0/0 gradients: max 71, clipped 1/50 updates Step 4,350 | last 50 updates / 1,600 crops | lr 2.85e-05 points (35/curve, tolerance 2 vox): 48,989 right / 2,527 wrong | 95.1% correct | 4,484 unknown loss 0.9845 | geometry 0.2788 | confidence 1.4115 | mean error 0.686 vox speed: 1381 ms/update | data wait 1 ms/update | 23.0 crops/s | 23.0 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1591 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 231, from the previous prediction 1360 confidence-labeled crops: terminal 7/1537 (0.5%) | recoverable 107/1537 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 480/1600 (30.0%), 1: 288/1600 (18.0%), 2: 320/1600 (20.0%), 3: 224/1600 (14.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 656 rows | mean depth 13 decisions | mean travel 106 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 480 56% 42% 27%>2% 25% 6 293 88%/8%/1%/4% 12% 175/2/0/29 0175A_5mm_v1 288 58% 42% 27%>0% 27% 0 168 88%/9%/0%/3% 3% 122/4/0/0 1447_5mm_v1 320 57% 43% 23%>0% 23% 0 149 90%/7%/0%/3% 2% 138/0/20/1 paris4_afv_central 224 56% 44% 25%>0% 25% 0 135 90%/6%/0%/4% 6% 96/1/0/2 PHerc0125_3mm_v2 288 57% 43% 27%>0% 27% 0 157 89%/4%/1%/6% 12% 125/3/0/0 gradients: max 78, clipped 3/50 updates Step 4,400 | last 50 updates / 1,600 crops | lr 2.84e-05 points (35/curve, tolerance 2 vox): 49,128 right / 1,940 wrong | 96.2% correct | 4,932 unknown loss 0.7994 | geometry 0.2435 | confidence 1.1119 | mean error 0.653 vox speed: 1387 ms/update | data wait 1 ms/update | 22.9 crops/s | 22.9 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 230, from the previous prediction 1370 confidence-labeled crops: terminal 3/1533 (0.2%) | recoverable 107/1533 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 192/1600 (12.0%), 2: 480/1600 (30.0%), 3: 224/1600 (14.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 665 rows | mean depth 17 decisions | mean travel 140 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 51% 49% 18%>0% 18% 0 136 92%/7%/0%/1% 10% 135/0/0/7 0175A_5mm_v1 192 55% 45% 24%>0% 24% 0 166 90%/6%/0%/4% 4% 86/0/0/0 1447_5mm_v1 480 55% 45% 26%>0% 26% 0 148 90%/7%/0%/3% 2% 188/6/9/28 paris4_afv_central 224 56% 44% 29%>0% 29% 0 169 84%/7%/0%/8% 6% 98/1/0/0 PHerc0125_3mm_v2 416 55% 45% 26%>0% 26% 0 143 88%/6%/0%/6% 11% 158/7/12/29 gradients: max 45, clipped 0/50 updates Step 4,450 | last 50 updates / 1,600 crops | lr 2.84e-05 points (35/curve, tolerance 2 vox): 50,200 right / 1,800 wrong | 96.5% correct | 4,000 unknown loss 0.8153 | geometry 0.2418 | confidence 1.1470 | mean error 0.631 vox speed: 1342 ms/update | data wait 1 ms/update | 23.7 crops/s | 23.7 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 309, from the previous prediction 1291 confidence-labeled crops: terminal 2/1552 (0.1%) | recoverable 107/1552 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 224/1600 (14.0%), 2: 480/1600 (30.0%), 3: 480/1600 (30.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 612 rows | mean depth 16 decisions | mean travel 132 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 54% 46% 24%>0% 24% 0 144 90%/10%/0%/0% 8% 89/0/0/0 0175A_5mm_v1 224 51% 49% 26%>0% 26% 0 180 91%/7%/0%/2% 4% 109/4/2/0 1447_5mm_v1 480 56% 44% 24%>0% 24% 0 134 89%/6%/0%/4% 4% 158/3/0/51 paris4_afv_central 480 56% 44% 24%>0% 24% 0 142 91%/6%/0%/3% 5% 160/2/43/49 PHerc0125_3mm_v2 224 48% 52% 21%>0% 21% 0 136 89%/5%/0%/5% 12% 96/2/0/20 gradients: max 61, clipped 1/50 updates Step 4,500 | last 50 updates / 1,600 crops | lr 2.84e-05 points (35/curve, tolerance 2 vox): 49,656 right / 2,084 wrong | 96.0% correct | 4,260 unknown loss 0.9199 | geometry 0.2771 | confidence 1.2856 | mean error 0.654 vox speed: 1377 ms/update | data wait 1 ms/update | 23.1 crops/s | 23.1 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 378, from the previous prediction 1222 confidence-labeled crops: terminal 4/1541 (0.3%) | recoverable 115/1541 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 224/1600 (14.0%), 2: 384/1600 (24.0%), 3: 544/1600 (34.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 535 rows | mean depth 16 decisions | mean travel 126 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 53% 47% 23%>0% 23% 0 165 89%/8%/0%/3% 8% 120/1/0/0 0175A_5mm_v1 224 53% 47% 23%>0% 23% 0 168 90%/7%/0%/3% 6% 105/2/0/0 1447_5mm_v1 384 55% 45% 23%>0% 23% 0 140 88%/8%/0%/4% 4% 144/2/0/30 paris4_afv_central 544 53% 47% 21%>0% 21% 0 137 91%/6%/0%/3% 6% 86/1/0/171 PHerc0125_3mm_v2 192 58% 42% 23%>0% 23% 0 152 85%/9%/1%/6% 12% 80/3/0/0 gradients: max 67, clipped 1/50 updates Step 4,550 | last 50 updates / 1,600 crops | lr 2.83e-05 points (35/curve, tolerance 2 vox): 48,703 right / 2,323 wrong | 95.4% correct | 4,974 unknown loss 0.9121 | geometry 0.2734 | confidence 1.2775 | mean error 0.689 vox speed: 1377 ms/update | data wait 1 ms/update | 23.1 crops/s | 23.1 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1594 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 185, from the previous prediction 1409 confidence-labeled crops: terminal 4/1533 (0.3%) | recoverable 104/1533 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 320/1600 (20.0%), 2: 192/1600 (12.0%), 3: 416/1600 (26.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 1638 chunk references / 1 windows live chains: 682 rows | mean depth 18 decisions | mean travel 147 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 51% 48% 24%>2% 23% 5 272 92%/5%/0%/3% 8% 168/0/0/0 0175A_5mm_v1 320 54% 46% 27%>0% 27% 0 154 90%/7%/0%/3% 3% 130/0/98/18 1447_5mm_v1 192 57% 43% 28%>0% 28% 0 145 91%/5%/0%/4% 2% 82/5/0/0 paris4_afv_central 416 57% 43% 26%>0% 26% 0 135 87%/8%/1%/4% 7% 164/2/0/15 PHerc0125_3mm_v2 320 57% 43% 31%>0% 31% 0 161 87%/7%/0%/6% 13% 138/4/0/0 gradients: max 91, clipped 2/50 updates Step 4,554 | dagger_states: 7594 | dagger_source_step: 4000 | dagger_caches: 1 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/paris4_afv_central/decisions_004000.npz" | dagger_fibers: 384 | dagger_supply: {"states": 7594, "traces": 384, "fibers": 384, "rows": {"terminal": 481, "premature_stop": 76, "recoverable": 92, "pre_excursion": 450, "ordinary": 6481}, "events": {"terminal": 62, "premature_stop": 76, "recoverable": 27, "pre_excursion": 82, "ordinary": 368}, "supervision": {"following": 7004, "recoverable": 80, "terminal": 481, "unknown": 29}, "departed_traces": 68, "terminal_traces": 62, "stop_reasons": {"after_failure": 59, "confidence": 79, "unannotated": 246}, "traced_voxels": 111256.58964721445, "seconds": 776.7688442394137, "voxels_per_second": 143.22998466313774} | dagger_coverage: {"requested": 384, "seeds": 384, "skipped": [], "epoch": 0} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 0 | dagger_launch_interval: null | dagger_publication_age: 554 | dataset: "paris4_afv_central" Step 4,600 | last 50 updates / 1,600 crops | lr 2.83e-05 points (35/curve, tolerance 2 vox): 49,127 right / 2,289 wrong | 95.5% correct | 4,584 unknown loss 0.9488 | geometry 0.3015 | confidence 1.2946 | mean error 0.707 vox speed: 888 ms/update | data wait 1 ms/update | 35.8 crops/s | 35.8 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1599 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 281, from the previous prediction 1318 confidence-labeled crops: terminal 3/1534 (0.2%) | recoverable 129/1534 (8.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 256/1600 (16.0%), 2: 224/1600 (14.0%), 3: 416/1600 (26.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 600 rows | mean depth 15 decisions | mean travel 121 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 55% 44% 27%>0% 27% 1 179 89%/9%/0%/3% 8% 96/0/96/17 0175A_5mm_v1 256 56% 44% 28%>0% 28% 0 143 86%/10%/0%/3% 5% 97/1/30/16 1447_5mm_v1 224 54% 46% 21%>0% 21% 0 181 94%/4%/0%/2% 1% 104/2/0/0 paris4_afv_central 416 61% 39% 27%>0% 27% 0 137 87%/8%/0%/4% 5% 135/5/0/26 PHerc0125_3mm_v2 448 56% 44% 25%>0% 25% 0 154 85%/8%/0%/6% 13% 168/3/16/30 gradients: max 48, clipped 0/50 updates Step 4,650 | last 50 updates / 1,600 crops | lr 2.83e-05 points (35/curve, tolerance 2 vox): 49,418 right / 2,297 wrong | 95.6% correct | 4,285 unknown loss 0.9334 | geometry 0.2977 | confidence 1.2715 | mean error 0.681 vox speed: 887 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1592 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 261, from the previous prediction 1331 confidence-labeled crops: terminal 1/1540 (0.1%) | recoverable 113/1540 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 416/1600 (26.0%), 2: 352/1600 (22.0%), 3: 224/1600 (14.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 633 rows | mean depth 15 decisions | mean travel 122 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 57% 41% 29%>2% 27% 6 304 89%/8%/0%/3% 12% 161/1/32/10 0175A_5mm_v1 416 53% 47% 22%>0% 22% 0 137 88%/8%/0%/4% 3% 130/3/0/65 1447_5mm_v1 352 52% 48% 24%>0% 24% 0 165 89%/7%/0%/4% 4% 167/3/32/1 paris4_afv_central 224 56% 44% 24%>0% 24% 0 151 93%/3%/0%/4% 7% 98/0/0/0 PHerc0125_3mm_v2 192 53% 47% 25%>0% 25% 0 135 86%/8%/0%/6% 14% 77/2/0/14 gradients: max 57, clipped 0/50 updates Step 4,700 | last 50 updates / 1,600 crops | lr 2.82e-05 points (35/curve, tolerance 2 vox): 49,568 right / 2,035 wrong | 96.1% correct | 4,397 unknown loss 0.8488 | geometry 0.2508 | confidence 1.1960 | mean error 0.645 vox speed: 890 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1596 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 257, from the previous prediction 1339 confidence-labeled crops: terminal 3/1539 (0.2%) | recoverable 103/1539 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 416/1600 (26.0%), 2: 320/1600 (20.0%), 3: 416/1600 (26.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 653 rows | mean depth 18 decisions | mean travel 148 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 53% 45% 25%>2% 23% 2 339 94%/6%/0%/1% 9% 70/0/0/17 0175A_5mm_v1 416 58% 42% 26%>0% 26% 0 137 89%/6%/0%/5% 4% 156/1/0/20 1447_5mm_v1 320 55% 45% 24%>0% 24% 0 142 88%/8%/1%/3% 2% 111/1/80/32 paris4_afv_central 416 53% 47% 27%>0% 27% 0 141 89%/8%/0%/3% 7% 197/4/0/0 PHerc0125_3mm_v2 256 54% 46% 25%>0% 25% 0 144 90%/4%/0%/5% 12% 119/5/0/0 gradients: max 89, clipped 1/50 updates Step 4,750 | last 50 updates / 1,600 crops | lr 2.82e-05 points (35/curve, tolerance 2 vox): 49,593 right / 2,348 wrong | 95.5% correct | 4,059 unknown loss 0.9893 | geometry 0.2743 | confidence 1.4301 | mean error 0.696 vox speed: 890 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1597 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 261, from the previous prediction 1336 confidence-labeled crops: terminal 3/1543 (0.2%) | recoverable 121/1543 (7.8%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 192/1600 (12.0%), 2: 192/1600 (12.0%), 3: 384/1600 (24.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 635 rows | mean depth 15 decisions | mean travel 126 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 57% 42% 26%>1% 25% 2 224 90%/8%/0%/2% 8% 161/1/0/0 0175A_5mm_v1 192 52% 48% 24%>0% 24% 0 139 89%/8%/0%/4% 4% 85/1/0/8 1447_5mm_v1 192 57% 43% 22%>0% 22% 0 150 88%/9%/1%/3% 2% 53/2/0/29 paris4_afv_central 384 54% 46% 25%>0% 25% 0 137 89%/8%/0%/3% 6% 149/1/126/27 PHerc0125_3mm_v2 448 55% 45% 23%>0% 23% 0 137 88%/6%/0%/6% 12% 187/4/0/16 gradients: max 82, clipped 1/50 updates Step 4,800 | last 50 updates / 1,600 crops | lr 2.81e-05 points (35/curve, tolerance 2 vox): 48,734 right / 2,352 wrong | 95.4% correct | 4,914 unknown loss 0.9063 | geometry 0.2853 | confidence 1.2421 | mean error 0.705 vox speed: 887 ms/update | data wait 1 ms/update | 35.8 crops/s | 35.8 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1595 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 305, from the previous prediction 1290 confidence-labeled crops: terminal 3/1521 (0.2%) | recoverable 122/1521 (8.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 384/1600 (24.0%), 2: 320/1600 (20.0%), 3: 224/1600 (14.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 598 rows | mean depth 17 decisions | mean travel 137 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 50% 49% 27%>2% 25% 3 278 88%/9%/0%/3% 10% 140/0/0/0 0175A_5mm_v1 384 54% 46% 24%>0% 24% 0 139 87%/9%/0%/3% 3% 177/2/0/0 1447_5mm_v1 320 56% 44% 24%>0% 24% 0 146 90%/5%/0%/5% 4% 141/5/0/0 paris4_afv_central 224 55% 45% 25%>0% 25% 0 139 84%/8%/0%/8% 7% 72/0/0/29 PHerc0125_3mm_v2 384 52% 48% 24%>0% 24% 0 133 85%/7%/1%/7% 12% 68/2/0/116 gradients: max 58, clipped 0/50 updates Step 4,850 | last 50 updates / 1,600 crops | lr 2.81e-05 points (35/curve, tolerance 2 vox): 49,354 right / 2,173 wrong | 95.8% correct | 4,473 unknown loss 0.8927 | geometry 0.2862 | confidence 1.2130 | mean error 0.667 vox speed: 890 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1595 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 272, from the previous prediction 1323 confidence-labeled crops: terminal 1/1540 (0.1%) | recoverable 104/1540 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 160/1600 (10.0%), 1: 480/1600 (30.0%), 2: 224/1600 (14.0%), 3: 384/1600 (24.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 630 rows | mean depth 18 decisions | mean travel 144 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 160 49% 48% 24%>3% 21% 4 373 89%/9%/0%/2% 11% 77/0/0/0 0175A_5mm_v1 480 52% 48% 23%>0% 23% 0 144 89%/7%/0%/4% 5% 144/2/29/87 1447_5mm_v1 224 54% 46% 25%>0% 25% 0 175 87%/8%/0%/5% 2% 104/4/0/0 paris4_afv_central 384 53% 47% 28%>0% 28% 0 136 90%/7%/0%/4% 5% 178/4/0/2 PHerc0125_3mm_v2 352 58% 42% 24%>0% 24% 0 135 91%/5%/0%/4% 13% 127/1/0/22 gradients: max 45, clipped 0/50 updates Step 4,900 | last 50 updates / 1,600 crops | lr 2.81e-05 points (35/curve, tolerance 2 vox): 49,663 right / 1,911 wrong | 96.3% correct | 4,426 unknown loss 0.8582 | geometry 0.2442 | confidence 1.2281 | mean error 0.634 vox speed: 889 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1595 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 190, from the previous prediction 1405 confidence-labeled crops: terminal 5/1537 (0.3%) | recoverable 106/1537 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 256/1600 (16.0%), 2: 384/1600 (24.0%), 3: 416/1600 (26.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 702 rows | mean depth 14 decisions | mean travel 112 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 50% 48% 28%>2% 26% 3 308 89%/8%/0%/2% 10% 122/0/0/0 0175A_5mm_v1 256 57% 43% 24%>0% 24% 0 134 90%/6%/0%/4% 4% 109/2/0/0 1447_5mm_v1 384 57% 43% 27%>0% 27% 0 156 92%/5%/0%/3% 1% 148/1/1/16 paris4_afv_central 416 56% 44% 25%>0% 25% 0 147 88%/6%/0%/5% 6% 183/1/0/0 PHerc0125_3mm_v2 288 51% 49% 23%>0% 23% 0 156 86%/8%/1%/5% 14% 140/3/0/0 gradients: max 51, clipped 0/50 updates Step 4,950 | last 50 updates / 1,600 crops | lr 2.80e-05 points (35/curve, tolerance 2 vox): 49,817 right / 2,084 wrong | 96.0% correct | 4,099 unknown loss 0.8717 | geometry 0.2651 | confidence 1.2131 | mean error 0.659 vox speed: 888 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1595 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 206, from the previous prediction 1389 confidence-labeled crops: terminal 0/1546 (0.0%) | recoverable 117/1546 (7.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 384/1600 (24.0%), 2: 288/1600 (18.0%), 3: 256/1600 (16.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 674 rows | mean depth 18 decisions | mean travel 142 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 56% 42% 30%>2% 28% 4 310 90%/7%/0%/3% 10% 121/0/0/0 0175A_5mm_v1 384 54% 46% 22%>0% 22% 0 136 87%/9%/0%/4% 4% 153/3/0/23 1447_5mm_v1 288 52% 48% 23%>0% 23% 0 159 92%/6%/0%/2% 3% 137/1/0/0 paris4_afv_central 256 59% 41% 29%>0% 29% 0 145 88%/7%/0%/5% 6% 91/3/0/13 PHerc0125_3mm_v2 384 55% 45% 28%>0% 28% 0 147 89%/7%/0%/4% 11% 172/5/0/0 gradients: max 45, clipped 0/50 updates Step 5,000 | last 50 updates / 1,600 crops | lr 2.80e-05 points (35/curve, tolerance 2 vox): 49,694 right / 2,073 wrong | 96.0% correct | 4,233 unknown loss 0.9240 | geometry 0.2539 | confidence 1.3403 | mean error 0.659 vox speed: 888 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.13 attempts/decision | peak allocated VRAM 41.09 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1590 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 197, from the previous prediction 1393 confidence-labeled crops: terminal 1/1547 (0.1%) | recoverable 82/1547 (5.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 416/1600 (26.0%), 2: 192/1600 (12.0%), 3: 320/1600 (20.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 720 rows | mean depth 16 decisions | mean travel 134 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 54% 44% 25%>3% 22% 6 375 93%/5%/0%/2% 8% 165/0/42/3 0175A_5mm_v1 416 52% 48% 25%>0% 25% 0 145 92%/6%/0%/2% 4% 200/4/0/0 1447_5mm_v1 192 56% 44% 18%>0% 18% 0 139 93%/4%/0%/4% 3% 84/0/0/0 paris4_afv_central 320 55% 45% 24%>0% 24% 0 145 92%/6%/0%/2% 5% 143/1/0/0 PHerc0125_3mm_v2 288 53% 47% 23%>0% 23% 0 151 87%/5%/0%/8% 13% 128/2/48/8 gradients: max 49, clipped 0/50 updates Step 5,000 | dagger_launched: true Step 5,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 249, "stopped": 0, "mean_example_error": 0.9246636754082095, "geometry_examples": 31, "examples": 32, "inference_and_measurement_seconds": 1.949739545583725, "total_diagnostic_seconds": 3.764785584062338} Step 5,000 | monitor rollout @ 0.50 coverage 76.7% | precision 80.7% | diverged 25.0% Step 5,000 | monitor rollout @ 0.50 coverage 93.9% | precision 100.0% | diverged 0.0% Step 5,000 | monitor rollout @ 0.50 coverage 99.2% | precision 100.0% | diverged 0.0% Step 5,000 | monitor rollout @ 0.50 coverage 94.4% | precision 95.1% | diverged 6.2% Step 5,000 | monitor rollout @ 0.50 coverage 82.6% | precision 71.4% | diverged 21.9% Step 5,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 6/8 (75.0%) | false stops 1 1.5-3: recovered 6/8 (75.0%) | false stops 0 3-4.5: recovered 6/8 (75.0%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 0 entire proposed 8-point prefix correct (distance): 21/28 (75.0%) gate @ 0.50: false stops 0/22 (0.0%) | accepted wrong 5/26 (19.2%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.540 1.511 following 16 464 1.071 1.076 recoverable 12 348 2.166 2.090 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 1.071 1.076 332 32 812 1.540 1.511 improved/worsened/compared states: 1/2/28 Step 5,000 | batch_diagnostics_seconds: 3.7671072147786617 | diagnostics_seconds: 761.4237646497786 | recovery_seconds: 14.966433130204678 Step 5,001 | dagger_states: 4981 | dagger_source_step: 5000 | dagger_caches: 1 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/PHerc0125_3mm_v2/decisions_005000.npz" | dagger_fibers: 384 | dagger_supply: {"states": 4981, "traces": 384, "fibers": 384, "rows": {"terminal": 658, "premature_stop": 118, "recoverable": 135, "pre_excursion": 667, "ordinary": 3358}, "events": {"terminal": 93, "premature_stop": 118, "recoverable": 52, "pre_excursion": 127, "ordinary": 341}, "supervision": {"following": 4140, "recoverable": 101, "terminal": 658, "unknown": 82}, "departed_traces": 111, "terminal_traces": 93, "stop_reasons": {"after_failure": 81, "confidence": 130, "unannotated": 173}, "traced_voxels": 65115.22489237283, "seconds": 355.73699866235256, "voxels_per_second": 183.04316148508602} | dagger_coverage: {"requested": 384, "seeds": 384, "skipped": [], "epoch": 0} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 0 | dagger_launch_interval: null | dagger_publication_age: 1 | dataset: "PHerc0125_3mm_v2" Step 5,050 | last 50 updates / 1,600 crops | lr 2.80e-05 points (35/curve, tolerance 2 vox): 49,452 right / 2,013 wrong | 96.1% correct | 4,535 unknown loss 0.8397 | geometry 0.2622 | confidence 1.1551 | mean error 0.650 vox speed: 890 ms/update | data wait 1 ms/update | 1.9 crops/s | 1.9 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1599 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 276, from the previous prediction 1323 confidence-labeled crops: terminal 3/1546 (0.2%) | recoverable 102/1546 (6.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 160/1600 (10.0%), 2: 320/1600 (20.0%), 3: 512/1600 (32.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 597 rows | mean depth 13 decisions | mean travel 107 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 57% 43% 24%>0% 23% 1 182 90%/8%/0%/2% 8% 81/0/0/84 0175A_5mm_v1 160 56% 44% 27%>0% 27% 0 154 89%/8%/0%/4% 4% 71/1/0/0 1447_5mm_v1 320 56% 44% 31%>0% 31% 0 166 88%/10%/0%/2% 3% 141/0/0/0 paris4_afv_central 512 54% 46% 22%>0% 22% 0 140 93%/4%/0%/4% 6% 208/1/2/26 PHerc0125_3mm_v2 224 57% 43% 24%>0% 24% 0 157 88%/4%/1%/7% 12% 96/1/0/0 gradients: max 80, clipped 1/50 updates Step 5,100 | last 50 updates / 1,600 crops | lr 2.79e-05 points (35/curve, tolerance 2 vox): 49,510 right / 2,139 wrong | 95.9% correct | 4,351 unknown loss 0.8790 | geometry 0.2489 | confidence 1.2602 | mean error 0.653 vox speed: 887 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1589 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 168, from the previous prediction 1421 confidence-labeled crops: terminal 6/1544 (0.4%) | recoverable 107/1544 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 416/1600 (26.0%), 2: 288/1600 (18.0%), 3: 224/1600 (14.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 741 rows | mean depth 19 decisions | mean travel 150 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 51% 46% 28%>3% 25% 6 408 88%/9%/1%/2% 11% 152/0/0/9 0175A_5mm_v1 416 50% 50% 21%>0% 21% 0 152 92%/4%/0%/4% 4% 207/2/37/2 1447_5mm_v1 288 52% 48% 23%>0% 23% 0 147 90%/6%/0%/4% 2% 139/3/0/0 paris4_afv_central 224 55% 45% 24%>0% 24% 0 140 88%/9%/0%/3% 5% 100/1/0/1 PHerc0125_3mm_v2 320 55% 45% 24%>0% 24% 0 158 88%/6%/1%/6% 13% 143/2/0/0 gradients: max 57, clipped 0/50 updates Step 5,150 | last 50 updates / 1,600 crops | lr 2.79e-05 points (35/curve, tolerance 2 vox): 48,257 right / 2,775 wrong | 94.6% correct | 4,968 unknown loss 1.0088 | geometry 0.2897 | confidence 1.4382 | mean error 0.708 vox speed: 885 ms/update | data wait 1 ms/update | 35.8 crops/s | 35.8 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1599 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 280, from the previous prediction 1319 confidence-labeled crops: terminal 4/1534 (0.3%) | recoverable 98/1534 (6.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 192/1600 (12.0%), 2: 352/1600 (22.0%), 3: 256/1600 (16.0%), 4: 480/1600 (30.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 615 rows | mean depth 17 decisions | mean travel 140 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 60% 40% 23%>0% 23% 1 172 91%/6%/0%/3% 10% 127/0/0/0 0175A_5mm_v1 192 59% 41% 28%>0% 28% 0 154 90%/4%/0%/6% 4% 78/2/0/0 1447_5mm_v1 352 55% 45% 25%>0% 25% 0 145 90%/7%/0%/3% 2% 128/2/16/30 paris4_afv_central 256 56% 44% 29%>0% 29% 0 162 90%/7%/0%/2% 6% 113/5/0/0 PHerc0125_3mm_v2 480 52% 48% 24%>0% 24% 0 146 87%/6%/0%/7% 17% 169/5/0/62 gradients: max 69, clipped 2/50 updates Step 5,200 | last 50 updates / 1,600 crops | lr 2.78e-05 points (35/curve, tolerance 2 vox): 49,819 right / 2,301 wrong | 95.6% correct | 3,880 unknown loss 0.8374 | geometry 0.2569 | confidence 1.1610 | mean error 0.665 vox speed: 902 ms/update | data wait 1 ms/update | 35.1 crops/s | 35.1 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1596 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 160, from the previous prediction 1436 confidence-labeled crops: terminal 4/1548 (0.3%) | recoverable 108/1548 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 288/1600 (18.0%), 2: 256/1600 (16.0%), 3: 384/1600 (24.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 37 chunks / 74.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 730 rows | mean depth 14 decisions | mean travel 118 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 52% 47% 24%>1% 23% 4 237 91%/8%/0%/2% 8% 179/0/0/0 0175A_5mm_v1 288 58% 42% 25%>0% 25% 0 172 89%/6%/0%/5% 3% 122/1/0/0 1447_5mm_v1 256 53% 47% 24%>0% 24% 0 139 92%/5%/0%/3% 3% 113/0/0/7 paris4_afv_central 384 52% 48% 23%>0% 23% 0 154 88%/8%/0%/4% 6% 184/3/0/0 PHerc0125_3mm_v2 288 54% 46% 26%>0% 26% 0 138 89%/6%/1%/4% 13% 132/3/0/0 gradients: max 77, clipped 1/50 updates Step 5,250 | last 50 updates / 1,600 crops | lr 2.78e-05 points (35/curve, tolerance 2 vox): 48,822 right / 2,281 wrong | 95.5% correct | 4,897 unknown loss 0.9637 | geometry 0.2595 | confidence 1.4085 | mean error 0.673 vox speed: 886 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 229, from the previous prediction 1371 confidence-labeled crops: terminal 3/1540 (0.2%) | recoverable 108/1540 (7.0%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 320/1600 (20.0%), 2: 192/1600 (12.0%), 3: 352/1600 (22.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 675 rows | mean depth 19 decisions | mean travel 155 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 57% 43% 27%>0% 27% 0 141 88%/10%/0%/2% 9% 154/2/0/12 0175A_5mm_v1 320 51% 49% 22%>0% 22% 0 159 93%/4%/0%/3% 3% 156/1/0/0 1447_5mm_v1 192 55% 45% 25%>0% 25% 0 161 90%/7%/0%/3% 2% 86/1/0/0 paris4_afv_central 352 51% 49% 26%>0% 26% 0 139 90%/6%/1%/3% 7% 129/2/23/42 PHerc0125_3mm_v2 352 57% 43% 28%>0% 28% 0 140 86%/6%/0%/7% 13% 150/0/0/0 gradients: max 65, clipped 1/50 updates Step 5,300 | last 50 updates / 1,600 crops | lr 2.77e-05 points (35/curve, tolerance 2 vox): 49,178 right / 2,283 wrong | 95.6% correct | 4,539 unknown loss 0.9224 | geometry 0.2828 | confidence 1.2791 | mean error 0.668 vox speed: 893 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 229, from the previous prediction 1371 confidence-labeled crops: terminal 3/1542 (0.2%) | recoverable 106/1542 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 160/1600 (10.0%), 1: 352/1600 (22.0%), 2: 384/1600 (24.0%), 3: 288/1600 (18.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 665 rows | mean depth 15 decisions | mean travel 125 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 160 54% 46% 24%>0% 24% 0 150 94%/3%/0%/2% 8% 73/0/0/0 0175A_5mm_v1 352 54% 46% 23%>0% 23% 0 139 89%/8%/0%/3% 5% 132/0/14/30 1447_5mm_v1 384 57% 43% 24%>0% 24% 0 162 89%/7%/0%/4% 3% 166/2/0/0 paris4_afv_central 288 53% 47% 24%>0% 24% 0 136 91%/5%/0%/3% 6% 115/3/0/21 PHerc0125_3mm_v2 416 54% 46% 25%>0% 25% 0 140 87%/8%/1%/5% 14% 179/2/0/13 gradients: max 66, clipped 1/50 updates Step 5,350 | last 50 updates / 1,600 crops | lr 2.77e-05 points (35/curve, tolerance 2 vox): 50,039 right / 1,689 wrong | 96.7% correct | 4,272 unknown loss 0.7591 | geometry 0.2182 | confidence 1.0819 | mean error 0.610 vox speed: 887 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1600 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 287, from the previous prediction 1313 confidence-labeled crops: terminal 3/1543 (0.2%) | recoverable 103/1543 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 384/1600 (24.0%), 2: 384/1600 (24.0%), 3: 352/1600 (22.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 607 rows | mean depth 16 decisions | mean travel 133 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 52% 48% 25%>0% 25% 0 188 93%/7%/0%/1% 6% 93/0/0/0 0175A_5mm_v1 384 54% 46% 27%>0% 27% 0 132 89%/6%/0%/5% 3% 75/0/0/101 1447_5mm_v1 384 53% 47% 21%>0% 21% 0 143 90%/7%/0%/3% 1% 163/6/17/17 paris4_afv_central 352 55% 45% 28%>0% 28% 0 141 86%/8%/0%/5% 6% 158/3/0/0 PHerc0125_3mm_v2 288 56% 44% 20%>0% 20% 0 137 92%/3%/1%/3% 11% 118/0/0/9 gradients: max 51, clipped 0/50 updates Step 5,400 | last 50 updates / 1,600 crops | lr 2.77e-05 points (35/curve, tolerance 2 vox): 49,120 right / 2,344 wrong | 95.4% correct | 4,536 unknown loss 0.9324 | geometry 0.2646 | confidence 1.3357 | mean error 0.671 vox speed: 886 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1587 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 282, from the previous prediction 1305 confidence-labeled crops: terminal 8/1532 (0.5%) | recoverable 100/1532 (6.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 288/1600 (18.0%), 2: 320/1600 (20.0%), 3: 224/1600 (14.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 628 rows | mean depth 20 decisions | mean travel 161 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 55% 41% 26%>4% 23% 9 520 91%/5%/1%/3% 10% 144/0/9/0 0175A_5mm_v1 288 47% 53% 19%>0% 19% 0 148 90%/7%/0%/4% 4% 142/3/0/11 1447_5mm_v1 320 57% 43% 26%>0% 26% 0 136 88%/8%/0%/3% 2% 96/1/0/43 paris4_afv_central 224 54% 46% 24%>0% 24% 0 153 90%/6%/0%/4% 5% 102/4/0/0 PHerc0125_3mm_v2 416 55% 45% 27%>0% 27% 0 140 87%/6%/1%/7% 13% 144/4/0/43 gradients: max 63, clipped 1/50 updates Step 5,450 | last 50 updates / 1,600 crops | lr 2.76e-05 points (35/curve, tolerance 2 vox): 49,756 right / 2,244 wrong | 95.7% correct | 4,000 unknown loss 0.8555 | geometry 0.2432 | confidence 1.2246 | mean error 0.647 vox speed: 886 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1591 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 197, from the previous prediction 1394 confidence-labeled crops: terminal 8/1542 (0.5%) | recoverable 109/1542 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 192/1600 (12.0%), 2: 416/1600 (26.0%), 3: 352/1600 (22.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 687 rows | mean depth 16 decisions | mean travel 128 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 56% 41% 29%>3% 26% 5 496 87%/9%/1%/3% 10% 106/1/1/12 0175A_5mm_v1 192 47% 53% 22%>0% 22% 0 163 92%/7%/0%/1% 3% 101/2/0/0 1447_5mm_v1 416 53% 47% 22%>0% 22% 0 137 89%/6%/0%/5% 2% 165/1/0/32 paris4_afv_central 352 58% 42% 26%>0% 26% 0 140 90%/7%/0%/4% 6% 148/2/23/1 PHerc0125_3mm_v2 352 53% 47% 26%>0% 26% 0 144 88%/7%/1%/5% 13% 167/7/0/0 gradients: max 65, clipped 2/50 updates Step 5,500 | last 50 updates / 1,600 crops | lr 2.76e-05 points (35/curve, tolerance 2 vox): 49,914 right / 1,586 wrong | 96.9% correct | 4,500 unknown loss 0.8536 | geometry 0.2445 | confidence 1.2182 | mean error 0.628 vox speed: 885 ms/update | data wait 1 ms/update | 35.8 crops/s | 35.8 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1594 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 226, from the previous prediction 1368 confidence-labeled crops: terminal 1/1541 (0.1%) | recoverable 107/1541 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 320/1600 (20.0%), 2: 288/1600 (18.0%), 3: 352/1600 (22.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 636 rows | mean depth 17 decisions | mean travel 134 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 56% 41% 26%>3% 22% 4 496 93%/6%/0%/2% 5% 70/0/0/8 0175A_5mm_v1 320 55% 45% 25%>0% 25% 0 157 90%/7%/0%/3% 4% 144/2/34/1 1447_5mm_v1 288 59% 41% 23%>0% 23% 0 146 90%/8%/0%/3% 3% 118/3/0/0 paris4_afv_central 352 54% 46% 29%>0% 29% 0 140 88%/7%/0%/5% 5% 160/1/16/1 PHerc0125_3mm_v2 448 55% 45% 23%>0% 23% 0 135 89%/6%/0%/5% 10% 144/4/0/58 gradients: max 55, clipped 0/50 updates Step 5,550 | last 50 updates / 1,600 crops | lr 2.75e-05 points (35/curve, tolerance 2 vox): 50,033 right / 2,076 wrong | 96.0% correct | 3,891 unknown loss 0.9013 | geometry 0.2750 | confidence 1.2526 | mean error 0.668 vox speed: 889 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.12 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1594 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 243, from the previous prediction 1351 confidence-labeled crops: terminal 1/1552 (0.1%) | recoverable 107/1552 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 288/1600 (18.0%), 2: 288/1600 (18.0%), 3: 320/1600 (20.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 656 rows | mean depth 20 decisions | mean travel 163 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 56% 43% 25%>1% 24% 4 281 91%/7%/0%/2% 7% 193/0/0/0 0175A_5mm_v1 288 52% 48% 24%>0% 24% 0 140 90%/8%/0%/2% 5% 137/3/16/1 1447_5mm_v1 288 50% 50% 25%>0% 25% 0 142 91%/7%/0%/3% 2% 143/2/0/0 paris4_afv_central 320 58% 42% 26%>0% 26% 0 137 90%/5%/0%/4% 7% 70/0/17/63 PHerc0125_3mm_v2 256 56% 44% 26%>0% 26% 0 146 88%/7%/0%/4% 15% 113/3/0/0 gradients: max 89, clipped 1/50 updates Step 5,600 | last 50 updates / 1,600 crops | lr 2.75e-05 points (35/curve, tolerance 2 vox): 49,634 right / 2,008 wrong | 96.1% correct | 4,358 unknown loss 0.9160 | geometry 0.2709 | confidence 1.2902 | mean error 0.679 vox speed: 892 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1593 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 281, from the previous prediction 1312 confidence-labeled crops: terminal 2/1544 (0.1%) | recoverable 99/1544 (6.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 224/1600 (14.0%), 2: 288/1600 (18.0%), 3: 256/1600 (16.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 633 rows | mean depth 17 decisions | mean travel 137 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 51% 47% 23%>2% 21% 7 407 91%/6%/0%/3% 8% 108/0/0/88 0175A_5mm_v1 224 52% 48% 22%>0% 22% 0 143 91%/4%/0%/5% 4% 100/1/51/7 1447_5mm_v1 288 54% 46% 23%>0% 23% 0 144 92%/6%/0%/3% 4% 132/2/0/0 paris4_afv_central 256 59% 41% 26%>0% 26% 0 145 88%/7%/0%/4% 6% 106/0/0/0 PHerc0125_3mm_v2 416 55% 45% 25%>0% 25% 0 148 88%/7%/0%/5% 9% 187/6/0/0 gradients: max 65, clipped 1/50 updates Step 5,650 | last 50 updates / 1,600 crops | lr 2.75e-05 points (35/curve, tolerance 2 vox): 50,075 right / 1,778 wrong | 96.6% correct | 4,147 unknown loss 0.7931 | geometry 0.2442 | confidence 1.0977 | mean error 0.630 vox speed: 892 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1592 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 296, from the previous prediction 1296 confidence-labeled crops: terminal 1/1549 (0.1%) | recoverable 119/1549 (7.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 224/1600 (14.0%), 2: 576/1600 (36.0%), 3: 288/1600 (18.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 569 rows | mean depth 17 decisions | mean travel 137 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 60% 38% 31%>3% 28% 5 650 88%/10%/0%/2% 7% 63/0/0/45 0175A_5mm_v1 224 58% 42% 26%>0% 26% 0 150 93%/6%/0%/1% 4% 95/0/0/0 1447_5mm_v1 576 56% 44% 24%>0% 24% 0 137 88%/8%/0%/4% 3% 184/3/42/69 paris4_afv_central 288 57% 43% 28%>0% 28% 0 176 90%/8%/0%/2% 7% 125/1/0/0 PHerc0125_3mm_v2 224 54% 46% 26%>0% 26% 0 146 89%/4%/0%/6% 13% 102/3/37/2 gradients: max 57, clipped 0/50 updates Step 5,700 | last 50 updates / 1,600 crops | lr 2.74e-05 points (35/curve, tolerance 2 vox): 49,689 right / 2,038 wrong | 96.1% correct | 4,273 unknown loss 0.9034 | geometry 0.2618 | confidence 1.2832 | mean error 0.654 vox speed: 888 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1598 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 214, from the previous prediction 1384 confidence-labeled crops: terminal 2/1555 (0.1%) | recoverable 91/1555 (5.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 320/1600 (20.0%), 2: 256/1600 (16.0%), 3: 320/1600 (20.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 689 rows | mean depth 21 decisions | mean travel 169 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 55% 45% 23%>1% 23% 2 236 94%/5%/0%/1% 7% 114/0/0/0 0175A_5mm_v1 320 55% 45% 23%>0% 23% 0 161 92%/6%/0%/2% 2% 143/3/0/0 1447_5mm_v1 256 51% 49% 24%>0% 24% 0 142 92%/6%/0%/2% 2% 126/1/0/0 paris4_afv_central 320 50% 50% 20%>0% 20% 0 151 94%/3%/0%/2% 7% 159/6/0/0 PHerc0125_3mm_v2 448 60% 40% 26%>0% 26% 0 141 87%/7%/0%/6% 13% 147/5/0/34 gradients: max 59, clipped 0/50 updates Step 5,750 | last 50 updates / 1,600 crops | lr 2.74e-05 points (35/curve, tolerance 2 vox): 49,989 right / 1,937 wrong | 96.3% correct | 4,074 unknown loss 0.8309 | geometry 0.2706 | confidence 1.1205 | mean error 0.648 vox speed: 891 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.13 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1594 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 335, from the previous prediction 1259 confidence-labeled crops: terminal 0/1548 (0.0%) | recoverable 114/1548 (7.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 384/1600 (24.0%), 2: 352/1600 (22.0%), 3: 352/1600 (22.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 572 rows | mean depth 15 decisions | mean travel 125 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 56% 42% 23%>2% 21% 6 392 93%/6%/0%/1% 8% 122/0/0/0 0175A_5mm_v1 384 53% 47% 23%>0% 23% 0 145 90%/7%/0%/3% 4% 170/1/0/9 1447_5mm_v1 352 53% 47% 26%>0% 26% 0 136 88%/9%/0%/3% 3% 100/1/0/66 paris4_afv_central 352 51% 49% 24%>0% 24% 0 142 89%/7%/0%/3% 6% 81/1/51/93 PHerc0125_3mm_v2 224 56% 44% 25%>0% 25% 0 145 86%/6%/0%/8% 11% 99/2/0/0 gradients: max 50, clipped 0/50 updates Step 5,800 | last 50 updates / 1,600 crops | lr 2.73e-05 points (35/curve, tolerance 2 vox): 49,292 right / 1,782 wrong | 96.5% correct | 4,926 unknown loss 0.8427 | geometry 0.2682 | confidence 1.1490 | mean error 0.663 vox speed: 888 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1595 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 238, from the previous prediction 1357 confidence-labeled crops: terminal 1/1535 (0.1%) | recoverable 120/1535 (7.8%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 192/1600 (12.0%), 2: 320/1600 (20.0%), 3: 416/1600 (26.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 655 rows | mean depth 19 decisions | mean travel 152 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 53% 45% 23%>1% 22% 4 293 91%/7%/0%/2% 7% 158/0/53/1 0175A_5mm_v1 192 56% 44% 23%>0% 23% 0 142 88%/7%/0%/5% 3% 84/1/0/0 1447_5mm_v1 320 55% 45% 27%>0% 27% 0 152 87%/9%/0%/4% 3% 145/2/0/0 paris4_afv_central 416 56% 44% 24%>0% 24% 0 137 87%/7%/0%/6% 6% 121/2/0/61 PHerc0125_3mm_v2 320 54% 46% 24%>0% 24% 0 159 89%/7%/0%/4% 9% 147/2/0/0 gradients: max 56, clipped 0/50 updates Step 5,850 | last 50 updates / 1,600 crops | lr 2.73e-05 points (35/curve, tolerance 2 vox): 48,802 right / 2,619 wrong | 94.9% correct | 4,579 unknown loss 1.0173 | geometry 0.2845 | confidence 1.4657 | mean error 0.709 vox speed: 889 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1594 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 238, from the previous prediction 1356 confidence-labeled crops: terminal 2/1540 (0.1%) | recoverable 115/1540 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 256/1600 (16.0%), 2: 288/1600 (18.0%), 3: 256/1600 (16.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 680 rows | mean depth 20 decisions | mean travel 164 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 53% 45% 21%>2% 19% 5 376 89%/8%/0%/3% 9% 112/1/47/46 0175A_5mm_v1 256 53% 47% 23%>0% 23% 0 141 92%/5%/0%/3% 2% 121/3/0/0 1447_5mm_v1 288 53% 47% 25%>0% 25% 0 154 90%/7%/0%/3% 3% 135/0/0/0 paris4_afv_central 256 55% 45% 24%>0% 24% 0 141 88%/7%/0%/5% 5% 114/4/0/0 PHerc0125_3mm_v2 448 56% 44% 23%>0% 23% 0 147 87%/8%/0%/4% 15% 198/6/2/0 gradients: max 72, clipped 1/50 updates Step 5,900 | last 50 updates / 1,600 crops | lr 2.72e-05 points (35/curve, tolerance 2 vox): 49,552 right / 1,950 wrong | 96.2% correct | 4,498 unknown loss 0.8403 | geometry 0.2424 | confidence 1.1958 | mean error 0.645 vox speed: 887 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1598 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 180, from the previous prediction 1418 confidence-labeled crops: terminal 4/1542 (0.3%) | recoverable 107/1542 (6.9%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 320/1600 (20.0%), 2: 320/1600 (20.0%), 3: 416/1600 (26.0%), 4: 160/1600 (10.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 701 rows | mean depth 15 decisions | mean travel 121 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 55% 45% 25%>1% 24% 2 197 93%/5%/0%/2% 9% 162/0/0/10 0175A_5mm_v1 320 56% 44% 25%>0% 25% 0 150 91%/7%/0%/2% 5% 141/1/0/0 1447_5mm_v1 320 58% 42% 27%>0% 27% 0 153 84%/10%/1%/5% 4% 133/2/0/0 paris4_afv_central 416 56% 44% 25%>0% 25% 0 141 88%/6%/0%/5% 5% 181/8/0/0 PHerc0125_3mm_v2 160 48% 52% 21%>0% 21% 0 135 91%/6%/0%/3% 12% 84/1/0/0 gradients: max 71, clipped 2/50 updates Step 5,950 | last 50 updates / 1,600 crops | lr 2.72e-05 points (35/curve, tolerance 2 vox): 49,611 right / 2,231 wrong | 95.7% correct | 4,158 unknown loss 0.8356 | geometry 0.2579 | confidence 1.1552 | mean error 0.654 vox speed: 888 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1589 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 240, from the previous prediction 1349 confidence-labeled crops: terminal 8/1550 (0.5%) | recoverable 112/1550 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 256/1600 (16.0%), 2: 448/1600 (28.0%), 3: 288/1600 (18.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 655 rows | mean depth 20 decisions | mean travel 163 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 52% 45% 22%>3% 18% 6 538 91%/5%/2%/2% 10% 121/1/0/23 0175A_5mm_v1 256 52% 48% 23%>0% 23% 0 144 90%/7%/0%/4% 4% 123/2/0/0 1447_5mm_v1 448 52% 48% 25%>0% 25% 0 134 90%/7%/0%/3% 2% 168/2/126/45 paris4_afv_central 288 55% 45% 26%>0% 26% 0 143 89%/8%/0%/3% 6% 130/2/0/0 PHerc0125_3mm_v2 288 61% 39% 32%>0% 32% 0 156 86%/8%/0%/5% 13% 113/2/0/0 gradients: max 58, clipped 0/50 updates Step 6,000 | last 50 updates / 1,600 crops | lr 2.71e-05 points (35/curve, tolerance 2 vox): 48,769 right / 2,502 wrong | 95.1% correct | 4,729 unknown loss 0.9341 | geometry 0.2692 | confidence 1.3298 | mean error 0.671 vox speed: 890 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1583 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 204, from the previous prediction 1379 confidence-labeled crops: terminal 14/1538 (0.9%) | recoverable 100/1538 (6.5%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 320/1600 (20.0%), 2: 192/1600 (12.0%), 3: 416/1600 (26.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 671 rows | mean depth 20 decisions | mean travel 160 vox | 8 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 51% 44% 29%>5% 23% 9 665 88%/7%/2%/3% 12% 140/1/0/0 0175A_5mm_v1 320 57% 43% 28%>0% 28% 0 140 89%/6%/0%/5% 3% 109/0/9/28 1447_5mm_v1 192 48% 52% 24%>0% 24% 0 138 89%/5%/0%/6% 2% 80/1/0/20 paris4_afv_central 416 56% 44% 24%>0% 24% 0 145 88%/7%/1%/4% 5% 183/5/1/0 PHerc0125_3mm_v2 352 55% 45% 23%>0% 23% 0 154 88%/6%/2%/4% 15% 159/3/0/0 gradients: max 53, clipped 0/50 updates Step 6,000 | dagger_launched: true Step 6,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 252, "stopped": 0, "mean_example_error": 0.6057353187352419, "geometry_examples": 32, "examples": 32, "inference_and_measurement_seconds": 2.1008139178156853, "total_diagnostic_seconds": 4.202193308621645} Step 6,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 6/8 (75.0%) | false stops 1 1.5-3: recovered 6/8 (75.0%) | false stops 0 3-4.5: recovered 6/8 (75.0%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 0 entire proposed 8-point prefix correct (distance): 21/28 (75.0%) gate @ 0.50: false stops 0/22 (0.0%) | accepted wrong 5/26 (19.2%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.506 1.468 following 16 464 1.065 1.048 recoverable 12 348 2.095 2.028 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 1.065 1.048 332 32 812 1.506 1.468 improved/worsened/compared states: 2/1/28 Step 6,000 | batch_diagnostics_seconds: 4.20473025739193 | recovery_seconds: 14.751649636775255 Step 6,050 | last 50 updates / 1,600 crops | lr 2.71e-05 points (35/curve, tolerance 2 vox): 50,007 right / 1,915 wrong | 96.3% correct | 4,078 unknown loss 0.8336 | geometry 0.2608 | confidence 1.1456 | mean error 0.655 vox speed: 889 ms/update | data wait 1 ms/update | 24.3 crops/s | 24.3 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1587 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 190, from the previous prediction 1397 confidence-labeled crops: terminal 2/1544 (0.1%) | recoverable 105/1544 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 320/1600 (20.0%), 2: 384/1600 (24.0%), 3: 288/1600 (18.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 700 rows | mean depth 16 decisions | mean travel 130 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 51% 45% 24%>4% 20% 9 508 89%/8%/0%/2% 9% 159/1/0/0 0175A_5mm_v1 320 50% 50% 19%>0% 19% 0 143 92%/5%/0%/2% 3% 155/2/0/4 1447_5mm_v1 384 56% 44% 28%>0% 28% 0 146 88%/8%/0%/4% 3% 169/1/0/0 paris4_afv_central 288 56% 44% 22%>0% 22% 0 138 89%/7%/0%/4% 3% 104/3/21/24 PHerc0125_3mm_v2 256 56% 44% 25%>0% 25% 0 168 90%/4%/1%/5% 11% 113/4/0/0 gradients: max 62, clipped 1/50 updates Step 6,100 | last 50 updates / 1,600 crops | lr 2.70e-05 points (35/curve, tolerance 2 vox): 49,889 right / 1,849 wrong | 96.4% correct | 4,262 unknown loss 0.8243 | geometry 0.2634 | confidence 1.1219 | mean error 0.661 vox speed: 1418 ms/update | data wait 1 ms/update | 22.4 crops/s | 22.4 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1595 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 263, from the previous prediction 1332 confidence-labeled crops: terminal 1/1556 (0.1%) | recoverable 119/1556 (7.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 576/1600 (36.0%), 1: 224/1600 (14.0%), 2: 352/1600 (22.0%), 3: 192/1600 (12.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 595 rows | mean depth 20 decisions | mean travel 163 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 576 54% 45% 24%>1% 23% 5 279 91%/7%/0%/2% 7% 169/1/5/91 0175A_5mm_v1 224 63% 37% 28%>0% 28% 0 139 89%/8%/0%/4% 5% 82/3/0/0 1447_5mm_v1 352 58% 42% 26%>0% 26% 0 149 91%/7%/0%/3% 1% 148/2/0/0 paris4_afv_central 192 49% 51% 22%>0% 22% 0 149 88%/9%/0%/3% 5% 97/1/0/0 PHerc0125_3mm_v2 256 61% 39% 33%>0% 33% 0 160 88%/8%/0%/4% 12% 99/3/0/0 gradients: max 55, clipped 0/50 updates Step 6,150 | last 50 updates / 1,600 crops | lr 2.70e-05 points (35/curve, tolerance 2 vox): 49,412 right / 2,177 wrong | 95.8% correct | 4,411 unknown loss 0.9341 | geometry 0.2705 | confidence 1.3273 | mean error 0.671 vox speed: 1356 ms/update | data wait 1 ms/update | 23.4 crops/s | 23.4 decisions/s | 1.14 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1597 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 179, from the previous prediction 1418 confidence-labeled crops: terminal 1/1551 (0.1%) | recoverable 112/1551 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 384/1600 (24.0%), 2: 256/1600 (16.0%), 3: 384/1600 (24.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 702 rows | mean depth 20 decisions | mean travel 165 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 54% 45% 26%>1% 25% 3 307 91%/8%/0%/1% 7% 85/0/0/15 0175A_5mm_v1 384 52% 48% 23%>0% 23% 0 142 90%/5%/0%/5% 2% 183/3/0/0 1447_5mm_v1 256 58% 42% 21%>0% 21% 0 138 92%/7%/0%/1% 2% 108/0/0/0 paris4_afv_central 384 56% 44% 29%>0% 29% 0 155 89%/8%/0%/3% 6% 168/3/0/0 PHerc0125_3mm_v2 352 55% 45% 26%>0% 26% 0 154 89%/7%/0%/3% 15% 158/5/0/0 gradients: max 66, clipped 3/50 updates Step 6,200 | last 50 updates / 1,600 crops | lr 2.69e-05 points (35/curve, tolerance 2 vox): 49,488 right / 2,419 wrong | 95.3% correct | 4,093 unknown loss 0.8998 | geometry 0.2812 | confidence 1.2372 | mean error 0.678 vox speed: 1322 ms/update | data wait 1 ms/update | 24.0 crops/s | 24.0 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1591 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 259, from the previous prediction 1332 confidence-labeled crops: terminal 7/1536 (0.5%) | recoverable 112/1536 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 224/1600 (14.0%), 2: 384/1600 (24.0%), 3: 320/1600 (20.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 609 rows | mean depth 17 decisions | mean travel 137 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 55% 42% 27%>4% 24% 7 540 87%/9%/1%/3% 10% 107/0/0/0 0175A_5mm_v1 224 55% 45% 28%>0% 28% 0 137 91%/6%/0%/4% 4% 81/1/0/19 1447_5mm_v1 384 57% 43% 23%>0% 23% 0 139 90%/7%/0%/3% 2% 166/2/0/0 paris4_afv_central 320 56% 44% 28%>0% 28% 0 160 88%/8%/1%/4% 6% 142/1/0/0 PHerc0125_3mm_v2 416 58% 42% 27%>0% 27% 0 139 88%/6%/0%/6% 14% 113/1/19/62 gradients: max 50, clipped 0/50 updates Step 6,250 | last 50 updates / 1,600 crops | lr 2.69e-05 points (35/curve, tolerance 2 vox): 48,517 right / 2,331 wrong | 95.4% correct | 5,152 unknown loss 1.0009 | geometry 0.2891 | confidence 1.4236 | mean error 0.699 vox speed: 1409 ms/update | data wait 1 ms/update | 22.6 crops/s | 22.6 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1586 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 313, from the previous prediction 1273 confidence-labeled crops: terminal 3/1535 (0.2%) | recoverable 118/1535 (7.7%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 256/1600 (16.0%), 2: 192/1600 (12.0%), 3: 352/1600 (22.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 579 rows | mean depth 21 decisions | mean travel 172 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 54% 43% 28%>4% 25% 9 571 88%/8%/1%/3% 11% 164/1/0/0 0175A_5mm_v1 256 54% 46% 29%>0% 29% 0 165 90%/7%/0%/3% 5% 119/1/0/0 1447_5mm_v1 192 56% 44% 30%>0% 30% 0 140 88%/9%/0%/3% 1% 49/3/59/36 paris4_afv_central 352 59% 41% 25%>0% 25% 0 137 87%/7%/0%/6% 7% 131/4/22/13 PHerc0125_3mm_v2 416 54% 46% 26%>0% 26% 0 132 88%/7%/0%/5% 11% 116/3/0/75 gradients: max 59, clipped 0/50 updates Step 6,300 | last 50 updates / 1,600 crops | lr 2.69e-05 points (35/curve, tolerance 2 vox): 49,625 right / 2,489 wrong | 95.2% correct | 3,886 unknown loss 0.9260 | geometry 0.2712 | confidence 1.3095 | mean error 0.675 vox speed: 1340 ms/update | data wait 1 ms/update | 23.6 crops/s | 23.6 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1596 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 202, from the previous prediction 1394 confidence-labeled crops: terminal 6/1550 (0.4%) | recoverable 117/1550 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 192/1600 (12.0%), 2: 320/1600 (20.0%), 3: 416/1600 (26.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 700 rows | mean depth 22 decisions | mean travel 175 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 55% 44% 27%>1% 26% 4 253 89%/8%/0%/3% 10% 190/2/86/9 0175A_5mm_v1 192 51% 49% 20%>0% 20% 0 161 91%/7%/0%/2% 4% 94/1/0/0 1447_5mm_v1 320 54% 46% 25%>0% 25% 0 152 90%/8%/0%/3% 3% 147/5/0/0 paris4_afv_central 416 54% 46% 26%>0% 26% 0 139 89%/7%/1%/3% 6% 175/3/0/15 PHerc0125_3mm_v2 224 58% 42% 26%>0% 26% 0 146 87%/6%/1%/6% 15% 94/3/0/1 gradients: max 66, clipped 2/50 updates Step 6,350 | last 50 updates / 1,600 crops | lr 2.68e-05 points (35/curve, tolerance 2 vox): 49,516 right / 1,962 wrong | 96.2% correct | 4,522 unknown loss 0.8325 | geometry 0.2513 | confidence 1.1624 | mean error 0.652 vox speed: 1296 ms/update | data wait 1 ms/update | 24.5 crops/s | 24.5 decisions/s | 1.13 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1599 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 253, from the previous prediction 1346 confidence-labeled crops: terminal 3/1537 (0.2%) | recoverable 107/1537 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 320/1600 (20.0%), 2: 448/1600 (28.0%), 3: 288/1600 (18.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 637 rows | mean depth 17 decisions | mean travel 136 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 52% 48% 28%>0% 27% 1 217 90%/8%/0%/1% 8% 61/1/0/77 0175A_5mm_v1 320 60% 40% 30%>0% 30% 0 179 88%/8%/0%/5% 4% 128/0/0/0 1447_5mm_v1 448 51% 49% 22%>0% 22% 0 140 92%/5%/0%/2% 2% 219/4/0/0 paris4_afv_central 288 47% 53% 22%>0% 22% 0 134 89%/6%/0%/5% 6% 125/4/0/28 PHerc0125_3mm_v2 256 59% 41% 29%>0% 29% 0 160 86%/7%/0%/7% 14% 104/2/0/0 gradients: max 52, clipped 0/50 updates Step 6,400 | last 50 updates / 1,600 crops | lr 2.68e-05 points (35/curve, tolerance 2 vox): 49,369 right / 2,081 wrong | 96.0% correct | 4,550 unknown loss 0.8985 | geometry 0.2648 | confidence 1.2676 | mean error 0.663 vox speed: 1422 ms/update | data wait 1 ms/update | 22.2 crops/s | 22.2 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1599 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 296, from the previous prediction 1303 confidence-labeled crops: terminal 1/1534 (0.1%) | recoverable 123/1534 (8.0%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 256/1600 (16.0%), 2: 416/1600 (26.0%), 3: 352/1600 (22.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 592 rows | mean depth 21 decisions | mean travel 174 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 56% 43% 30%>0% 29% 1 201 89%/9%/0%/2% 11% 111/1/0/0 0175A_5mm_v1 256 53% 47% 24%>0% 24% 0 142 91%/7%/0%/2% 5% 120/5/0/0 1447_5mm_v1 416 54% 46% 25%>0% 25% 0 137 87%/7%/0%/6% 3% 75/0/0/117 paris4_afv_central 352 56% 44% 24%>0% 24% 0 138 89%/6%/0%/5% 6% 146/8/0/8 PHerc0125_3mm_v2 320 56% 44% 28%>0% 28% 0 148 85%/10%/0%/4% 12% 140/5/0/0 gradients: max 46, clipped 0/50 updates Step 6,450 | last 50 updates / 1,600 crops | lr 2.67e-05 points (35/curve, tolerance 2 vox): 50,085 right / 1,496 wrong | 97.1% correct | 4,419 unknown loss 0.7474 | geometry 0.2136 | confidence 1.0676 | mean error 0.591 vox speed: 1310 ms/update | data wait 1 ms/update | 24.3 crops/s | 24.3 decisions/s | 1.10 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1596 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 280, from the previous prediction 1316 confidence-labeled crops: terminal 2/1545 (0.1%) | recoverable 109/1545 (7.1%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 448/1600 (28.0%), 2: 416/1600 (26.0%), 3: 288/1600 (18.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 586 rows | mean depth 22 decisions | mean travel 178 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 53% 45% 26%>2% 23% 4 389 93%/5%/0%/3% 6% 86/1/0/0 0175A_5mm_v1 448 55% 45% 26%>0% 26% 0 136 91%/6%/0%/3% 4% 126/2/22/74 1447_5mm_v1 416 57% 43% 27%>0% 27% 0 142 89%/7%/0%/4% 3% 175/1/0/5 paris4_afv_central 288 54% 46% 24%>0% 24% 0 137 87%/8%/1%/5% 5% 100/4/0/32 PHerc0125_3mm_v2 256 57% 43% 28%>0% 28% 0 141 88%/8%/0%/4% 13% 99/3/49/10 gradients: max 45, clipped 0/50 updates Step 6,500 | last 50 updates / 1,600 crops | lr 2.67e-05 points (35/curve, tolerance 2 vox): 48,357 right / 2,112 wrong | 95.8% correct | 5,531 unknown loss 0.8975 | geometry 0.2680 | confidence 1.2590 | mean error 0.674 vox speed: 1342 ms/update | data wait 1 ms/update | 23.6 crops/s | 23.6 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1597 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 313, from the previous prediction 1284 confidence-labeled crops: terminal 1/1523 (0.1%) | recoverable 103/1523 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 160/1600 (10.0%), 1: 224/1600 (14.0%), 2: 352/1600 (22.0%), 3: 352/1600 (22.0%), 4: 512/1600 (32.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 582 rows | mean depth 18 decisions | mean travel 143 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 160 52% 46% 25%>2% 23% 3 392 89%/8%/0%/3% 12% 73/2/64/0 0175A_5mm_v1 224 57% 43% 29%>0% 29% 0 140 89%/8%/0%/3% 4% 83/2/0/14 1447_5mm_v1 352 57% 43% 23%>0% 23% 0 144 90%/6%/0%/4% 2% 152/1/0/0 paris4_afv_central 352 58% 42% 28%>0% 28% 0 145 87%/9%/0%/4% 7% 147/1/0/0 PHerc0125_3mm_v2 512 55% 45% 23%>0% 23% 0 132 88%/4%/0%/7% 12% 127/2/62/101 gradients: max 73, clipped 1/50 updates Step 6,550 | last 50 updates / 1,600 crops | lr 2.66e-05 points (35/curve, tolerance 2 vox): 49,690 right / 2,198 wrong | 95.8% correct | 4,112 unknown loss 0.9634 | geometry 0.2904 | confidence 1.3460 | mean error 0.671 vox speed: 1396 ms/update | data wait 1 ms/update | 22.8 crops/s | 22.8 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1595 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 265, from the previous prediction 1330 confidence-labeled crops: terminal 4/1547 (0.3%) | recoverable 114/1547 (7.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 480/1600 (30.0%), 1: 192/1600 (12.0%), 2: 352/1600 (22.0%), 3: 192/1600 (12.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 642 rows | mean depth 19 decisions | mean travel 156 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 480 56% 42% 26%>1% 25% 5 288 91%/8%/0%/1% 8% 180/2/68/24 0175A_5mm_v1 192 55% 45% 27%>0% 27% 0 160 90%/8%/0%/3% 4% 86/1/0/0 1447_5mm_v1 352 53% 47% 25%>0% 25% 0 150 91%/5%/0%/4% 3% 167/2/0/0 paris4_afv_central 192 54% 46% 22%>0% 22% 0 148 90%/8%/1%/1% 5% 88/2/0/0 PHerc0125_3mm_v2 384 54% 46% 22%>0% 22% 0 133 85%/8%/0%/7% 12% 121/3/0/57 gradients: max 52, clipped 0/50 updates Step 6,600 | last 50 updates / 1,600 crops | lr 2.66e-05 points (35/curve, tolerance 2 vox): 48,768 right / 2,069 wrong | 95.9% correct | 5,163 unknown loss 0.9479 | geometry 0.2910 | confidence 1.3138 | mean error 0.688 vox speed: 1281 ms/update | data wait 1 ms/update | 24.8 crops/s | 24.8 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1595 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 352, from the previous prediction 1243 confidence-labeled crops: terminal 3/1528 (0.2%) | recoverable 117/1528 (7.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 416/1600 (26.0%), 2: 160/1600 (10.0%), 3: 160/1600 (10.0%), 4: 480/1600 (30.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 50 chunk references / 1 windows live chains: 545 rows | mean depth 21 decisions | mean travel 173 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 55% 43% 27%>1% 25% 5 465 86%/10%/0%/4% 8% 77/0/0/90 0175A_5mm_v1 416 57% 43% 26%>0% 26% 0 152 90%/7%/0%/3% 4% 180/2/0/0 1447_5mm_v1 160 49% 51% 27%>0% 27% 0 160 87%/9%/0%/4% 3% 82/1/0/0 paris4_afv_central 160 59% 41% 25%>0% 25% 0 182 91%/6%/0%/4% 7% 65/0/0/0 PHerc0125_3mm_v2 480 54% 46% 23%>0% 23% 0 133 87%/5%/0%/8% 11% 141/4/0/82 gradients: max 56, clipped 0/50 updates Step 6,650 | last 50 updates / 1,600 crops | lr 2.65e-05 points (35/curve, tolerance 2 vox): 49,866 right / 2,034 wrong | 96.1% correct | 4,100 unknown loss 0.9087 | geometry 0.2458 | confidence 1.3258 | mean error 0.651 vox speed: 1253 ms/update | data wait 1 ms/update | 25.3 crops/s | 25.3 decisions/s | 1.14 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1597 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 299, from the previous prediction 1298 confidence-labeled crops: terminal 5/1547 (0.3%) | recoverable 106/1547 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 384/1600 (24.0%), 2: 192/1600 (12.0%), 3: 288/1600 (18.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 588 rows | mean depth 16 decisions | mean travel 134 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 55% 45% 24%>1% 24% 3 433 90%/7%/0%/3% 8% 52/0/0/105 0175A_5mm_v1 384 55% 45% 27%>0% 27% 0 152 89%/8%/0%/3% 3% 156/2/12/18 1447_5mm_v1 192 57% 43% 24%>0% 24% 0 174 91%/6%/0%/3% 3% 82/0/3/0 paris4_afv_central 288 55% 45% 28%>0% 28% 0 178 90%/8%/0%/3% 5% 131/2/1/0 PHerc0125_3mm_v2 384 57% 43% 24%>0% 24% 0 139 89%/4%/1%/5% 12% 167/1/0/0 gradients: max 67, clipped 1/50 updates Step 6,700 | last 50 updates / 1,600 crops | lr 2.65e-05 points (35/curve, tolerance 2 vox): 49,652 right / 1,694 wrong | 96.7% correct | 4,654 unknown loss 0.7903 | geometry 0.2348 | confidence 1.1110 | mean error 0.627 vox speed: 1379 ms/update | data wait 1 ms/update | 23.0 crops/s | 23.0 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1592 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 392, from the previous prediction 1200 confidence-labeled crops: terminal 3/1541 (0.2%) | recoverable 108/1541 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 224/1600 (14.0%), 2: 320/1600 (20.0%), 3: 416/1600 (26.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 476 rows | mean depth 20 decisions | mean travel 164 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 53% 44% 28%>3% 25% 7 520 93%/6%/0%/1% 8% 113/2/0/0 0175A_5mm_v1 224 57% 43% 27%>0% 27% 0 134 92%/5%/0%/3% 4% 72/1/0/24 1447_5mm_v1 320 62% 38% 28%>0% 28% 0 -- 88%/8%/0%/4% 3% 0/0/119/121 paris4_afv_central 416 55% 45% 24%>0% 24% 0 160 89%/6%/0%/4% 5% 143/4/75/46 PHerc0125_3mm_v2 384 52% 48% 24%>0% 24% 0 140 87%/7%/0%/5% 11% 148/2/0/36 gradients: max 62, clipped 1/50 updates Step 6,750 | last 50 updates / 1,600 crops | lr 2.64e-05 points (35/curve, tolerance 2 vox): 49,682 right / 2,192 wrong | 95.8% correct | 4,126 unknown loss 0.9441 | geometry 0.2830 | confidence 1.3223 | mean error 0.664 vox speed: 1375 ms/update | data wait 1 ms/update | 23.1 crops/s | 23.1 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1594 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 301, from the previous prediction 1293 confidence-labeled crops: terminal 3/1555 (0.2%) | recoverable 95/1555 (6.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 224/1600 (14.0%), 2: 352/1600 (22.0%), 3: 448/1600 (28.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 604 rows | mean depth 24 decisions | mean travel 191 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 54% 44% 26%>2% 23% 6 452 93%/5%/0%/1% 10% 113/0/0/0 0175A_5mm_v1 224 51% 49% 23%>0% 23% 0 155 90%/6%/0%/4% 4% 109/3/0/0 1447_5mm_v1 352 53% 47% 24%>0% 24% 0 149 89%/7%/0%/4% 3% 127/1/39/38 paris4_afv_central 448 58% 42% 28%>0% 28% 0 133 91%/6%/0%/2% 6% 104/1/0/85 PHerc0125_3mm_v2 320 53% 47% 20%>0% 20% 0 140 92%/4%/1%/3% 14% 151/7/0/0 gradients: max 67, clipped 1/50 updates Step 6,800 | last 50 updates / 1,600 crops | lr 2.64e-05 points (35/curve, tolerance 2 vox): 48,482 right / 2,208 wrong | 95.6% correct | 5,310 unknown loss 0.9548 | geometry 0.2698 | confidence 1.3700 | mean error 0.672 vox speed: 1266 ms/update | data wait 1 ms/update | 25.1 crops/s | 25.1 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1599 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 226, from the previous prediction 1373 confidence-labeled crops: terminal 3/1527 (0.2%) | recoverable 105/1527 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 384/1600 (24.0%), 2: 160/1600 (10.0%), 3: 224/1600 (14.0%), 4: 480/1600 (30.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 653 rows | mean depth 14 decisions | mean travel 116 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 53% 47% 22%>0% 22% 1 187 90%/8%/0%/1% 8% 151/1/84/13 0175A_5mm_v1 384 60% 40% 26%>0% 26% 0 169 89%/5%/0%/6% 3% 154/4/0/0 1447_5mm_v1 160 59% 41% 32%>0% 32% 0 145 88%/6%/0%/6% 2% 50/0/0/16 paris4_afv_central 224 55% 45% 25%>0% 25% 0 146 87%/9%/0%/4% 8% 101/3/0/0 PHerc0125_3mm_v2 480 56% 44% 24%>0% 24% 0 133 88%/5%/1%/6% 12% 197/5/50/13 gradients: max 50, clipped 0/50 updates Step 6,850 | last 50 updates / 1,600 crops | lr 2.63e-05 points (35/curve, tolerance 2 vox): 49,244 right / 2,058 wrong | 96.0% correct | 4,698 unknown loss 0.9740 | geometry 0.2732 | confidence 1.4017 | mean error 0.677 vox speed: 1232 ms/update | data wait 1 ms/update | 25.8 crops/s | 25.8 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1591 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 295, from the previous prediction 1296 confidence-labeled crops: terminal 1/1550 (0.1%) | recoverable 106/1550 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 160/1600 (10.0%), 2: 288/1600 (18.0%), 3: 416/1600 (26.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 612 rows | mean depth 15 decisions | mean travel 125 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 53% 45% 25%>2% 23% 9 620 92%/6%/0%/2% 10% 96/0/0/75 0175A_5mm_v1 160 53% 47% 26%>0% 26% 0 172 91%/7%/0%/2% 3% 75/2/0/0 1447_5mm_v1 288 55% 45% 27%>0% 27% 0 139 92%/6%/0%/2% 3% 130/3/0/0 paris4_afv_central 416 52% 48% 23%>0% 23% 0 141 89%/7%/0%/3% 6% 201/5/0/0 PHerc0125_3mm_v2 352 55% 45% 25%>0% 25% 0 144 88%/7%/0%/5% 12% 110/1/0/49 gradients: max 59, clipped 0/50 updates Step 6,900 | last 50 updates / 1,600 crops | lr 2.63e-05 points (35/curve, tolerance 2 vox): 48,120 right / 2,648 wrong | 94.8% correct | 5,232 unknown loss 0.9920 | geometry 0.3122 | confidence 1.3597 | mean error 0.702 vox speed: 1257 ms/update | data wait 1 ms/update | 25.2 crops/s | 25.2 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1586 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 215, from the previous prediction 1371 confidence-labeled crops: terminal 7/1534 (0.5%) | recoverable 134/1534 (8.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 256/1600 (16.0%), 2: 256/1600 (16.0%), 3: 320/1600 (20.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 675 rows | mean depth 22 decisions | mean travel 175 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 53% 43% 29%>4% 25% 9 672 88%/9%/1%/2% 8% 137/0/0/0 0175A_5mm_v1 256 56% 44% 28%>0% 28% 0 160 88%/7%/0%/5% 6% 112/0/0/0 1447_5mm_v1 256 58% 42% 26%>0% 26% 0 153 85%/10%/0%/5% 6% 99/1/0/8 paris4_afv_central 320 56% 44% 26%>0% 26% 0 141 90%/7%/0%/3% 7% 141/1/0/0 PHerc0125_3mm_v2 448 56% 44% 25%>0% 25% 0 140 85%/9%/1%/6% 14% 186/7/0/9 gradients: max 76, clipped 2/50 updates Step 6,950 | last 50 updates / 1,600 crops | lr 2.62e-05 points (35/curve, tolerance 2 vox): 50,109 right / 1,843 wrong | 96.5% correct | 4,048 unknown loss 0.8474 | geometry 0.2362 | confidence 1.2225 | mean error 0.626 vox speed: 1281 ms/update | data wait 1 ms/update | 24.8 crops/s | 24.8 decisions/s | 1.11 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1588 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 210, from the previous prediction 1378 confidence-labeled crops: terminal 6/1546 (0.4%) | recoverable 114/1546 (7.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 512/1600 (32.0%), 2: 224/1600 (14.0%), 3: 192/1600 (12.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 664 rows | mean depth 14 decisions | mean travel 114 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 54% 45% 28%>1% 27% 4 312 89%/9%/0%/2% 6% 129/1/0/0 0175A_5mm_v1 512 60% 39% 28%>2% 26% 8 331 89%/8%/0%/3% 4% 199/2/0/0 1447_5mm_v1 224 55% 45% 26%>0% 26% 0 156 89%/7%/0%/4% 3% 100/1/0/0 paris4_afv_central 192 56% 44% 27%>0% 27% 0 149 88%/9%/0%/3% 4% 58/1/0/26 PHerc0125_3mm_v2 384 54% 46% 22%>0% 22% 0 145 90%/4%/1%/5% 11% 178/8/0/0 gradients: max 67, clipped 1/50 updates Step 7,000 | last 50 updates / 1,600 crops | lr 2.61e-05 points (35/curve, tolerance 2 vox): 48,707 right / 2,272 wrong | 95.5% correct | 5,021 unknown loss 0.8670 | geometry 0.2431 | confidence 1.2478 | mean error 0.630 vox speed: 1242 ms/update | data wait 1 ms/update | 25.6 crops/s | 25.6 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1589 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 299, from the previous prediction 1290 confidence-labeled crops: terminal 13/1531 (0.8%) | recoverable 100/1531 (6.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 352/1600 (22.0%), 2: 384/1600 (24.0%), 3: 224/1600 (14.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 612 rows | mean depth 17 decisions | mean travel 134 vox | 9 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 52% 46% 27%>2% 25% 5 381 88%/9%/0%/3% 8% 117/0/0/0 0175A_5mm_v1 352 57% 41% 26%>2% 24% 6 449 86%/7%/1%/5% 5% 87/0/40/59 1447_5mm_v1 384 52% 48% 21%>0% 21% 0 140 93%/4%/1%/2% 4% 145/4/26/40 paris4_afv_central 224 52% 48% 27%>0% 27% 0 160 86%/9%/1%/4% 7% 107/4/0/0 PHerc0125_3mm_v2 384 51% 49% 22%>0% 22% 0 139 87%/4%/1%/8% 11% 156/2/0/33 gradients: max 61, clipped 1/50 updates Step 7,000 | dagger_skipped_busy: true Step 7,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 243, "stopped": 1, "mean_example_error": 0.6308608015378316, "geometry_examples": 30, "examples": 32, "inference_and_measurement_seconds": 2.8335970044136047, "total_diagnostic_seconds": 4.678591758012772} Step 7,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 6/8 (75.0%) | false stops 1 1.5-3: recovered 6/8 (75.0%) | false stops 0 3-4.5: recovered 6/8 (75.0%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 0 entire proposed 8-point prefix correct (distance): 21/28 (75.0%) gate @ 0.50: false stops 0/22 (0.0%) | accepted wrong 5/26 (19.2%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.508 1.451 following 16 464 1.049 1.039 recoverable 12 348 2.120 2.002 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 1.049 1.039 332 32 812 1.508 1.451 improved/worsened/compared states: 2/1/28 Step 7,000 | batch_diagnostics_seconds: 4.681775543838739 | recovery_seconds: 16.3581470772624 Step 7,001 | dagger_states: 9579 | dagger_source_step: 6000 | dagger_caches: 2 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/decisions_006000.npz" | dagger_fibers: 128 | dagger_supply: {"states": 9579, "traces": 128, "fibers": 128, "rows": {"terminal": 49, "premature_stop": 65, "recoverable": 383, "pre_excursion": 470, "ordinary": 8599}, "events": {"terminal": 10, "premature_stop": 65, "recoverable": 89, "pre_excursion": 92, "ordinary": 123}, "supervision": {"following": 9128, "recoverable": 370, "terminal": 49, "unknown": 32}, "departed_traces": 46, "terminal_traces": 10, "stop_reasons": {"after_failure": 8, "confidence": 68, "max_len": 4, "unannotated": 48}, "traced_voxels": 164583.99359560476, "seconds": 1342.2911552712321, "voxels_per_second": 122.61422788139272} | dagger_coverage: {"requested": 128, "seeds": 128, "skipped": [], "epoch": 0} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 1 | dagger_launch_interval: 5000 | dagger_publication_age: 1001 | dataset: "paris4" Step 7,050 | last 50 updates / 1,600 crops | lr 2.61e-05 points (35/curve, tolerance 2 vox): 49,257 right / 2,052 wrong | 96.0% correct | 4,691 unknown loss 0.8519 | geometry 0.2498 | confidence 1.2044 | mean error 0.645 vox speed: 892 ms/update | data wait 1 ms/update | 23.8 crops/s | 23.8 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1592 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 277, from the previous prediction 1315 confidence-labeled crops: terminal 5/1543 (0.3%) | recoverable 116/1543 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 352/1600 (22.0%), 2: 352/1600 (22.0%), 3: 320/1600 (20.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 603 rows | mean depth 23 decisions | mean travel 184 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 57% 42% 24%>1% 23% 2 230 90%/8%/0%/2% 9% 132/0/32/2 0175A_5mm_v1 352 53% 46% 26%>2% 24% 6 362 89%/6%/1%/4% 5% 123/5/0/38 1447_5mm_v1 352 55% 45% 26%>0% 26% 0 134 88%/9%/0%/3% 4% 90/1/0/70 paris4_afv_central 320 58% 42% 27%>0% 27% 0 182 89%/6%/0%/5% 5% 133/3/0/0 PHerc0125_3mm_v2 256 51% 49% 21%>0% 21% 0 146 87%/7%/1%/6% 13% 125/3/0/0 gradients: max 48, clipped 0/50 updates Step 7,100 | last 50 updates / 1,600 crops | lr 2.60e-05 points (35/curve, tolerance 2 vox): 50,258 right / 1,647 wrong | 96.8% correct | 4,095 unknown loss 0.8023 | geometry 0.2315 | confidence 1.1416 | mean error 0.628 vox speed: 891 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1599 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 318, from the previous prediction 1281 confidence-labeled crops: terminal 1/1548 (0.1%) | recoverable 104/1548 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 480/1600 (30.0%), 1: 224/1600 (14.0%), 2: 416/1600 (26.0%), 3: 288/1600 (18.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 575 rows | mean depth 15 decisions | mean travel 125 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 480 59% 41% 24%>0% 23% 1 173 93%/5%/0%/2% 8% 154/1/21/44 0175A_5mm_v1 224 54% 46% 30%>0% 30% 0 155 91%/7%/0%/2% 2% 102/1/0/0 1447_5mm_v1 416 56% 44% 25%>0% 25% 0 134 88%/7%/0%/5% 2% 88/3/0/93 paris4_afv_central 288 52% 48% 27%>0% 27% 0 152 89%/9%/0%/2% 6% 137/3/0/0 PHerc0125_3mm_v2 192 51% 49% 23%>0% 23% 0 154 90%/4%/1%/5% 9% 94/4/0/0 gradients: max 50, clipped 0/50 updates Step 7,150 | last 50 updates / 1,600 crops | lr 2.60e-05 points (35/curve, tolerance 2 vox): 49,522 right / 2,386 wrong | 95.4% correct | 4,092 unknown loss 0.9158 | geometry 0.2894 | confidence 1.2529 | mean error 0.693 vox speed: 888 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1591 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 303, from the previous prediction 1288 confidence-labeled crops: terminal 6/1551 (0.4%) | recoverable 104/1551 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 352/1600 (22.0%), 2: 128/1600 (8.0%), 3: 288/1600 (18.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 580 rows | mean depth 20 decisions | mean travel 165 vox | 6 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 52% 46% 23%>2% 21% 9 738 92%/6%/0%/2% 7% 80/0/0/111 0175A_5mm_v1 352 56% 44% 23%>0% 23% 0 161 90%/7%/0%/3% 4% 156/1/0/0 1447_5mm_v1 128 62% 38% 25%>0% 25% 0 147 93%/4%/0%/3% 4% 49/3/0/0 paris4_afv_central 288 56% 44% 25%>0% 25% 0 138 90%/7%/1%/2% 6% 101/0/1/27 PHerc0125_3mm_v2 416 53% 47% 25%>0% 25% 0 156 88%/7%/1%/5% 12% 194/7/0/0 gradients: max 83, clipped 1/50 updates Step 7,200 | last 50 updates / 1,600 crops | lr 2.59e-05 points (35/curve, tolerance 2 vox): 48,968 right / 2,355 wrong | 95.4% correct | 4,677 unknown loss 0.9535 | geometry 0.2811 | confidence 1.3446 | mean error 0.675 vox speed: 888 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1569 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 188, from the previous prediction 1381 confidence-labeled crops: terminal 8/1535 (0.5%) | recoverable 136/1535 (8.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 320/1600 (20.0%), 2: 288/1600 (18.0%), 3: 320/1600 (20.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 675 rows | mean depth 18 decisions | mean travel 146 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 53% 44% 28%>2% 26% 9 491 86%/8%/0%/5% 10% 146/2/0/24 0175A_5mm_v1 320 49% 44% 32%>7% 26% 16 854 82%/13%/1%/4% 8% 135/3/48/7 1447_5mm_v1 288 53% 47% 24%>0% 24% 0 176 89%/8%/1%/2% 3% 134/3/0/0 paris4_afv_central 320 56% 44% 24%>0% 24% 0 137 89%/6%/0%/5% 5% 125/1/0/17 PHerc0125_3mm_v2 288 53% 47% 21%>0% 21% 0 149 88%/6%/0%/5% 10% 135/6/12/0 gradients: max 67, clipped 2/50 updates Step 7,250 | last 50 updates / 1,600 crops | lr 2.59e-05 points (35/curve, tolerance 2 vox): 50,090 right / 2,049 wrong | 96.1% correct | 3,861 unknown loss 0.8645 | geometry 0.2490 | confidence 1.2309 | mean error 0.646 vox speed: 889 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1588 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 228, from the previous prediction 1360 confidence-labeled crops: terminal 3/1552 (0.2%) | recoverable 125/1552 (8.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 288/1600 (18.0%), 2: 352/1600 (22.0%), 3: 448/1600 (28.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 647 rows | mean depth 17 decisions | mean travel 137 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 56% 43% 24%>1% 24% 2 239 91%/7%/0%/1% 9% 124/0/0/0 0175A_5mm_v1 288 60% 40% 27%>0% 27% 0 142 89%/8%/0%/3% 4% 116/1/0/0 1447_5mm_v1 352 54% 43% 32%>3% 29% 9 418 86%/9%/1%/4% 4% 143/3/100/8 paris4_afv_central 448 55% 45% 25%>0% 25% 0 135 89%/7%/0%/3% 6% 141/1/0/62 PHerc0125_3mm_v2 224 45% 55% 24%>0% 24% 0 144 89%/7%/0%/4% 14% 123/3/0/0 gradients: max 51, clipped 0/50 updates Step 7,300 | last 50 updates / 1,600 crops | lr 2.58e-05 points (35/curve, tolerance 2 vox): 49,365 right / 1,787 wrong | 96.5% correct | 4,848 unknown loss 0.7665 | geometry 0.2259 | confidence 1.0812 | mean error 0.630 vox speed: 888 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.14 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1589 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 328, from the previous prediction 1261 confidence-labeled crops: terminal 1/1533 (0.1%) | recoverable 97/1533 (6.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 320/1600 (20.0%), 2: 352/1600 (22.0%), 3: 288/1600 (18.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 560 rows | mean depth 19 decisions | mean travel 158 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 51% 46% 27%>3% 25% 11 486 88%/8%/0%/3% 7% 189/1/36/2 0175A_5mm_v1 320 55% 45% 26%>0% 26% 0 137 94%/3%/0%/3% 4% 131/1/0/14 1447_5mm_v1 352 55% 45% 25%>0% 25% 0 134 89%/5%/0%/6% 4% 35/3/0/124 paris4_afv_central 288 61% 39% 27%>0% 27% 0 137 88%/7%/0%/4% 5% 103/1/0/10 PHerc0125_3mm_v2 224 54% 46% 23%>0% 23% 0 149 88%/5%/0%/7% 10% 102/1/0/0 gradients: max 54, clipped 0/50 updates Step 7,350 | last 50 updates / 1,600 crops | lr 2.58e-05 points (35/curve, tolerance 2 vox): 50,137 right / 1,767 wrong | 96.6% correct | 4,096 unknown loss 0.7402 | geometry 0.2096 | confidence 1.0612 | mean error 0.602 vox speed: 887 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.12 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1594 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 236, from the previous prediction 1358 confidence-labeled crops: terminal 5/1548 (0.3%) | recoverable 79/1548 (5.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 448/1600 (28.0%), 2: 384/1600 (24.0%), 3: 224/1600 (14.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 693 rows | mean depth 17 decisions | mean travel 141 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 53% 45% 24%>2% 22% 6 380 92%/5%/0%/2% 8% 159/0/0/0 0175A_5mm_v1 448 53% 47% 25%>0% 25% 0 143 92%/5%/0%/4% 4% 184/2/0/25 1447_5mm_v1 384 52% 48% 21%>0% 21% 0 141 93%/2%/0%/4% 2% 161/4/0/23 paris4_afv_central 224 53% 47% 25%>0% 25% 0 141 88%/7%/1%/4% 4% 101/2/7/4 PHerc0125_3mm_v2 192 54% 46% 26%>0% 26% 0 180 90%/8%/1%/2% 8% 88/1/0/0 gradients: max 55, clipped 0/50 updates Step 7,400 | last 50 updates / 1,600 crops | lr 2.57e-05 points (35/curve, tolerance 2 vox): 49,583 right / 2,381 wrong | 95.4% correct | 4,036 unknown loss 0.8842 | geometry 0.2460 | confidence 1.2763 | mean error 0.647 vox speed: 885 ms/update | data wait 1 ms/update | 35.8 crops/s | 35.8 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1581 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 159, from the previous prediction 1422 confidence-labeled crops: terminal 11/1549 (0.7%) | recoverable 118/1549 (7.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 224/1600 (14.0%), 2: 352/1600 (22.0%), 3: 288/1600 (18.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 743 rows | mean depth 16 decisions | mean travel 129 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 52% 42% 29%>6% 24% 10 875 87%/8%/2%/2% 11% 120/1/11/1 0175A_5mm_v1 224 51% 47% 22%>1% 21% 3 291 90%/8%/0%/2% 4% 101/0/0/5 1447_5mm_v1 352 54% 46% 24%>0% 24% 0 143 91%/6%/0%/3% 1% 163/2/0/0 paris4_afv_central 288 48% 52% 22%>0% 22% 0 146 90%/8%/1%/1% 6% 150/2/0/0 PHerc0125_3mm_v2 448 53% 47% 23%>0% 23% 0 155 86%/7%/0%/6% 11% 209/4/0/0 gradients: max 62, clipped 1/50 updates Step 7,450 | last 50 updates / 1,600 crops | lr 2.57e-05 points (35/curve, tolerance 2 vox): 49,526 right / 1,909 wrong | 96.3% correct | 4,565 unknown loss 0.7787 | geometry 0.2384 | confidence 1.0806 | mean error 0.643 vox speed: 891 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.12 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1583 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 276, from the previous prediction 1307 confidence-labeled crops: terminal 3/1541 (0.2%) | recoverable 109/1541 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 288/1600 (18.0%), 2: 384/1600 (24.0%), 3: 224/1600 (14.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 627 rows | mean depth 18 decisions | mean travel 145 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 55% 43% 30%>2% 28% 7 418 88%/9%/0%/3% 7% 178/1/0/0 0175A_5mm_v1 288 55% 42% 26%>3% 23% 9 511 91%/6%/0%/2% 5% 121/1/0/0 1447_5mm_v1 384 54% 46% 26%>0% 26% 0 136 86%/7%/0%/7% 4% 124/2/0/52 paris4_afv_central 224 54% 46% 21%>0% 21% 0 155 92%/7%/0%/1% 5% 102/1/3/0 PHerc0125_3mm_v2 288 53% 47% 21%>0% 21% 0 136 90%/5%/0%/5% 10% 102/1/8/33 gradients: max 71, clipped 1/50 updates Step 7,500 | last 50 updates / 1,600 crops | lr 2.56e-05 points (35/curve, tolerance 2 vox): 50,229 right / 1,639 wrong | 96.8% correct | 4,132 unknown loss 0.7715 | geometry 0.2209 | confidence 1.1011 | mean error 0.600 vox speed: 889 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.13 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1591 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 274, from the previous prediction 1317 confidence-labeled crops: terminal 3/1546 (0.2%) | recoverable 109/1546 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 160/1600 (10.0%), 1: 640/1600 (40.0%), 2: 288/1600 (18.0%), 3: 256/1600 (16.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 39 chunks / 78.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 657 rows | mean depth 17 decisions | mean travel 142 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 160 48% 51% 20%>2% 18% 3 385 93%/6%/0%/1% 6% 81/3/0/0 0175A_5mm_v1 640 53% 46% 26%>1% 25% 4 272 88%/9%/0%/3% 5% 235/2/13/58 1447_5mm_v1 288 52% 48% 23%>0% 23% 0 146 89%/5%/0%/6% 2% 137/6/0/0 paris4_afv_central 256 54% 46% 21%>0% 21% 0 157 89%/7%/1%/3% 4% 117/5/0/0 PHerc0125_3mm_v2 256 53% 47% 21%>0% 21% 0 140 91%/5%/0%/4% 14% 87/0/0/34 gradients: max 49, clipped 0/50 updates Step 7,550 | last 50 updates / 1,600 crops | lr 2.56e-05 points (35/curve, tolerance 2 vox): 49,291 right / 2,139 wrong | 95.8% correct | 4,570 unknown loss 0.8994 | geometry 0.2671 | confidence 1.2646 | mean error 0.635 vox speed: 887 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1566 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 234, from the previous prediction 1332 confidence-labeled crops: terminal 7/1540 (0.5%) | recoverable 124/1540 (8.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 288/1600 (18.0%), 2: 384/1600 (24.0%), 3: 384/1600 (24.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 294 chunk references / 1 windows live chains: 657 rows | mean depth 18 decisions | mean travel 144 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 55% 42% 28%>2% 25% 8 520 88%/10%/0%/2% 7% 136/0/0/0 0175A_5mm_v1 288 44% 49% 20%>7% 13% 21 1411 85%/8%/1%/5% 12% 84/5/0/57 1447_5mm_v1 384 59% 41% 27%>0% 27% 0 139 90%/7%/0%/3% 2% 157/3/0/0 paris4_afv_central 384 54% 46% 25%>0% 25% 0 141 89%/7%/1%/3% 8% 175/1/0/0 PHerc0125_3mm_v2 224 53% 47% 26%>0% 26% 0 160 88%/6%/0%/6% 10% 105/3/0/0 gradients: max 55, clipped 0/50 updates Step 7,600 | last 50 updates / 1,600 crops | lr 2.55e-05 points (35/curve, tolerance 2 vox): 48,863 right / 2,217 wrong | 95.7% correct | 4,920 unknown loss 0.8809 | geometry 0.2777 | confidence 1.2064 | mean error 0.683 vox speed: 888 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1564 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 260, from the previous prediction 1304 confidence-labeled crops: terminal 4/1531 (0.3%) | recoverable 98/1531 (6.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 384/1600 (24.0%), 2: 224/1600 (14.0%), 3: 288/1600 (18.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 644 rows | mean depth 18 decisions | mean travel 149 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 51% 47% 25%>2% 23% 11 471 89%/7%/0%/3% 8% 205/0/3/4 0175A_5mm_v1 384 48% 46% 23%>6% 17% 25 1003 91%/5%/1%/3% 7% 132/6/0/44 1447_5mm_v1 224 51% 49% 20%>0% 20% 0 154 92%/5%/0%/3% 2% 110/3/0/0 paris4_afv_central 288 55% 45% 26%>0% 26% 0 137 85%/8%/0%/7% 7% 86/1/21/45 PHerc0125_3mm_v2 256 57% 43% 22%>0% 22% 0 179 88%/5%/0%/7% 11% 111/1/31/0 gradients: max 53, clipped 0/50 updates Step 7,650 | last 50 updates / 1,600 crops | lr 2.54e-05 points (35/curve, tolerance 2 vox): 48,503 right / 2,486 wrong | 95.1% correct | 5,011 unknown loss 0.9102 | geometry 0.2713 | confidence 1.2777 | mean error 0.668 vox speed: 887 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1569 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 263, from the previous prediction 1306 confidence-labeled crops: terminal 11/1540 (0.7%) | recoverable 105/1540 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 320/1600 (20.0%), 2: 320/1600 (20.0%), 3: 320/1600 (20.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 615 rows | mean depth 19 decisions | mean travel 153 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 53% 46% 24%>0% 24% 1 219 93%/5%/0%/2% 9% 72/1/0/76 0175A_5mm_v1 320 46% 45% 25%>9% 16% 28 1106 88%/7%/2%/4% 10% 143/3/0/0 1447_5mm_v1 320 56% 44% 24%>0% 24% 0 151 91%/7%/0%/2% 3% 141/2/0/0 paris4_afv_central 320 54% 46% 26%>0% 26% 0 139 86%/8%/1%/5% 6% 122/0/0/26 PHerc0125_3mm_v2 320 57% 43% 27%>0% 28% 0 140 87%/6%/1%/7% 12% 137/2/31/0 gradients: max 49, clipped 0/50 updates Step 7,700 | last 50 updates / 1,600 crops | lr 2.54e-05 points (35/curve, tolerance 2 vox): 49,684 right / 2,044 wrong | 96.0% correct | 4,272 unknown loss 0.8371 | geometry 0.2698 | confidence 1.1347 | mean error 0.665 vox speed: 889 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1581 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 207, from the previous prediction 1374 confidence-labeled crops: terminal 6/1543 (0.4%) | recoverable 112/1543 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 320/1600 (20.0%), 2: 320/1600 (20.0%), 3: 384/1600 (24.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 669 rows | mean depth 19 decisions | mean travel 150 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 55% 43% 26%>2% 24% 7 415 90%/9%/0%/2% 8% 166/0/0/0 0175A_5mm_v1 320 51% 45% 28%>4% 25% 11 573 88%/8%/1%/3% 6% 145/5/0/0 1447_5mm_v1 320 58% 42% 29%>0% 29% 0 150 91%/7%/0%/2% 2% 135/3/0/0 paris4_afv_central 384 59% 41% 29%>0% 29% 0 150 89%/5%/1%/5% 6% 158/4/0/0 PHerc0125_3mm_v2 192 52% 48% 23%>0% 23% 0 138 86%/6%/0%/7% 12% 65/0/11/28 gradients: max 58, clipped 0/50 updates Step 7,750 | last 50 updates / 1,600 crops | lr 2.53e-05 points (35/curve, tolerance 2 vox): 48,017 right / 2,586 wrong | 94.9% correct | 5,397 unknown loss 1.0089 | geometry 0.3070 | confidence 1.4038 | mean error 0.719 vox speed: 888 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.26 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1586 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 355, from the previous prediction 1231 confidence-labeled crops: terminal 8/1529 (0.5%) | recoverable 102/1529 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 160/1600 (10.0%), 2: 192/1600 (12.0%), 3: 448/1600 (28.0%), 4: 480/1600 (30.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 1193 chunk references / 1 windows live chains: 575 rows | mean depth 16 decisions | mean travel 130 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 50% 48% 26%>2% 24% 6 392 89%/8%/0%/2% 8% 155/1/0/0 0175A_5mm_v1 160 53% 42% 28%>5% 23% 6 759 89%/4%/2%/5% 7% 67/1/0/0 1447_5mm_v1 192 51% 49% 22%>0% 22% 0 163 92%/3%/0%/5% 1% 94/2/0/0 paris4_afv_central 448 51% 49% 21%>0% 21% 0 138 89%/6%/0%/4% 7% 128/1/0/92 PHerc0125_3mm_v2 480 58% 42% 29%>0% 29% 0 135 86%/7%/0%/6% 14% 131/1/0/71 gradients: max 58, clipped 0/50 updates Step 7,800 | last 50 updates / 1,600 crops | lr 2.53e-05 points (35/curve, tolerance 2 vox): 48,663 right / 2,152 wrong | 95.8% correct | 5,185 unknown loss 0.8273 | geometry 0.2479 | confidence 1.1589 | mean error 0.631 vox speed: 892 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1580 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 234, from the previous prediction 1346 confidence-labeled crops: terminal 8/1524 (0.5%) | recoverable 114/1524 (7.5%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 256/1600 (16.0%), 2: 416/1600 (26.0%), 3: 416/1600 (26.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 672 rows | mean depth 20 decisions | mean travel 162 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 51% 47% 24%>1% 22% 3 337 90%/8%/0%/2% 8% 106/0/4/0 0175A_5mm_v1 256 53% 43% 27%>5% 23% 11 721 86%/8%/2%/4% 7% 109/1/0/0 1447_5mm_v1 416 54% 44% 24%>1% 23% 5 272 88%/7%/0%/5% 3% 185/1/0/0 paris4_afv_central 416 54% 46% 25%>0% 25% 0 140 86%/7%/0%/6% 6% 149/3/0/41 PHerc0125_3mm_v2 288 57% 43% 24%>0% 24% 0 141 88%/6%/0%/6% 13% 123/1/0/0 gradients: max 42, clipped 0/50 updates Step 7,850 | last 50 updates / 1,600 crops | lr 2.52e-05 points (35/curve, tolerance 2 vox): 49,503 right / 1,839 wrong | 96.4% correct | 4,658 unknown loss 0.8322 | geometry 0.2598 | confidence 1.1447 | mean error 0.645 vox speed: 885 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.14 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1558 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 213, from the previous prediction 1345 confidence-labeled crops: terminal 4/1543 (0.3%) | recoverable 111/1543 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 288/1600 (18.0%), 2: 288/1600 (18.0%), 3: 288/1600 (18.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 698 rows | mean depth 17 decisions | mean travel 138 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 54% 44% 25%>2% 23% 7 396 90%/8%/0%/1% 6% 198/1/26/1 0175A_5mm_v1 288 46% 48% 26%>7% 19% 16 869 90%/8%/1%/2% 8% 137/3/0/0 1447_5mm_v1 288 53% 41% 27%>5% 22% 13 671 89%/7%/1%/3% 5% 119/1/0/0 paris4_afv_central 288 52% 48% 23%>0% 23% 0 140 89%/6%/0%/5% 3% 121/4/0/18 PHerc0125_3mm_v2 288 53% 47% 26%>0% 26% 0 141 86%/5%/0%/8% 9% 123/5/0/11 gradients: max 74, clipped 1/50 updates Step 7,900 | last 50 updates / 1,600 crops | lr 2.52e-05 points (35/curve, tolerance 2 vox): 49,321 right / 1,981 wrong | 96.1% correct | 4,698 unknown loss 0.7960 | geometry 0.2531 | confidence 1.0858 | mean error 0.647 vox speed: 887 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1584 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 209, from the previous prediction 1375 confidence-labeled crops: terminal 4/1535 (0.3%) | recoverable 129/1535 (8.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 320/1600 (20.0%), 2: 224/1600 (14.0%), 3: 416/1600 (26.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 670 rows | mean depth 17 decisions | mean travel 141 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 58% 40% 29%>2% 28% 5 415 87%/11%/0%/3% 7% 140/2/20/14 0175A_5mm_v1 320 55% 45% 24%>0% 24% 0 147 86%/8%/0%/6% 4% 129/3/79/16 1447_5mm_v1 224 58% 38% 33%>4% 29% 10 655 86%/9%/1%/4% 5% 84/1/0/0 paris4_afv_central 416 55% 45% 20%>0% 20% 0 146 89%/6%/0%/6% 6% 189/6/0/0 PHerc0125_3mm_v2 256 50% 50% 23%>0% 23% 0 144 90%/7%/0%/2% 9% 128/1/0/0 gradients: max 65, clipped 1/50 updates Step 7,950 | last 50 updates / 1,600 crops | lr 2.51e-05 points (35/curve, tolerance 2 vox): 49,551 right / 2,160 wrong | 95.8% correct | 4,289 unknown loss 0.8586 | geometry 0.2509 | confidence 1.2155 | mean error 0.644 vox speed: 885 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1586 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 205, from the previous prediction 1381 confidence-labeled crops: terminal 8/1549 (0.5%) | recoverable 121/1549 (7.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 224/1600 (14.0%), 2: 352/1600 (22.0%), 3: 384/1600 (24.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 671 rows | mean depth 18 decisions | mean travel 148 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 55% 43% 26%>2% 24% 5 431 88%/8%/0%/4% 7% 110/3/0/29 0175A_5mm_v1 224 52% 46% 24%>3% 21% 6 481 89%/4%/1%/5% 7% 96/1/35/6 1447_5mm_v1 352 53% 46% 23%>1% 22% 3 238 89%/8%/1%/2% 3% 163/5/0/0 paris4_afv_central 384 53% 47% 25%>0% 25% 0 149 89%/9%/0%/2% 5% 181/5/14/0 PHerc0125_3mm_v2 320 58% 42% 26%>0% 26% 0 139 87%/8%/1%/4% 15% 121/1/8/13 gradients: max 51, clipped 0/50 updates Step 8,000 | last 50 updates / 1,600 crops | lr 2.50e-05 points (35/curve, tolerance 2 vox): 49,457 right / 1,718 wrong | 96.6% correct | 4,825 unknown loss 0.8213 | geometry 0.2316 | confidence 1.1794 | mean error 0.617 vox speed: 886 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1586 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 302, from the previous prediction 1284 confidence-labeled crops: terminal 3/1534 (0.2%) | recoverable 104/1534 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 384/1600 (24.0%), 2: 256/1600 (16.0%), 3: 192/1600 (12.0%), 4: 480/1600 (30.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 593 rows | mean depth 19 decisions | mean travel 156 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 55% 45% 24%>0% 24% 0 143 92%/6%/0%/2% 6% 129/0/0/0 0175A_5mm_v1 384 52% 46% 24%>2% 23% 6 369 90%/6%/1%/4% 3% 149/2/6/28 1447_5mm_v1 256 50% 46% 25%>3% 22% 7 463 89%/7%/0%/4% 4% 119/6/0/0 paris4_afv_central 192 55% 45% 20%>0% 20% 0 148 92%/5%/1%/3% 5% 86/2/0/0 PHerc0125_3mm_v2 480 59% 41% 26%>0% 26% 0 134 86%/7%/0%/7% 12% 110/4/0/89 gradients: max 66, clipped 1/50 updates Step 8,000 | dagger_launched: true Step 8,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 244, "stopped": 0, "mean_example_error": 0.4406377955206803, "geometry_examples": 28, "examples": 32, "inference_and_measurement_seconds": 2.0720834732055664, "total_diagnostic_seconds": 3.826139021664858} Step 8,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 6/8 (75.0%) | false stops 0 1.5-3: recovered 6/8 (75.0%) | false stops 0 3-4.5: recovered 6/8 (75.0%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 1 entire proposed 8-point prefix correct (distance): 21/28 (75.0%) gate @ 0.50: false stops 0/23 (0.0%) | accepted wrong 5/26 (19.2%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.542 1.490 following 16 464 1.095 1.104 recoverable 12 348 2.138 2.004 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 1.095 1.104 332 32 812 1.542 1.490 improved/worsened/compared states: 2/2/28 Step 8,000 | batch_diagnostics_seconds: 3.8289678767323494 | recovery_seconds: 15.909884851425886 Step 8,050 | last 50 updates / 1,600 crops | lr 2.50e-05 points (35/curve, tolerance 2 vox): 49,363 right / 2,029 wrong | 96.1% correct | 4,608 unknown loss 0.8697 | geometry 0.2463 | confidence 1.2468 | mean error 0.659 vox speed: 1320 ms/update | data wait 1 ms/update | 18.3 crops/s | 18.3 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1561 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 197, from the previous prediction 1364 confidence-labeled crops: terminal 8/1531 (0.5%) | recoverable 105/1531 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 384/1600 (24.0%), 2: 224/1600 (14.0%), 3: 448/1600 (28.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 664 rows | mean depth 18 decisions | mean travel 147 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 52% 46% 27%>2% 25% 5 497 91%/7%/0%/2% 7% 118/2/0/0 0175A_5mm_v1 384 54% 43% 25%>3% 22% 19 1177 89%/7%/1%/4% 5% 121/1/0/44 1447_5mm_v1 224 50% 47% 24%>3% 21% 7 456 87%/5%/1%/7% 4% 106/1/0/0 paris4_afv_central 448 57% 43% 27%>0% 27% 0 152 89%/7%/0%/4% 5% 192/4/0/0 PHerc0125_3mm_v2 288 54% 46% 27%>0% 27% 0 140 86%/7%/0%/7% 9% 127/2/0/5 gradients: max 73, clipped 3/50 updates Step 8,100 | last 50 updates / 1,600 crops | lr 2.49e-05 points (35/curve, tolerance 2 vox): 48,509 right / 2,173 wrong | 95.7% correct | 5,318 unknown loss 0.8382 | geometry 0.2575 | confidence 1.1614 | mean error 0.662 vox speed: 1347 ms/update | data wait 1 ms/update | 23.6 crops/s | 23.6 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1587 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 279, from the previous prediction 1308 confidence-labeled crops: terminal 9/1522 (0.6%) | recoverable 117/1522 (7.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 288/1600 (18.0%), 2: 320/1600 (20.0%), 3: 576/1600 (36.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 626 rows | mean depth 19 decisions | mean travel 152 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 56% 44% 29%>0% 29% 0 163 88%/10%/0%/2% 6% 84/0/0/0 0175A_5mm_v1 288 50% 49% 26%>1% 25% 3 303 84%/10%/1%/5% 5% 105/1/0/36 1447_5mm_v1 320 49% 48% 21%>3% 18% 8 463 90%/6%/1%/3% 5% 153/1/0/0 paris4_afv_central 576 57% 43% 24%>0% 24% 0 136 87%/6%/0%/6% 7% 180/3/38/65 PHerc0125_3mm_v2 224 54% 46% 24%>0% 24% 0 161 88%/5%/0%/7% 14% 104/0/0/0 gradients: max 56, clipped 0/50 updates Step 8,150 | last 50 updates / 1,600 crops | lr 2.49e-05 points (35/curve, tolerance 2 vox): 49,737 right / 1,711 wrong | 96.7% correct | 4,552 unknown loss 0.7313 | geometry 0.2186 | confidence 1.0254 | mean error 0.602 vox speed: 1308 ms/update | data wait 1 ms/update | 24.2 crops/s | 24.2 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1567 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 296, from the previous prediction 1271 confidence-labeled crops: terminal 8/1540 (0.5%) | recoverable 94/1540 (6.1%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 384/1600 (24.0%), 2: 256/1600 (16.0%), 3: 480/1600 (30.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 589 rows | mean depth 19 decisions | mean travel 152 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 47% 49% 28%>4% 24% 10 692 92%/7%/0%/1% 8% 126/1/0/0 0175A_5mm_v1 384 56% 41% 27%>3% 24% 11 552 90%/5%/1%/3% 4% 148/1/0/11 1447_5mm_v1 256 48% 48% 22%>5% 17% 12 603 92%/4%/1%/3% 4% 122/1/0/0 paris4_afv_central 480 53% 47% 26%>0% 26% 0 134 89%/7%/0%/4% 4% 102/3/0/124 PHerc0125_3mm_v2 224 59% 41% 26%>0% 26% 0 157 86%/6%/0%/8% 12% 91/2/0/0 gradients: max 54, clipped 0/50 updates Step 8,200 | last 50 updates / 1,600 crops | lr 2.48e-05 points (35/curve, tolerance 2 vox): 49,533 right / 2,564 wrong | 95.1% correct | 3,903 unknown loss 0.9224 | geometry 0.2826 | confidence 1.2797 | mean error 0.687 vox speed: 1352 ms/update | data wait 1 ms/update | 23.3 crops/s | 23.3 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1560 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 173, from the previous prediction 1387 confidence-labeled crops: terminal 11/1552 (0.7%) | recoverable 103/1552 (6.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 288/1600 (18.0%), 2: 224/1600 (14.0%), 3: 320/1600 (20.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 751 rows | mean depth 18 decisions | mean travel 146 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 50% 47% 24%>2% 22% 10 469 93%/5%/0%/2% 6% 212/0/0/0 0175A_5mm_v1 288 50% 43% 25%>7% 18% 14 944 86%/8%/2%/4% 11% 124/4/0/0 1447_5mm_v1 224 51% 47% 25%>1% 24% 3 281 89%/7%/0%/4% 3% 106/4/0/0 paris4_afv_central 320 48% 49% 21%>2% 18% 8 344 90%/5%/1%/4% 7% 158/2/0/0 PHerc0125_3mm_v2 320 53% 47% 25%>0% 25% 0 157 88%/8%/0%/3% 12% 151/0/0/0 gradients: max 56, clipped 0/50 updates Step 8,250 | last 50 updates / 1,600 crops | lr 2.47e-05 points (35/curve, tolerance 2 vox): 49,037 right / 2,465 wrong | 95.2% correct | 4,498 unknown loss 0.9161 | geometry 0.2920 | confidence 1.2482 | mean error 0.690 vox speed: 1329 ms/update | data wait 1 ms/update | 23.9 crops/s | 23.9 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1564 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 351, from the previous prediction 1213 confidence-labeled crops: terminal 7/1539 (0.5%) | recoverable 101/1539 (6.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 384/1600 (24.0%), 2: 352/1600 (22.0%), 3: 224/1600 (14.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 557 rows | mean depth 19 decisions | mean travel 150 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 56% 43% 21%>1% 20% 5 636 91%/7%/0%/2% 8% 66/0/43/84 0175A_5mm_v1 384 48% 48% 22%>4% 18% 15 688 88%/6%/1%/5% 6% 163/4/26/20 1447_5mm_v1 352 53% 43% 24%>4% 20% 14 632 89%/7%/0%/4% 4% 140/4/68/11 paris4_afv_central 224 56% 44% 19%>0% 19% 0 167 93%/5%/0%/2% 5% 99/1/0/0 PHerc0125_3mm_v2 288 50% 50% 24%>0% 24% 0 147 87%/7%/0%/6% 14% 89/3/89/56 gradients: max 73, clipped 2/50 updates Step 8,300 | last 50 updates / 1,600 crops | lr 2.47e-05 points (35/curve, tolerance 2 vox): 48,957 right / 2,048 wrong | 96.0% correct | 4,995 unknown loss 0.8288 | geometry 0.2659 | confidence 1.1256 | mean error 0.668 vox speed: 1341 ms/update | data wait 1 ms/update | 23.7 crops/s | 23.7 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1588 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 265, from the previous prediction 1323 confidence-labeled crops: terminal 7/1526 (0.5%) | recoverable 93/1526 (6.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 192/1600 (12.0%), 2: 320/1600 (20.0%), 3: 416/1600 (26.0%), 4: 480/1600 (30.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 645 rows | mean depth 18 decisions | mean travel 149 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 42% 55% 21%>3% 18% 5 472 93%/5%/0%/2% 4% 105/0/0/1 0175A_5mm_v1 192 49% 47% 27%>4% 23% 6 577 86%/8%/2%/4% 4% 91/1/0/0 1447_5mm_v1 320 50% 50% 25%>0% 25% 0 142 90%/6%/0%/4% 2% 161/0/0/0 paris4_afv_central 416 55% 45% 22%>0% 22% 0 154 92%/6%/0%/2% 5% 184/5/42/2 PHerc0125_3mm_v2 480 56% 44% 26%>0% 26% 0 140 85%/5%/0%/9% 12% 104/3/0/106 gradients: max 55, clipped 0/50 updates Step 8,350 | last 50 updates / 1,600 crops | lr 2.46e-05 points (35/curve, tolerance 2 vox): 49,067 right / 2,238 wrong | 95.6% correct | 4,695 unknown loss 0.8678 | geometry 0.2617 | confidence 1.2122 | mean error 0.656 vox speed: 1319 ms/update | data wait 1 ms/update | 24.1 crops/s | 24.1 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1557 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 258, from the previous prediction 1299 confidence-labeled crops: terminal 9/1532 (0.6%) | recoverable 107/1532 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 416/1600 (26.0%), 2: 416/1600 (26.0%), 3: 160/1600 (10.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 615 rows | mean depth 21 decisions | mean travel 168 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 59% 40% 25%>0% 25% 2 235 88%/7%/0%/5% 7% 167/0/0/0 0175A_5mm_v1 416 46% 48% 26%>6% 20% 21 953 88%/6%/2%/5% 8% 182/4/3/17 1447_5mm_v1 416 53% 45% 26%>2% 24% 10 685 91%/6%/0%/3% 5% 117/0/0/69 paris4_afv_central 160 61% 39% 26%>0% 26% 0 148 87%/9%/1%/3% 7% 62/1/0/0 PHerc0125_3mm_v2 192 55% 45% 24%>0% 24% 0 151 87%/7%/1%/6% 11% 87/1/0/0 gradients: max 76, clipped 5/50 updates Step 8,400 | last 50 updates / 1,600 crops | lr 2.46e-05 points (35/curve, tolerance 2 vox): 49,653 right / 2,056 wrong | 96.0% correct | 4,291 unknown loss 0.8587 | geometry 0.3032 | confidence 1.1109 | mean error 0.695 vox speed: 1349 ms/update | data wait 1 ms/update | 23.5 crops/s | 23.5 decisions/s | 1.13 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1575 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 245, from the previous prediction 1330 confidence-labeled crops: terminal 3/1547 (0.2%) | recoverable 108/1547 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 512/1600 (32.0%), 1: 160/1600 (10.0%), 2: 352/1600 (22.0%), 3: 320/1600 (20.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 660 rows | mean depth 17 decisions | mean travel 141 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 512 52% 47% 23%>1% 21% 6 357 91%/8%/0%/2% 6% 193/1/0/49 0175A_5mm_v1 160 42% 46% 27%>12% 15% 17 1435 89%/8%/1%/2% 11% 74/2/0/0 1447_5mm_v1 352 58% 42% 27%>0% 27% 0 136 89%/7%/0%/5% 2% 126/1/0/21 paris4_afv_central 320 52% 48% 22%>0% 22% 0 168 90%/6%/0%/4% 4% 155/3/0/0 PHerc0125_3mm_v2 256 56% 44% 26%>0% 26% 0 158 90%/5%/0%/5% 14% 112/3/0/0 gradients: max 61, clipped 1/50 updates Step 8,450 | last 50 updates / 1,600 crops | lr 2.45e-05 points (35/curve, tolerance 2 vox): 48,014 right / 2,895 wrong | 94.3% correct | 5,091 unknown loss 1.0277 | geometry 0.3125 | confidence 1.4304 | mean error 0.701 vox speed: 1362 ms/update | data wait 1 ms/update | 23.3 crops/s | 23.3 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1516 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 269, from the previous prediction 1247 confidence-labeled crops: terminal 11/1538 (0.7%) | recoverable 125/1538 (8.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 416/1600 (26.0%), 2: 192/1600 (12.0%), 3: 288/1600 (18.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 622 rows | mean depth 20 decisions | mean travel 161 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 51% 49% 26%>0% 25% 1 237 89%/8%/0%/3% 9% 73/0/0/67 0175A_5mm_v1 416 38% 48% 25%>14% 10% 43 1954 87%/6%/2%/5% 10% 161/4/18/38 1447_5mm_v1 192 43% 47% 26%>9% 17% 14 1013 86%/7%/1%/5% 9% 91/0/0/0 paris4_afv_central 288 56% 44% 24%>0% 24% 0 164 86%/9%/0%/4% 5% 128/7/0/0 PHerc0125_3mm_v2 416 59% 41% 29%>0% 29% 0 159 87%/8%/0%/4% 14% 169/5/47/2 gradients: max 57, clipped 0/50 updates Step 8,500 | last 50 updates / 1,600 crops | lr 2.44e-05 points (35/curve, tolerance 2 vox): 48,444 right / 2,538 wrong | 95.0% correct | 5,018 unknown loss 0.9103 | geometry 0.2579 | confidence 1.3048 | mean error 0.653 vox speed: 1334 ms/update | data wait 1 ms/update | 23.8 crops/s | 23.8 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1542 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 239, from the previous prediction 1303 confidence-labeled crops: terminal 13/1534 (0.8%) | recoverable 124/1534 (8.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 384/1600 (24.0%), 2: 448/1600 (28.0%), 3: 192/1600 (12.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 666 rows | mean depth 20 decisions | mean travel 163 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 55% 44% 27%>0% 26% 1 244 90%/8%/0%/1% 7% 99/0/0/0 0175A_5mm_v1 384 43% 45% 24%>12% 12% 39 1693 84%/8%/3%/5% 11% 157/4/0/17 1447_5mm_v1 448 55% 44% 25%>1% 23% 5 312 88%/8%/0%/4% 4% 190/0/12/5 paris4_afv_central 192 54% 46% 26%>0% 26% 0 163 87%/12%/0%/1% 4% 89/1/0/0 PHerc0125_3mm_v2 352 49% 51% 22%>0% 22% 0 145 88%/5%/1%/6% 13% 131/0/0/47 gradients: max 54, clipped 0/50 updates Step 8,550 | last 50 updates / 1,600 crops | lr 2.44e-05 points (35/curve, tolerance 2 vox): 49,590 right / 2,702 wrong | 94.8% correct | 3,708 unknown loss 0.9850 | geometry 0.3059 | confidence 1.3582 | mean error 0.703 vox speed: 1347 ms/update | data wait 1 ms/update | 23.5 crops/s | 23.5 decisions/s | 1.14 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1571 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 231, from the previous prediction 1340 confidence-labeled crops: terminal 10/1558 (0.6%) | recoverable 126/1558 (8.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 512/1600 (32.0%), 1: 256/1600 (16.0%), 2: 224/1600 (14.0%), 3: 352/1600 (22.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 644 rows | mean depth 17 decisions | mean travel 141 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 512 54% 45% 23%>0% 23% 2 223 90%/8%/0%/1% 9% 232/2/0/0 0175A_5mm_v1 256 43% 48% 28%>9% 18% 25 2077 86%/8%/2%/4% 11% 62/0/0/60 1447_5mm_v1 224 54% 46% 25%>0% 25% 0 144 88%/9%/0%/3% 2% 97/6/43/5 paris4_afv_central 352 57% 43% 28%>0% 28% 0 155 90%/8%/0%/2% 5% 150/1/0/0 PHerc0125_3mm_v2 256 55% 45% 22%>0% 22% 0 138 88%/6%/1%/5% 13% 103/3/0/13 gradients: max 89, clipped 2/50 updates Step 8,600 | last 50 updates / 1,600 crops | lr 2.43e-05 points (35/curve, tolerance 2 vox): 49,254 right / 2,415 wrong | 95.3% correct | 4,331 unknown loss 0.9464 | geometry 0.2960 | confidence 1.3009 | mean error 0.694 vox speed: 1344 ms/update | data wait 1 ms/update | 23.6 crops/s | 23.6 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1561 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 326, from the previous prediction 1235 confidence-labeled crops: terminal 4/1537 (0.3%) | recoverable 103/1537 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 544/1600 (34.0%), 1: 128/1600 (8.0%), 2: 288/1600 (18.0%), 3: 352/1600 (22.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 567 rows | mean depth 16 decisions | mean travel 128 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 544 50% 48% 22%>2% 20% 11 828 92%/6%/0%/2% 8% 109/0/0/151 0175A_5mm_v1 128 48% 46% 26%>6% 20% 8 927 88%/5%/1%/6% 7% 59/3/0/0 1447_5mm_v1 288 47% 46% 25%>7% 18% 18 908 87%/7%/1%/5% 6% 122/1/21/11 paris4_afv_central 352 57% 43% 24%>0% 24% 0 144 90%/6%/0%/4% 4% 150/1/48/3 PHerc0125_3mm_v2 288 56% 44% 24%>0% 24% 0 139 86%/7%/0%/7% 13% 127/2/0/0 gradients: max 73, clipped 2/50 updates Step 8,650 | last 50 updates / 1,600 crops | lr 2.43e-05 points (35/curve, tolerance 2 vox): 49,452 right / 2,127 wrong | 95.9% correct | 4,421 unknown loss 0.8346 | geometry 0.2520 | confidence 1.1651 | mean error 0.640 vox speed: 1335 ms/update | data wait 1 ms/update | 23.8 crops/s | 23.8 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1566 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 253, from the previous prediction 1313 confidence-labeled crops: terminal 12/1539 (0.8%) | recoverable 120/1539 (7.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 320/1600 (20.0%), 2: 352/1600 (22.0%), 3: 320/1600 (20.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 605 rows | mean depth 23 decisions | mean travel 189 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 55% 45% 26%>0% 25% 2 237 90%/7%/0%/3% 7% 144/0/0/42 0175A_5mm_v1 320 49% 45% 23%>6% 18% 16 892 86%/8%/2%/4% 8% 145/4/0/0 1447_5mm_v1 352 55% 41% 27%>4% 23% 14 623 85%/8%/1%/6% 6% 146/0/0/0 paris4_afv_central 320 56% 44% 32%>0% 32% 0 145 90%/7%/0%/2% 6% 89/4/36/52 PHerc0125_3mm_v2 192 58% 42% 30%>0% 30% 0 166 89%/6%/1%/5% 9% 81/2/0/0 gradients: max 62, clipped 2/50 updates Step 8,652 | dagger_states: 7869 | dagger_source_step: 8000 | dagger_caches: 2 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/0175A_5mm_v1/decisions_008000.npz" | dagger_fibers: 384 | dagger_supply: {"states": 7869, "traces": 384, "fibers": 384, "rows": {"terminal": 240, "premature_stop": 53, "recoverable": 8, "pre_excursion": 201, "ordinary": 7360}, "events": {"terminal": 30, "premature_stop": 53, "recoverable": 6, "pre_excursion": 36, "ordinary": 379}, "supervision": {"following": 7614, "recoverable": 8, "terminal": 240, "unknown": 7}, "departed_traces": 34, "terminal_traces": 30, "stop_reasons": {"after_failure": 30, "confidence": 55, "unannotated": 299}, "traced_voxels": 120957.68751552247, "seconds": 891.8082026094198, "voxels_per_second": 135.63195220855982} | dagger_coverage: {"requested": 384, "seeds": 384, "skipped": [], "epoch": 0} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 0 | dagger_launch_interval: 6000 | dagger_publication_age: 652 | dataset: "0175A_5mm_v1" Step 8,700 | last 50 updates / 1,600 crops | lr 2.42e-05 points (35/curve, tolerance 2 vox): 48,912 right / 2,452 wrong | 95.2% correct | 4,636 unknown loss 0.9551 | geometry 0.2828 | confidence 1.3447 | mean error 0.675 vox speed: 887 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.14 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1542 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 172, from the previous prediction 1370 confidence-labeled crops: terminal 7/1535 (0.5%) | recoverable 93/1535 (6.1%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 288/1600 (18.0%), 2: 224/1600 (14.0%), 3: 384/1600 (24.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 717 rows | mean depth 19 decisions | mean travel 152 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 53% 46% 26%>1% 26% 2 224 91%/7%/0%/2% 7% 178/0/0/0 0175A_5mm_v1 288 43% 49% 24%>8% 16% 27 1331 92%/4%/1%/3% 7% 131/2/8/9 1447_5mm_v1 224 48% 43% 25%>9% 17% 18 1098 88%/6%/0%/6% 6% 97/2/0/0 paris4_afv_central 384 56% 44% 26%>0% 26% 0 144 91%/5%/1%/3% 5% 169/3/0/0 PHerc0125_3mm_v2 320 53% 44% 30%>2% 28% 7 328 86%/7%/0%/8% 14% 142/3/0/0 gradients: max 65, clipped 1/50 updates Step 8,750 | last 50 updates / 1,600 crops | lr 2.41e-05 points (35/curve, tolerance 2 vox): 48,864 right / 2,336 wrong | 95.4% correct | 4,800 unknown loss 0.9548 | geometry 0.2505 | confidence 1.4086 | mean error 0.659 vox speed: 886 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1542 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 240, from the previous prediction 1302 confidence-labeled crops: terminal 12/1528 (0.8%) | recoverable 129/1528 (8.4%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 256/1600 (16.0%), 2: 288/1600 (18.0%), 3: 448/1600 (28.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 629 rows | mean depth 19 decisions | mean travel 158 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 49% 48% 22%>3% 18% 10 614 92%/6%/0%/2% 9% 152/0/0/0 0175A_5mm_v1 256 43% 43% 28%>14% 14% 41 2793 80%/11%/4%/5% 13% 71/5/102/39 1447_5mm_v1 288 59% 41% 31%>0% 31% 0 154 86%/9%/0%/5% 2% 118/1/0/0 paris4_afv_central 448 56% 44% 26%>0% 26% 0 142 88%/8%/0%/4% 5% 161/2/0/36 PHerc0125_3mm_v2 288 55% 45% 19%>0% 19% 0 150 86%/6%/0%/8% 11% 127/1/48/3 gradients: max 64, clipped 2/50 updates Step 8,800 | last 50 updates / 1,600 crops | lr 2.41e-05 points (35/curve, tolerance 2 vox): 48,311 right / 2,387 wrong | 95.3% correct | 5,302 unknown loss 0.9255 | geometry 0.2852 | confidence 1.2806 | mean error 0.662 vox speed: 886 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.14 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1550 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 245, from the previous prediction 1305 confidence-labeled crops: terminal 10/1523 (0.7%) | recoverable 97/1523 (6.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 192/1600 (12.0%), 2: 448/1600 (28.0%), 3: 224/1600 (14.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 670 rows | mean depth 21 decisions | mean travel 168 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 53% 46% 21%>1% 20% 4 319 92%/5%/0%/3% 7% 178/1/0/0 0175A_5mm_v1 192 43% 53% 21%>4% 17% 7 664 88%/7%/2%/3% 7% 94/2/0/7 1447_5mm_v1 448 50% 45% 25%>4% 21% 16 626 89%/6%/0%/5% 5% 203/2/10/0 paris4_afv_central 224 58% 37% 34%>4% 30% 9 636 83%/8%/1%/7% 7% 83/2/0/0 PHerc0125_3mm_v2 352 52% 46% 22%>1% 20% 9 409 87%/5%/0%/7% 14% 112/2/0/51 gradients: max 47, clipped 0/50 updates Step 8,850 | last 50 updates / 1,600 crops | lr 2.40e-05 points (35/curve, tolerance 2 vox): 48,856 right / 1,968 wrong | 96.1% correct | 5,176 unknown loss 0.9310 | geometry 0.2650 | confidence 1.3319 | mean error 0.660 vox speed: 886 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1580 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 254, from the previous prediction 1326 confidence-labeled crops: terminal 0/1538 (0.0%) | recoverable 103/1538 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 288/1600 (18.0%), 2: 192/1600 (12.0%), 3: 320/1600 (20.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 660 rows | mean depth 17 decisions | mean travel 140 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 54% 45% 26%>1% 25% 3 283 91%/7%/0%/2% 8% 158/1/0/0 0175A_5mm_v1 288 52% 42% 24%>6% 18% 14 977 91%/6%/0%/4% 6% 121/0/0/0 1447_5mm_v1 192 53% 47% 32%>0% 32% 0 140 88%/10%/0%/2% 3% 76/1/2/14 paris4_afv_central 320 51% 49% 20%>0% 20% 0 144 92%/5%/0%/2% 5% 156/3/0/0 PHerc0125_3mm_v2 448 54% 46% 23%>0% 23% 0 139 87%/6%/0%/7% 11% 149/4/0/56 gradients: max 52, clipped 0/50 updates Step 8,900 | last 50 updates / 1,600 crops | lr 2.39e-05 points (35/curve, tolerance 2 vox): 49,252 right / 2,240 wrong | 95.6% correct | 4,508 unknown loss 0.8916 | geometry 0.2479 | confidence 1.2874 | mean error 0.617 vox speed: 894 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1551 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 217, from the previous prediction 1334 confidence-labeled crops: terminal 15/1541 (1.0%) | recoverable 101/1541 (6.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 288/1600 (18.0%), 2: 448/1600 (28.0%), 3: 288/1600 (18.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 635 rows | mean depth 20 decisions | mean travel 165 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 52% 46% 23%>2% 20% 6 480 93%/5%/0%/2% 6% 132/0/0/0 0175A_5mm_v1 288 52% 40% 24%>8% 16% 22 1287 87%/6%/3%/4% 10% 114/0/1/0 1447_5mm_v1 448 54% 42% 25%>3% 22% 16 754 89%/7%/1%/3% 6% 165/1/0/25 paris4_afv_central 288 56% 44% 28%>0% 28% 0 154 90%/7%/0%/3% 6% 126/1/0/0 PHerc0125_3mm_v2 288 61% 39% 29%>0% 29% 0 136 85%/8%/1%/6% 15% 98/2/0/15 gradients: max 48, clipped 0/50 updates Step 8,950 | last 50 updates / 1,600 crops | lr 2.39e-05 points (35/curve, tolerance 2 vox): 48,126 right / 2,732 wrong | 94.6% correct | 5,142 unknown loss 0.9847 | geometry 0.2788 | confidence 1.4118 | mean error 0.677 vox speed: 890 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1539 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 196, from the previous prediction 1343 confidence-labeled crops: terminal 14/1538 (0.9%) | recoverable 109/1538 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 320/1600 (20.0%), 2: 320/1600 (20.0%), 3: 352/1600 (22.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 674 rows | mean depth 21 decisions | mean travel 174 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 48% 51% 22%>1% 21% 2 265 91%/7%/0%/2% 8% 131/0/0/0 0175A_5mm_v1 320 51% 44% 24%>5% 19% 14 1013 90%/5%/1%/4% 8% 108/1/19/32 1447_5mm_v1 320 53% 39% 31%>8% 24% 22 1089 88%/8%/1%/3% 5% 123/2/0/3 paris4_afv_central 352 56% 44% 26%>0% 26% 0 159 86%/10%/0%/4% 5% 155/4/0/0 PHerc0125_3mm_v2 352 50% 45% 26%>5% 20% 19 544 87%/4%/2%/7% 14% 157/3/0/0 gradients: max 64, clipped 1/50 updates Step 9,000 | last 50 updates / 1,600 crops | lr 2.38e-05 points (35/curve, tolerance 2 vox): 48,992 right / 1,933 wrong | 96.2% correct | 5,075 unknown loss 0.8495 | geometry 0.2382 | confidence 1.2226 | mean error 0.627 vox speed: 886 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1570 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 203, from the previous prediction 1367 confidence-labeled crops: terminal 5/1530 (0.3%) | recoverable 111/1530 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 224/1600 (14.0%), 2: 384/1600 (24.0%), 3: 320/1600 (20.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 696 rows | mean depth 17 decisions | mean travel 138 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 50% 46% 27%>3% 23% 8 661 87%/9%/0%/4% 9% 119/0/0/0 0175A_5mm_v1 224 53% 47% 23%>0% 23% 0 139 93%/4%/0%/2% 4% 100/1/0/5 1447_5mm_v1 384 51% 46% 25%>3% 23% 10 499 90%/6%/0%/4% 3% 150/6/0/27 paris4_afv_central 320 53% 43% 26%>4% 22% 12 527 86%/8%/1%/6% 6% 137/5/0/0 PHerc0125_3mm_v2 416 54% 46% 26%>0% 26% 0 142 87%/7%/0%/6% 12% 190/9/0/0 gradients: max 61, clipped 1/50 updates Step 9,000 | dagger_launched: true Step 9,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 251, "stopped": 0, "mean_example_error": 0.7956011699092004, "geometry_examples": 31, "examples": 32, "inference_and_measurement_seconds": 1.90380085632205, "total_diagnostic_seconds": 3.8446890339255333} Step 9,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 6/8 (75.0%) | false stops 0 1.5-3: recovered 6/8 (75.0%) | false stops 0 3-4.5: recovered 6/8 (75.0%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 1 entire proposed 8-point prefix correct (distance): 21/28 (75.0%) gate @ 0.50: false stops 0/23 (0.0%) | accepted wrong 4/26 (15.4%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.542 1.515 following 16 464 1.137 1.157 recoverable 12 348 2.083 1.992 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 1.137 1.157 332 32 812 1.542 1.515 improved/worsened/compared states: 2/2/28 Step 9,000 | batch_diagnostics_seconds: 3.84701494500041 | recovery_seconds: 15.570384796708822 Step 9,050 | last 50 updates / 1,600 crops | lr 2.38e-05 points (35/curve, tolerance 2 vox): 49,053 right / 2,152 wrong | 95.8% correct | 4,795 unknown loss 0.8107 | geometry 0.2457 | confidence 1.1300 | mean error 0.622 vox speed: 1276 ms/update | data wait 1 ms/update | 18.8 crops/s | 18.8 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1538 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 257, from the previous prediction 1281 confidence-labeled crops: terminal 13/1535 (0.8%) | recoverable 118/1535 (7.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 256/1600 (16.0%), 2: 416/1600 (26.0%), 3: 384/1600 (24.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 600 rows | mean depth 19 decisions | mean travel 161 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 54% 44% 29%>2% 27% 6 464 89%/8%/0%/2% 7% 155/0/0/0 0175A_5mm_v1 256 50% 43% 26%>7% 19% 14 1090 86%/8%/1%/5% 6% 107/2/0/4 1447_5mm_v1 416 49% 46% 22%>5% 18% 26 1238 87%/7%/2%/4% 7% 130/1/0/61 paris4_afv_central 384 57% 41% 29%>2% 26% 9 466 89%/7%/0%/4% 6% 126/3/33/31 PHerc0125_3mm_v2 192 57% 43% 28%>0% 28% 0 149 86%/6%/1%/7% 10% 82/0/0/0 gradients: max 74, clipped 1/50 updates Step 9,100 | last 50 updates / 1,600 crops | lr 2.37e-05 points (35/curve, tolerance 2 vox): 48,512 right / 2,454 wrong | 95.2% correct | 5,034 unknown loss 0.8608 | geometry 0.2569 | confidence 1.2077 | mean error 0.644 vox speed: 1352 ms/update | data wait 1 ms/update | 23.5 crops/s | 23.5 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1564 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 223, from the previous prediction 1341 confidence-labeled crops: terminal 16/1531 (1.0%) | recoverable 102/1531 (6.7%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 160/1600 (10.0%), 2: 448/1600 (28.0%), 3: 352/1600 (22.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 672 rows | mean depth 23 decisions | mean travel 187 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 49% 51% 24%>0% 24% 0 152 91%/6%/0%/2% 7% 130/0/0/0 0175A_5mm_v1 160 45% 44% 24%>11% 12% 15 1568 86%/6%/4%/4% 10% 70/0/0/0 1447_5mm_v1 448 50% 47% 26%>3% 23% 12 589 88%/6%/1%/4% 5% 144/2/0/66 paris4_afv_central 352 53% 45% 22%>2% 21% 6 332 88%/7%/1%/4% 4% 158/4/0/0 PHerc0125_3mm_v2 384 55% 45% 29%>0% 29% 0 141 86%/6%/1%/7% 14% 170/3/2/4 gradients: max 58, clipped 0/50 updates Step 9,150 | last 50 updates / 1,600 crops | lr 2.36e-05 points (35/curve, tolerance 2 vox): 48,635 right / 2,603 wrong | 94.9% correct | 4,762 unknown loss 1.0042 | geometry 0.2999 | confidence 1.4085 | mean error 0.695 vox speed: 1358 ms/update | data wait 1 ms/update | 23.3 crops/s | 23.3 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1553 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 278, from the previous prediction 1275 confidence-labeled crops: terminal 11/1535 (0.7%) | recoverable 130/1535 (8.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 256/1600 (16.0%), 2: 160/1600 (10.0%), 3: 384/1600 (24.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 580 rows | mean depth 18 decisions | mean travel 144 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 55% 44% 25%>1% 24% 6 499 88%/9%/0%/3% 10% 128/0/118/54 0175A_5mm_v1 256 46% 42% 23%>12% 11% 28 1695 88%/7%/2%/3% 8% 107/1/0/0 1447_5mm_v1 160 58% 42% 31%>0% 31% 0 139 90%/6%/0%/4% 2% 67/2/0/0 paris4_afv_central 384 57% 43% 25%>0% 25% 0 138 89%/7%/0%/4% 6% 129/2/0/37 PHerc0125_3mm_v2 384 58% 39% 29%>3% 26% 11 418 83%/9%/2%/6% 12% 149/4/0/0 gradients: max 58, clipped 0/50 updates Step 9,200 | last 50 updates / 1,600 crops | lr 2.36e-05 points (35/curve, tolerance 2 vox): 48,193 right / 2,786 wrong | 94.5% correct | 5,021 unknown loss 0.9293 | geometry 0.2910 | confidence 1.2764 | mean error 0.679 vox speed: 1346 ms/update | data wait 1 ms/update | 23.6 crops/s | 23.6 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1518 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 313, from the previous prediction 1205 confidence-labeled crops: terminal 18/1537 (1.2%) | recoverable 112/1537 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 224/1600 (14.0%), 2: 288/1600 (18.0%), 3: 384/1600 (24.0%), 4: 480/1600 (30.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 594 rows | mean depth 19 decisions | mean travel 153 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 55% 44% 26%>1% 25% 2 301 88%/9%/0%/2% 7% 98/0/0/1 0175A_5mm_v1 224 42% 46% 24%>12% 12% 23 1575 88%/7%/2%/3% 8% 103/4/0/0 1447_5mm_v1 288 56% 40% 25%>4% 21% 11 733 89%/8%/1%/2% 4% 116/1/0/0 paris4_afv_central 384 49% 46% 26%>5% 21% 18 749 86%/6%/2%/6% 7% 140/2/0/35 PHerc0125_3mm_v2 480 48% 48% 22%>4% 19% 21 714 87%/6%/1%/6% 14% 137/4/0/95 gradients: max 53, clipped 0/50 updates Step 9,250 | last 50 updates / 1,600 crops | lr 2.35e-05 points (35/curve, tolerance 2 vox): 48,202 right / 2,382 wrong | 95.3% correct | 5,416 unknown loss 0.8983 | geometry 0.2885 | confidence 1.2197 | mean error 0.695 vox speed: 1342 ms/update | data wait 1 ms/update | 23.5 crops/s | 23.5 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1549 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 271, from the previous prediction 1278 confidence-labeled crops: terminal 11/1528 (0.7%) | recoverable 94/1528 (6.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 288/1600 (18.0%), 2: 288/1600 (18.0%), 3: 384/1600 (24.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 653 rows | mean depth 25 decisions | mean travel 204 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 58% 41% 22%>1% 21% 3 341 92%/7%/0%/1% 5% 132/0/0/0 0175A_5mm_v1 288 45% 49% 22%>6% 15% 16 975 90%/4%/1%/5% 7% 138/2/32/2 1447_5mm_v1 288 45% 47% 25%>8% 18% 19 1020 84%/9%/1%/6% 6% 135/3/0/0 paris4_afv_central 384 49% 49% 23%>2% 21% 8 412 92%/3%/1%/4% 7% 139/4/0/49 PHerc0125_3mm_v2 320 52% 48% 22%>0% 22% 0 136 86%/7%/0%/7% 12% 109/1/0/45 gradients: max 54, clipped 0/50 updates Step 9,300 | last 50 updates / 1,600 crops | lr 2.34e-05 points (35/curve, tolerance 2 vox): 48,532 right / 2,229 wrong | 95.6% correct | 5,239 unknown loss 0.8377 | geometry 0.2431 | confidence 1.1892 | mean error 0.627 vox speed: 1338 ms/update | data wait 1 ms/update | 23.7 crops/s | 23.7 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1521 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 231, from the previous prediction 1290 confidence-labeled crops: terminal 13/1533 (0.8%) | recoverable 111/1533 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 224/1600 (14.0%), 2: 576/1600 (36.0%), 3: 224/1600 (14.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 631 rows | mean depth 17 decisions | mean travel 142 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 54% 45% 25%>1% 24% 2 310 92%/6%/0%/2% 7% 116/0/0/0 0175A_5mm_v1 224 44% 45% 27%>11% 16% 27 1932 87%/7%/3%/3% 10% 89/0/8/12 1447_5mm_v1 576 49% 45% 27%>6% 20% 35 1264 86%/9%/1%/5% 6% 184/3/105/74 paris4_afv_central 224 55% 45% 24%>0% 24% 0 163 91%/5%/0%/4% 5% 101/2/0/0 PHerc0125_3mm_v2 320 54% 44% 25%>2% 23% 6 315 88%/6%/0%/6% 11% 141/3/0/0 gradients: max 49, clipped 0/50 updates Step 9,350 | last 50 updates / 1,600 crops | lr 2.34e-05 points (35/curve, tolerance 2 vox): 48,543 right / 2,671 wrong | 94.8% correct | 4,786 unknown loss 0.9732 | geometry 0.3132 | confidence 1.3200 | mean error 0.684 vox speed: 1333 ms/update | data wait 1 ms/update | 23.8 crops/s | 23.8 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1526 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 285, from the previous prediction 1241 confidence-labeled crops: terminal 14/1553 (0.9%) | recoverable 100/1553 (6.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 224/1600 (14.0%), 2: 448/1600 (28.0%), 3: 256/1600 (16.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 620 rows | mean depth 19 decisions | mean travel 156 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 55% 44% 24%>1% 23% 2 285 92%/7%/0%/1% 6% 128/2/0/0 0175A_5mm_v1 224 45% 41% 24%>14% 10% 27 2439 85%/8%/2%/4% 11% 70/1/0/22 1447_5mm_v1 448 52% 43% 23%>6% 17% 29 1476 91%/5%/1%/3% 6% 112/1/0/79 paris4_afv_central 256 52% 48% 22%>0% 22% 0 158 93%/6%/0%/1% 7% 123/5/0/0 PHerc0125_3mm_v2 384 49% 49% 24%>2% 22% 8 313 88%/6%/1%/5% 14% 187/4/0/0 gradients: max 47, clipped 0/50 updates Step 9,400 | last 50 updates / 1,600 crops | lr 2.33e-05 points (35/curve, tolerance 2 vox): 49,509 right / 2,090 wrong | 95.9% correct | 4,401 unknown loss 0.8031 | geometry 0.2395 | confidence 1.1272 | mean error 0.636 vox speed: 1341 ms/update | data wait 1 ms/update | 23.6 crops/s | 23.6 decisions/s | 1.13 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1563 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 288, from the previous prediction 1275 confidence-labeled crops: terminal 12/1544 (0.8%) | recoverable 113/1544 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 352/1600 (22.0%), 2: 416/1600 (26.0%), 3: 256/1600 (16.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 620 rows | mean depth 25 decisions | mean travel 203 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 51% 49% 20%>0% 20% 1 197 91%/7%/0%/2% 6% 152/1/50/4 0175A_5mm_v1 352 49% 44% 25%>7% 18% 20 1162 86%/9%/2%/4% 7% 154/5/0/0 1447_5mm_v1 416 50% 48% 23%>1% 22% 5 477 90%/7%/0%/2% 3% 103/2/0/97 paris4_afv_central 256 57% 41% 31%>2% 29% 5 397 87%/7%/1%/5% 7% 106/3/0/0 PHerc0125_3mm_v2 256 57% 43% 24%>0% 24% 0 153 88%/5%/1%/6% 11% 105/8/49/6 gradients: max 63, clipped 1/50 updates Step 9,450 | last 50 updates / 1,600 crops | lr 2.32e-05 points (35/curve, tolerance 2 vox): 48,006 right / 2,779 wrong | 94.5% correct | 5,215 unknown loss 0.9767 | geometry 0.2919 | confidence 1.3696 | mean error 0.699 vox speed: 1367 ms/update | data wait 1 ms/update | 23.2 crops/s | 23.2 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1518 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 286, from the previous prediction 1232 confidence-labeled crops: terminal 13/1525 (0.9%) | recoverable 117/1525 (7.7%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 224/1600 (14.0%), 2: 288/1600 (18.0%), 3: 288/1600 (18.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 627 rows | mean depth 18 decisions | mean travel 142 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 51% 49% 23%>1% 23% 2 281 92%/7%/0%/1% 7% 111/0/0/60 0175A_5mm_v1 224 40% 45% 29%>15% 14% 27 1938 81%/10%/1%/8% 8% 101/1/0/0 1447_5mm_v1 288 46% 45% 25%>9% 17% 22 1169 85%/8%/1%/6% 6% 130/2/0/0 paris4_afv_central 288 49% 48% 23%>3% 19% 10 530 90%/6%/1%/3% 6% 137/4/0/0 PHerc0125_3mm_v2 448 54% 44% 27%>2% 25% 12 459 86%/7%/1%/6% 15% 148/6/0/47 gradients: max 74, clipped 2/50 updates Step 9,500 | last 50 updates / 1,600 crops | lr 2.32e-05 points (35/curve, tolerance 2 vox): 48,212 right / 2,608 wrong | 94.9% correct | 5,180 unknown loss 0.8909 | geometry 0.2658 | confidence 1.2501 | mean error 0.650 vox speed: 1350 ms/update | data wait 1 ms/update | 23.5 crops/s | 23.5 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1523 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 268, from the previous prediction 1255 confidence-labeled crops: terminal 20/1537 (1.3%) | recoverable 115/1537 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 224/1600 (14.0%), 2: 320/1600 (20.0%), 3: 544/1600 (34.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 609 rows | mean depth 20 decisions | mean travel 164 vox | 8 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 52% 48% 28%>1% 28% 1 239 89%/9%/0%/2% 6% 92/0/0/0 0175A_5mm_v1 224 51% 39% 24%>10% 14% 23 2342 88%/7%/1%/4% 9% 61/2/11/27 1447_5mm_v1 320 53% 42% 26%>5% 21% 12 774 88%/7%/2%/3% 5% 136/2/0/0 paris4_afv_central 544 48% 47% 23%>5% 18% 24 826 87%/6%/1%/6% 7% 178/2/108/79 PHerc0125_3mm_v2 320 52% 45% 27%>3% 24% 9 399 87%/8%/2%/4% 15% 142/4/0/1 gradients: max 50, clipped 0/50 updates Step 9,550 | last 50 updates / 1,600 crops | lr 2.31e-05 points (35/curve, tolerance 2 vox): 49,417 right / 2,505 wrong | 95.2% correct | 4,078 unknown loss 0.9867 | geometry 0.2890 | confidence 1.3954 | mean error 0.694 vox speed: 1352 ms/update | data wait 1 ms/update | 23.5 crops/s | 23.5 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1566 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 237, from the previous prediction 1329 confidence-labeled crops: terminal 7/1547 (0.5%) | recoverable 103/1547 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 512/1600 (32.0%), 1: 256/1600 (16.0%), 2: 192/1600 (12.0%), 3: 224/1600 (14.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 657 rows | mean depth 25 decisions | mean travel 206 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 512 59% 40% 27%>0% 27% 2 225 92%/6%/0%/2% 6% 206/0/0/0 0175A_5mm_v1 256 49% 46% 23%>5% 18% 11 825 91%/4%/1%/4% 5% 117/2/0/1 1447_5mm_v1 192 52% 48% 27%>0% 27% 0 163 90%/7%/0%/3% 2% 93/0/0/0 paris4_afv_central 224 47% 50% 22%>3% 19% 7 483 89%/6%/1%/4% 8% 103/0/0/8 PHerc0125_3mm_v2 416 51% 46% 24%>2% 22% 13 512 87%/8%/0%/5% 13% 138/8/0/55 gradients: max 57, clipped 0/50 updates Step 9,600 | last 50 updates / 1,600 crops | lr 2.30e-05 points (35/curve, tolerance 2 vox): 48,825 right / 2,773 wrong | 94.6% correct | 4,402 unknown loss 0.8763 | geometry 0.2579 | confidence 1.2367 | mean error 0.647 vox speed: 1360 ms/update | data wait 1 ms/update | 23.2 crops/s | 23.2 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1509 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 213, from the previous prediction 1296 confidence-labeled crops: terminal 20/1550 (1.3%) | recoverable 128/1550 (8.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 384/1600 (24.0%), 2: 256/1600 (16.0%), 3: 256/1600 (16.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 662 rows | mean depth 18 decisions | mean travel 147 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 56% 43% 25%>1% 25% 4 361 92%/7%/0%/1% 8% 150/3/0/44 0175A_5mm_v1 384 55% 40% 28%>6% 22% 22 1089 88%/7%/2%/4% 8% 152/1/0/0 1447_5mm_v1 256 36% 47% 27%>17% 9% 29 1868 82%/11%/3%/5% 14% 120/4/0/0 paris4_afv_central 256 45% 50% 22%>5% 17% 12 639 87%/9%/2%/2% 7% 128/2/0/0 PHerc0125_3mm_v2 256 53% 44% 27%>3% 23% 8 443 86%/8%/0%/6% 14% 112/1/0/0 gradients: max 47, clipped 0/50 updates Step 9,650 | last 50 updates / 1,600 crops | lr 2.30e-05 points (35/curve, tolerance 2 vox): 48,124 right / 2,880 wrong | 94.4% correct | 4,996 unknown loss 1.0130 | geometry 0.2952 | confidence 1.4355 | mean error 0.695 vox speed: 1343 ms/update | data wait 1 ms/update | 23.6 crops/s | 23.6 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1506 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 297, from the previous prediction 1209 confidence-labeled crops: terminal 16/1541 (1.0%) | recoverable 120/1541 (7.8%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 448/1600 (28.0%), 2: 160/1600 (10.0%), 3: 352/1600 (22.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 577 rows | mean depth 22 decisions | mean travel 186 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 55% 44% 24%>1% 23% 4 425 89%/8%/0%/2% 8% 113/0/0/57 0175A_5mm_v1 448 44% 45% 25%>10% 15% 40 2374 84%/9%/1%/5% 10% 134/4/0/68 1447_5mm_v1 160 52% 44% 27%>4% 22% 7 753 90%/7%/1%/2% 7% 70/0/0/0 paris4_afv_central 352 54% 41% 26%>5% 21% 17 726 88%/6%/2%/4% 7% 144/2/0/0 PHerc0125_3mm_v2 256 51% 45% 24%>4% 20% 9 474 88%/5%/1%/5% 14% 116/3/0/0 gradients: max 63, clipped 1/50 updates Step 9,700 | last 50 updates / 1,600 crops | lr 2.29e-05 points (35/curve, tolerance 2 vox): 49,034 right / 2,413 wrong | 95.3% correct | 4,553 unknown loss 0.9070 | geometry 0.2610 | confidence 1.2919 | mean error 0.642 vox speed: 1363 ms/update | data wait 1 ms/update | 23.3 crops/s | 23.3 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1553 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 230, from the previous prediction 1323 confidence-labeled crops: terminal 15/1540 (1.0%) | recoverable 148/1540 (9.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 320/1600 (20.0%), 2: 384/1600 (24.0%), 3: 288/1600 (18.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 626 rows | mean depth 24 decisions | mean travel 195 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 54% 45% 24%>1% 23% 3 307 86%/12%/0%/3% 8% 160/0/0/0 0175A_5mm_v1 320 48% 45% 26%>8% 18% 24 2911 82%/8%/3%/7% 11% 53/1/0/90 1447_5mm_v1 384 51% 47% 23%>2% 21% 9 473 89%/8%/1%/3% 4% 179/3/0/0 paris4_afv_central 288 60% 40% 30%>0% 30% 0 145 85%/11%/0%/3% 4% 115/1/18/0 PHerc0125_3mm_v2 256 52% 46% 23%>2% 22% 4 308 88%/6%/1%/5% 15% 119/4/0/0 gradients: max 56, clipped 0/50 updates Step 9,742 | dagger_states: 9012 | dagger_source_step: 9000 | dagger_caches: 2 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/1447_5mm_v1/decisions_009000.npz" | dagger_fibers: 384 | dagger_supply: {"states": 9012, "traces": 384, "fibers": 384, "rows": {"terminal": 200, "premature_stop": 62, "recoverable": 29, "pre_excursion": 195, "ordinary": 8511}, "events": {"terminal": 27, "premature_stop": 62, "recoverable": 10, "pre_excursion": 34, "ordinary": 379}, "supervision": {"following": 8768, "recoverable": 13, "terminal": 200, "unknown": 31}, "departed_traces": 31, "terminal_traces": 27, "stop_reasons": {"after_failure": 25, "confidence": 64, "unannotated": 295}, "traced_voxels": 138979.12155237122, "seconds": 1016.2276714742184, "voxels_per_second": 136.75982799283292} | dagger_coverage: {"requested": 384, "seeds": 384, "skipped": [], "epoch": 0} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 0 | dagger_launch_interval: 6000 | dagger_publication_age: 742 | dataset: "1447_5mm_v1" Step 9,750 | last 50 updates / 1,600 crops | lr 2.28e-05 points (35/curve, tolerance 2 vox): 46,724 right / 3,330 wrong | 93.3% correct | 5,946 unknown loss 0.9914 | geometry 0.3042 | confidence 1.3743 | mean error 0.678 vox speed: 1215 ms/update | data wait 1 ms/update | 26.1 crops/s | 26.1 decisions/s | 1.27 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1424 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 286, from the previous prediction 1138 confidence-labeled crops: terminal 30/1523 (2.0%) | recoverable 113/1523 (7.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 288/1600 (18.0%), 2: 448/1600 (28.0%), 3: 384/1600 (24.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 991 chunk references / 1 windows live chains: 571 rows | mean depth 18 decisions | mean travel 150 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 57% 43% 27%>1% 26% 1 268 90%/9%/0%/1% 7% 82/1/0/0 0175A_5mm_v1 288 44% 47% 25%>9% 16% 20 1372 88%/6%/1%/5% 8% 136/1/0/0 1447_5mm_v1 448 41% 43% 26%>16% 10% 53 3013 83%/6%/4%/7% 10% 102/2/128/91 paris4_afv_central 384 44% 41% 29%>14% 14% 55 1900 85%/7%/2%/6% 12% 134/4/32/25 PHerc0125_3mm_v2 288 57% 41% 26%>3% 23% 8 460 86%/9%/1%/3% 15% 117/3/0/0 gradients: max 46, clipped 0/50 updates Step 9,800 | last 50 updates / 1,600 crops | lr 2.28e-05 points (35/curve, tolerance 2 vox): 47,536 right / 3,076 wrong | 93.9% correct | 5,388 unknown loss 0.9572 | geometry 0.2805 | confidence 1.3533 | mean error 0.684 vox speed: 886 ms/update | data wait 1 ms/update | 35.7 crops/s | 35.7 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1516 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 200, from the previous prediction 1316 confidence-labeled crops: terminal 18/1534 (1.2%) | recoverable 96/1534 (6.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 288/1600 (18.0%), 2: 320/1600 (20.0%), 3: 288/1600 (18.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 691 rows | mean depth 19 decisions | mean travel 163 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 48% 51% 20%>1% 20% 2 300 92%/6%/0%/2% 7% 130/1/0/0 0175A_5mm_v1 288 49% 43% 25%>8% 17% 20 1291 89%/5%/3%/3% 9% 124/1/0/0 1447_5mm_v1 320 47% 42% 26%>11% 15% 32 1534 86%/7%/2%/5% 8% 132/2/0/4 paris4_afv_central 288 50% 44% 27%>7% 20% 16 881 89%/5%/0%/6% 8% 126/3/0/0 PHerc0125_3mm_v2 448 53% 46% 23%>1% 22% 6 299 87%/6%/0%/6% 12% 179/1/57/27 gradients: max 49, clipped 0/50 updates Step 9,850 | last 50 updates / 1,600 crops | lr 2.27e-05 points (35/curve, tolerance 2 vox): 46,686 right / 3,317 wrong | 93.4% correct | 5,997 unknown loss 1.0325 | geometry 0.2872 | confidence 1.4905 | mean error 0.692 vox speed: 889 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1466 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 316, from the previous prediction 1150 confidence-labeled crops: terminal 30/1514 (2.0%) | recoverable 98/1514 (6.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 256/1600 (16.0%), 2: 320/1600 (20.0%), 3: 320/1600 (20.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 554 rows | mean depth 27 decisions | mean travel 223 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 55% 44% 25%>1% 25% 2 244 91%/6%/0%/3% 7% 170/0/0/0 0175A_5mm_v1 256 44% 39% 30%>18% 13% 35 2865 80%/8%/5%/6% 13% 84/2/42/15 1447_5mm_v1 320 45% 44% 26%>11% 15% 33 2128 87%/6%/2%/5% 8% 92/0/0/50 paris4_afv_central 320 41% 44% 26%>15% 12% 41 1572 83%/6%/3%/8% 12% 142/3/0/0 PHerc0125_3mm_v2 320 54% 46% 26%>0% 26% 0 140 88%/5%/0%/7% 13% 66/1/0/80 gradients: max 66, clipped 1/50 updates Step 9,900 | last 50 updates / 1,600 crops | lr 2.26e-05 points (35/curve, tolerance 2 vox): 46,870 right / 3,218 wrong | 93.6% correct | 5,912 unknown loss 0.9440 | geometry 0.2816 | confidence 1.3248 | mean error 0.675 vox speed: 888 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1489 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 214, from the previous prediction 1275 confidence-labeled crops: terminal 28/1509 (1.9%) | recoverable 111/1509 (7.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 384/1600 (24.0%), 2: 352/1600 (22.0%), 3: 256/1600 (16.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 677 rows | mean depth 18 decisions | mean travel 146 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 56% 44% 29%>0% 29% 0 141 88%/10%/0%/3% 4% 72/1/94/12 0175A_5mm_v1 384 40% 45% 23%>15% 8% 41 2110 84%/5%/4%/7% 13% 170/1/30/2 1447_5mm_v1 352 49% 45% 27%>6% 21% 19 953 86%/6%/2%/6% 5% 157/3/0/0 paris4_afv_central 256 51% 44% 25%>5% 20% 14 789 87%/8%/1%/4% 9% 112/0/0/0 PHerc0125_3mm_v2 416 49% 46% 24%>4% 19% 17 599 84%/7%/1%/8% 14% 166/3/0/27 gradients: max 51, clipped 0/50 updates Step 9,950 | last 50 updates / 1,600 crops | lr 2.26e-05 points (35/curve, tolerance 2 vox): 47,484 right / 2,820 wrong | 94.4% correct | 5,696 unknown loss 1.0428 | geometry 0.3233 | confidence 1.4391 | mean error 0.716 vox speed: 889 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1508 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 268, from the previous prediction 1240 confidence-labeled crops: terminal 11/1533 (0.7%) | recoverable 125/1533 (8.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 256/1600 (16.0%), 2: 192/1600 (12.0%), 3: 320/1600 (20.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 632 rows | mean depth 20 decisions | mean travel 164 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 55% 44% 24%>1% 23% 3 313 90%/8%/0%/1% 8% 153/1/0/45 0175A_5mm_v1 256 41% 49% 22%>10% 12% 26 2304 86%/9%/1%/4% 13% 91/3/32/34 1447_5mm_v1 192 41% 44% 21%>16% 5% 28 1935 85%/6%/1%/9% 7% 84/4/0/0 paris4_afv_central 320 57% 41% 32%>2% 30% 7 461 86%/7%/0%/7% 5% 123/1/19/9 PHerc0125_3mm_v2 384 48% 47% 21%>4% 17% 16 550 86%/8%/2%/4% 14% 181/3/0/0 gradients: max 68, clipped 1/50 updates Step 10,000 | last 50 updates / 1,600 crops | lr 2.25e-05 points (35/curve, tolerance 2 vox): 48,325 right / 2,867 wrong | 94.4% correct | 4,808 unknown loss 0.9919 | geometry 0.3088 | confidence 1.3661 | mean error 0.702 vox speed: 888 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1481 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 297, from the previous prediction 1184 confidence-labeled crops: terminal 17/1531 (1.1%) | recoverable 112/1531 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 480/1600 (30.0%), 1: 160/1600 (10.0%), 2: 320/1600 (20.0%), 3: 352/1600 (22.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 616 rows | mean depth 25 decisions | mean travel 203 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 480 48% 50% 20%>1% 19% 5 395 94%/5%/0%/1% 7% 204/1/0/38 0175A_5mm_v1 160 51% 39% 32%>10% 22% 14 1733 86%/7%/1%/6% 8% 63/3/0/0 1447_5mm_v1 320 39% 43% 26%>17% 9% 41 2090 85%/8%/1%/6% 9% 139/2/0/0 paris4_afv_central 352 45% 48% 26%>7% 19% 23 1089 85%/8%/2%/5% 10% 124/2/0/45 PHerc0125_3mm_v2 288 55% 40% 27%>6% 21% 18 970 84%/8%/1%/7% 13% 86/3/0/28 gradients: max 54, clipped 0/50 updates Step 10,000 | dagger_launched: true Step 10,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 256, "stopped": 0, "mean_example_error": 0.39318711334659207, "geometry_examples": 31, "examples": 32, "inference_and_measurement_seconds": 1.9852206781506538, "total_diagnostic_seconds": 3.9135759957134724} Step 10,000 | monitor rollout @ 0.50 coverage 75.8% | precision 78.7% | diverged 28.1% Step 10,000 | monitor rollout @ 0.50 coverage 95.0% | precision 100.0% | diverged 0.0% Step 10,000 | monitor rollout @ 0.50 coverage 99.3% | precision 100.0% | diverged 0.0% Step 10,000 | monitor rollout @ 0.50 coverage 94.2% | precision 95.0% | diverged 6.2% Step 10,000 | monitor rollout @ 0.50 coverage 86.3% | precision 71.4% | diverged 25.0% Step 10,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 6/8 (75.0%) | false stops 1 1.5-3: recovered 6/8 (75.0%) | false stops 0 3-4.5: recovered 6/8 (75.0%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 1 entire proposed 8-point prefix correct (distance): 20/28 (71.4%) gate @ 0.50: false stops 1/22 (4.5%) | accepted wrong 5/25 (20.0%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.567 1.520 following 16 464 1.153 1.172 recoverable 12 348 2.118 1.984 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 1.153 1.172 332 32 812 1.567 1.520 improved/worsened/compared states: 2/2/28 Step 10,000 | batch_diagnostics_seconds: 3.9162592589855194 | diagnostics_seconds: 778.6000851094723 | recovery_seconds: 14.409001540392637 Step 10,001 | dagger_states: 7243 | dagger_source_step: 10000 | dagger_caches: 2 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/paris4_afv_central/decisions_010000.npz" | dagger_fibers: 384 | dagger_supply: {"states": 7243, "traces": 384, "fibers": 384, "rows": {"terminal": 525, "premature_stop": 75, "recoverable": 50, "pre_excursion": 474, "ordinary": 6108}, "events": {"terminal": 68, "premature_stop": 75, "recoverable": 27, "pre_excursion": 86, "ordinary": 365}, "supervision": {"following": 6654, "recoverable": 45, "terminal": 525, "unknown": 19}, "departed_traces": 79, "terminal_traces": 68, "stop_reasons": {"after_failure": 64, "confidence": 79, "unannotated": 241}, "traced_voxels": 105375.8138078321, "seconds": 537.4492693208158, "voxels_per_second": 196.06653096951345} | dagger_coverage: {"requested": 384, "seeds": 384, "skipped": [], "epoch": 0} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 0 | dagger_launch_interval: 6000 | dagger_publication_age: 1 | dataset: "paris4_afv_central" Step 10,050 | last 50 updates / 1,600 crops | lr 2.24e-05 points (35/curve, tolerance 2 vox): 47,401 right / 3,208 wrong | 93.7% correct | 5,391 unknown loss 0.9496 | geometry 0.2649 | confidence 1.3695 | mean error 0.667 vox speed: 885 ms/update | data wait 1 ms/update | 1.9 crops/s | 1.9 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1493 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 236, from the previous prediction 1257 confidence-labeled crops: terminal 27/1531 (1.8%) | recoverable 109/1531 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 192/1600 (12.0%), 2: 384/1600 (24.0%), 3: 288/1600 (18.0%), 4: 480/1600 (30.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 641 rows | mean depth 16 decisions | mean travel 133 vox | 7 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 55% 43% 26%>2% 24% 5 845 91%/7%/0%/2% 7% 58/0/0/53 0175A_5mm_v1 192 43% 45% 23%>12% 11% 18 1824 83%/9%/3%/5% 12% 86/1/0/0 1447_5mm_v1 384 43% 44% 24%>12% 12% 40 1754 86%/8%/2%/3% 9% 157/1/42/13 paris4_afv_central 288 49% 46% 23%>5% 18% 14 731 89%/5%/1%/5% 7% 126/1/0/7 PHerc0125_3mm_v2 480 49% 47% 25%>4% 22% 17 502 85%/6%/2%/7% 12% 214/7/0/12 gradients: max 55, clipped 0/50 updates Step 10,100 | last 50 updates / 1,600 crops | lr 2.24e-05 points (35/curve, tolerance 2 vox): 47,686 right / 2,604 wrong | 94.8% correct | 5,710 unknown loss 0.8753 | geometry 0.2640 | confidence 1.2227 | mean error 0.652 vox speed: 902 ms/update | data wait 1 ms/update | 35.1 crops/s | 35.1 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1478 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 272, from the previous prediction 1206 confidence-labeled crops: terminal 24/1529 (1.6%) | recoverable 130/1529 (8.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 384/1600 (24.0%), 2: 384/1600 (24.0%), 3: 224/1600 (14.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 603 rows | mean depth 24 decisions | mean travel 194 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 58% 42% 25%>0% 25% 1 193 88%/9%/0%/2% 7% 175/0/0/0 0175A_5mm_v1 384 50% 43% 26%>7% 19% 23 1327 87%/8%/2%/4% 8% 163/5/17/1 1447_5mm_v1 384 45% 45% 25%>9% 15% 43 3075 83%/10%/2%/5% 9% 76/1/54/98 paris4_afv_central 224 38% 51% 21%>11% 10% 22 1260 90%/4%/3%/3% 7% 107/3/0/8 PHerc0125_3mm_v2 192 51% 43% 29%>6% 22% 12 786 79%/8%/1%/12% 14% 82/4/0/0 gradients: max 50, clipped 0/50 updates Step 10,150 | last 50 updates / 1,600 crops | lr 2.23e-05 points (35/curve, tolerance 2 vox): 49,022 right / 2,549 wrong | 95.1% correct | 4,429 unknown loss 0.8623 | geometry 0.2641 | confidence 1.1965 | mean error 0.656 vox speed: 893 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.13 attempts/decision | peak allocated VRAM 41.10 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1524 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 230, from the previous prediction 1294 confidence-labeled crops: terminal 15/1545 (1.0%) | recoverable 120/1545 (7.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 256/1600 (16.0%), 2: 448/1600 (28.0%), 3: 448/1600 (28.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 40 chunks / 80.0 MiB | promoted 5 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 666 rows | mean depth 23 decisions | mean travel 186 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 46% 54% 28%>0% 27% 1 217 89%/9%/0%/2% 7% 120/0/0/0 0175A_5mm_v1 256 50% 48% 23%>2% 21% 5 554 91%/6%/0%/2% 5% 92/2/0/32 1447_5mm_v1 448 48% 46% 27%>6% 21% 30 1277 88%/7%/1%/4% 4% 169/4/0/35 paris4_afv_central 448 53% 44% 29%>3% 25% 13 573 88%/7%/1%/4% 6% 196/1/0/0 PHerc0125_3mm_v2 224 50% 40% 33%>10% 23% 21 1148 83%/9%/1%/7% 21% 89/3/0/0 gradients: max 67, clipped 2/50 updates /ephemeral/sean/fiber_sequence/cache/python/cpython-3.14.4-linux-x86_64-gnu/lib/python3.14/multiprocessing/resource_tracker.py:396: UserWarning: resource_tracker: There appear to be 262 leaked semaphore objects to clean up at shutdown: {'/mp-t7xgs10r', '/mp-icn0af6g', '/mp-9oizbcr1', '/mp-uhglz26h', '/mp-h3coy4cl', '/mp-mqbc65h1', '/mp-vq43h0tr', '/mp-jbm13a3_', '/mp-lnz0esvi', '/mp-c46ryzzz', '/mp-b5skjgt7', '/mp-er96if5i', '/mp-avf2lq7b', '/mp-5xh4zuy0', '/mp-f5383k_f', '/mp-tfy7s_6w', '/mp-9poexn0m', '/mp-w18z02s5', '/mp-cy8_azyn', '/mp-pqoeij73', '/mp-u81s5klx', '/mp-u3vqwpab', '/mp-m8pm6mh9', '/mp-hrbq4z3l', '/mp-brdjyakt', '/mp-u0k5jf32', '/mp-g1ui_674', '/mp-9996jzgd', '/mp-_s4mf15b', '/mp-6r6xxf5t', '/mp-m6thdfhq', '/mp-vfu39tly', '/mp-71oxovgn', '/mp-efs4_0o9', '/mp-miwrg97g', '/mp-cjnxqazy', '/mp-988no4ja', '/mp-i372teee', '/mp-coltq93n', '/mp-2aqnm0_f', '/mp-ealtadu2', '/mp-jeyvf2a7', '/mp-cl5w3e7d', '/mp-5uz5o7fj', '/mp-hk8c9j3i', '/mp-mvorp7mg', '/mp-b1zfglgk', '/mp-fst5kv5u', '/mp-e0x_i9jk', '/mp-diqyhh8d', '/mp-526nngpx', '/mp-i98y2yom', '/mp-fqdk1b54', '/mp-9aq0k9c6', '/mp-qoisv5b2', '/mp-q32tdbb1', '/mp-k1rnc1wl', '/mp-2y8hysvq', '/mp-ck3svcur', '/mp-wd7jo7n8', '/mp-8xck53_2', '/mp-m873gi4n', '/mp-tq4cbi_y', '/mp-tgh8dcoi', '/mp-66od6loh', '/mp-m9aeufpd', '/mp-pc7m757x', '/mp-xu4tztyb', '/mp-7rce4ioc', '/mp-j9mnheso', '/mp-t7y8xitg', '/mp-9saa2pxz', '/mp-9mvvt0lk', '/mp-ybevimeq', '/mp-_sgbmaqk', '/mp-cpnqqaq7', '/mp-bncho_w_', '/mp-0fyect4a', '/mp-ff2kmwh_', '/mp-haw7gwyg', '/mp-ptdm3ro_', '/mp-zskfazrp', '/mp-rdud7ior', '/mp-jya8gxwq', '/mp-_0d0zidv', '/mp-gpasdyf5', '/mp-72gmvdvm', '/mp-ftalu3z1', '/mp-dxa4vth7', '/mp-2tbna1s2', '/mp-h6jkyyca', '/mp-ch835s7i', '/mp-eg0m5afd', '/mp-ecxtk27l', '/mp-9k36rewz', '/mp-_fmd91bo', '/mp-27yzuczh', '/mp-bbzl2pe9', '/mp-twae7t3r', '/mp-okd9z7tk', '/mp-25vovthz', '/mp-bbl5mnjo', '/mp-wy73zhk1', '/mp-od9hqrbw', '/mp-0urfmped', '/mp-zfj99bi5', '/mp-23kzq0po', '/mp-rf7kyico', '/mp-_iepokg3', '/mp-4m3k1uae', '/mp-hkniu7a4', '/mp-5orlq1j1', '/mp-z25vndb1', '/mp-haeunqym', '/mp-bt4pbi97', '/mp-bfpj9kon', '/mp-brjxmf25', '/mp-vhem2hn8', '/mp-al164uhz', '/mp-18hmv_o9', '/mp-p52iitiz', '/mp-wjky11fq', '/mp-b5etyq8h', '/mp-jo7y33fs', '/mp-45ocpmms', '/mp-lvm4ljzo', '/mp-7acpzdj7', '/mp-62bdv3s_', '/mp-dkf0a3uu', '/mp-mps389pc', '/mp-lbya173l', '/mp-wbfunu2r', '/mp-yiepee1w', '/mp-d59lb3we', '/mp-y9_p2dpy', '/mp-1uwgj9mc', '/mp-_44n0f53', '/mp-b3m9lacz', '/mp-jj__5s5s', '/mp-h9k5m0sa', '/mp-a6v_wslt', '/mp-tl0enpph', '/mp-r4pmm8fj', '/mp-0iycgfqy', '/mp-13c8w_j8', '/mp-vupxws83', '/mp-c1zz45ee', '/mp-czzjoq_z', '/mp-in8s1re3', '/mp-3p802c_a', '/mp-ol4xcqe4', '/mp-ht36tib5', '/mp-j5wff337', '/mp-i9es15cz', '/mp-fpy0ehne', '/mp-d8wo7vj7', '/mp-e_vz3d19', '/mp-srfkkw9c', '/mp-ldy5ufgs', '/mp-dkvo0ene', '/mp-6qgxupai', '/mp-j3yqdxbt', '/mp-2i5hy163', '/mp-1l6a7zh3', '/mp-2w_yhlgd', '/mp-f8l6m2k3', '/mp-xr03rzhz', '/mp-05ozmgqb', '/mp-advrfojv', '/mp-z1riqmgl', '/mp-acxpquo7', '/mp-s0g9e908', '/mp-c4huryy2', '/mp-ex7mxe7x', '/mp-7qqnsx7h', '/mp-edutnopl', '/mp-n2z5istv', '/mp-etda0pdp', '/mp-rwxevz4x', '/mp-08ugo1vc', '/mp-iepg_q3s', '/mp-ns_0bny7', '/mp-u5n073pu', '/mp-28ba56hn', '/mp-ym77kef0', '/mp-1xfrxt1j', '/mp-ypv3ays0', '/mp-sw1kcm1x', '/mp-bba5v3ys', '/mp-m47yw5te', '/mp-_fg9u7qs', '/mp-c2ezzsgr', '/mp-9l8qr81l', '/mp-9ba0_o8m', '/mp-99h0bqzu', '/mp-4lre072t', '/mp-anhyuf7k', '/mp-0nejjdjs', '/mp-su0h3bxx', '/mp-kfsk25bt', '/mp-lnyse_no', '/mp-5mo9ieeg', '/mp-m8y3cbh0', '/mp-qule0rv9', '/mp-en3_wju2', '/mp-fgvnv7fv', '/mp-swvpjcqx', '/mp-08ndi5cc', '/mp-q4xzrdo9', '/mp-vvpda6wc', '/mp-a1ir2slc', '/mp-5y32ye5b', '/mp-5f8o1z8p', '/mp-29mz7yzy', '/mp-d7nv4t9r', '/mp-dq_oyizk', '/mp-6b8ecgi5', '/mp-_1ngjvcc', '/mp-p70a3i4z', '/mp-7lotqq4e', '/mp-glzec5p_', '/mp-kia5efp7', '/mp-mcnxg40u', '/mp-g8dfjm_l', '/mp-nin_k4qm', '/mp-0mxg1h70', '/mp-fhj4iqcr', '/mp-ykfn_a9q', '/mp-qkib343f', '/mp-m_61cycp', '/mp-5g_18tqz', '/mp-a3wtlw34', '/mp-ngt2e3ob', '/mp-hopw0wh_', '/mp-y4545mlh', '/mp-7su14s7u', '/mp-wwp5sxf3', '/mp-8xm0ogph', '/mp-tq5yujbu', '/mp-g9fpa5kc', '/mp-pq639u0o', '/mp-e2zta6gm', '/mp-9zz373c2', '/mp-6nroui_z', '/mp-ok7r57wb', '/mp-rma6r73_', '/mp-7ci82iio', '/mp-1vfnejlb', '/mp-bgmd709p', '/mp-umberfks', '/mp-011a6_s9', '/mp-pqe8_94w', '/mp-ztoi2cmu', '/mp-jwwszs6f', '/mp-zak1d8oh', '/mp-4cfh9g1h', '/mp-tn3uss5s', '/mp-4i3yfqhf', '/mp-bqgg4188', '/mp-se8k1wq0', '/mp-gje3shqa', '/mp-5138e698'} warnings.warn( [mixed_ct_afv_unified_scoring_v2_resenc +0.0s] Starting training (regression): device=cuda, workers=64 [mixed_ct_afv_unified_scoring_v2_resenc +2.2s] Resuming: init_weights only applies to a new run and is ignored [mixed_ct_afv_unified_scoring_v2_resenc +2.2s] Loading manifest and fiber annotations [mixed_ct_afv_unified_scoring_v2_resenc +46.3s] Loaded 563 training fibers and 188 validation fibers [mixed_ct_afv_unified_scoring_v2_resenc +46.4s] Preparing per-crop CT z-score normalization CT per-crop z-score, epsilon 1e-06: s3://vesuvius-challenge-open-data/PHercParis4/volumes/20260411134726-2.400um-0.2m-78keV-masked.zarr::2 CT per-crop z-score, epsilon 1e-06: s3://vesuvius-challenge-open-data/PHerc0175A/volumes/20250521115057-8.640um-1.2m-116keV-masked.zarr::0 CT per-crop z-score, epsilon 1e-06: s3://vesuvius-challenge-open-data/PHerc1447/volumes/20250521151220-8.640um-1.2m-116keV-masked.zarr::0 CT per-crop z-score, epsilon 1e-06: s3://vesuvius-challenge-open-data/PHercParis4/volumes/20260411134726-2.400um-0.2m-78keV-masked.zarr::2 CT per-crop z-score, epsilon 1e-06: s3://vesuvius-challenge-open-data/PHerc0125/volumes/20250821151825-9.362um-1.2m-113keV-masked.zarr::0 [mixed_ct_afv_unified_scoring_v2_resenc +46.4s] Preparing monitor recovery fixture [mixed_ct_afv_unified_scoring_v2_resenc +46.4s] Initializing models and optimizer [mixed_ct_afv_unified_scoring_v2_resenc +46.4s] Gradient clipping at global norm 60 (0 disables clipping) [mixed_ct_afv_unified_scoring_v2_resenc +46.4s] Refinement loss: every pass for every decision; initial proposal full weight, passes x1 of their mean; selection and the attempt metric follow confidence 0.7 [mixed_ct_afv_unified_scoring_v2_resenc +48.1s] Compiling training operations; first forward/backward passes may take several minutes [mixed_ct_afv_unified_scoring_v2_resenc +48.1s] Loading replay banks and preparing data loader [mixed_ct_afv_unified_scoring_v2_resenc +48.2s] 32 independent decisions per batch [mixed_ct_afv_unified_scoring_v2_resenc +48.4s] Checking mixed-source datasets and AFV checksums [mixed_ct_afv_unified_scoring_v2_resenc +75.4s] Dataset sampling: paris4=20.0%, 0175A_5mm_v1=20.0%, 1447_5mm_v1=20.0%, paris4_afv_central=20.0%, PHerc0125_3mm_v2=20.0%; AFV training draws and AFV collection order proportional to length**3 [mixed_ct_afv_unified_scoring_v2_resenc +75.4s] Task budget per source: fresh 30%, live 45%, dagger_pre_excursion 6%, dagger_recoverable 8%, dagger_terminal 8%, dagger_premature_stop 3%, dagger_ordinary 0%, synthetic_terminal 0%; replay age ceiling 3000, event cap 64, synthetic terminal fallback cap 0% [mixed_ct_afv_unified_scoring_v2_resenc +75.4s] paris4: fresh locations weighted (bend x2, compression x2) for 563/563 fibers from paris4_3cfaec19b98d_v1.npz [mixed_ct_afv_unified_scoring_v2_resenc +75.4s] paris4: restored replay draw counts for 378 events [mixed_ct_afv_unified_scoring_v2_resenc +75.7s] 0175A_5mm_v1: fresh locations weighted (bend x2, compression x2) for 60894/60894 fibers from 0175A_5mm_v1_d4374c646e4f_v1.npz [mixed_ct_afv_unified_scoring_v2_resenc +75.7s] 0175A_5mm_v1: restored replay draw counts for 282 events [mixed_ct_afv_unified_scoring_v2_resenc +76.2s] 1447_5mm_v1: fresh locations weighted (bend x2, compression x2) for 80356/80356 fibers from 1447_5mm_v1_e235b5235b3d_v1.npz [mixed_ct_afv_unified_scoring_v2_resenc +76.2s] 1447_5mm_v1: restored replay draw counts for 223 events [mixed_ct_afv_unified_scoring_v2_resenc +76.6s] paris4_afv_central: fresh locations weighted (bend x2, compression x2) for 83004/83004 fibers from paris4_afv_central_bf27920bc12f_v1.npz [mixed_ct_afv_unified_scoring_v2_resenc +76.6s] paris4_afv_central: restored replay draw counts for 245 events [mixed_ct_afv_unified_scoring_v2_resenc +76.8s] PHerc0125_3mm_v2: fresh locations weighted (bend x2, compression x2) for 41893/41893 fibers from PHerc0125_3mm_v2_432a7fb7d155_v1.npz [mixed_ct_afv_unified_scoring_v2_resenc +76.8s] PHerc0125_3mm_v2: restored replay draw counts for 381 events Step 10,000 | event: "ct_normalization" | calibration: {"method": "crop_zscore_v1", "volumes": {"s3://vesuvius-challenge-open-data/PHercParis4/volumes/20260411134726-2.400um-0.2m-78keV-masked.zarr::2": {"method": "crop_zscore_v1", "volume": "s3://vesuvius-challenge-open-data/PHercParis4/volumes/20260411134726-2.400um-0.2m-78keV-masked.zarr::2", "epsilon": 1e-06, "shape": [18946, 8174, 8174], "chunks": [128, 128, 128], "dtype": "|u1"}, "s3://vesuvius-challenge-open-data/PHerc0175A/volumes/20250521115057-8.640um-1.2m-116keV-masked.zarr::0": {"method": "crop_zscore_v1", "volume": "s3://vesuvius-challenge-open-data/PHerc0175A/volumes/20250521115057-8.640um-1.2m-116keV-masked.zarr::0", "epsilon": 1e-06, "shape": [12748, 9363, 9363], "chunks": [128, 128, 128], "dtype": "|u1"}, "s3://vesuvius-challenge-open-data/PHerc1447/volumes/20250521151220-8.640um-1.2m-116keV-masked.zarr::0": {"method": "crop_zscore_v1", "volume": "s3://vesuvius-challenge-open-data/PHerc1447/volumes/20250521151220-8.640um-1.2m-116keV-masked.zarr::0", "epsilon": 1e-06, "shape": [24297, 8343, 8343], "chunks": [128, 128, 128], "dtype": "|u1"}, "s3://vesuvius-challenge-open-data/PHerc0125/volumes/20250821151825-9.362um-1.2m-113keV-masked.zarr::0": {"method": "crop_zscore_v1", "volume": "s3://vesuvius-challenge-open-data/PHerc0125/volumes/20250821151825-9.362um-1.2m-113keV-masked.zarr::0", "epsilon": 1e-06, "shape": [20840, 8387, 8387], "chunks": [128, 128, 128], "dtype": "|u1"}}} Step 10,000 | event: "dataset_configuration" | datasets: [{"name": "paris4", "kind": "paris4", "fibers": 563}, {"name": "0175A_5mm_v1", "kind": "afv", "fibers": 60894, "excluded_fibers": 406, "sha256": "d4374c646e4fb9020ea0b76b8254bc1586cdedbdaa170278ce97fceea41ffe4e", "coordinate_space": "PHerc0175A/20250521115057", "validation": {"strategy": "fiber_hash", "seed": 7349, "count": 406}, "grid_scale": 2.0}, {"name": "1447_5mm_v1", "kind": "afv", "fibers": 80356, "excluded_fibers": 406, "sha256": "e235b5235b3dfef879037a0267e7a50f5a2ba791ffdcc16c2fade3a306705fa2", "coordinate_space": "PHerc1447/20250521151220", "validation": {"strategy": "fiber_hash", "seed": 7349, "count": 406}, "grid_scale": 2.0}, {"name": "paris4_afv_central", "kind": "afv", "fibers": 83004, "excluded_fibers": 406, "sha256": "bf27920bc12f96baebfbe17c82bf4655efe67e48809c87bddcd0a1e2e01c68f6", "coordinate_space": "PHercParis4/20260411134726", "validation": {"strategy": "fiber_hash", "seed": 7349, "count": 406}, "grid_scale": 8.0}, {"name": "PHerc0125_3mm_v2", "kind": "afv", "fibers": 41893, "excluded_fibers": 406, "sha256": "432a7fb7d155046d319da87d27b05dff1cb81baae623a50af82357c1ac02649a", "coordinate_space": "PHerc0125/20250821151825", "validation": {"strategy": "fiber_hash", "seed": 7349, "count": 406}, "grid_scale": 2.0}] | names: ["paris4", "0175A_5mm_v1", "1447_5mm_v1", "paris4_afv_central", "PHerc0125_3mm_v2"] | probabilities: [0.2, 0.2, 0.2, 0.2, 0.2] | dataset_config_sha256: "754d0a401da58c93f53f8520c61439d7d5752afb1f241b3c44e839f7da921588" | input_mode: "ct" | evaluation_scope: "Source-specific held-out fibers; Paris 4 recovery fixture" Step 10,000 | resumed /ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/ckpt_010000.pt commit 8 | batch 32 / grad steps 1 | causal survival confidence Step 10,000 | event: "optimizer_configuration" | reset: false | data_parallel_processes: 1 | lr_restart_step: 0 | optimizer_state_entries: 277 | groups: [{"parameters": 277}] | trainable_parameters: 22466691 Step 10,000 | identity sampling (settings in the run configuration) [mixed_ct_afv_unified_scoring_v2_resenc +76.8s] Remote CT prefetch: 48 concurrent fetches, 512 requests per priority queue, 16 future batch plans per remote source/worker, separate async process Step 10,000 | event: "remote_prefetch_configuration" | enabled: true | connections: 48 | queue_size: 512 | lookahead: 16 | timeout: 120.0 | sources: 5 [mixed_ct_afv_unified_scoring_v2_resenc +76.8s] Starting data loader; batch 32 × grad steps 1 = 32 decisions per update; starting update 10001 [mixed_ct_afv_unified_scoring_v2_resenc +431.8s] Update 10001: data ready in 23.8s; running forward/backward and optimizer [mixed_ct_afv_unified_scoring_v2_resenc +453.6s] Update 10001 complete in 21.8s; loss=1.19615 Step 10,001 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 244, "stopped": 0, "mean_example_error": 0.77965940590258, "geometry_examples": 27, "examples": 32, "inference_and_measurement_seconds": 10.101836174726486, "total_diagnostic_seconds": 12.252800017595291} Step 10,001 | batch_diagnostics_seconds: 12.255876667797565 [mixed_ct_afv_unified_scoring_v2_resenc +465.9s] Update 10002: waiting for data [mixed_ct_afv_unified_scoring_v2_resenc +465.9s] Update 10002: data ready in 0.0s; running forward/backward and optimizer [mixed_ct_afv_unified_scoring_v2_resenc +467.2s] Update 10002 complete in 1.3s; loss=1.11359 [mixed_ct_afv_unified_scoring_v2_resenc +467.2s] Update 10003: waiting for data [mixed_ct_afv_unified_scoring_v2_resenc +467.2s] Update 10003: data ready in 0.0s; running forward/backward and optimizer [mixed_ct_afv_unified_scoring_v2_resenc +468.4s] Update 10003 complete in 1.2s; loss=1.68521 [mixed_ct_afv_unified_scoring_v2_resenc +468.4s] Update 10004: waiting for data [mixed_ct_afv_unified_scoring_v2_resenc +468.4s] Update 10004: data ready in 0.0s; running forward/backward and optimizer [mixed_ct_afv_unified_scoring_v2_resenc +469.3s] Update 10004 complete in 0.9s; loss=1.66704 [mixed_ct_afv_unified_scoring_v2_resenc +469.3s] Update 10005: waiting for data [mixed_ct_afv_unified_scoring_v2_resenc +469.3s] Update 10005: data ready in 0.0s; running forward/backward and optimizer [mixed_ct_afv_unified_scoring_v2_resenc +470.2s] Update 10005 complete in 0.9s; loss=0.95313 [mixed_ct_afv_unified_scoring_v2_resenc +470.2s] Startup progress complete; regular metrics every 50 updates Step 10,050 | last 50 updates / 1,600 crops | lr 2.24e-05 points (35/curve, tolerance 2 vox): 41,965 right / 5,521 wrong | 88.4% correct | 8,514 unknown loss 1.3244 | geometry 0.4702 | confidence 1.7085 | mean error 0.834 vox speed: 1323 ms/update | data wait 477 ms/update | 3.7 crops/s | 3.7 decisions/s | 1.62 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1260 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 777, from the previous prediction 483 confidence-labeled crops: terminal 45/1493 (3.0%) | recoverable 142/1493 (9.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 224/1600 (14.0%), 2: 416/1600 (26.0%), 3: 320/1600 (20.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 24 chunks / 48.0 MiB | promoted 9 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 55% 45% 25%>0% 25% 0 -- 91%/8%/1%/1% 8% 0/0/0/174 0175A_5mm_v1 224 38% 49% 18%>13% 4% 35 2028 88%/4%/4%/4% 10% 0/0/0/110 1447_5mm_v1 416 31% 46% 24%>24% 0% 64 1021 72%/10%/5%/13% 24% 0/0/0/190 paris4_afv_central 320 26% 46% 28%>28% 0% 86 32 72%/14%/4%/9% 34% 0/0/0/147 PHerc0125_3mm_v2 256 55% 45% 26%>0% 26% 0 -- 87%/6%/0%/7% 12% 0/0/0/116 gradients: max 66, clipped 2/50 updates Step 10,100 | last 50 updates / 1,600 crops | lr 2.24e-05 points (35/curve, tolerance 2 vox): 41,983 right / 4,957 wrong | 89.4% correct | 9,060 unknown loss 1.2696 | geometry 0.4148 | confidence 1.7096 | mean error 0.814 vox speed: 896 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.65 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1289 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 792, from the previous prediction 497 confidence-labeled crops: terminal 40/1472 (2.7%) | recoverable 120/1472 (8.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 256/1600 (16.0%), 2: 288/1600 (18.0%), 3: 320/1600 (20.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 27 chunks / 54.0 MiB | promoted 9 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 54% 46% 25%>0% 25% 0 -- 90%/6%/0%/5% 7% 0/0/0/133 0175A_5mm_v1 256 37% 47% 24%>16% 7% 46 2087 84%/4%/5%/7% 14% 0/0/0/120 1447_5mm_v1 288 31% 45% 25%>25% 0% 50 1080 77%/9%/5%/9% 20% 0/0/0/129 paris4_afv_central 320 32% 43% 25%>25% 0% 84 78 71%/14%/4%/11% 30% 0/0/0/138 PHerc0125_3mm_v2 448 51% 49% 22%>0% 22% 0 -- 85%/5%/0%/10% 13% 0/0/0/221 gradients: max 53, clipped 0/50 updates Step 10,150 | last 50 updates / 1,600 crops | lr 2.23e-05 points (35/curve, tolerance 2 vox): 41,733 right / 5,026 wrong | 89.3% correct | 9,241 unknown loss 1.2229 | geometry 0.3991 | confidence 1.6476 | mean error 0.792 vox speed: 893 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.72 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1282 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 619, from the previous prediction 663 confidence-labeled crops: terminal 43/1487 (2.9%) | recoverable 133/1487 (8.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 192/1600 (12.0%), 2: 512/1600 (32.0%), 3: 288/1600 (18.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 27 chunks / 54.0 MiB | promoted 9 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 173 rows | mean depth 1 decisions | mean travel 35 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 57% 43% 29%>0% 29% 0 129 86%/11%/0%/3% 7% 24/0/0/59 0175A_5mm_v1 192 34% 54% 18%>12% 6% 35 2120 86%/4%/2%/8% 13% 0/0/0/103 1447_5mm_v1 512 34% 47% 19%>19% 0% 71 788 80%/7%/4%/8% 20% 79/4/0/161 paris4_afv_central 288 28% 47% 25%>25% 0% 84 125 75%/11%/6%/8% 28% 20/1/0/115 PHerc0125_3mm_v2 416 58% 42% 25%>0% 25% 0 132 83%/8%/0%/9% 11% 50/1/0/123 gradients: max 70, clipped 1/50 updates Step 10,200 | last 50 updates / 1,600 crops | lr 2.22e-05 points (35/curve, tolerance 2 vox): 44,482 right / 4,442 wrong | 90.9% correct | 7,076 unknown loss 1.1121 | geometry 0.3199 | confidence 1.5845 | mean error 0.715 vox speed: 892 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.57 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1320 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 310, from the previous prediction 1010 confidence-labeled crops: terminal 53/1518 (3.5%) | recoverable 121/1518 (8.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 448/1600 (28.0%), 2: 288/1600 (18.0%), 3: 384/1600 (24.0%), 4: 96/1600 (6.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 27 chunks / 54.0 MiB | promoted 9 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 542 rows | mean depth 1 decisions | mean travel 16 vox | 7 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 52% 48% 22%>0% 22% 0 138 90%/8%/0%/2% 7% 183/4/0/0 0175A_5mm_v1 448 39% 46% 25%>16% 9% 60 1193 81%/8%/5%/6% 13% 98/0/0/106 1447_5mm_v1 288 39% 44% 20%>17% 3% 47 522 85%/5%/3%/6% 11% 97/2/0/31 paris4_afv_central 384 30% 43% 26%>26% 0% 84 151 78%/10%/5%/7% 25% 114/2/0/53 PHerc0125_3mm_v2 96 48% 52% 17%>0% 17% 0 143 86%/4%/1%/8% 10% 50/1/0/0 gradients: max 55, clipped 0/50 updates Step 10,250 | last 50 updates / 1,600 crops | lr 2.22e-05 points (35/curve, tolerance 2 vox): 43,793 right / 4,873 wrong | 90.0% correct | 7,334 unknown loss 1.1318 | geometry 0.3323 | confidence 1.5989 | mean error 0.740 vox speed: 894 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.54 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1315 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 383, from the previous prediction 932 confidence-labeled crops: terminal 58/1515 (3.8%) | recoverable 113/1515 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 384/1600 (24.0%), 2: 224/1600 (14.0%), 3: 384/1600 (24.0%), 4: 160/1600 (10.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 31 chunks / 62.0 MiB | promoted 9 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 450 rows | mean depth 1 decisions | mean travel 22 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 53% 47% 22%>0% 22% 0 137 89%/8%/0%/3% 7% 102/1/0/110 0175A_5mm_v1 384 33% 50% 25%>17% 8% 58 1112 84%/5%/7%/4% 15% 101/2/0/90 1447_5mm_v1 224 29% 46% 30%>25% 6% 43 584 80%/6%/5%/8% 15% 86/1/0/18 paris4_afv_central 384 28% 44% 28%>28% 0% 93 190 75%/10%/5%/10% 24% 85/1/0/85 PHerc0125_3mm_v2 160 52% 48% 18%>0% 18% 0 166 92%/6%/0%/2% 10% 76/0/18/1 gradients: max 52, clipped 0/50 updates Step 10,300 | last 50 updates / 1,600 crops | lr 2.21e-05 points (35/curve, tolerance 2 vox): 44,196 right / 4,378 wrong | 91.0% correct | 7,426 unknown loss 0.9843 | geometry 0.2705 | confidence 1.4276 | mean error 0.687 vox speed: 893 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.48 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1410 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 278, from the previous prediction 1132 confidence-labeled crops: terminal 57/1497 (3.8%) | recoverable 99/1497 (6.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 320/1600 (20.0%), 2: 352/1600 (22.0%), 3: 224/1600 (14.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 32 chunks / 64.0 MiB | promoted 10 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 583 rows | mean depth 2 decisions | mean travel 24 vox | 15 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 56% 44% 26%>0% 26% 0 134 89%/8%/0%/2% 9% 99/1/0/41 0175A_5mm_v1 320 37% 44% 26%>18% 8% 49 1228 81%/4%/7%/8% 16% 75/2/0/67 1447_5mm_v1 352 39% 47% 24%>15% 9% 37 415 84%/6%/5%/6% 10% 164/4/0/0 paris4_afv_central 224 33% 41% 27%>26% 0% 51 192 79%/5%/6%/9% 20% 92/0/0/0 PHerc0125_3mm_v2 384 57% 43% 25%>0% 25% 0 152 84%/7%/2%/8% 13% 153/2/24/13 gradients: max 54, clipped 0/50 updates Step 10,350 | last 50 updates / 1,600 crops | lr 2.20e-05 points (35/curve, tolerance 2 vox): 45,214 right / 4,166 wrong | 91.6% correct | 6,620 unknown loss 0.9954 | geometry 0.3039 | confidence 1.3830 | mean error 0.689 vox speed: 895 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.49 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1371 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 289, from the previous prediction 1082 confidence-labeled crops: terminal 53/1504 (3.5%) | recoverable 120/1504 (8.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 224/1600 (14.0%), 2: 416/1600 (26.0%), 3: 480/1600 (30.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 44 chunks / 88.0 MiB | promoted 12 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 567 rows | mean depth 2 decisions | mean travel 23 vox | 7 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 58% 42% 28%>0% 28% 0 146 88%/9%/0%/2% 7% 122/3/0/0 0175A_5mm_v1 224 42% 38% 30%>20% 10% 37 894 77%/7%/7%/9% 13% 86/4/0/0 1447_5mm_v1 416 42% 47% 23%>11% 12% 35 419 84%/7%/4%/5% 9% 154/2/0/43 paris4_afv_central 480 29% 48% 27%>24% 3% 99 230 79%/8%/4%/9% 19% 147/9/0/81 PHerc0125_3mm_v2 192 53% 47% 24%>0% 24% 0 132 90%/5%/0%/5% 12% 58/3/0/33 gradients: max 46, clipped 0/50 updates Step 10,400 | last 50 updates / 1,600 crops | lr 2.20e-05 points (35/curve, tolerance 2 vox): 45,017 right / 3,762 wrong | 92.3% correct | 7,221 unknown loss 0.9836 | geometry 0.2677 | confidence 1.4318 | mean error 0.675 vox speed: 894 ms/update | data wait 1 ms/update | 35.1 crops/s | 35.1 decisions/s | 1.41 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1429 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 402, from the previous prediction 1027 confidence-labeled crops: terminal 42/1503 (2.8%) | recoverable 107/1503 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 320/1600 (20.0%), 2: 416/1600 (26.0%), 3: 288/1600 (18.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 44 chunks / 88.0 MiB | promoted 12 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 465 rows | mean depth 2 decisions | mean travel 24 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 54% 46% 25%>0% 25% 0 158 88%/8%/0%/4% 9% 117/0/0/0 0175A_5mm_v1 320 40% 48% 23%>12% 11% 32 709 85%/7%/3%/6% 11% 127/2/74/26 1447_5mm_v1 416 44% 43% 25%>13% 12% 49 722 82%/7%/4%/7% 9% 70/1/0/109 paris4_afv_central 288 34% 47% 25%>18% 6% 55 229 82%/7%/5%/6% 16% 93/2/0/43 PHerc0125_3mm_v2 320 52% 48% 25%>0% 25% 0 131 88%/5%/1%/7% 10% 58/1/0/95 gradients: max 56, clipped 0/50 updates Step 10,450 | last 50 updates / 1,600 crops | lr 2.19e-05 points (35/curve, tolerance 2 vox): 44,748 right / 3,936 wrong | 91.9% correct | 7,316 unknown loss 0.9671 | geometry 0.2903 | confidence 1.3535 | mean error 0.683 vox speed: 890 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.44 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1400 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 369, from the previous prediction 1031 confidence-labeled crops: terminal 51/1509 (3.4%) | recoverable 106/1509 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 384/1600 (24.0%), 2: 320/1600 (20.0%), 3: 448/1600 (28.0%), 4: 160/1600 (10.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 44 chunks / 88.0 MiB | promoted 12 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 515 rows | mean depth 3 decisions | mean travel 24 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 50% 50% 20%>0% 20% 0 141 93%/5%/0%/2% 8% 137/1/102/6 0175A_5mm_v1 384 43% 47% 21%>10% 12% 30 978 83%/6%/3%/7% 9% 68/4/54/114 1447_5mm_v1 320 39% 44% 31%>17% 14% 41 544 80%/8%/5%/6% 10% 121/4/0/19 paris4_afv_central 448 35% 49% 22%>17% 6% 80 264 81%/7%/5%/7% 16% 120/1/0/99 PHerc0125_3mm_v2 160 57% 43% 23%>0% 23% 0 164 88%/6%/0%/6% 14% 69/1/0/0 gradients: max 52, clipped 0/50 updates Step 10,500 | last 50 updates / 1,600 crops | lr 2.18e-05 points (35/curve, tolerance 2 vox): 45,340 right / 3,841 wrong | 92.2% correct | 6,819 unknown loss 0.9750 | geometry 0.2899 | confidence 1.3702 | mean error 0.700 vox speed: 893 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.41 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1446 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 268, from the previous prediction 1178 confidence-labeled crops: terminal 44/1512 (2.9%) | recoverable 127/1512 (8.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 416/1600 (26.0%), 2: 320/1600 (20.0%), 3: 256/1600 (16.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 44 chunks / 88.0 MiB | promoted 12 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 618 rows | mean depth 3 decisions | mean travel 24 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 56% 44% 26%>0% 26% 0 142 87%/9%/0%/4% 9% 98/1/0/0 0175A_5mm_v1 416 40% 48% 24%>12% 12% 50 1221 79%/8%/6%/8% 12% 79/1/0/120 1447_5mm_v1 320 40% 47% 26%>13% 12% 28 439 82%/7%/4%/7% 10% 149/3/0/0 paris4_afv_central 256 38% 45% 23%>17% 7% 40 227 83%/10%/2%/5% 11% 116/2/0/0 PHerc0125_3mm_v2 384 54% 46% 25%>0% 25% 0 158 89%/7%/1%/4% 13% 176/6/35/1 gradients: max 69, clipped 1/50 updates Step 10,550 | last 50 updates / 1,600 crops | lr 2.17e-05 points (35/curve, tolerance 2 vox): 46,360 right / 3,870 wrong | 92.3% correct | 5,770 unknown loss 0.9430 | geometry 0.2828 | confidence 1.3205 | mean error 0.686 vox speed: 895 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.36 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1473 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 238, from the previous prediction 1235 confidence-labeled crops: terminal 42/1522 (2.8%) | recoverable 89/1522 (5.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 192/1600 (12.0%), 2: 448/1600 (28.0%), 3: 352/1600 (22.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 44 chunks / 88.0 MiB | promoted 12 | preempted 0 | deferred 0 | errors 0 | lookahead 1391 chunk references / 1 windows live chains: 659 rows | mean depth 3 decisions | mean travel 26 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 49% 51% 22%>0% 22% 0 137 92%/5%/0%/2% 8% 143/4/18/20 0175A_5mm_v1 192 43% 46% 26%>11% 15% 18 669 80%/7%/7%/6% 14% 77/2/0/11 1447_5mm_v1 448 48% 43% 23%>9% 14% 30 376 89%/4%/3%/4% 7% 192/6/38/1 paris4_afv_central 352 37% 44% 24%>19% 6% 54 261 81%/7%/5%/8% 15% 156/3/13/0 PHerc0125_3mm_v2 288 56% 44% 23%>0% 23% 0 143 89%/6%/0%/6% 13% 91/1/0/37 gradients: max 54, clipped 0/50 updates Step 10,600 | last 50 updates / 1,600 crops | lr 2.17e-05 points (35/curve, tolerance 2 vox): 46,558 right / 3,521 wrong | 93.0% correct | 5,921 unknown loss 0.9663 | geometry 0.2754 | confidence 1.3818 | mean error 0.696 vox speed: 893 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.32 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1485 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 267, from the previous prediction 1218 confidence-labeled crops: terminal 38/1530 (2.5%) | recoverable 100/1530 (6.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 288/1600 (18.0%), 2: 288/1600 (18.0%), 3: 320/1600 (20.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 55 chunks / 110.0 MiB | promoted 12 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 638 rows | mean depth 3 decisions | mean travel 28 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 55% 45% 24%>0% 24% 0 141 90%/7%/0%/2% 10% 103/2/0/12 0175A_5mm_v1 288 42% 45% 30%>14% 16% 26 708 81%/7%/6%/6% 12% 128/2/18/1 1447_5mm_v1 288 48% 44% 24%>9% 16% 21 373 85%/7%/1%/7% 4% 126/2/0/0 paris4_afv_central 320 34% 50% 23%>16% 7% 45 239 86%/6%/4%/3% 13% 160/5/0/0 PHerc0125_3mm_v2 448 52% 48% 23%>0% 23% 0 129 91%/4%/0%/4% 11% 121/1/0/95 gradients: max 61, clipped 1/50 updates Step 10,650 | last 50 updates / 1,600 crops | lr 2.16e-05 points (35/curve, tolerance 2 vox): 46,160 right / 3,745 wrong | 92.5% correct | 6,095 unknown loss 1.0249 | geometry 0.3199 | confidence 1.4101 | mean error 0.715 vox speed: 889 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.34 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1449 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 249, from the previous prediction 1200 confidence-labeled crops: terminal 36/1527 (2.4%) | recoverable 128/1527 (8.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 256/1600 (16.0%), 2: 256/1600 (16.0%), 3: 416/1600 (26.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 55 chunks / 110.0 MiB | promoted 12 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 608 rows | mean depth 3 decisions | mean travel 29 vox | 7 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 57% 43% 28%>0% 28% 0 135 91%/8%/0%/1% 8% 105/1/0/32 0175A_5mm_v1 256 49% 41% 24%>10% 14% 22 631 87%/7%/2%/4% 7% 102/0/73/4 1447_5mm_v1 256 50% 41% 29%>9% 20% 21 409 83%/9%/2%/6% 6% 104/2/0/0 paris4_afv_central 416 33% 46% 26%>21% 5% 81 347 80%/7%/6%/6% 19% 139/4/31/51 PHerc0125_3mm_v2 352 54% 46% 27%>0% 27% 0 127 86%/8%/0%/5% 12% 158/3/0/3 gradients: max 48, clipped 0/50 updates Step 10,700 | last 50 updates / 1,600 crops | lr 2.15e-05 points (35/curve, tolerance 2 vox): 47,274 right / 3,135 wrong | 93.8% correct | 5,591 unknown loss 0.9938 | geometry 0.2958 | confidence 1.3960 | mean error 0.686 vox speed: 892 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.30 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1484 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 333, from the previous prediction 1151 confidence-labeled crops: terminal 25/1525 (1.6%) | recoverable 129/1525 (8.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 384/1600 (24.0%), 2: 256/1600 (16.0%), 3: 224/1600 (14.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 55 chunks / 110.0 MiB | promoted 12 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 599 rows | mean depth 3 decisions | mean travel 30 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 54% 46% 26%>0% 26% 0 141 90%/8%/0%/2% 10% 127/2/0/35 0175A_5mm_v1 384 43% 49% 21%>7% 14% 28 754 87%/6%/2%/4% 7% 100/2/0/89 1447_5mm_v1 256 43% 47% 24%>10% 14% 21 408 83%/8%/4%/5% 7% 121/1/0/0 paris4_afv_central 224 29% 46% 31%>26% 5% 47 336 76%/13%/2%/8% 15% 98/0/0/4 PHerc0125_3mm_v2 384 52% 48% 24%>0% 24% 0 131 87%/8%/0%/5% 13% 153/2/0/33 gradients: max 67, clipped 1/50 updates Step 10,750 | last 50 updates / 1,600 crops | lr 2.15e-05 points (35/curve, tolerance 2 vox): 46,398 right / 3,873 wrong | 92.3% correct | 5,729 unknown loss 0.9270 | geometry 0.2927 | confidence 1.2685 | mean error 0.681 vox speed: 892 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.36 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1442 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 257, from the previous prediction 1185 confidence-labeled crops: terminal 51/1530 (3.3%) | recoverable 116/1530 (7.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 480/1600 (30.0%), 1: 352/1600 (22.0%), 2: 320/1600 (20.0%), 3: 384/1600 (24.0%), 4: 64/1600 (4.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 58 chunks / 116.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 619 rows | mean depth 4 decisions | mean travel 34 vox | 6 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 480 50% 50% 23%>0% 23% 0 135 89%/7%/0%/4% 9% 160/1/0/78 0175A_5mm_v1 352 47% 42% 27%>11% 16% 31 726 82%/8%/5%/5% 11% 128/3/0/19 1447_5mm_v1 320 40% 46% 26%>14% 12% 38 544 84%/8%/3%/5% 9% 146/1/0/0 paris4_afv_central 384 39% 42% 25%>19% 6% 57 325 82%/7%/6%/5% 16% 161/5/0/0 PHerc0125_3mm_v2 64 62% 38% 27%>0% 27% 0 149 89%/6%/0%/5% 13% 24/1/0/0 gradients: max 77, clipped 4/50 updates Step 10,800 | last 50 updates / 1,600 crops | lr 2.14e-05 points (35/curve, tolerance 2 vox): 46,097 right / 3,940 wrong | 92.1% correct | 5,963 unknown loss 0.9699 | geometry 0.2897 | confidence 1.3604 | mean error 0.683 vox speed: 892 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.38 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1456 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 244, from the previous prediction 1212 confidence-labeled crops: terminal 44/1522 (2.9%) | recoverable 124/1522 (8.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 384/1600 (24.0%), 2: 320/1600 (20.0%), 3: 288/1600 (18.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 58 chunks / 116.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 635 rows | mean depth 4 decisions | mean travel 35 vox | 6 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 49% 51% 22%>0% 22% 0 139 91%/7%/0%/2% 9% 95/3/0/36 0175A_5mm_v1 384 45% 46% 23%>10% 13% 37 884 84%/8%/2%/6% 8% 115/5/0/60 1447_5mm_v1 320 43% 45% 26%>12% 15% 28 481 82%/7%/5%/5% 8% 144/2/0/0 paris4_afv_central 288 34% 45% 26%>21% 5% 55 348 80%/8%/6%/7% 17% 130/3/0/0 PHerc0125_3mm_v2 352 57% 43% 30%>0% 30% 0 162 86%/9%/1%/5% 14% 151/2/0/0 gradients: max 47, clipped 0/50 updates Step 10,850 | last 50 updates / 1,600 crops | lr 2.13e-05 points (35/curve, tolerance 2 vox): 46,369 right / 3,255 wrong | 93.4% correct | 6,376 unknown loss 0.9797 | geometry 0.2544 | confidence 1.4506 | mean error 0.674 vox speed: 893 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.36 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1495 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 202, from the previous prediction 1293 confidence-labeled crops: terminal 32/1520 (2.1%) | recoverable 110/1520 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 320/1600 (20.0%), 2: 352/1600 (22.0%), 3: 224/1600 (14.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 61 chunks / 122.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 668 rows | mean depth 4 decisions | mean travel 36 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 53% 47% 22%>0% 22% 0 144 91%/7%/0%/2% 8% 134/0/0/0 0175A_5mm_v1 320 46% 45% 26%>8% 18% 24 560 85%/9%/3%/2% 10% 145/2/0/0 1447_5mm_v1 352 46% 45% 24%>9% 15% 27 492 86%/5%/2%/7% 8% 119/0/14/38 paris4_afv_central 224 33% 46% 29%>21% 8% 41 356 77%/8%/6%/9% 15% 104/1/12/0 PHerc0125_3mm_v2 416 60% 40% 29%>0% 29% 0 151 88%/6%/0%/5% 12% 166/3/10/0 gradients: max 58, clipped 0/50 updates Step 10,900 | last 50 updates / 1,600 crops | lr 2.12e-05 points (35/curve, tolerance 2 vox): 48,317 right / 2,335 wrong | 95.4% correct | 5,348 unknown loss 0.8365 | geometry 0.2268 | confidence 1.2192 | mean error 0.616 vox speed: 896 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1516 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 283, from the previous prediction 1233 confidence-labeled crops: terminal 20/1533 (1.3%) | recoverable 106/1533 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 384/1600 (24.0%), 2: 352/1600 (22.0%), 3: 224/1600 (14.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 63 chunks / 126.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 627 rows | mean depth 5 decisions | mean travel 38 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 52% 48% 24%>0% 24% 0 151 91%/8%/0%/1% 8% 124/2/0/0 0175A_5mm_v1 384 45% 49% 22%>5% 16% 19 410 89%/6%/1%/3% 5% 188/6/0/2 1447_5mm_v1 352 47% 46% 24%>7% 18% 22 375 88%/7%/2%/4% 3% 141/2/0/21 paris4_afv_central 224 40% 42% 22%>18% 4% 36 376 85%/5%/4%/7% 13% 94/4/0/0 PHerc0125_3mm_v2 384 52% 48% 23%>0% 23% 0 133 87%/7%/1%/6% 11% 80/2/52/104 gradients: max 67, clipped 2/50 updates Step 10,950 | last 50 updates / 1,600 crops | lr 2.12e-05 points (35/curve, tolerance 2 vox): 46,094 right / 4,188 wrong | 91.7% correct | 5,718 unknown loss 0.9267 | geometry 0.2706 | confidence 1.3122 | mean error 0.685 vox speed: 890 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.32 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1456 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 245, from the previous prediction 1211 confidence-labeled crops: terminal 52/1520 (3.4%) | recoverable 103/1520 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 320/1600 (20.0%), 2: 384/1600 (24.0%), 3: 288/1600 (18.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 63 chunks / 126.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 676 rows | mean depth 5 decisions | mean travel 41 vox | 7 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 53% 47% 22%>0% 22% 0 143 90%/6%/0%/4% 8% 141/1/90/9 0175A_5mm_v1 320 37% 53% 22%>11% 12% 36 954 85%/5%/4%/6% 9% 105/3/0/64 1447_5mm_v1 384 47% 43% 29%>11% 18% 28 510 84%/7%/5%/4% 10% 156/1/0/8 paris4_afv_central 288 32% 47% 27%>20% 6% 52 395 79%/7%/7%/7% 17% 136/5/0/0 PHerc0125_3mm_v2 288 52% 48% 24%>0% 24% 0 144 88%/6%/0%/6% 13% 138/5/0/0 gradients: max 49, clipped 0/50 updates Step 11,000 | last 50 updates / 1,600 crops | lr 2.11e-05 points (35/curve, tolerance 2 vox): 46,353 right / 3,086 wrong | 93.8% correct | 6,561 unknown loss 0.8896 | geometry 0.2459 | confidence 1.2875 | mean error 0.656 vox speed: 891 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.34 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1462 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 211, from the previous prediction 1251 confidence-labeled crops: terminal 31/1518 (2.0%) | recoverable 109/1518 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 416/1600 (26.0%), 2: 320/1600 (20.0%), 3: 192/1600 (12.0%), 4: 480/1600 (30.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 63 chunks / 126.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 1626 chunk references / 1 windows live chains: 630 rows | mean depth 5 decisions | mean travel 39 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 50% 50% 26%>0% 26% 0 146 90%/10%/0%/0% 10% 96/1/0/0 0175A_5mm_v1 416 43% 47% 26%>10% 16% 35 679 87%/6%/3%/5% 8% 181/1/0/13 1447_5mm_v1 320 50% 38% 31%>12% 18% 31 594 82%/9%/3%/5% 7% 121/4/0/0 paris4_afv_central 192 31% 41% 32%>29% 3% 50 507 80%/6%/5%/9% 17% 78/2/0/0 PHerc0125_3mm_v2 480 58% 42% 30%>0% 30% 0 129 89%/5%/0%/6% 10% 154/9/0/49 gradients: max 49, clipped 0/50 updates Step 11,000 | dagger_launched: true Step 11,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 251, "stopped": 0, "mean_example_error": 0.8874736964702606, "geometry_examples": 25, "examples": 32, "inference_and_measurement_seconds": 1.8011422455310822, "total_diagnostic_seconds": 3.7631509862840176} Step 11,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 6/8 (75.0%) | false stops 1 1.5-3: recovered 6/8 (75.0%) | false stops 0 3-4.5: recovered 6/8 (75.0%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 1 entire proposed 8-point prefix correct (distance): 21/28 (75.0%) gate @ 0.50: false stops 0/23 (0.0%) | accepted wrong 5/26 (19.2%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.467 1.422 following 16 464 1.038 1.036 recoverable 12 348 2.040 1.936 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 1.038 1.036 332 32 812 1.467 1.422 improved/worsened/compared states: 3/1/28 Step 11,000 | batch_diagnostics_seconds: 3.7700930386781693 | recovery_seconds: 16.000037763267756 Step 11,050 | last 50 updates / 1,600 crops | lr 2.10e-05 points (35/curve, tolerance 2 vox): 46,292 right / 3,536 wrong | 92.9% correct | 6,172 unknown loss 1.0347 | geometry 0.2719 | confidence 1.5256 | mean error 0.697 vox speed: 995 ms/update | data wait 1 ms/update | 22.4 crops/s | 22.4 decisions/s | 1.35 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1483 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 287, from the previous prediction 1196 confidence-labeled crops: terminal 34/1513 (2.2%) | recoverable 84/1513 (5.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 352/1600 (22.0%), 2: 192/1600 (12.0%), 3: 224/1600 (14.0%), 4: 576/1600 (36.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 63 chunks / 126.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 578 rows | mean depth 5 decisions | mean travel 41 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 55% 45% 18%>0% 18% 0 153 92%/6%/0%/2% 9% 114/0/0/0 0175A_5mm_v1 352 42% 45% 24%>14% 11% 36 822 83%/5%/5%/7% 11% 157/2/0/0 1447_5mm_v1 192 46% 41% 28%>13% 15% 21 604 85%/7%/4%/4% 9% 79/0/0/0 paris4_afv_central 224 42% 38% 29%>20% 9% 43 464 83%/7%/3%/8% 14% 86/1/1/0 PHerc0125_3mm_v2 576 54% 46% 23%>0% 23% 0 128 90%/4%/0%/6% 12% 142/3/0/122 gradients: max 60, clipped 0/50 updates Step 11,100 | last 50 updates / 1,600 crops | lr 2.10e-05 points (35/curve, tolerance 2 vox): 46,380 right / 3,529 wrong | 92.9% correct | 6,091 unknown loss 0.9093 | geometry 0.2786 | confidence 1.2614 | mean error 0.681 vox speed: 1425 ms/update | data wait 1 ms/update | 22.3 crops/s | 22.3 decisions/s | 1.33 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1465 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 227, from the previous prediction 1238 confidence-labeled crops: terminal 38/1515 (2.5%) | recoverable 110/1515 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 224/1600 (14.0%), 2: 320/1600 (20.0%), 3: 416/1600 (26.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 65 chunks / 130.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 633 rows | mean depth 6 decisions | mean travel 49 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 62% 38% 30%>0% 30% 0 173 89%/9%/0%/2% 7% 84/0/0/0 0175A_5mm_v1 224 47% 47% 26%>5% 21% 11 449 85%/10%/2%/4% 5% 106/0/0/0 1447_5mm_v1 320 38% 46% 25%>16% 9% 38 657 84%/5%/4%/7% 9% 148/2/0/0 paris4_afv_central 416 38% 44% 23%>17% 6% 63 410 83%/5%/5%/7% 12% 184/2/0/0 PHerc0125_3mm_v2 416 58% 42% 25%>0% 25% 0 135 87%/7%/0%/6% 14% 111/3/0/64 gradients: max 60, clipped 0/50 updates Step 11,144 | dagger_states: 1604 | dagger_source_step: 11000 | dagger_caches: 3 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/decisions_011000.npz" | dagger_fibers: 128 | dagger_supply: {"states": 1604, "traces": 128, "fibers": 128, "rows": {"terminal": 17, "premature_stop": 119, "recoverable": 35, "pre_excursion": 47, "ordinary": 1381}, "events": {"terminal": 3, "premature_stop": 119, "recoverable": 10, "pre_excursion": 12, "ordinary": 120}, "supervision": {"following": 1545, "recoverable": 26, "terminal": 17, "unknown": 16}, "departed_traces": 10, "terminal_traces": 3, "stop_reasons": {"after_failure": 2, "confidence": 121, "unannotated": 5}, "traced_voxels": 22687.66027531304, "seconds": 195.68388195335865, "voxels_per_second": 115.94036283847156} | dagger_coverage: {"requested": 128, "seeds": 128, "skipped": [], "epoch": 0} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 0 | dagger_launch_interval: null | dagger_publication_age: 144 | dataset: "paris4" Step 11,150 | last 50 updates / 1,600 crops | lr 2.09e-05 points (35/curve, tolerance 2 vox): 46,903 right / 3,326 wrong | 93.4% correct | 5,771 unknown loss 0.9521 | geometry 0.3003 | confidence 1.3035 | mean error 0.698 vox speed: 1311 ms/update | data wait 1 ms/update | 24.2 crops/s | 24.2 decisions/s | 1.30 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1484 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 211, from the previous prediction 1273 confidence-labeled crops: terminal 30/1527 (2.0%) | recoverable 101/1527 (6.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 384/1600 (24.0%), 2: 288/1600 (18.0%), 3: 384/1600 (24.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 66 chunks / 132.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 651 rows | mean depth 6 decisions | mean travel 47 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 58% 42% 26%>0% 26% 0 171 92%/6%/0%/3% 12% 92/2/42/1 0175A_5mm_v1 384 49% 48% 21%>3% 18% 11 326 91%/5%/1%/2% 6% 183/2/0/0 1447_5mm_v1 288 48% 44% 26%>8% 18% 20 463 87%/7%/2%/3% 5% 126/2/0/0 paris4_afv_central 384 41% 44% 25%>15% 10% 70 571 80%/7%/5%/9% 15% 102/2/10/66 PHerc0125_3mm_v2 320 54% 46% 25%>0% 25% 0 150 88%/7%/0%/5% 13% 148/7/0/0 gradients: max 62, clipped 1/50 updates Step 11,200 | last 50 updates / 1,600 crops | lr 2.08e-05 points (35/curve, tolerance 2 vox): 47,520 right / 3,203 wrong | 93.7% correct | 5,277 unknown loss 0.9209 | geometry 0.2591 | confidence 1.3236 | mean error 0.662 vox speed: 893 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.29 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1458 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 242, from the previous prediction 1216 confidence-labeled crops: terminal 31/1525 (2.0%) | recoverable 107/1525 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 288/1600 (18.0%), 2: 320/1600 (20.0%), 3: 384/1600 (24.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 66 chunks / 132.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 605 rows | mean depth 6 decisions | mean travel 50 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 58% 42% 31%>0% 31% 0 144 89%/9%/0%/2% 8% 130/1/8/19 0175A_5mm_v1 288 51% 41% 25%>8% 17% 20 646 85%/8%/2%/5% 10% 117/7/0/0 1447_5mm_v1 320 50% 42% 27%>9% 19% 22 503 85%/8%/2%/5% 5% 133/1/0/0 paris4_afv_central 384 36% 46% 24%>18% 6% 73 617 82%/6%/4%/8% 16% 105/1/0/73 PHerc0125_3mm_v2 256 53% 47% 23%>0% 23% 0 155 94%/3%/0%/4% 9% 120/2/0/0 gradients: max 75, clipped 1/50 updates Step 11,250 | last 50 updates / 1,600 crops | lr 2.07e-05 points (35/curve, tolerance 2 vox): 46,866 right / 3,240 wrong | 93.5% correct | 5,894 unknown loss 0.9396 | geometry 0.2751 | confidence 1.3290 | mean error 0.682 vox speed: 892 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.35 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1467 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 254, from the previous prediction 1213 confidence-labeled crops: terminal 27/1511 (1.8%) | recoverable 118/1511 (7.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 416/1600 (26.0%), 2: 192/1600 (12.0%), 3: 288/1600 (18.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 66 chunks / 132.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 603 rows | mean depth 7 decisions | mean travel 57 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 56% 44% 26%>0% 26% 0 133 87%/8%/0%/5% 8% 95/0/0/47 0175A_5mm_v1 416 45% 44% 25%>11% 14% 34 910 85%/6%/3%/6% 7% 135/4/68/49 1447_5mm_v1 192 46% 42% 30%>12% 18% 20 633 85%/6%/3%/6% 7% 80/0/0/0 paris4_afv_central 288 33% 44% 30%>22% 7% 59 504 80%/10%/3%/7% 14% 128/4/0/0 PHerc0125_3mm_v2 384 56% 44% 25%>0% 25% 0 137 89%/7%/0%/4% 13% 165/6/80/5 gradients: max 57, clipped 0/50 updates Step 11,300 | last 50 updates / 1,600 crops | lr 2.07e-05 points (35/curve, tolerance 2 vox): 47,354 right / 3,072 wrong | 93.9% correct | 5,574 unknown loss 0.8813 | geometry 0.2573 | confidence 1.2482 | mean error 0.646 vox speed: 891 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.32 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1466 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 324, from the previous prediction 1142 confidence-labeled crops: terminal 37/1519 (2.4%) | recoverable 89/1519 (5.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 544/1600 (34.0%), 2: 192/1600 (12.0%), 3: 352/1600 (22.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 66 chunks / 132.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 572 rows | mean depth 7 decisions | mean travel 56 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 57% 43% 23%>0% 23% 0 137 91%/7%/0%/2% 7% 92/1/0/33 0175A_5mm_v1 544 44% 47% 24%>10% 14% 42 1031 86%/5%/3%/7% 9% 147/4/0/106 1447_5mm_v1 192 49% 47% 24%>4% 20% 6 324 92%/6%/1%/2% 2% 90/4/0/0 paris4_afv_central 352 39% 42% 26%>19% 7% 56 501 83%/5%/6%/7% 16% 148/2/0/0 PHerc0125_3mm_v2 224 58% 42% 29%>0% 29% 0 147 86%/7%/0%/7% 12% 95/5/0/0 gradients: max 75, clipped 1/50 updates Step 11,350 | last 50 updates / 1,600 crops | lr 2.06e-05 points (35/curve, tolerance 2 vox): 47,815 right / 2,591 wrong | 94.9% correct | 5,594 unknown loss 0.8584 | geometry 0.2599 | confidence 1.1969 | mean error 0.637 vox speed: 892 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.26 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1498 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 267, from the previous prediction 1231 confidence-labeled crops: terminal 19/1512 (1.3%) | recoverable 91/1512 (6.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 320/1600 (20.0%), 2: 448/1600 (28.0%), 3: 192/1600 (12.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 66 chunks / 132.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 633 rows | mean depth 7 decisions | mean travel 55 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 51% 44% 25%>5% 20% 9 185 89%/5%/0%/5% 10% 113/1/0/0 0175A_5mm_v1 320 50% 44% 23%>5% 18% 15 507 87%/8%/2%/3% 5% 129/2/0/13 1447_5mm_v1 448 49% 41% 25%>10% 15% 34 604 89%/5%/2%/4% 6% 167/8/35/18 paris4_afv_central 192 42% 43% 24%>15% 9% 28 450 81%/8%/2%/9% 10% 83/0/0/0 PHerc0125_3mm_v2 384 55% 45% 22%>0% 22% 0 136 89%/4%/0%/7% 10% 141/2/47/31 gradients: max 52, clipped 0/50 updates Step 11,400 | last 50 updates / 1,600 crops | lr 2.05e-05 points (35/curve, tolerance 2 vox): 47,256 right / 3,433 wrong | 93.2% correct | 5,311 unknown loss 0.8995 | geometry 0.2623 | confidence 1.2745 | mean error 0.653 vox speed: 896 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.31 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1467 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 288, from the previous prediction 1179 confidence-labeled crops: terminal 42/1519 (2.8%) | recoverable 104/1519 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 224/1600 (14.0%), 2: 320/1600 (20.0%), 3: 352/1600 (22.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 66 chunks / 132.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 610 rows | mean depth 8 decisions | mean travel 65 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 49% 51% 21%>0% 21% 0 154 93%/6%/0%/1% 7% 129/1/18/1 0175A_5mm_v1 224 45% 46% 24%>9% 14% 19 696 87%/5%/4%/5% 7% 102/0/0/0 1447_5mm_v1 320 42% 48% 24%>10% 14% 33 795 87%/4%/2%/7% 6% 95/4/0/58 paris4_afv_central 352 34% 45% 27%>20% 7% 63 535 81%/8%/8%/4% 17% 160/0/0/0 PHerc0125_3mm_v2 448 59% 41% 26%>0% 26% 0 136 85%/8%/0%/7% 9% 124/2/0/59 gradients: max 56, clipped 0/50 updates Step 11,450 | last 50 updates / 1,600 crops | lr 2.05e-05 points (35/curve, tolerance 2 vox): 48,302 right / 2,732 wrong | 94.6% correct | 4,966 unknown loss 0.8668 | geometry 0.2276 | confidence 1.2783 | mean error 0.618 vox speed: 894 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.25 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1505 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 208, from the previous prediction 1297 confidence-labeled crops: terminal 23/1530 (1.5%) | recoverable 117/1530 (7.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 448/1600 (28.0%), 2: 352/1600 (22.0%), 3: 256/1600 (16.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 70 chunks / 140.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 679 rows | mean depth 7 decisions | mean travel 59 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 56% 44% 29%>0% 29% 0 141 89%/9%/0%/2% 7% 113/1/0/0 0175A_5mm_v1 448 46% 49% 21%>5% 17% 19 438 88%/7%/1%/3% 7% 219/3/42/2 1447_5mm_v1 352 46% 44% 25%>10% 15% 36 774 83%/8%/2%/7% 8% 106/1/0/48 paris4_afv_central 256 39% 48% 21%>13% 8% 32 422 84%/5%/4%/6% 11% 122/3/0/0 PHerc0125_3mm_v2 288 58% 42% 28%>0% 28% 0 139 89%/7%/0%/4% 10% 119/4/0/3 gradients: max 57, clipped 0/50 updates Step 11,500 | last 50 updates / 1,600 crops | lr 2.04e-05 points (35/curve, tolerance 2 vox): 46,876 right / 3,389 wrong | 93.3% correct | 5,735 unknown loss 0.9318 | geometry 0.2659 | confidence 1.3317 | mean error 0.682 vox speed: 897 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.30 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1471 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 313, from the previous prediction 1158 confidence-labeled crops: terminal 32/1516 (2.1%) | recoverable 103/1516 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 352/1600 (22.0%), 2: 288/1600 (18.0%), 3: 224/1600 (14.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 70 chunks / 140.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 564 rows | mean depth 8 decisions | mean travel 64 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 58% 42% 28%>0% 28% 0 140 90%/6%/0%/3% 7% 79/0/27/82 0175A_5mm_v1 352 44% 43% 26%>12% 14% 34 1126 85%/5%/4%/6% 9% 105/0/58/48 1447_5mm_v1 288 39% 46% 27%>15% 12% 32 699 84%/8%/3%/6% 9% 133/2/0/0 paris4_afv_central 224 40% 41% 27%>19% 8% 40 582 83%/8%/4%/5% 13% 92/1/0/0 PHerc0125_3mm_v2 352 56% 44% 23%>0% 23% 0 143 87%/6%/0%/7% 12% 155/1/0/0 gradients: max 68, clipped 1/50 updates Step 11,550 | last 50 updates / 1,600 crops | lr 2.03e-05 points (35/curve, tolerance 2 vox): 47,086 right / 3,798 wrong | 92.5% correct | 5,116 unknown loss 0.9660 | geometry 0.2779 | confidence 1.3762 | mean error 0.695 vox speed: 892 ms/update | data wait 1 ms/update | 35.1 crops/s | 35.1 decisions/s | 1.30 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1487 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 223, from the previous prediction 1264 confidence-labeled crops: terminal 38/1525 (2.5%) | recoverable 88/1525 (5.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 512/1600 (32.0%), 1: 256/1600 (16.0%), 2: 128/1600 (8.0%), 3: 320/1600 (20.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 72 chunks / 144.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 669 rows | mean depth 8 decisions | mean travel 63 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 512 55% 45% 23%>0% 23% 0 138 91%/6%/0%/3% 7% 201/1/0/31 0175A_5mm_v1 256 42% 49% 23%>9% 14% 19 692 87%/4%/3%/6% 7% 117/0/0/9 1447_5mm_v1 128 45% 42% 30%>12% 17% 12 702 84%/6%/6%/4% 10% 54/0/0/0 paris4_afv_central 320 33% 44% 28%>23% 5% 63 632 79%/7%/6%/7% 17% 140/1/0/0 PHerc0125_3mm_v2 384 55% 45% 25%>0% 25% 0 143 89%/4%/1%/6% 12% 157/0/0/14 gradients: max 65, clipped 1/50 updates Step 11,600 | last 50 updates / 1,600 crops | lr 2.02e-05 points (35/curve, tolerance 2 vox): 47,612 right / 3,122 wrong | 93.8% correct | 5,266 unknown loss 0.9071 | geometry 0.2541 | confidence 1.3060 | mean error 0.661 vox speed: 894 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.28 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1494 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 257, from the previous prediction 1237 confidence-labeled crops: terminal 33/1530 (2.2%) | recoverable 94/1530 (6.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 256/1600 (16.0%), 2: 352/1600 (22.0%), 3: 288/1600 (18.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 72 chunks / 144.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 620 rows | mean depth 8 decisions | mean travel 68 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 53% 46% 22%>1% 21% 3 148 91%/6%/0%/3% 8% 103/0/0/15 0175A_5mm_v1 256 50% 44% 25%>6% 18% 15 576 87%/7%/2%/4% 8% 112/1/0/0 1447_5mm_v1 352 49% 40% 26%>11% 15% 30 671 86%/6%/4%/5% 6% 142/0/0/0 paris4_afv_central 288 38% 45% 24%>17% 7% 44 535 84%/6%/4%/6% 11% 130/4/0/0 PHerc0125_3mm_v2 448 54% 46% 24%>0% 24% 0 134 90%/5%/0%/5% 12% 133/1/0/72 gradients: max 64, clipped 1/50 updates Step 11,650 | last 50 updates / 1,600 crops | lr 2.02e-05 points (35/curve, tolerance 2 vox): 46,624 right / 3,256 wrong | 93.5% correct | 6,120 unknown loss 1.0039 | geometry 0.3056 | confidence 1.3966 | mean error 0.712 vox speed: 893 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.33 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1506 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 257, from the previous prediction 1249 confidence-labeled crops: terminal 26/1518 (1.7%) | recoverable 109/1518 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 128/1600 (8.0%), 2: 256/1600 (16.0%), 3: 288/1600 (18.0%), 4: 576/1600 (36.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 72 chunks / 144.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 597 rows | mean depth 9 decisions | mean travel 72 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 59% 41% 25%>0% 25% 0 139 90%/7%/0%/3% 5% 143/0/0/0 0175A_5mm_v1 128 49% 41% 30%>10% 20% 12 872 86%/9%/2%/4% 9% 52/3/0/0 1447_5mm_v1 256 47% 44% 22%>9% 13% 19 568 86%/7%/3%/4% 6% 113/0/0/0 paris4_afv_central 288 36% 47% 26%>17% 9% 50 655 81%/6%/5%/7% 15% 109/2/46/25 PHerc0125_3mm_v2 576 56% 44% 26%>0% 26% 0 133 87%/6%/0%/7% 13% 180/8/0/72 gradients: max 55, clipped 0/50 updates Step 11,700 | last 50 updates / 1,600 crops | lr 2.01e-05 points (35/curve, tolerance 2 vox): 47,717 right / 3,002 wrong | 94.1% correct | 5,281 unknown loss 0.9169 | geometry 0.2892 | confidence 1.2554 | mean error 0.673 vox speed: 890 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.29 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1485 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 234, from the previous prediction 1251 confidence-labeled crops: terminal 21/1541 (1.4%) | recoverable 111/1541 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 448/1600 (28.0%), 2: 288/1600 (18.0%), 3: 320/1600 (20.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 72 chunks / 144.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 666 rows | mean depth 8 decisions | mean travel 68 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 58% 42% 24%>0% 24% 0 141 92%/7%/0%/1% 9% 94/3/0/0 0175A_5mm_v1 448 47% 47% 24%>6% 19% 22 538 88%/6%/1%/5% 6% 206/5/78/5 1447_5mm_v1 288 51% 43% 26%>6% 20% 15 517 88%/7%/1%/4% 4% 97/0/105/26 paris4_afv_central 320 38% 42% 23%>20% 3% 61 737 84%/7%/4%/4% 16% 116/1/0/20 PHerc0125_3mm_v2 320 51% 49% 25%>0% 25% 0 136 90%/7%/0%/3% 13% 153/1/0/3 gradients: max 52, clipped 0/50 updates Step 11,750 | last 50 updates / 1,600 crops | lr 2.00e-05 points (35/curve, tolerance 2 vox): 46,244 right / 3,467 wrong | 93.0% correct | 6,289 unknown loss 0.8794 | geometry 0.2767 | confidence 1.2053 | mean error 0.668 vox speed: 896 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.39 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1455 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 277, from the previous prediction 1178 confidence-labeled crops: terminal 38/1512 (2.5%) | recoverable 104/1512 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 448/1600 (28.0%), 2: 224/1600 (14.0%), 3: 320/1600 (20.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 72 chunks / 144.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 1263 chunk references / 1 windows live chains: 589 rows | mean depth 9 decisions | mean travel 73 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 55% 45% 23%>0% 23% 0 153 93%/7%/0%/1% 8% 113/0/50/3 0175A_5mm_v1 448 45% 43% 28%>11% 17% 42 1744 81%/6%/4%/8% 12% 75/0/0/119 1447_5mm_v1 224 44% 46% 29%>10% 19% 18 601 86%/8%/2%/4% 6% 102/2/0/2 paris4_afv_central 320 33% 48% 24%>19% 5% 53 619 82%/8%/4%/7% 16% 145/2/0/8 PHerc0125_3mm_v2 352 56% 44% 24%>0% 24% 0 138 89%/5%/0%/6% 10% 154/2/0/0 gradients: max 48, clipped 0/50 updates Step 11,800 | last 50 updates / 1,600 crops | lr 1.99e-05 points (35/curve, tolerance 2 vox): 46,808 right / 3,078 wrong | 93.8% correct | 6,114 unknown loss 0.8929 | geometry 0.2465 | confidence 1.2928 | mean error 0.646 vox speed: 892 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.30 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1473 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 305, from the previous prediction 1168 confidence-labeled crops: terminal 32/1519 (2.1%) | recoverable 102/1519 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 320/1600 (20.0%), 2: 288/1600 (18.0%), 3: 384/1600 (24.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 72 chunks / 144.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 588 rows | mean depth 8 decisions | mean travel 72 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 47% 53% 21%>0% 21% 0 159 92%/8%/0%/0% 5% 119/0/0/0 0175A_5mm_v1 320 44% 48% 25%>9% 16% 29 1549 88%/4%/2%/5% 8% 55/1/0/97 1447_5mm_v1 288 50% 42% 28%>8% 20% 21 591 84%/10%/2%/4% 5% 120/1/0/0 paris4_afv_central 384 39% 47% 23%>14% 9% 58 740 80%/7%/5%/8% 15% 116/1/0/64 PHerc0125_3mm_v2 384 54% 46% 21%>0% 21% 0 142 89%/5%/1%/6% 12% 178/5/14/0 gradients: max 69, clipped 1/50 updates Step 11,850 | last 50 updates / 1,600 crops | lr 1.99e-05 points (35/curve, tolerance 2 vox): 46,585 right / 3,450 wrong | 93.1% correct | 5,965 unknown loss 0.9664 | geometry 0.2815 | confidence 1.3698 | mean error 0.662 vox speed: 894 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.36 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1450 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 229, from the previous prediction 1221 confidence-labeled crops: terminal 36/1524 (2.4%) | recoverable 108/1524 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 288/1600 (18.0%), 2: 256/1600 (16.0%), 3: 416/1600 (26.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 72 chunks / 144.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 649 rows | mean depth 10 decisions | mean travel 78 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 53% 47% 24%>0% 24% 0 159 92%/6%/0%/2% 8% 136/2/0/0 0175A_5mm_v1 288 44% 44% 24%>11% 13% 26 898 87%/5%/4%/3% 10% 128/2/0/0 1447_5mm_v1 256 47% 44% 26%>9% 17% 21 593 86%/8%/2%/4% 7% 113/1/0/0 paris4_afv_central 416 35% 45% 25%>20% 5% 79 839 80%/8%/4%/8% 16% 134/0/0/54 PHerc0125_3mm_v2 352 50% 50% 25%>0% 25% 0 136 87%/7%/0%/6% 11% 138/2/51/39 gradients: max 51, clipped 0/50 updates Step 11,900 | last 50 updates / 1,600 crops | lr 1.98e-05 points (35/curve, tolerance 2 vox): 46,458 right / 3,546 wrong | 92.9% correct | 5,996 unknown loss 0.8770 | geometry 0.2695 | confidence 1.2150 | mean error 0.663 vox speed: 892 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.35 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1450 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 231, from the previous prediction 1219 confidence-labeled crops: terminal 39/1516 (2.6%) | recoverable 105/1516 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 288/1600 (18.0%), 2: 384/1600 (24.0%), 3: 416/1600 (26.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 72 chunks / 144.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 638 rows | mean depth 9 decisions | mean travel 74 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 57% 43% 25%>0% 25% 0 158 90%/7%/0%/3% 7% 137/0/0/0 0175A_5mm_v1 288 50% 43% 24%>6% 18% 18 620 86%/8%/2%/5% 7% 125/2/4/0 1447_5mm_v1 384 47% 41% 28%>12% 16% 38 968 84%/7%/4%/6% 7% 120/3/14/37 paris4_afv_central 416 37% 45% 25%>18% 7% 65 694 82%/7%/4%/7% 16% 170/4/0/16 PHerc0125_3mm_v2 192 55% 45% 24%>0% 24% 0 144 90%/4%/1%/6% 14% 86/2/0/0 gradients: max 55, clipped 0/50 updates Step 11,950 | last 50 updates / 1,600 crops | lr 1.97e-05 points (35/curve, tolerance 2 vox): 46,874 right / 3,493 wrong | 93.1% correct | 5,633 unknown loss 1.0515 | geometry 0.3072 | confidence 1.4887 | mean error 0.709 vox speed: 893 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.31 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1486 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 236, from the previous prediction 1250 confidence-labeled crops: terminal 34/1530 (2.2%) | recoverable 127/1530 (8.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 160/1600 (10.0%), 2: 192/1600 (12.0%), 3: 352/1600 (22.0%), 4: 512/1600 (32.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 72 chunks / 144.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 642 rows | mean depth 9 decisions | mean travel 77 vox | 7 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 56% 44% 28%>0% 28% 0 137 89%/9%/0%/2% 7% 113/0/24/57 0175A_5mm_v1 160 48% 44% 21%>8% 13% 12 743 84%/7%/3%/6% 6% 64/4/66/7 1447_5mm_v1 192 40% 45% 27%>15% 12% 29 1059 83%/10%/4%/3% 12% 73/1/0/14 paris4_afv_central 352 36% 47% 25%>16% 9% 57 705 82%/7%/5%/7% 15% 142/1/0/24 PHerc0125_3mm_v2 512 51% 49% 26%>0% 26% 0 146 87%/8%/1%/5% 11% 250/9/0/0 gradients: max 67, clipped 4/50 updates Step 12,000 | last 50 updates / 1,600 crops | lr 1.96e-05 points (35/curve, tolerance 2 vox): 47,008 right / 2,969 wrong | 94.1% correct | 6,023 unknown loss 0.8694 | geometry 0.2508 | confidence 1.2372 | mean error 0.655 vox speed: 894 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.29 attempts/decision | peak allocated VRAM 41.42 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1494 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 254, from the previous prediction 1240 confidence-labeled crops: terminal 30/1502 (2.0%) | recoverable 100/1502 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 480/1600 (30.0%), 1: 288/1600 (18.0%), 2: 256/1600 (16.0%), 3: 288/1600 (18.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 72 chunks / 144.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 623 rows | mean depth 10 decisions | mean travel 83 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 480 55% 45% 27%>0% 27% 0 141 89%/8%/0%/4% 7% 169/0/0/45 0175A_5mm_v1 288 47% 43% 26%>10% 16% 25 937 84%/5%/5%/7% 9% 116/0/34/8 1447_5mm_v1 256 46% 43% 30%>11% 20% 23 722 83%/8%/2%/7% 6% 110/0/0/0 paris4_afv_central 288 40% 43% 24%>17% 7% 45 677 84%/5%/3%/8% 13% 117/4/0/7 PHerc0125_3mm_v2 288 53% 47% 26%>0% 26% 0 139 87%/6%/0%/7% 12% 111/3/0/23 gradients: max 60, clipped 0/50 updates Step 12,000 | dagger_launched: true Step 12,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 248, "stopped": 1, "mean_example_error": 0.3875258091476656, "geometry_examples": 31, "examples": 32, "inference_and_measurement_seconds": 1.8864197134971619, "total_diagnostic_seconds": 3.7262198477983475} Step 12,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 6/8 (75.0%) | false stops 1 1.5-3: recovered 6/8 (75.0%) | false stops 0 3-4.5: recovered 6/8 (75.0%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 0 entire proposed 8-point prefix correct (distance): 21/28 (75.0%) gate @ 0.50: false stops 0/22 (0.0%) | accepted wrong 5/26 (19.2%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.416 1.383 following 16 464 0.957 0.957 recoverable 12 348 2.028 1.953 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 0.957 0.957 332 32 812 1.416 1.383 improved/worsened/compared states: 2/1/28 Step 12,000 | batch_diagnostics_seconds: 3.7289486788213253 | recovery_seconds: 15.986105751246214 Step 12,050 | last 50 updates / 1,600 crops | lr 1.96e-05 points (35/curve, tolerance 2 vox): 47,606 right / 3,322 wrong | 93.5% correct | 5,072 unknown loss 0.9621 | geometry 0.2930 | confidence 1.3381 | mean error 0.694 vox speed: 1324 ms/update | data wait 1 ms/update | 18.1 crops/s | 18.1 decisions/s | 1.28 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1485 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 255, from the previous prediction 1230 confidence-labeled crops: terminal 29/1535 (1.9%) | recoverable 105/1535 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 320/1600 (20.0%), 2: 224/1600 (14.0%), 3: 416/1600 (26.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 72 chunks / 144.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 650 rows | mean depth 10 decisions | mean travel 82 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 52% 48% 26%>0% 26% 0 134 95%/5%/0%/0% 9% 121/1/0/78 0175A_5mm_v1 320 45% 49% 23%>6% 17% 19 596 88%/6%/1%/5% 6% 156/0/0/0 1447_5mm_v1 224 50% 41% 25%>9% 16% 16 694 83%/9%/2%/5% 5% 92/1/0/0 paris4_afv_central 416 38% 45% 22%>17% 4% 66 678 81%/6%/5%/7% 15% 185/1/31/3 PHerc0125_3mm_v2 224 57% 43% 27%>0% 27% 0 138 88%/8%/0%/4% 9% 96/3/0/0 gradients: max 52, clipped 0/50 updates Step 12,100 | last 50 updates / 1,600 crops | lr 1.95e-05 points (35/curve, tolerance 2 vox): 47,075 right / 3,299 wrong | 93.5% correct | 5,626 unknown loss 0.9386 | geometry 0.2841 | confidence 1.3089 | mean error 0.660 vox speed: 1367 ms/update | data wait 1 ms/update | 23.2 crops/s | 23.2 decisions/s | 1.32 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1462 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 226, from the previous prediction 1236 confidence-labeled crops: terminal 36/1525 (2.4%) | recoverable 114/1525 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 224/1600 (14.0%), 2: 288/1600 (18.0%), 3: 448/1600 (28.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 72 chunks / 144.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 645 rows | mean depth 10 decisions | mean travel 82 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 54% 46% 27%>0% 27% 0 145 92%/7%/0%/2% 7% 175/0/0/0 0175A_5mm_v1 224 48% 41% 23%>11% 12% 23 1709 84%/5%/3%/8% 9% 45/1/0/47 1447_5mm_v1 288 47% 44% 24%>9% 15% 22 657 87%/7%/1%/5% 5% 127/1/0/0 paris4_afv_central 448 38% 45% 26%>17% 9% 65 758 80%/8%/6%/7% 16% 174/3/5/29 PHerc0125_3mm_v2 256 52% 48% 26%>0% 26% 0 154 88%/9%/1%/3% 11% 124/1/0/0 gradients: max 55, clipped 0/50 updates Step 12,150 | last 50 updates / 1,600 crops | lr 1.94e-05 points (35/curve, tolerance 2 vox): 46,550 right / 3,567 wrong | 92.9% correct | 5,883 unknown loss 1.0206 | geometry 0.3058 | confidence 1.4295 | mean error 0.706 vox speed: 1374 ms/update | data wait 1 ms/update | 23.1 crops/s | 23.1 decisions/s | 1.35 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1445 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 273, from the previous prediction 1172 confidence-labeled crops: terminal 31/1518 (2.0%) | recoverable 126/1518 (8.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 192/1600 (12.0%), 2: 320/1600 (20.0%), 3: 384/1600 (24.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 72 chunks / 144.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 569 rows | mean depth 10 decisions | mean travel 84 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 57% 43% 26%>0% 26% 0 143 89%/8%/0%/3% 9% 110/0/0/0 0175A_5mm_v1 192 44% 47% 24%>8% 16% 15 761 88%/7%/3%/3% 7% 91/1/0/0 1447_5mm_v1 320 47% 39% 30%>14% 16% 40 1089 84%/7%/4%/5% 9% 105/2/41/21 paris4_afv_central 384 38% 43% 26%>19% 7% 67 1177 80%/8%/3%/9% 17% 82/3/0/84 PHerc0125_3mm_v2 448 54% 46% 27%>0% 27% 0 148 86%/9%/0%/5% 10% 181/5/0/25 gradients: max 59, clipped 0/50 updates Step 12,200 | last 50 updates / 1,600 crops | lr 1.93e-05 points (35/curve, tolerance 2 vox): 47,316 right / 3,291 wrong | 93.5% correct | 5,393 unknown loss 0.7987 | geometry 0.2146 | confidence 1.1680 | mean error 0.616 vox speed: 1372 ms/update | data wait 1 ms/update | 23.1 crops/s | 23.1 decisions/s | 1.31 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1466 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 287, from the previous prediction 1179 confidence-labeled crops: terminal 44/1512 (2.9%) | recoverable 110/1512 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 320/1600 (20.0%), 2: 352/1600 (22.0%), 3: 384/1600 (24.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 72 chunks / 144.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 605 rows | mean depth 11 decisions | mean travel 89 vox | 10 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 59% 41% 26%>0% 26% 0 159 91%/6%/0%/3% 8% 106/1/0/0 0175A_5mm_v1 320 46% 41% 29%>13% 15% 35 1166 80%/9%/4%/7% 9% 130/3/0/0 1447_5mm_v1 352 48% 44% 24%>7% 16% 26 1024 84%/8%/2%/7% 4% 77/1/0/79 paris4_afv_central 384 37% 48% 23%>15% 8% 50 642 82%/7%/5%/6% 11% 183/3/0/0 PHerc0125_3mm_v2 288 51% 49% 26%>0% 26% 0 134 89%/5%/2%/5% 12% 109/3/0/31 gradients: max 46, clipped 0/50 updates Step 12,250 | last 50 updates / 1,600 crops | lr 1.93e-05 points (35/curve, tolerance 2 vox): 46,682 right / 4,144 wrong | 91.8% correct | 5,174 unknown loss 0.9700 | geometry 0.2763 | confidence 1.3873 | mean error 0.698 vox speed: 1378 ms/update | data wait 1 ms/update | 23.0 crops/s | 23.0 decisions/s | 1.35 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1471 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 179, from the previous prediction 1292 confidence-labeled crops: terminal 48/1534 (3.1%) | recoverable 103/1534 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 320/1600 (20.0%), 2: 160/1600 (10.0%), 3: 480/1600 (30.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 72 chunks / 144.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 675 rows | mean depth 9 decisions | mean travel 79 vox | 6 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 56% 44% 23%>0% 23% 0 148 93%/6%/0%/2% 8% 182/0/0/0 0175A_5mm_v1 320 48% 44% 30%>8% 21% 27 861 83%/8%/3%/6% 8% 136/2/15/5 1447_5mm_v1 160 52% 43% 26%>5% 21% 8 469 88%/8%/1%/3% 5% 69/1/0/0 paris4_afv_central 480 37% 44% 27%>19% 8% 74 783 80%/7%/8%/6% 20% 210/3/0/0 PHerc0125_3mm_v2 224 56% 44% 22%>0% 22% 0 139 90%/4%/0%/6% 10% 78/0/0/20 gradients: max 53, clipped 0/50 updates Step 12,269 | dagger_states: 3887 | dagger_source_step: 12000 | dagger_caches: 3 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/0175A_5mm_v1/decisions_012000.npz" | dagger_fibers: 384 | dagger_supply: {"states": 3887, "traces": 384, "fibers": 384, "rows": {"terminal": 43, "premature_stop": 266, "recoverable": 4, "pre_excursion": 62, "ordinary": 3506}, "events": {"terminal": 8, "premature_stop": 266, "recoverable": 4, "pre_excursion": 11, "ordinary": 379}, "supervision": {"following": 3834, "recoverable": 3, "terminal": 43, "unknown": 7}, "departed_traces": 10, "terminal_traces": 8, "stop_reasons": {"after_failure": 3, "confidence": 272, "unannotated": 109}, "traced_voxels": 56193.873138371, "seconds": 383.104881439358, "voxels_per_second": 146.68012823863214} | dagger_coverage: {"requested": 384, "seeds": 384, "skipped": [], "epoch": 0} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 0 | dagger_launch_interval: null | dagger_publication_age: 269 | dataset: "0175A_5mm_v1" Step 12,300 | last 50 updates / 1,600 crops | lr 1.92e-05 points (35/curve, tolerance 2 vox): 47,093 right / 3,319 wrong | 93.4% correct | 5,588 unknown loss 1.0122 | geometry 0.2652 | confidence 1.4938 | mean error 0.682 vox speed: 1050 ms/update | data wait 1 ms/update | 30.2 crops/s | 30.2 decisions/s | 1.30 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1495 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 189, from the previous prediction 1306 confidence-labeled crops: terminal 25/1526 (1.6%) | recoverable 119/1526 (7.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 256/1600 (16.0%), 2: 288/1600 (18.0%), 3: 256/1600 (16.0%), 4: 512/1600 (32.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 72 chunks / 144.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 679 rows | mean depth 11 decisions | mean travel 89 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 58% 42% 30%>0% 30% 0 153 88%/10%/0%/2% 9% 122/0/0/0 0175A_5mm_v1 256 43% 48% 23%>9% 15% 21 1092 84%/8%/2%/5% 7% 83/1/0/40 1447_5mm_v1 288 43% 48% 22%>9% 14% 23 639 84%/7%/2%/7% 5% 139/0/0/0 paris4_afv_central 256 36% 43% 29%>21% 8% 49 866 80%/8%/6%/7% 17% 107/1/0/3 PHerc0125_3mm_v2 512 55% 45% 23%>0% 23% 0 148 90%/6%/0%/4% 11% 228/3/0/0 gradients: max 59, clipped 0/50 updates Step 12,350 | last 50 updates / 1,600 crops | lr 1.91e-05 points (35/curve, tolerance 2 vox): 46,304 right / 3,593 wrong | 92.8% correct | 6,103 unknown loss 0.7934 | geometry 0.2544 | confidence 1.0780 | mean error 0.623 vox speed: 893 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.39 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1429 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 207, from the previous prediction 1222 confidence-labeled crops: terminal 53/1516 (3.5%) | recoverable 95/1516 (6.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 288/1600 (18.0%), 2: 448/1600 (28.0%), 3: 416/1600 (26.0%), 4: 160/1600 (10.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 72 chunks / 144.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 677 rows | mean depth 10 decisions | mean travel 85 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 57% 43% 23%>0% 23% 0 136 93%/6%/0%/1% 8% 103/0/128/20 0175A_5mm_v1 288 40% 50% 24%>10% 14% 26 1084 88%/6%/2%/5% 8% 106/0/0/38 1447_5mm_v1 448 40% 49% 26%>12% 14% 34 789 83%/6%/4%/6% 7% 211/3/0/8 paris4_afv_central 416 36% 43% 28%>21% 7% 75 864 82%/6%/6%/6% 17% 180/1/0/0 PHerc0125_3mm_v2 160 52% 48% 21%>0% 21% 0 144 82%/6%/1%/11% 11% 77/2/0/0 gradients: max 63, clipped 1/50 updates Step 12,400 | last 50 updates / 1,600 crops | lr 1.90e-05 points (35/curve, tolerance 2 vox): 47,108 right / 3,157 wrong | 93.7% correct | 5,735 unknown loss 0.9310 | geometry 0.2534 | confidence 1.3553 | mean error 0.656 vox speed: 891 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.29 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1478 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 317, from the previous prediction 1161 confidence-labeled crops: terminal 32/1517 (2.1%) | recoverable 105/1517 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 512/1600 (32.0%), 1: 256/1600 (16.0%), 2: 288/1600 (18.0%), 3: 288/1600 (18.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 72 chunks / 144.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 574 rows | mean depth 11 decisions | mean travel 91 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 512 51% 49% 24%>0% 24% 0 143 90%/8%/0%/2% 9% 175/1/0/75 0175A_5mm_v1 256 44% 46% 25%>10% 14% 22 923 86%/7%/2%/5% 8% 118/0/0/0 1447_5mm_v1 288 43% 47% 22%>10% 11% 32 1405 86%/5%/3%/7% 6% 59/1/0/75 paris4_afv_central 288 35% 45% 27%>20% 7% 51 835 81%/5%/6%/9% 14% 130/2/0/0 PHerc0125_3mm_v2 256 57% 43% 30%>0% 30% 0 152 86%/7%/1%/5% 11% 92/0/9/18 gradients: max 66, clipped 1/50 updates Step 12,450 | last 50 updates / 1,600 crops | lr 1.90e-05 points (35/curve, tolerance 2 vox): 47,046 right / 3,276 wrong | 93.5% correct | 5,678 unknown loss 0.9159 | geometry 0.2580 | confidence 1.3158 | mean error 0.652 vox speed: 893 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.29 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1482 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 208, from the previous prediction 1274 confidence-labeled crops: terminal 34/1519 (2.2%) | recoverable 113/1519 (7.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 256/1600 (16.0%), 2: 352/1600 (22.0%), 3: 416/1600 (26.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 72 chunks / 144.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 692 rows | mean depth 11 decisions | mean travel 93 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 51% 49% 23%>0% 22% 1 149 92%/7%/0%/1% 10% 128/0/0/13 0175A_5mm_v1 256 51% 46% 25%>4% 22% 8 461 88%/6%/1%/5% 4% 117/1/0/0 1447_5mm_v1 352 43% 47% 23%>10% 13% 30 793 86%/8%/1%/5% 5% 144/2/0/23 paris4_afv_central 416 41% 42% 25%>17% 8% 61 815 79%/7%/6%/8% 15% 175/3/0/0 PHerc0125_3mm_v2 288 56% 44% 25%>0% 25% 0 164 86%/7%/0%/7% 13% 128/2/0/0 gradients: max 56, clipped 0/50 updates Step 12,500 | last 50 updates / 1,600 crops | lr 1.89e-05 points (35/curve, tolerance 2 vox): 47,636 right / 3,167 wrong | 93.8% correct | 5,197 unknown loss 0.9388 | geometry 0.2641 | confidence 1.3494 | mean error 0.657 vox speed: 896 ms/update | data wait 1 ms/update | 35.1 crops/s | 35.1 decisions/s | 1.28 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1469 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 200, from the previous prediction 1269 confidence-labeled crops: terminal 28/1523 (1.8%) | recoverable 113/1523 (7.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 192/1600 (12.0%), 2: 512/1600 (32.0%), 3: 320/1600 (20.0%), 4: 160/1600 (10.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 72 chunks / 144.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 712 rows | mean depth 10 decisions | mean travel 87 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 50% 46% 27%>4% 23% 13 274 88%/7%/0%/5% 12% 140/1/0/52 0175A_5mm_v1 192 43% 50% 21%>7% 14% 12 660 88%/7%/1%/4% 5% 96/1/0/0 1447_5mm_v1 512 42% 45% 26%>13% 13% 49 880 84%/8%/3%/5% 8% 232/5/0/0 paris4_afv_central 320 38% 50% 19%>11% 8% 31 564 84%/8%/4%/5% 10% 161/3/0/0 PHerc0125_3mm_v2 160 48% 52% 21%>0% 21% 0 179 89%/4%/0%/7% 10% 83/2/0/0 gradients: max 50, clipped 0/50 updates Step 12,550 | last 50 updates / 1,600 crops | lr 1.88e-05 points (35/curve, tolerance 2 vox): 45,860 right / 3,320 wrong | 93.2% correct | 6,820 unknown loss 0.9756 | geometry 0.2724 | confidence 1.4064 | mean error 0.679 vox speed: 893 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.34 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1474 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 252, from the previous prediction 1222 confidence-labeled crops: terminal 37/1502 (2.5%) | recoverable 102/1502 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 352/1600 (22.0%), 2: 224/1600 (14.0%), 3: 256/1600 (16.0%), 4: 576/1600 (36.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 72 chunks / 144.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 590 rows | mean depth 10 decisions | mean travel 85 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 56% 44% 27%>0% 27% 0 152 89%/9%/0%/2% 5% 84/0/0/0 0175A_5mm_v1 352 45% 44% 24%>11% 13% 32 1158 86%/5%/3%/6% 10% 138/2/76/16 1447_5mm_v1 224 45% 42% 31%>13% 17% 23 991 83%/6%/4%/7% 7% 92/1/0/1 paris4_afv_central 256 37% 42% 29%>21% 8% 45 929 81%/6%/6%/7% 16% 108/3/0/0 PHerc0125_3mm_v2 576 57% 43% 26%>0% 26% 0 148 86%/6%/1%/7% 11% 168/2/0/79 gradients: max 57, clipped 0/50 updates Step 12,600 | last 50 updates / 1,600 crops | lr 1.87e-05 points (35/curve, tolerance 2 vox): 46,754 right / 2,786 wrong | 94.4% correct | 6,460 unknown loss 0.8615 | geometry 0.2369 | confidence 1.2491 | mean error 0.634 vox speed: 892 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.26 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1504 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 328, from the previous prediction 1176 confidence-labeled crops: terminal 25/1509 (1.7%) | recoverable 106/1509 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 544/1600 (34.0%), 2: 224/1600 (14.0%), 3: 224/1600 (14.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 72 chunks / 144.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 558 rows | mean depth 12 decisions | mean travel 99 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 57% 43% 25%>0% 25% 0 172 90%/8%/0%/3% 6% 96/0/0/0 0175A_5mm_v1 544 49% 44% 24%>7% 17% 37 1072 85%/6%/3%/6% 9% 149/2/0/91 1447_5mm_v1 224 47% 45% 23%>8% 15% 16 685 84%/8%/1%/7% 3% 100/2/0/0 paris4_afv_central 224 35% 48% 25%>17% 8% 36 789 86%/4%/4%/6% 15% 107/2/0/0 PHerc0125_3mm_v2 384 58% 42% 26%>0% 26% 0 134 86%/7%/0%/7% 11% 106/2/0/57 gradients: max 48, clipped 0/50 updates Step 12,650 | last 50 updates / 1,600 crops | lr 1.87e-05 points (35/curve, tolerance 2 vox): 47,859 right / 2,684 wrong | 94.7% correct | 5,457 unknown loss 0.9080 | geometry 0.2865 | confidence 1.2429 | mean error 0.649 vox speed: 891 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.24 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1521 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 262, from the previous prediction 1259 confidence-labeled crops: terminal 24/1517 (1.6%) | recoverable 108/1517 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 384/1600 (24.0%), 2: 160/1600 (10.0%), 3: 320/1600 (20.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 72 chunks / 144.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 653 rows | mean depth 13 decisions | mean travel 107 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 52% 48% 20%>0% 20% 0 156 90%/6%/0%/4% 7% 212/0/83/5 0175A_5mm_v1 384 49% 45% 25%>6% 19% 19 960 88%/7%/1%/4% 7% 102/0/0/72 1447_5mm_v1 160 43% 53% 23%>4% 19% 6 398 86%/7%/1%/6% 1% 85/1/0/0 paris4_afv_central 320 38% 47% 24%>16% 9% 45 784 81%/6%/5%/8% 13% 149/1/0/0 PHerc0125_3mm_v2 288 53% 47% 24%>0% 24% 0 138 86%/7%/1%/6% 14% 105/1/0/30 gradients: max 51, clipped 0/50 updates Step 12,700 | last 50 updates / 1,600 crops | lr 1.86e-05 points (35/curve, tolerance 2 vox): 47,597 right / 2,874 wrong | 94.3% correct | 5,529 unknown loss 0.9343 | geometry 0.2843 | confidence 1.3000 | mean error 0.695 vox speed: 898 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.26 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1504 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 247, from the previous prediction 1257 confidence-labeled crops: terminal 21/1526 (1.4%) | recoverable 122/1526 (8.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 480/1600 (30.0%), 1: 256/1600 (16.0%), 2: 320/1600 (20.0%), 3: 288/1600 (18.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 73 chunks / 146.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 597 rows | mean depth 9 decisions | mean travel 79 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 480 55% 45% 25%>0% 25% 0 148 90%/7%/0%/3% 8% 131/4/0/83 0175A_5mm_v1 256 46% 49% 25%>5% 20% 17 938 86%/7%/1%/5% 8% 93/2/0/33 1447_5mm_v1 320 45% 48% 27%>8% 19% 22 669 85%/9%/2%/3% 6% 152/0/0/0 paris4_afv_central 288 40% 42% 24%>18% 6% 44 905 84%/6%/4%/7% 11% 121/0/0/0 PHerc0125_3mm_v2 256 61% 39% 32%>0% 32% 0 166 85%/9%/0%/6% 10% 100/0/0/0 gradients: max 54, clipped 0/50 updates Step 12,750 | last 50 updates / 1,600 crops | lr 1.85e-05 points (35/curve, tolerance 2 vox): 47,832 right / 2,664 wrong | 94.7% correct | 5,504 unknown loss 0.7902 | geometry 0.2273 | confidence 1.1259 | mean error 0.613 vox speed: 896 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.25 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1507 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 185, from the previous prediction 1322 confidence-labeled crops: terminal 32/1529 (2.1%) | recoverable 92/1529 (6.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 416/1600 (26.0%), 2: 288/1600 (18.0%), 3: 224/1600 (14.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 73 chunks / 146.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 737 rows | mean depth 12 decisions | mean travel 98 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 53% 47% 23%>0% 23% 0 139 93%/5%/0%/2% 7% 116/0/0/20 0175A_5mm_v1 416 48% 47% 23%>6% 17% 20 637 87%/7%/1%/5% 6% 194/2/0/0 1447_5mm_v1 288 44% 46% 24%>10% 14% 25 809 86%/5%/4%/4% 7% 132/3/0/0 paris4_afv_central 224 37% 46% 23%>18% 5% 35 879 83%/5%/6%/5% 12% 102/1/0/0 PHerc0125_3mm_v2 384 50% 50% 21%>0% 21% 0 151 89%/6%/0%/6% 10% 193/9/0/0 gradients: max 47, clipped 0/50 updates Step 12,800 | last 50 updates / 1,600 crops | lr 1.84e-05 points (35/curve, tolerance 2 vox): 47,814 right / 2,849 wrong | 94.4% correct | 5,337 unknown loss 0.8512 | geometry 0.2226 | confidence 1.2573 | mean error 0.612 vox speed: 895 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.27 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1462 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 249, from the previous prediction 1213 confidence-labeled crops: terminal 30/1531 (2.0%) | recoverable 107/1531 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 672/1600 (42.0%), 2: 256/1600 (16.0%), 3: 224/1600 (14.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 73 chunks / 146.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 618 rows | mean depth 11 decisions | mean travel 88 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 50% 50% 23%>0% 23% 0 149 89%/7%/0%/3% 9% 127/0/0/0 0175A_5mm_v1 672 46% 44% 25%>10% 16% 52 1242 88%/6%/2%/4% 8% 219/5/0/79 1447_5mm_v1 256 45% 47% 27%>9% 18% 20 815 88%/6%/1%/5% 5% 96/1/21/24 paris4_afv_central 224 36% 43% 31%>21% 9% 43 1034 80%/8%/5%/8% 13% 96/1/0/0 PHerc0125_3mm_v2 192 58% 42% 27%>0% 27% 0 144 87%/9%/1%/4% 11% 80/2/0/0 gradients: max 52, clipped 0/50 updates Step 12,850 | last 50 updates / 1,600 crops | lr 1.84e-05 points (35/curve, tolerance 2 vox): 46,697 right / 2,632 wrong | 94.7% correct | 6,671 unknown loss 0.8175 | geometry 0.2523 | confidence 1.1302 | mean error 0.641 vox speed: 891 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.32 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1484 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 331, from the previous prediction 1153 confidence-labeled crops: terminal 20/1506 (1.3%) | recoverable 104/1506 (6.9%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 384/1600 (24.0%), 2: 480/1600 (30.0%), 3: 128/1600 (8.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 73 chunks / 146.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 548 rows | mean depth 10 decisions | mean travel 86 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 61% 39% 27%>0% 27% 0 170 89%/8%/0%/4% 10% 87/1/0/0 0175A_5mm_v1 384 42% 51% 22%>7% 14% 23 1092 88%/5%/1%/6% 7% 129/1/0/66 1447_5mm_v1 480 42% 49% 22%>10% 13% 40 965 88%/5%/2%/6% 5% 158/3/0/75 paris4_afv_central 128 36% 40% 30%>24% 6% 35 2138 72%/12%/3%/12% 16% 13/0/39/38 PHerc0125_3mm_v2 384 55% 45% 27%>0% 27% 0 135 86%/7%/0%/6% 11% 161/1/65/13 gradients: max 55, clipped 0/50 updates Step 12,900 | last 50 updates / 1,600 crops | lr 1.83e-05 points (35/curve, tolerance 2 vox): 47,436 right / 2,683 wrong | 94.6% correct | 5,881 unknown loss 0.9282 | geometry 0.2521 | confidence 1.3523 | mean error 0.657 vox speed: 892 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.25 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1474 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 288, from the previous prediction 1186 confidence-labeled crops: terminal 22/1511 (1.5%) | recoverable 115/1511 (7.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 192/1600 (12.0%), 2: 384/1600 (24.0%), 3: 224/1600 (14.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 73 chunks / 146.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 564 rows | mean depth 14 decisions | mean travel 116 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 50% 47% 25%>3% 22% 10 259 88%/8%/0%/4% 10% 211/1/0/0 0175A_5mm_v1 192 43% 42% 35%>15% 20% 24 1409 81%/7%/3%/8% 11% 80/1/0/0 1447_5mm_v1 384 48% 46% 24%>6% 18% 26 942 86%/8%/2%/4% 5% 106/1/0/70 paris4_afv_central 224 40% 43% 25%>17% 8% 50 2032 78%/9%/3%/10% 15% 25/0/94/71 PHerc0125_3mm_v2 352 60% 40% 28%>0% 28% 0 149 90%/4%/0%/6% 9% 142/3/0/0 gradients: max 44, clipped 0/50 updates Step 12,950 | last 50 updates / 1,600 crops | lr 1.82e-05 points (35/curve, tolerance 2 vox): 46,820 right / 3,595 wrong | 92.9% correct | 5,585 unknown loss 0.9082 | geometry 0.2962 | confidence 1.2240 | mean error 0.701 vox speed: 891 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.32 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1453 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 247, from the previous prediction 1206 confidence-labeled crops: terminal 39/1522 (2.6%) | recoverable 111/1522 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 320/1600 (20.0%), 2: 416/1600 (26.0%), 3: 352/1600 (22.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 73 chunks / 146.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 652 rows | mean depth 12 decisions | mean travel 98 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 55% 44% 26%>2% 25% 4 240 89%/8%/0%/3% 11% 119/1/102/21 0175A_5mm_v1 320 44% 47% 26%>9% 17% 28 947 87%/7%/3%/4% 9% 149/5/0/0 1447_5mm_v1 416 46% 44% 20%>9% 11% 32 875 87%/6%/2%/5% 5% 161/4/0/23 paris4_afv_central 352 34% 48% 24%>18% 6% 56 1072 79%/7%/6%/7% 16% 146/3/0/24 PHerc0125_3mm_v2 192 53% 47% 23%>0% 23% 0 139 89%/7%/0%/5% 11% 77/1/4/14 gradients: max 54, clipped 0/50 updates Step 13,000 | last 50 updates / 1,600 crops | lr 1.81e-05 points (35/curve, tolerance 2 vox): 46,657 right / 2,872 wrong | 94.2% correct | 6,471 unknown loss 0.8610 | geometry 0.2820 | confidence 1.1580 | mean error 0.659 vox speed: 897 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.31 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1469 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 234, from the previous prediction 1235 confidence-labeled crops: terminal 21/1512 (1.4%) | recoverable 98/1512 (6.5%) | target connections crossing a neighbor 3 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 320/1600 (20.0%), 2: 512/1600 (32.0%), 3: 192/1600 (12.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 73 chunks / 146.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 622 rows | mean depth 9 decisions | mean travel 79 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 47% 49% 21%>3% 18% 4 263 89%/8%/0%/3% 11% 83/1/0/12 0175A_5mm_v1 320 53% 40% 25%>7% 18% 19 866 88%/7%/2%/4% 7% 127/4/0/0 1447_5mm_v1 512 44% 48% 22%>9% 14% 35 823 87%/5%/1%/6% 4% 216/4/0/28 paris4_afv_central 192 32% 49% 24%>19% 5% 49 1872 85%/5%/5%/6% 18% 35/0/0/59 PHerc0125_3mm_v2 384 58% 42% 26%>0% 26% 0 148 86%/6%/1%/7% 12% 161/2/0/0 gradients: max 60, clipped 1/50 updates Step 13,000 | dagger_launched: true Step 13,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 247, "stopped": 1, "mean_example_error": 1.5483533684164286, "geometry_examples": 32, "examples": 32, "inference_and_measurement_seconds": 1.9499311186373234, "total_diagnostic_seconds": 3.845841057598591} Step 13,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 6/8 (75.0%) | false stops 0 1.5-3: recovered 6/8 (75.0%) | false stops 1 3-4.5: recovered 6/8 (75.0%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 0 entire proposed 8-point prefix correct (distance): 21/28 (75.0%) gate @ 0.50: false stops 1/23 (4.3%) | accepted wrong 6/27 (22.2%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.447 1.349 following 16 464 0.940 0.853 recoverable 12 348 2.122 2.011 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 0.940 0.853 332 32 812 1.447 1.349 improved/worsened/compared states: 4/0/28 Step 13,000 | batch_diagnostics_seconds: 3.852463059127331 | recovery_seconds: 15.436366029083729 Step 13,050 | last 50 updates / 1,600 crops | lr 1.80e-05 points (35/curve, tolerance 2 vox): 46,569 right / 3,779 wrong | 92.5% correct | 5,652 unknown loss 0.9336 | geometry 0.2772 | confidence 1.3127 | mean error 0.677 vox speed: 1321 ms/update | data wait 1 ms/update | 18.3 crops/s | 18.3 decisions/s | 1.31 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1454 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 188, from the previous prediction 1266 confidence-labeled crops: terminal 39/1526 (2.6%) | recoverable 95/1526 (6.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 192/1600 (12.0%), 2: 352/1600 (22.0%), 3: 288/1600 (18.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 73 chunks / 146.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 696 rows | mean depth 13 decisions | mean travel 110 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 50% 48% 26%>2% 24% 7 209 93%/6%/0%/1% 9% 216/1/0/0 0175A_5mm_v1 192 40% 47% 22%>14% 9% 23 1238 82%/5%/8%/5% 12% 90/3/0/0 1447_5mm_v1 352 47% 43% 25%>10% 15% 30 990 86%/7%/2%/5% 6% 132/2/0/20 paris4_afv_central 288 33% 41% 30%>26% 4% 68 1259 78%/7%/6%/9% 17% 119/3/0/0 PHerc0125_3mm_v2 320 55% 45% 22%>0% 22% 0 142 90%/4%/0%/6% 9% 139/3/9/6 gradients: max 65, clipped 1/50 updates Step 13,100 | last 50 updates / 1,600 crops | lr 1.80e-05 points (35/curve, tolerance 2 vox): 46,128 right / 3,947 wrong | 92.1% correct | 5,925 unknown loss 1.0010 | geometry 0.2941 | confidence 1.4137 | mean error 0.686 vox speed: 1383 ms/update | data wait 1 ms/update | 22.9 crops/s | 22.9 decisions/s | 1.37 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1447 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 287, from the previous prediction 1160 confidence-labeled crops: terminal 47/1513 (3.1%) | recoverable 93/1513 (6.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 224/1600 (14.0%), 2: 224/1600 (14.0%), 3: 512/1600 (32.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 73 chunks / 146.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 594 rows | mean depth 10 decisions | mean travel 90 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 55% 44% 25%>0% 24% 1 157 89%/8%/0%/3% 8% 128/1/0/0 0175A_5mm_v1 224 46% 44% 26%>10% 16% 20 1082 87%/4%/4%/5% 8% 98/2/0/0 1447_5mm_v1 224 37% 49% 23%>14% 9% 26 1012 85%/4%/5%/6% 10% 110/0/0/0 paris4_afv_central 512 37% 45% 25%>18% 7% 81 1230 82%/6%/5%/6% 17% 165/4/0/65 PHerc0125_3mm_v2 352 59% 41% 24%>0% 24% 0 132 86%/6%/0%/8% 11% 93/4/0/52 gradients: max 48, clipped 0/50 updates Step 13,150 | last 50 updates / 1,600 crops | lr 1.79e-05 points (35/curve, tolerance 2 vox): 46,832 right / 3,391 wrong | 93.2% correct | 5,777 unknown loss 0.9864 | geometry 0.2734 | confidence 1.4261 | mean error 0.689 vox speed: 1366 ms/update | data wait 1 ms/update | 23.2 crops/s | 23.2 decisions/s | 1.29 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1481 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 286, from the previous prediction 1195 confidence-labeled crops: terminal 27/1523 (1.8%) | recoverable 111/1523 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 448/1600 (28.0%), 2: 256/1600 (16.0%), 3: 224/1600 (14.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 75 chunks / 150.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 590 rows | mean depth 12 decisions | mean travel 96 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 51% 46% 23%>3% 20% 12 372 89%/6%/0%/5% 10% 111/0/0/82 0175A_5mm_v1 448 48% 45% 24%>7% 16% 26 848 86%/6%/2%/6% 9% 202/3/0/0 1447_5mm_v1 256 47% 46% 21%>7% 14% 16 698 87%/6%/1%/6% 4% 117/1/0/0 paris4_afv_central 224 39% 45% 28%>17% 12% 48 1746 78%/11%/6%/6% 18% 45/0/0/55 PHerc0125_3mm_v2 256 55% 45% 27%>0% 27% 0 144 90%/7%/0%/4% 10% 115/3/0/0 gradients: max 53, clipped 0/50 updates Step 13,200 | last 50 updates / 1,600 crops | lr 1.78e-05 points (35/curve, tolerance 2 vox): 46,705 right / 3,547 wrong | 92.9% correct | 5,748 unknown loss 0.8118 | geometry 0.2577 | confidence 1.1081 | mean error 0.637 vox speed: 1373 ms/update | data wait 1 ms/update | 23.1 crops/s | 23.1 decisions/s | 1.30 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1450 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 376, from the previous prediction 1074 confidence-labeled crops: terminal 46/1520 (3.0%) | recoverable 125/1520 (8.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 544/1600 (34.0%), 2: 256/1600 (16.0%), 3: 384/1600 (24.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 75 chunks / 150.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 485 rows | mean depth 13 decisions | mean travel 107 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 56% 44% 22%>0% 22% 0 141 88%/9%/0%/3% 7% 98/1/0/0 0175A_5mm_v1 544 46% 45% 23%>10% 13% 47 2327 87%/6%/3%/4% 9% 82/1/128/161 1447_5mm_v1 256 50% 41% 30%>9% 21% 20 899 83%/9%/1%/7% 5% 104/2/10/0 paris4_afv_central 384 41% 44% 25%>15% 10% 51 1256 79%/9%/6%/5% 14% 110/3/0/59 PHerc0125_3mm_v2 192 53% 47% 22%>0% 22% 0 154 86%/8%/1%/6% 13% 91/2/0/0 gradients: max 51, clipped 0/50 updates Step 13,250 | last 50 updates / 1,600 crops | lr 1.77e-05 points (35/curve, tolerance 2 vox): 47,083 right / 3,462 wrong | 93.2% correct | 5,455 unknown loss 0.8728 | geometry 0.2484 | confidence 1.2487 | mean error 0.644 vox speed: 1379 ms/update | data wait 1 ms/update | 23.0 crops/s | 23.0 decisions/s | 1.36 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1452 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 209, from the previous prediction 1243 confidence-labeled crops: terminal 43/1524 (2.8%) | recoverable 113/1524 (7.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 128/1600 (8.0%), 2: 544/1600 (34.0%), 3: 256/1600 (16.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 668 rows | mean depth 13 decisions | mean travel 109 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 54% 46% 25%>0% 25% 0 157 90%/8%/0%/2% 8% 146/1/0/0 0175A_5mm_v1 128 38% 49% 20%>13% 6% 16 1228 88%/3%/3%/5% 9% 63/1/0/0 1447_5mm_v1 544 43% 47% 22%>10% 11% 37 1149 84%/6%/4%/5% 7% 191/5/67/63 paris4_afv_central 256 30% 43% 32%>27% 5% 60 1334 79%/9%/5%/7% 17% 111/1/0/0 PHerc0125_3mm_v2 352 55% 45% 26%>0% 26% 0 156 86%/8%/1%/6% 12% 157/1/0/0 gradients: max 52, clipped 0/50 updates Step 13,289 | dagger_states: 4266 | dagger_source_step: 13000 | dagger_caches: 3 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/1447_5mm_v1/decisions_013000.npz" | dagger_fibers: 384 | dagger_supply: {"states": 4266, "traces": 384, "fibers": 384, "rows": {"terminal": 62, "premature_stop": 278, "recoverable": 2, "pre_excursion": 70, "ordinary": 3853}, "events": {"terminal": 9, "premature_stop": 278, "recoverable": 2, "pre_excursion": 12, "ordinary": 378}, "supervision": {"following": 4199, "recoverable": 4, "terminal": 62, "unknown": 1}, "departed_traces": 12, "terminal_traces": 9, "stop_reasons": {"after_failure": 6, "confidence": 281, "unannotated": 97}, "traced_voxels": 61793.140096645, "seconds": 408.24456445872784, "voxels_per_second": 151.3630443030481} | dagger_coverage: {"requested": 384, "seeds": 384, "skipped": [], "epoch": 0} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 0 | dagger_launch_interval: null | dagger_publication_age: 289 | dataset: "1447_5mm_v1" Step 13,300 | last 50 updates / 1,600 crops | lr 1.77e-05 points (35/curve, tolerance 2 vox): 46,555 right / 3,849 wrong | 92.4% correct | 5,596 unknown loss 1.0112 | geometry 0.2789 | confidence 1.4648 | mean error 0.681 vox speed: 1217 ms/update | data wait 1 ms/update | 26.1 crops/s | 26.1 decisions/s | 1.31 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1470 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 216, from the previous prediction 1254 confidence-labeled crops: terminal 45/1525 (3.0%) | recoverable 98/1525 (6.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 256/1600 (16.0%), 2: 256/1600 (16.0%), 3: 416/1600 (26.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 688 rows | mean depth 9 decisions | mean travel 75 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 52% 46% 25%>2% 23% 6 244 89%/6%/0%/4% 10% 146/0/0/0 0175A_5mm_v1 256 43% 48% 21%>9% 12% 21 924 88%/5%/2%/4% 9% 124/0/0/0 1447_5mm_v1 256 41% 49% 22%>11% 11% 20 1305 84%/6%/4%/5% 7% 69/3/0/56 paris4_afv_central 416 37% 45% 27%>18% 9% 63 1045 81%/6%/7%/7% 15% 188/4/18/1 PHerc0125_3mm_v2 352 53% 47% 22%>0% 22% 0 143 91%/7%/0%/3% 11% 161/6/17/4 gradients: max 45, clipped 0/50 updates Step 13,350 | last 50 updates / 1,600 crops | lr 1.76e-05 points (35/curve, tolerance 2 vox): 46,182 right / 3,651 wrong | 92.7% correct | 6,167 unknown loss 0.9321 | geometry 0.2813 | confidence 1.3016 | mean error 0.658 vox speed: 890 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.34 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1460 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 266, from the previous prediction 1194 confidence-labeled crops: terminal 41/1508 (2.7%) | recoverable 100/1508 (6.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 512/1600 (32.0%), 2: 192/1600 (12.0%), 3: 288/1600 (18.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 635 rows | mean depth 14 decisions | mean travel 110 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 54% 43% 25%>2% 23% 5 262 91%/7%/0%/3% 10% 137/0/36/2 0175A_5mm_v1 512 48% 42% 27%>10% 16% 38 1187 83%/7%/4%/6% 10% 213/1/0/0 1447_5mm_v1 192 36% 50% 26%>14% 12% 22 1075 85%/4%/3%/8% 8% 96/2/0/0 paris4_afv_central 288 38% 44% 25%>18% 7% 45 1065 81%/8%/4%/7% 14% 127/0/0/0 PHerc0125_3mm_v2 288 52% 48% 24%>0% 24% 0 135 88%/5%/0%/7% 13% 62/0/0/75 gradients: max 62, clipped 1/50 updates Step 13,400 | last 50 updates / 1,600 crops | lr 1.75e-05 points (35/curve, tolerance 2 vox): 46,101 right / 3,628 wrong | 92.7% correct | 6,271 unknown loss 0.9539 | geometry 0.2714 | confidence 1.3649 | mean error 0.658 vox speed: 893 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.33 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1438 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 272, from the previous prediction 1166 confidence-labeled crops: terminal 38/1513 (2.5%) | recoverable 112/1513 (7.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 448/1600 (28.0%), 2: 224/1600 (14.0%), 3: 448/1600 (28.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 1337 chunk references / 1 windows live chains: 581 rows | mean depth 12 decisions | mean travel 100 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 50% 49% 25%>0% 25% 1 156 92%/7%/0%/1% 10% 132/2/38/10 0175A_5mm_v1 448 47% 44% 25%>9% 16% 30 1127 84%/7%/4%/6% 8% 179/5/0/19 1447_5mm_v1 224 54% 39% 29%>8% 22% 15 857 85%/8%/1%/6% 3% 87/0/0/0 paris4_afv_central 448 40% 42% 23%>18% 5% 86 1816 80%/7%/4%/8% 17% 96/1/9/92 PHerc0125_3mm_v2 192 55% 45% 23%>0% 23% 0 142 90%/6%/0%/4% 10% 87/0/0/0 gradients: max 62, clipped 1/50 updates Step 13,450 | last 50 updates / 1,600 crops | lr 1.74e-05 points (35/curve, tolerance 2 vox): 46,657 right / 3,534 wrong | 93.0% correct | 5,809 unknown loss 0.9557 | geometry 0.2776 | confidence 1.3562 | mean error 0.679 vox speed: 895 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.30 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1462 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 206, from the previous prediction 1256 confidence-labeled crops: terminal 37/1532 (2.4%) | recoverable 113/1532 (7.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 224/1600 (14.0%), 2: 416/1600 (26.0%), 3: 352/1600 (22.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 926 chunk references / 1 windows live chains: 719 rows | mean depth 10 decisions | mean travel 87 vox | 6 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 41% 58% 18%>0% 17% 1 158 92%/5%/0%/2% 8% 139/0/0/10 0175A_5mm_v1 224 46% 41% 30%>12% 18% 27 1499 86%/6%/3%/5% 11% 81/1/0/11 1447_5mm_v1 416 44% 47% 21%>9% 12% 28 865 89%/6%/2%/2% 6% 196/5/16/1 paris4_afv_central 352 38% 41% 31%>20% 11% 60 1222 75%/12%/5%/8% 16% 145/2/0/1 PHerc0125_3mm_v2 352 55% 45% 23%>0% 23% 0 159 90%/5%/1%/4% 11% 158/0/0/0 gradients: max 75, clipped 1/50 updates Step 13,500 | last 50 updates / 1,600 crops | lr 1.73e-05 points (35/curve, tolerance 2 vox): 45,494 right / 3,357 wrong | 93.1% correct | 7,149 unknown loss 0.8108 | geometry 0.2565 | confidence 1.1087 | mean error 0.631 vox speed: 896 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.34 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1411 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 326, from the previous prediction 1085 confidence-labeled crops: terminal 45/1497 (3.0%) | recoverable 106/1497 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 96/1600 (6.0%), 2: 416/1600 (26.0%), 3: 544/1600 (34.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 502 rows | mean depth 14 decisions | mean travel 117 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 56% 44% 26%>0% 26% 0 135 91%/7%/0%/1% 8% 99/0/0/57 0175A_5mm_v1 96 47% 42% 25%>11% 14% 10 1314 88%/3%/3%/6% 7% 40/1/0/0 1447_5mm_v1 416 44% 42% 27%>13% 14% 40 1603 82%/6%/4%/8% 8% 124/0/26/52 paris4_afv_central 544 38% 46% 22%>16% 6% 84 1551 79%/7%/4%/9% 14% 157/2/0/95 PHerc0125_3mm_v2 192 57% 43% 22%>0% 22% 0 173 85%/8%/0%/7% 11% 82/2/0/0 gradients: max 70, clipped 1/50 updates Step 13,550 | last 50 updates / 1,600 crops | lr 1.73e-05 points (35/curve, tolerance 2 vox): 47,049 right / 3,638 wrong | 92.8% correct | 5,313 unknown loss 1.0056 | geometry 0.2839 | confidence 1.4434 | mean error 0.681 vox speed: 893 ms/update | data wait 1 ms/update | 35.1 crops/s | 35.1 decisions/s | 1.29 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1476 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 285, from the previous prediction 1191 confidence-labeled crops: terminal 39/1537 (2.5%) | recoverable 115/1537 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 288/1600 (18.0%), 2: 256/1600 (16.0%), 3: 288/1600 (18.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 595 rows | mean depth 12 decisions | mean travel 101 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 53% 47% 23%>0% 23% 0 136 91%/7%/0%/1% 9% 133/0/0/77 0175A_5mm_v1 288 42% 46% 29%>12% 17% 30 1527 82%/8%/4%/5% 10% 113/1/45/19 1447_5mm_v1 256 40% 48% 23%>12% 10% 28 1431 85%/4%/5%/6% 11% 89/0/0/33 paris4_afv_central 288 41% 42% 24%>17% 7% 40 1113 84%/7%/4%/5% 15% 121/2/0/0 PHerc0125_3mm_v2 320 57% 43% 28%>0% 28% 0 157 85%/9%/0%/6% 12% 139/2/0/0 gradients: max 74, clipped 3/50 updates Step 13,600 | last 50 updates / 1,600 crops | lr 1.72e-05 points (35/curve, tolerance 2 vox): 47,362 right / 2,848 wrong | 94.3% correct | 5,790 unknown loss 0.8971 | geometry 0.2478 | confidence 1.2986 | mean error 0.663 vox speed: 897 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.26 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1502 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 249, from the previous prediction 1253 confidence-labeled crops: terminal 25/1516 (1.6%) | recoverable 110/1516 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 512/1600 (32.0%), 1: 352/1600 (22.0%), 2: 160/1600 (10.0%), 3: 352/1600 (22.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 678 rows | mean depth 11 decisions | mean travel 94 vox | 6 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 512 53% 47% 24%>0% 24% 0 135 89%/8%/0%/4% 6% 198/3/0/45 0175A_5mm_v1 352 47% 46% 21%>7% 14% 24 1013 86%/7%/1%/5% 9% 137/3/0/25 1447_5mm_v1 160 48% 44% 24%>8% 16% 12 843 88%/7%/2%/3% 5% 70/0/0/0 paris4_afv_central 352 36% 48% 24%>17% 7% 55 1037 81%/7%/4%/9% 13% 168/5/0/0 PHerc0125_3mm_v2 224 53% 47% 23%>0% 23% 0 154 88%/4%/1%/6% 10% 105/3/0/0 gradients: max 48, clipped 0/50 updates Step 13,650 | last 50 updates / 1,600 crops | lr 1.71e-05 points (35/curve, tolerance 2 vox): 47,008 right / 3,048 wrong | 93.9% correct | 5,944 unknown loss 0.8708 | geometry 0.2690 | confidence 1.2036 | mean error 0.653 vox speed: 896 ms/update | data wait 1 ms/update | 35.1 crops/s | 35.1 decisions/s | 1.25 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1477 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 332, from the previous prediction 1145 confidence-labeled crops: terminal 30/1517 (2.0%) | recoverable 111/1517 (7.3%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 352/1600 (22.0%), 2: 320/1600 (20.0%), 3: 288/1600 (18.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 550 rows | mean depth 14 decisions | mean travel 114 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 52% 48% 21%>0% 20% 1 163 90%/9%/0%/1% 8% 112/0/0/11 0175A_5mm_v1 352 48% 44% 24%>8% 16% 34 2525 83%/9%/2%/6% 6% 53/3/0/102 1447_5mm_v1 320 45% 46% 25%>9% 16% 23 912 86%/7%/3%/4% 10% 147/3/2/0 paris4_afv_central 288 41% 41% 23%>18% 5% 49 1212 81%/7%/4%/8% 12% 118/1/0/0 PHerc0125_3mm_v2 384 55% 45% 24%>0% 24% 0 137 90%/4%/0%/6% 12% 120/3/67/51 gradients: max 53, clipped 0/50 updates Step 13,700 | last 50 updates / 1,600 crops | lr 1.70e-05 points (35/curve, tolerance 2 vox): 47,123 right / 2,894 wrong | 94.2% correct | 5,983 unknown loss 0.8431 | geometry 0.2463 | confidence 1.1935 | mean error 0.640 vox speed: 902 ms/update | data wait 1 ms/update | 34.9 crops/s | 34.9 decisions/s | 1.27 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1486 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 229, from the previous prediction 1257 confidence-labeled crops: terminal 29/1514 (1.9%) | recoverable 111/1514 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 256/1600 (16.0%), 2: 224/1600 (14.0%), 3: 448/1600 (28.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 620 rows | mean depth 11 decisions | mean travel 99 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 59% 39% 29%>2% 27% 6 262 88%/8%/0%/5% 7% 164/1/0/0 0175A_5mm_v1 256 52% 42% 29%>6% 23% 16 982 87%/7%/1%/5% 5% 89/1/0/19 1447_5mm_v1 224 46% 47% 21%>7% 15% 14 718 88%/5%/3%/3% 5% 105/4/0/0 paris4_afv_central 448 41% 44% 25%>15% 10% 68 1317 81%/7%/4%/8% 15% 151/5/100/47 PHerc0125_3mm_v2 256 56% 44% 23%>0% 23% 0 137 87%/7%/0%/6% 10% 111/5/0/1 gradients: max 50, clipped 0/50 updates Step 13,750 | last 50 updates / 1,600 crops | lr 1.70e-05 points (35/curve, tolerance 2 vox): 47,385 right / 3,197 wrong | 93.7% correct | 5,418 unknown loss 0.8587 | geometry 0.2843 | confidence 1.1487 | mean error 0.647 vox speed: 897 ms/update | data wait 1 ms/update | 35.1 crops/s | 35.1 decisions/s | 1.29 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1453 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 285, from the previous prediction 1168 confidence-labeled crops: terminal 39/1524 (2.6%) | recoverable 113/1524 (7.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 192/1600 (12.0%), 2: 288/1600 (18.0%), 3: 416/1600 (26.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 1762 chunk references / 1 windows live chains: 572 rows | mean depth 12 decisions | mean travel 97 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 59% 38% 28%>3% 25% 12 365 90%/8%/0%/2% 6% 144/1/0/0 0175A_5mm_v1 192 50% 45% 19%>5% 14% 9 744 88%/5%/3%/5% 8% 86/0/0/0 1447_5mm_v1 288 41% 43% 28%>16% 12% 38 1775 84%/8%/3%/5% 8% 85/0/59/39 paris4_afv_central 416 40% 43% 26%>17% 9% 63 1355 80%/7%/6%/7% 14% 150/1/0/28 PHerc0125_3mm_v2 320 56% 44% 26%>0% 26% 0 136 88%/7%/0%/5% 10% 107/1/0/33 gradients: max 45, clipped 0/50 updates Step 13,800 | last 50 updates / 1,600 crops | lr 1.69e-05 points (35/curve, tolerance 2 vox): 45,988 right / 3,916 wrong | 92.2% correct | 6,096 unknown loss 0.9997 | geometry 0.2768 | confidence 1.4458 | mean error 0.677 vox speed: 897 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.36 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1430 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 299, from the previous prediction 1131 confidence-labeled crops: terminal 48/1515 (3.2%) | recoverable 116/1515 (7.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 256/1600 (16.0%), 2: 416/1600 (26.0%), 3: 320/1600 (20.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 569 rows | mean depth 13 decisions | mean travel 110 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 51% 44% 26%>5% 21% 6 445 89%/6%/0%/5% 9% 82/2/39/2 0175A_5mm_v1 256 48% 43% 26%>9% 16% 20 1172 81%/10%/5%/5% 11% 110/1/1/0 1447_5mm_v1 416 43% 45% 25%>12% 14% 39 1554 84%/8%/3%/5% 8% 126/3/0/63 paris4_afv_central 320 39% 40% 28%>21% 7% 66 2023 76%/8%/7%/8% 20% 75/1/0/53 PHerc0125_3mm_v2 416 52% 48% 24%>0% 24% 0 140 90%/5%/0%/4% 10% 176/3/0/22 gradients: max 55, clipped 0/50 updates Step 13,850 | last 50 updates / 1,600 crops | lr 1.68e-05 points (35/curve, tolerance 2 vox): 47,640 right / 2,649 wrong | 94.7% correct | 5,711 unknown loss 0.8468 | geometry 0.2406 | confidence 1.2124 | mean error 0.649 vox speed: 898 ms/update | data wait 1 ms/update | 35.1 crops/s | 35.1 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1506 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 264, from the previous prediction 1242 confidence-labeled crops: terminal 22/1514 (1.5%) | recoverable 112/1514 (7.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 224/1600 (14.0%), 2: 384/1600 (24.0%), 3: 224/1600 (14.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 631 rows | mean depth 13 decisions | mean travel 108 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 51% 47% 25%>2% 24% 6 252 90%/7%/0%/3% 8% 196/2/0/0 0175A_5mm_v1 224 50% 45% 30%>5% 25% 11 735 83%/11%/0%/6% 6% 95/1/68/5 1447_5mm_v1 384 47% 48% 20%>6% 14% 27 1042 88%/7%/2%/4% 5% 125/2/0/58 paris4_afv_central 224 35% 44% 32%>21% 12% 41 1325 78%/6%/6%/9% 14% 99/1/0/0 PHerc0125_3mm_v2 352 57% 43% 25%>0% 25% 0 133 88%/6%/0%/7% 9% 116/2/0/35 gradients: max 51, clipped 0/50 updates Step 13,900 | last 50 updates / 1,600 crops | lr 1.67e-05 points (35/curve, tolerance 2 vox): 47,456 right / 2,634 wrong | 94.7% correct | 5,910 unknown loss 0.7494 | geometry 0.2128 | confidence 1.0732 | mean error 0.605 vox speed: 895 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.25 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1486 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 269, from the previous prediction 1217 confidence-labeled crops: terminal 28/1508 (1.9%) | recoverable 111/1508 (7.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 352/1600 (22.0%), 2: 448/1600 (28.0%), 3: 352/1600 (22.0%), 4: 160/1600 (10.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 589 rows | mean depth 11 decisions | mean travel 89 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 50% 46% 26%>4% 22% 10 380 87%/7%/0%/6% 8% 132/0/0/0 0175A_5mm_v1 352 50% 43% 28%>7% 22% 20 1174 84%/9%/1%/6% 6% 109/3/60/43 1447_5mm_v1 448 50% 46% 27%>5% 22% 20 774 87%/6%/1%/6% 3% 134/2/0/70 paris4_afv_central 352 40% 43% 24%>16% 8% 53 1170 82%/6%/6%/7% 15% 153/6/0/0 PHerc0125_3mm_v2 160 62% 38% 29%>0% 29% 0 145 89%/7%/0%/4% 7% 61/0/0/0 gradients: max 48, clipped 0/50 updates Step 13,950 | last 50 updates / 1,600 crops | lr 1.66e-05 points (35/curve, tolerance 2 vox): 46,648 right / 3,419 wrong | 93.2% correct | 5,933 unknown loss 0.9734 | geometry 0.2897 | confidence 1.3674 | mean error 0.689 vox speed: 895 ms/update | data wait 1 ms/update | 34.8 crops/s | 34.8 decisions/s | 1.27 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1492 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 332, from the previous prediction 1160 confidence-labeled crops: terminal 33/1529 (2.2%) | recoverable 104/1529 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 352/1600 (22.0%), 2: 256/1600 (16.0%), 3: 256/1600 (16.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 567 rows | mean depth 15 decisions | mean travel 123 vox | 7 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 50% 47% 26%>2% 23% 9 367 90%/6%/0%/4% 10% 121/1/47/91 0175A_5mm_v1 352 44% 47% 22%>9% 13% 30 1660 85%/6%/3%/6% 8% 98/2/0/68 1447_5mm_v1 256 54% 43% 27%>4% 24% 9 503 85%/9%/2%/4% 4% 110/1/0/0 paris4_afv_central 256 37% 42% 27%>21% 7% 47 1406 84%/6%/4%/6% 16% 108/0/0/0 PHerc0125_3mm_v2 288 55% 45% 23%>0% 23% 0 159 89%/6%/2%/3% 12% 130/2/0/0 gradients: max 47, clipped 0/50 updates Step 14,000 | last 50 updates / 1,600 crops | lr 1.66e-05 points (35/curve, tolerance 2 vox): 47,437 right / 2,861 wrong | 94.3% correct | 5,702 unknown loss 0.8660 | geometry 0.2397 | confidence 1.2526 | mean error 0.646 vox speed: 890 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.26 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1500 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 242, from the previous prediction 1258 confidence-labeled crops: terminal 26/1520 (1.7%) | recoverable 109/1520 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 288/1600 (18.0%), 2: 288/1600 (18.0%), 3: 256/1600 (16.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 636 rows | mean depth 14 decisions | mean travel 115 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 56% 41% 26%>2% 24% 7 285 88%/7%/0%/5% 8% 165/0/0/7 0175A_5mm_v1 288 49% 44% 26%>7% 18% 22 1540 84%/7%/2%/7% 7% 73/3/0/53 1447_5mm_v1 288 47% 44% 24%>9% 15% 23 970 87%/6%/3%/5% 6% 127/2/0/0 paris4_afv_central 256 37% 47% 27%>16% 11% 39 1145 82%/9%/5%/5% 13% 120/1/0/0 PHerc0125_3mm_v2 352 57% 43% 24%>0% 24% 0 154 89%/6%/0%/5% 11% 151/4/0/0 gradients: max 50, clipped 0/50 updates Step 14,000 | dagger_launched: true Step 14,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 248, "stopped": 1, "mean_example_error": 0.38995412376619154, "geometry_examples": 31, "examples": 32, "inference_and_measurement_seconds": 1.926209393888712, "total_diagnostic_seconds": 3.7135002724826336} Step 14,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 6/8 (75.0%) | false stops 0 1.5-3: recovered 6/8 (75.0%) | false stops 0 3-4.5: recovered 6/8 (75.0%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 0 entire proposed 8-point prefix correct (distance): 22/28 (78.6%) gate @ 0.50: false stops 0/23 (0.0%) | accepted wrong 5/27 (18.5%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.471 1.425 following 16 464 0.993 0.988 recoverable 12 348 2.110 2.009 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 0.993 0.988 332 32 812 1.471 1.425 improved/worsened/compared states: 4/0/28 Step 14,000 | batch_diagnostics_seconds: 3.7222886718809605 | recovery_seconds: 17.080261901021004 Step 14,050 | last 50 updates / 1,600 crops | lr 1.65e-05 points (35/curve, tolerance 2 vox): 46,817 right / 2,794 wrong | 94.4% correct | 6,389 unknown loss 0.9644 | geometry 0.2714 | confidence 1.3861 | mean error 0.696 vox speed: 1382 ms/update | data wait 1 ms/update | 17.4 crops/s | 17.4 decisions/s | 1.27 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1510 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 246, from the previous prediction 1264 confidence-labeled crops: terminal 19/1509 (1.3%) | recoverable 101/1509 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 480/1600 (30.0%), 1: 416/1600 (26.0%), 2: 192/1600 (12.0%), 3: 160/1600 (10.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 602 rows | mean depth 12 decisions | mean travel 100 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 480 52% 47% 24%>2% 23% 8 298 89%/8%/0%/4% 9% 187/0/0/37 0175A_5mm_v1 416 48% 45% 25%>7% 18% 27 981 85%/6%/1%/7% 6% 180/4/0/7 1447_5mm_v1 192 38% 51% 22%>11% 11% 20 1053 89%/5%/2%/4% 6% 97/1/0/0 paris4_afv_central 160 49% 34% 30%>17% 13% 26 1440 79%/6%/6%/9% 13% 55/1/0/0 PHerc0125_3mm_v2 352 61% 39% 30%>0% 30% 0 138 88%/6%/0%/7% 13% 83/0/67/56 gradients: max 61, clipped 1/50 updates Step 14,100 | last 50 updates / 1,600 crops | lr 1.64e-05 points (35/curve, tolerance 2 vox): 47,909 right / 2,798 wrong | 94.5% correct | 5,293 unknown loss 0.9343 | geometry 0.2604 | confidence 1.3478 | mean error 0.673 vox speed: 1422 ms/update | data wait 1 ms/update | 22.3 crops/s | 22.3 decisions/s | 1.24 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1536 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 286, from the previous prediction 1250 confidence-labeled crops: terminal 18/1533 (1.2%) | recoverable 108/1533 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 480/1600 (30.0%), 1: 160/1600 (10.0%), 2: 384/1600 (24.0%), 3: 128/1600 (8.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 602 rows | mean depth 13 decisions | mean travel 108 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 480 56% 44% 26%>0% 26% 0 139 89%/8%/0%/3% 7% 192/1/0/18 0175A_5mm_v1 160 55% 40% 29%>5% 24% 8 813 87%/8%/1%/4% 6% 64/0/0/0 1447_5mm_v1 384 42% 48% 24%>10% 14% 31 1091 87%/5%/2%/5% 6% 155/3/128/30 paris4_afv_central 128 45% 44% 23%>11% 12% 16 1415 83%/8%/5%/5% 14% 41/0/58/15 PHerc0125_3mm_v2 448 56% 44% 24%>0% 24% 0 135 89%/6%/0%/4% 10% 150/3/0/48 gradients: max 56, clipped 0/50 updates Step 14,150 | last 50 updates / 1,600 crops | lr 1.63e-05 points (35/curve, tolerance 2 vox): 46,861 right / 3,167 wrong | 93.7% correct | 5,972 unknown loss 0.9376 | geometry 0.2576 | confidence 1.3601 | mean error 0.669 vox speed: 1418 ms/update | data wait 1 ms/update | 22.4 crops/s | 22.4 decisions/s | 1.26 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1491 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 314, from the previous prediction 1177 confidence-labeled crops: terminal 26/1512 (1.7%) | recoverable 111/1512 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 288/1600 (18.0%), 2: 224/1600 (14.0%), 3: 288/1600 (18.0%), 4: 544/1600 (34.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 565 rows | mean depth 15 decisions | mean travel 119 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 49% 46% 25%>5% 20% 10 421 89%/7%/0%/3% 6% 119/0/0/0 0175A_5mm_v1 288 46% 44% 26%>10% 17% 23 1219 85%/6%/3%/5% 9% 128/0/15/0 1447_5mm_v1 224 48% 44% 28%>8% 19% 17 951 87%/8%/1%/3% 4% 98/0/0/0 paris4_afv_central 288 40% 45% 26%>15% 11% 45 1464 80%/9%/5%/7% 14% 104/2/101/27 PHerc0125_3mm_v2 544 53% 47% 24%>0% 24% 0 133 87%/6%/0%/8% 11% 116/3/0/142 gradients: max 52, clipped 0/50 updates Step 14,200 | last 50 updates / 1,600 crops | lr 1.63e-05 points (35/curve, tolerance 2 vox): 47,568 right / 2,953 wrong | 94.2% correct | 5,479 unknown loss 0.8126 | geometry 0.2267 | confidence 1.1718 | mean error 0.620 vox speed: 1418 ms/update | data wait 1 ms/update | 22.4 crops/s | 22.4 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1493 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 276, from the previous prediction 1217 confidence-labeled crops: terminal 30/1525 (2.0%) | recoverable 98/1525 (6.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 160/1600 (10.0%), 1: 288/1600 (18.0%), 2: 512/1600 (32.0%), 3: 256/1600 (16.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 619 rows | mean depth 11 decisions | mean travel 92 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 160 54% 46% 24%>0% 24% 0 162 94%/4%/0%/2% 7% 74/0/0/0 0175A_5mm_v1 288 43% 47% 25%>10% 15% 21 1219 85%/6%/4%/4% 10% 134/4/0/0 1447_5mm_v1 512 51% 42% 24%>7% 17% 32 1125 88%/6%/1%/4% 5% 151/4/0/63 paris4_afv_central 256 40% 45% 23%>16% 8% 37 1317 84%/7%/4%/5% 11% 101/2/22/13 PHerc0125_3mm_v2 384 54% 46% 22%>0% 22% 0 138 87%/6%/1%/7% 9% 159/3/0/17 gradients: max 65, clipped 1/50 updates Step 14,246 | dagger_states: 4161 | dagger_source_step: 14000 | dagger_caches: 3 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/paris4_afv_central/decisions_014000.npz" | dagger_fibers: 384 | dagger_supply: {"states": 4161, "traces": 384, "fibers": 384, "rows": {"terminal": 192, "premature_stop": 257, "recoverable": 31, "pre_excursion": 204, "ordinary": 3471}, "events": {"terminal": 31, "premature_stop": 257, "recoverable": 13, "pre_excursion": 39, "ordinary": 366}, "supervision": {"following": 3931, "recoverable": 26, "terminal": 192, "unknown": 12}, "departed_traces": 33, "terminal_traces": 31, "stop_reasons": {"after_failure": 19, "confidence": 271, "unannotated": 94}, "traced_voxels": 58061.743459399106, "seconds": 363.27879174798727, "voxels_per_second": 159.82695598612742} | dagger_coverage: {"requested": 384, "seeds": 384, "skipped": [], "epoch": 0} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 0 | dagger_launch_interval: null | dagger_publication_age: 246 | dataset: "paris4_afv_central" Step 14,250 | last 50 updates / 1,600 crops | lr 1.62e-05 points (35/curve, tolerance 2 vox): 47,115 right / 3,015 wrong | 94.0% correct | 5,870 unknown loss 0.8935 | geometry 0.2580 | confidence 1.2711 | mean error 0.639 vox speed: 1339 ms/update | data wait 1 ms/update | 23.7 crops/s | 23.7 decisions/s | 1.29 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1487 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 320, from the previous prediction 1167 confidence-labeled crops: terminal 30/1513 (2.0%) | recoverable 109/1513 (7.2%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 320/1600 (20.0%), 2: 352/1600 (22.0%), 3: 256/1600 (16.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 576 rows | mean depth 14 decisions | mean travel 114 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 61% 39% 26%>0% 26% 0 186 89%/8%/0%/3% 6% 99/0/10/0 0175A_5mm_v1 320 43% 48% 28%>8% 19% 21 1393 85%/6%/3%/6% 9% 104/9/47/50 1447_5mm_v1 352 42% 51% 24%>7% 18% 18 821 86%/7%/1%/6% 3% 148/3/0/32 paris4_afv_central 256 39% 43% 27%>18% 9% 57 2182 80%/7%/6%/7% 18% 63/1/0/47 PHerc0125_3mm_v2 416 56% 44% 26%>0% 26% 0 140 87%/6%/1%/6% 11% 162/5/0/20 gradients: max 59, clipped 0/50 updates Step 14,300 | last 50 updates / 1,600 crops | lr 1.61e-05 points (35/curve, tolerance 2 vox): 47,235 right / 3,130 wrong | 93.8% correct | 5,635 unknown loss 0.9246 | geometry 0.2528 | confidence 1.3436 | mean error 0.648 vox speed: 897 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.26 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1496 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 209, from the previous prediction 1287 confidence-labeled crops: terminal 33/1530 (2.2%) | recoverable 85/1530 (5.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 480/1600 (30.0%), 1: 256/1600 (16.0%), 2: 288/1600 (18.0%), 3: 224/1600 (14.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 675 rows | mean depth 13 decisions | mean travel 108 vox | 6 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 480 55% 45% 22%>0% 22% 0 150 93%/5%/0%/2% 6% 198/0/62/19 0175A_5mm_v1 256 48% 40% 25%>11% 13% 26 1493 89%/4%/3%/4% 11% 103/2/0/0 1447_5mm_v1 288 47% 45% 21%>9% 12% 21 978 84%/7%/1%/8% 6% 128/1/0/1 paris4_afv_central 224 36% 44% 28%>21% 7% 46 1646 79%/6%/7%/8% 17% 87/1/0/11 PHerc0125_3mm_v2 352 55% 45% 24%>0% 24% 0 148 91%/5%/2%/3% 12% 159/2/0/0 gradients: max 60, clipped 1/50 updates Step 14,350 | last 50 updates / 1,600 crops | lr 1.60e-05 points (35/curve, tolerance 2 vox): 47,624 right / 2,842 wrong | 94.4% correct | 5,534 unknown loss 0.7960 | geometry 0.2272 | confidence 1.1377 | mean error 0.593 vox speed: 894 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.26 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1495 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 193, from the previous prediction 1302 confidence-labeled crops: terminal 36/1535 (2.3%) | recoverable 98/1535 (6.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 160/1600 (10.0%), 1: 416/1600 (26.0%), 2: 416/1600 (26.0%), 3: 256/1600 (16.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 694 rows | mean depth 11 decisions | mean travel 90 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 160 64% 36% 26%>0% 26% 0 140 94%/5%/0%/1% 8% 57/0/0/1 0175A_5mm_v1 416 45% 48% 25%>6% 19% 23 907 88%/7%/3%/3% 9% 184/2/0/17 1447_5mm_v1 416 44% 48% 22%>8% 15% 31 937 87%/6%/2%/5% 6% 187/2/0/12 paris4_afv_central 256 39% 43% 26%>18% 8% 41 1361 83%/6%/5%/6% 13% 110/0/0/0 PHerc0125_3mm_v2 352 56% 44% 27%>0% 27% 0 138 89%/7%/0%/5% 8% 156/2/0/0 gradients: max 47, clipped 0/50 updates Step 14,400 | last 50 updates / 1,600 crops | lr 1.59e-05 points (35/curve, tolerance 2 vox): 46,776 right / 3,232 wrong | 93.5% correct | 5,992 unknown loss 0.9169 | geometry 0.2619 | confidence 1.3100 | mean error 0.664 vox speed: 896 ms/update | data wait 1 ms/update | 35.1 crops/s | 35.1 decisions/s | 1.24 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1470 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 257, from the previous prediction 1213 confidence-labeled crops: terminal 31/1521 (2.0%) | recoverable 116/1521 (7.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 288/1600 (18.0%), 2: 288/1600 (18.0%), 3: 352/1600 (22.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 1370 chunk references / 1 windows live chains: 638 rows | mean depth 13 decisions | mean travel 105 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 52% 46% 24%>2% 22% 6 297 89%/7%/0%/4% 8% 155/3/0/23 0175A_5mm_v1 288 46% 46% 24%>8% 16% 23 1150 86%/9%/1%/3% 6% 124/3/0/8 1447_5mm_v1 288 47% 43% 27%>10% 17% 27 1164 88%/5%/3%/5% 5% 124/2/0/0 paris4_afv_central 352 39% 42% 30%>19% 12% 63 1833 77%/9%/5%/9% 17% 102/3/0/46 PHerc0125_3mm_v2 288 49% 51% 20%>0% 20% 0 144 89%/6%/0%/5% 11% 133/3/49/13 gradients: max 68, clipped 1/50 updates Step 14,450 | last 50 updates / 1,600 crops | lr 1.59e-05 points (35/curve, tolerance 2 vox): 46,156 right / 3,299 wrong | 93.3% correct | 6,545 unknown loss 0.8432 | geometry 0.2511 | confidence 1.1842 | mean error 0.639 vox speed: 898 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.34 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1437 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 314, from the previous prediction 1123 confidence-labeled crops: terminal 40/1503 (2.7%) | recoverable 117/1503 (7.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 352/1600 (22.0%), 2: 256/1600 (16.0%), 3: 480/1600 (30.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 555 rows | mean depth 13 decisions | mean travel 111 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 60% 40% 26%>0% 26% 0 144 88%/8%/0%/4% 4% 76/1/0/0 0175A_5mm_v1 352 42% 48% 26%>10% 16% 28 1787 82%/8%/3%/6% 9% 101/4/0/68 1447_5mm_v1 256 50% 39% 27%>11% 16% 25 1308 86%/6%/4%/5% 7% 100/1/0/0 paris4_afv_central 480 38% 45% 25%>16% 9% 75 2022 80%/8%/3%/8% 14% 126/2/0/92 PHerc0125_3mm_v2 320 52% 48% 23%>0% 23% 0 163 88%/6%/1%/5% 9% 152/2/0/0 gradients: max 67, clipped 1/50 updates Step 14,500 | last 50 updates / 1,600 crops | lr 1.58e-05 points (35/curve, tolerance 2 vox): 47,493 right / 3,085 wrong | 93.9% correct | 5,422 unknown loss 0.9599 | geometry 0.2816 | confidence 1.3567 | mean error 0.687 vox speed: 891 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.27 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1494 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 205, from the previous prediction 1289 confidence-labeled crops: terminal 21/1537 (1.4%) | recoverable 108/1537 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 352/1600 (22.0%), 2: 352/1600 (22.0%), 3: 256/1600 (16.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 671 rows | mean depth 12 decisions | mean travel 101 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 49% 47% 23%>4% 20% 8 399 88%/8%/0%/3% 12% 106/0/0/0 0175A_5mm_v1 352 51% 42% 25%>7% 18% 20 1053 88%/6%/1%/4% 7% 149/5/0/0 1447_5mm_v1 352 47% 45% 23%>8% 14% 28 974 89%/6%/1%/4% 5% 159/5/0/0 paris4_afv_central 256 40% 45% 25%>15% 9% 36 1718 83%/8%/5%/5% 14% 77/0/0/37 PHerc0125_3mm_v2 416 57% 43% 26%>0% 26% 0 152 90%/6%/0%/4% 12% 180/3/0/0 gradients: max 58, clipped 0/50 updates Step 14,550 | last 50 updates / 1,600 crops | lr 1.57e-05 points (35/curve, tolerance 2 vox): 46,760 right / 3,444 wrong | 93.1% correct | 5,796 unknown loss 0.9600 | geometry 0.2823 | confidence 1.3555 | mean error 0.693 vox speed: 894 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.31 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1471 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 184, from the previous prediction 1287 confidence-labeled crops: terminal 27/1529 (1.8%) | recoverable 109/1529 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 224/1600 (14.0%), 2: 256/1600 (16.0%), 3: 448/1600 (28.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 690 rows | mean depth 14 decisions | mean travel 114 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 56% 44% 25%>1% 25% 2 193 89%/7%/0%/4% 6% 140/1/0/0 0175A_5mm_v1 224 47% 46% 24%>7% 17% 15 964 88%/7%/2%/4% 8% 103/0/0/0 1447_5mm_v1 256 41% 51% 23%>9% 14% 22 937 89%/5%/1%/5% 5% 130/2/0/0 paris4_afv_central 448 36% 48% 23%>17% 6% 75 1706 81%/9%/4%/6% 16% 162/2/0/51 PHerc0125_3mm_v2 352 56% 44% 23%>0% 23% 0 144 90%/5%/0%/4% 12% 155/2/0/0 gradients: max 49, clipped 0/50 updates Step 14,600 | last 50 updates / 1,600 crops | lr 1.56e-05 points (35/curve, tolerance 2 vox): 46,740 right / 3,400 wrong | 93.2% correct | 5,860 unknown loss 0.9389 | geometry 0.2745 | confidence 1.3288 | mean error 0.676 vox speed: 895 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.31 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1484 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 273, from the previous prediction 1211 confidence-labeled crops: terminal 30/1516 (2.0%) | recoverable 104/1516 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 256/1600 (16.0%), 2: 160/1600 (10.0%), 3: 384/1600 (24.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 1014 chunk references / 1 windows live chains: 631 rows | mean depth 12 decisions | mean travel 99 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 56% 44% 25%>0% 25% 0 150 90%/7%/0%/4% 8% 116/0/20/54 0175A_5mm_v1 256 46% 45% 25%>8% 17% 18 1136 85%/7%/0%/7% 6% 116/4/0/0 1447_5mm_v1 160 52% 37% 27%>11% 16% 16 1424 84%/7%/4%/5% 8% 55/2/41/4 paris4_afv_central 384 35% 45% 27%>20% 7% 62 1535 82%/7%/5%/6% 15% 171/2/0/0 PHerc0125_3mm_v2 416 50% 50% 25%>0% 25% 0 134 88%/5%/1%/6% 13% 173/2/11/37 gradients: max 42, clipped 0/50 updates Step 14,650 | last 50 updates / 1,600 crops | lr 1.56e-05 points (35/curve, tolerance 2 vox): 47,421 right / 3,312 wrong | 93.5% correct | 5,267 unknown loss 0.8755 | geometry 0.2469 | confidence 1.2570 | mean error 0.637 vox speed: 893 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.26 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1473 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 222, from the previous prediction 1251 confidence-labeled crops: terminal 35/1522 (2.3%) | recoverable 99/1522 (6.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 320/1600 (20.0%), 2: 544/1600 (34.0%), 3: 192/1600 (12.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 674 rows | mean depth 13 decisions | mean travel 103 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 50% 47% 27%>3% 24% 5 341 90%/7%/0%/3% 11% 116/1/0/5 0175A_5mm_v1 320 48% 44% 24%>8% 16% 22 1143 86%/7%/2%/5% 7% 142/5/0/0 1447_5mm_v1 544 45% 46% 24%>9% 15% 44 1196 87%/6%/2%/4% 6% 225/6/0/25 paris4_afv_central 192 39% 40% 31%>21% 10% 38 1812 79%/6%/7%/8% 18% 69/2/0/8 PHerc0125_3mm_v2 288 58% 42% 28%>0% 28% 0 137 89%/5%/1%/6% 11% 122/1/0/0 gradients: max 59, clipped 0/50 updates Step 14,700 | last 50 updates / 1,600 crops | lr 1.55e-05 points (35/curve, tolerance 2 vox): 46,201 right / 3,159 wrong | 93.6% correct | 6,640 unknown loss 0.8942 | geometry 0.2595 | confidence 1.2693 | mean error 0.642 vox speed: 889 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.26 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1445 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 276, from the previous prediction 1169 confidence-labeled crops: terminal 35/1515 (2.3%) | recoverable 111/1515 (7.3%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 288/1600 (18.0%), 2: 416/1600 (26.0%), 3: 352/1600 (22.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 593 rows | mean depth 15 decisions | mean travel 121 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 52% 39% 26%>10% 16% 10 875 88%/6%/1%/6% 10% 74/1/0/0 0175A_5mm_v1 288 49% 40% 29%>11% 18% 31 2140 84%/7%/4%/5% 9% 84/0/84/30 1447_5mm_v1 416 47% 45% 26%>7% 19% 30 1420 87%/7%/0%/5% 5% 119/1/0/70 paris4_afv_central 352 36% 46% 24%>18% 7% 53 1424 83%/6%/6%/5% 14% 162/3/0/0 PHerc0125_3mm_v2 352 56% 44% 30%>0% 30% 0 143 86%/8%/0%/6% 10% 154/6/0/0 gradients: max 49, clipped 0/50 updates Step 14,750 | last 50 updates / 1,600 crops | lr 1.54e-05 points (35/curve, tolerance 2 vox): 48,156 right / 2,709 wrong | 94.7% correct | 5,135 unknown loss 0.7433 | geometry 0.2298 | confidence 1.0271 | mean error 0.598 vox speed: 890 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1493 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 262, from the previous prediction 1231 confidence-labeled crops: terminal 35/1533 (2.3%) | recoverable 115/1533 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 320/1600 (20.0%), 2: 384/1600 (24.0%), 3: 320/1600 (20.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 76 chunks / 152.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 619 rows | mean depth 14 decisions | mean travel 111 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 56% 44% 25%>0% 25% 0 175 87%/10%/0%/3% 7% 99/0/0/0 0175A_5mm_v1 320 52% 43% 27%>6% 21% 15 914 83%/8%/3%/5% 7% 137/3/0/0 1447_5mm_v1 384 47% 44% 21%>9% 12% 29 1664 89%/5%/2%/4% 7% 98/2/0/72 paris4_afv_central 320 43% 41% 27%>16% 10% 46 1449 84%/7%/6%/3% 13% 131/2/0/0 PHerc0125_3mm_v2 352 56% 44% 24%>0% 24% 0 142 88%/7%/0%/5% 6% 154/8/0/0 gradients: max 49, clipped 0/50 updates Step 14,800 | last 50 updates / 1,600 crops | lr 1.53e-05 points (35/curve, tolerance 2 vox): 47,372 right / 3,028 wrong | 94.0% correct | 5,600 unknown loss 0.8905 | geometry 0.2605 | confidence 1.2600 | mean error 0.661 vox speed: 897 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.26 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1484 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 223, from the previous prediction 1261 confidence-labeled crops: terminal 26/1519 (1.7%) | recoverable 92/1519 (6.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 416/1600 (26.0%), 2: 256/1600 (16.0%), 3: 320/1600 (20.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 77 chunks / 154.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 687 rows | mean depth 13 decisions | mean travel 108 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 48% 44% 30%>8% 22% 15 728 88%/7%/0%/5% 12% 126/0/0/0 0175A_5mm_v1 416 47% 47% 22%>6% 16% 22 946 90%/5%/1%/4% 6% 180/0/0/15 1447_5mm_v1 256 42% 50% 20%>8% 12% 18 908 88%/4%/3%/5% 6% 129/4/0/0 paris4_afv_central 320 39% 48% 23%>13% 10% 43 1497 81%/8%/4%/7% 14% 116/3/41/38 PHerc0125_3mm_v2 320 52% 48% 21%>0% 21% 0 143 90%/6%/0%/5% 10% 136/3/26/17 gradients: max 44, clipped 0/50 updates Step 14,850 | last 50 updates / 1,600 crops | lr 1.52e-05 points (35/curve, tolerance 2 vox): 46,837 right / 2,894 wrong | 94.2% correct | 6,269 unknown loss 0.8940 | geometry 0.2538 | confidence 1.2805 | mean error 0.642 vox speed: 894 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.25 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1492 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 301, from the previous prediction 1191 confidence-labeled crops: terminal 28/1517 (1.8%) | recoverable 113/1517 (7.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 384/1600 (24.0%), 2: 256/1600 (16.0%), 3: 320/1600 (20.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 77 chunks / 154.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 579 rows | mean depth 15 decisions | mean travel 127 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 52% 46% 28%>2% 26% 9 410 89%/8%/0%/3% 10% 130/0/53/33 0175A_5mm_v1 384 48% 46% 23%>5% 18% 27 1917 88%/5%/1%/5% 7% 77/2/0/100 1447_5mm_v1 256 52% 43% 25%>5% 20% 12 769 85%/9%/2%/4% 6% 109/1/0/0 paris4_afv_central 320 38% 44% 26%>18% 8% 48 1496 78%/8%/6%/8% 14% 140/5/0/0 PHerc0125_3mm_v2 288 57% 43% 23%>0% 23% 0 137 89%/6%/0%/6% 9% 123/7/0/2 gradients: max 55, clipped 0/50 updates Step 14,900 | last 50 updates / 1,600 crops | lr 1.52e-05 points (35/curve, tolerance 2 vox): 47,529 right / 2,835 wrong | 94.4% correct | 5,636 unknown loss 0.8647 | geometry 0.2522 | confidence 1.2249 | mean error 0.650 vox speed: 890 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1494 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 225, from the previous prediction 1269 confidence-labeled crops: terminal 27/1514 (1.8%) | recoverable 98/1514 (6.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 256/1600 (16.0%), 2: 480/1600 (30.0%), 3: 256/1600 (16.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 77 chunks / 154.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 654 rows | mean depth 14 decisions | mean travel 111 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 53% 44% 28%>3% 25% 6 418 90%/4%/0%/6% 8% 112/1/0/15 0175A_5mm_v1 256 52% 42% 29%>7% 22% 15 1173 85%/7%/1%/6% 6% 94/4/0/13 1447_5mm_v1 480 48% 43% 26%>9% 17% 34 1336 85%/7%/2%/6% 7% 180/1/76/28 paris4_afv_central 256 42% 45% 21%>13% 9% 31 1205 86%/4%/5%/5% 11% 115/1/0/0 PHerc0125_3mm_v2 320 52% 48% 24%>0% 24% 0 142 88%/7%/0%/4% 11% 153/6/0/0 gradients: max 80, clipped 2/50 updates Step 14,950 | last 50 updates / 1,600 crops | lr 1.51e-05 points (35/curve, tolerance 2 vox): 47,239 right / 3,413 wrong | 93.3% correct | 5,348 unknown loss 1.0229 | geometry 0.2889 | confidence 1.4679 | mean error 0.694 vox speed: 890 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.27 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1506 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 242, from the previous prediction 1264 confidence-labeled crops: terminal 23/1532 (1.5%) | recoverable 110/1532 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 384/1600 (24.0%), 2: 288/1600 (18.0%), 3: 224/1600 (14.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 77 chunks / 154.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 645 rows | mean depth 14 decisions | mean travel 115 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 54% 43% 27%>3% 24% 14 833 88%/8%/0%/3% 13% 106/0/0/74 0175A_5mm_v1 384 47% 48% 25%>6% 19% 19 874 88%/5%/2%/5% 8% 183/4/0/0 1447_5mm_v1 288 49% 49% 25%>3% 22% 8 481 88%/9%/0%/3% 4% 126/0/11/14 paris4_afv_central 224 37% 46% 25%>18% 7% 35 1510 83%/5%/6%/5% 16% 102/1/0/0 PHerc0125_3mm_v2 288 56% 44% 22%>0% 22% 0 142 88%/6%/1%/5% 13% 128/5/0/0 gradients: max 65, clipped 1/50 updates Step 15,000 | last 50 updates / 1,600 crops | lr 1.50e-05 points (35/curve, tolerance 2 vox): 47,044 right / 3,369 wrong | 93.3% correct | 5,587 unknown loss 0.9719 | geometry 0.2893 | confidence 1.3653 | mean error 0.686 vox speed: 892 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.26 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1480 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 278, from the previous prediction 1202 confidence-labeled crops: terminal 34/1531 (2.2%) | recoverable 105/1531 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 480/1600 (30.0%), 1: 256/1600 (16.0%), 2: 192/1600 (12.0%), 3: 384/1600 (24.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 77 chunks / 154.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 602 rows | mean depth 15 decisions | mean travel 123 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 480 51% 46% 28%>2% 26% 13 567 90%/7%/0%/2% 10% 135/2/0/86 0175A_5mm_v1 256 52% 43% 27%>6% 21% 15 971 89%/7%/2%/3% 6% 109/3/0/0 1447_5mm_v1 192 47% 42% 31%>11% 20% 20 1864 84%/5%/4%/7% 7% 62/3/0/18 paris4_afv_central 384 40% 44% 22%>16% 6% 56 1486 82%/6%/5%/7% 14% 162/2/35/8 PHerc0125_3mm_v2 288 53% 47% 27%>0% 27% 0 138 88%/6%/1%/5% 10% 134/2/14/0 gradients: max 59, clipped 0/50 updates Step 15,000 | dagger_launched: true Step 15,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 256, "stopped": 0, "mean_example_error": 0.6709755207261732, "geometry_examples": 31, "examples": 32, "inference_and_measurement_seconds": 1.8562705367803574, "total_diagnostic_seconds": 3.6554433964192867} Step 15,000 | monitor rollout @ 0.50 coverage 58.6% | precision 81.1% | diverged 21.9% Step 15,000 | monitor rollout @ 0.50 coverage 85.2% | precision 100.0% | diverged 0.0% Step 15,000 | monitor rollout @ 0.50 coverage 88.4% | precision 100.0% | diverged 0.0% Step 15,000 | monitor rollout @ 0.50 coverage 79.3% | precision 96.7% | diverged 3.1% Step 15,000 | monitor rollout @ 0.50 coverage 85.0% | precision 78.4% | diverged 18.8% Step 15,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 6/8 (75.0%) | false stops 0 1.5-3: recovered 6/8 (75.0%) | false stops 0 3-4.5: recovered 6/8 (75.0%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 0 entire proposed 8-point prefix correct (distance): 22/28 (78.6%) gate @ 0.50: false stops 0/23 (0.0%) | accepted wrong 5/27 (18.5%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.565 1.531 following 16 464 1.094 1.152 recoverable 12 348 2.194 2.037 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 1.094 1.152 332 32 812 1.565 1.531 improved/worsened/compared states: 3/1/28 Step 15,000 | batch_diagnostics_seconds: 3.6576227955520153 | diagnostics_seconds: 545.4110951535404 | recovery_seconds: 14.2286142334342 Step 15,001 | dagger_states: 3969 | dagger_source_step: 15000 | dagger_caches: 2 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/PHerc0125_3mm_v2/decisions_015000.npz" | dagger_fibers: 384 | dagger_supply: {"states": 3969, "traces": 384, "fibers": 384, "rows": {"terminal": 494, "premature_stop": 133, "recoverable": 77, "pre_excursion": 454, "ordinary": 2784}, "events": {"terminal": 75, "premature_stop": 133, "recoverable": 26, "pre_excursion": 90, "ordinary": 345}, "supervision": {"following": 3368, "recoverable": 52, "terminal": 494, "unknown": 55}, "departed_traces": 80, "terminal_traces": 75, "stop_reasons": {"after_failure": 58, "confidence": 150, "unannotated": 176}, "traced_voxels": 52315.38769277738, "seconds": 265.40903947502375, "voxels_per_second": 197.11230558030977} | dagger_coverage: {"requested": 384, "seeds": 384, "skipped": [], "epoch": 0} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 0 | dagger_launch_interval: null | dagger_publication_age: 1 | dataset: "PHerc0125_3mm_v2" Step 15,050 | last 50 updates / 1,600 crops | lr 1.49e-05 points (35/curve, tolerance 2 vox): 47,772 right / 2,937 wrong | 94.2% correct | 5,291 unknown loss 0.9749 | geometry 0.2833 | confidence 1.3832 | mean error 0.684 vox speed: 888 ms/update | data wait 1 ms/update | 2.6 crops/s | 2.6 decisions/s | 1.24 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1513 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 232, from the previous prediction 1281 confidence-labeled crops: terminal 18/1535 (1.2%) | recoverable 103/1535 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 416/1600 (26.0%), 2: 256/1600 (16.0%), 3: 256/1600 (16.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 77 chunks / 154.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 667 rows | mean depth 15 decisions | mean travel 124 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 48% 47% 23%>4% 18% 13 581 90%/6%/0%/4% 10% 138/0/0/43 0175A_5mm_v1 416 48% 45% 24%>7% 17% 28 1130 88%/6%/1%/4% 7% 179/5/62/7 1447_5mm_v1 256 51% 46% 21%>2% 18% 5 450 91%/7%/0%/2% 2% 119/2/0/0 paris4_afv_central 256 43% 44% 22%>13% 9% 30 1249 84%/5%/4%/6% 13% 113/2/0/0 PHerc0125_3mm_v2 288 51% 49% 19%>0% 19% 0 140 87%/8%/0%/5% 11% 118/3/0/22 gradients: max 56, clipped 0/50 updates Step 15,100 | last 50 updates / 1,600 crops | lr 1.48e-05 points (35/curve, tolerance 2 vox): 47,731 right / 2,373 wrong | 95.3% correct | 5,896 unknown loss 0.7228 | geometry 0.2139 | confidence 1.0179 | mean error 0.597 vox speed: 890 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.25 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1506 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 232, from the previous prediction 1274 confidence-labeled crops: terminal 23/1517 (1.5%) | recoverable 94/1517 (6.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 352/1600 (22.0%), 2: 384/1600 (24.0%), 3: 256/1600 (16.0%), 4: 160/1600 (10.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 77 chunks / 154.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 635 rows | mean depth 13 decisions | mean travel 106 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 54% 43% 24%>3% 20% 13 444 90%/5%/0%/5% 9% 179/3/0/14 0175A_5mm_v1 352 50% 44% 28%>6% 22% 17 1031 89%/5%/1%/5% 5% 135/3/12/20 1447_5mm_v1 384 51% 43% 25%>6% 18% 22 912 87%/7%/2%/4% 3% 165/3/0/0 paris4_afv_central 256 38% 49% 22%>13% 9% 30 1573 79%/6%/5%/9% 11% 85/3/0/40 PHerc0125_3mm_v2 160 56% 44% 28%>0% 28% 0 157 88%/8%/0%/5% 12% 71/2/0/0 gradients: max 50, clipped 0/50 updates Step 15,150 | last 50 updates / 1,600 crops | lr 1.48e-05 points (35/curve, tolerance 2 vox): 47,875 right / 3,054 wrong | 94.0% correct | 5,071 unknown loss 0.9265 | geometry 0.2562 | confidence 1.3406 | mean error 0.637 vox speed: 894 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1523 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 237, from the previous prediction 1286 confidence-labeled crops: terminal 32/1544 (2.1%) | recoverable 120/1544 (7.8%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 352/1600 (22.0%), 2: 224/1600 (14.0%), 3: 320/1600 (20.0%), 4: 480/1600 (30.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 80 chunks / 160.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 681 rows | mean depth 16 decisions | mean travel 129 vox | 9 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 52% 48% 24%>0% 24% 0 147 92%/6%/0%/2% 6% 108/1/0/0 0175A_5mm_v1 352 47% 49% 21%>5% 17% 15 917 87%/6%/2%/5% 6% 127/0/0/45 1447_5mm_v1 224 44% 50% 24%>6% 18% 12 850 87%/10%/2%/0% 7% 97/2/24/16 paris4_afv_central 320 39% 46% 24%>15% 9% 42 1375 83%/8%/6%/3% 15% 146/4/0/0 PHerc0125_3mm_v2 480 56% 44% 27%>0% 27% 0 148 88%/7%/0%/5% 11% 203/6/0/8 gradients: max 59, clipped 0/50 updates Step 15,200 | last 50 updates / 1,600 crops | lr 1.47e-05 points (35/curve, tolerance 2 vox): 47,185 right / 2,834 wrong | 94.3% correct | 5,981 unknown loss 0.8855 | geometry 0.2518 | confidence 1.2675 | mean error 0.637 vox speed: 895 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.27 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1513 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 362, from the previous prediction 1151 confidence-labeled crops: terminal 25/1506 (1.7%) | recoverable 104/1506 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 64/1600 (4.0%), 2: 640/1600 (40.0%), 3: 224/1600 (14.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 80 chunks / 160.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 499 rows | mean depth 14 decisions | mean travel 117 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 52% 44% 27%>4% 23% 8 466 91%/7%/0%/2% 11% 127/1/0/0 0175A_5mm_v1 64 52% 41% 27%>8% 19% 5 1283 84%/9%/2%/5% 10% 26/0/0/0 1447_5mm_v1 640 50% 44% 27%>6% 21% 39 1695 84%/7%/2%/6% 5% 138/2/0/143 paris4_afv_central 224 45% 45% 24%>10% 13% 23 1077 83%/5%/4%/8% 9% 101/2/0/0 PHerc0125_3mm_v2 384 56% 44% 18%>0% 18% 0 144 87%/5%/0%/8% 11% 107/1/0/61 gradients: max 71, clipped 1/50 updates Step 15,250 | last 50 updates / 1,600 crops | lr 1.46e-05 points (35/curve, tolerance 2 vox): 47,335 right / 2,847 wrong | 94.3% correct | 5,818 unknown loss 0.8540 | geometry 0.2470 | confidence 1.2139 | mean error 0.621 vox speed: 889 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1497 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 231, from the previous prediction 1266 confidence-labeled crops: terminal 30/1525 (2.0%) | recoverable 114/1525 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 448/1600 (28.0%), 2: 384/1600 (24.0%), 3: 256/1600 (16.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 80 chunks / 160.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 676 rows | mean depth 16 decisions | mean travel 129 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 48% 48% 21%>4% 17% 8 489 88%/6%/1%/5% 12% 123/0/0/0 0175A_5mm_v1 448 48% 47% 25%>5% 20% 22 893 88%/7%/2%/4% 6% 209/5/0/0 1447_5mm_v1 384 47% 45% 30%>8% 22% 26 1216 86%/9%/1%/4% 6% 139/4/0/34 paris4_afv_central 256 35% 50% 27%>15% 12% 31 1301 81%/7%/6%/6% 12% 129/2/6/0 PHerc0125_3mm_v2 256 57% 43% 25%>0% 25% 0 135 87%/6%/0%/7% 13% 76/2/0/34 gradients: max 52, clipped 0/50 updates Step 15,300 | last 50 updates / 1,600 crops | lr 1.45e-05 points (35/curve, tolerance 2 vox): 48,148 right / 2,842 wrong | 94.4% correct | 5,010 unknown loss 0.8855 | geometry 0.2642 | confidence 1.2427 | mean error 0.661 vox speed: 892 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1509 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 222, from the previous prediction 1287 confidence-labeled crops: terminal 21/1527 (1.4%) | recoverable 103/1527 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 352/1600 (22.0%), 2: 288/1600 (18.0%), 3: 288/1600 (18.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 80 chunks / 160.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 682 rows | mean depth 14 decisions | mean travel 117 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 48% 49% 23%>3% 20% 6 381 92%/5%/0%/3% 7% 156/2/0/0 0175A_5mm_v1 352 48% 46% 22%>6% 16% 17 999 89%/6%/2%/3% 8% 162/4/3/0 1447_5mm_v1 288 45% 51% 20%>4% 16% 11 832 88%/6%/1%/6% 4% 94/3/0/52 paris4_afv_central 288 41% 42% 31%>17% 14% 44 1599 80%/11%/4%/6% 10% 122/4/0/0 PHerc0125_3mm_v2 352 58% 42% 27%>0% 27% 0 147 88%/6%/0%/6% 11% 148/7/0/0 gradients: max 65, clipped 1/50 updates Step 15,350 | last 50 updates / 1,600 crops | lr 1.45e-05 points (35/curve, tolerance 2 vox): 46,599 right / 3,562 wrong | 92.9% correct | 5,839 unknown loss 0.8989 | geometry 0.2684 | confidence 1.2610 | mean error 0.659 vox speed: 893 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.30 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1475 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 279, from the previous prediction 1196 confidence-labeled crops: terminal 40/1518 (2.6%) | recoverable 108/1518 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 256/1600 (16.0%), 2: 256/1600 (16.0%), 3: 448/1600 (28.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 80 chunks / 160.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 562 rows | mean depth 14 decisions | mean travel 112 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 61% 39% 26%>0% 26% 0 159 90%/7%/0%/3% 7% 74/0/0/0 0175A_5mm_v1 256 54% 42% 25%>4% 21% 10 1639 86%/4%/1%/8% 6% 38/1/0/70 1447_5mm_v1 256 47% 46% 21%>7% 14% 14 917 89%/6%/3%/2% 6% 119/1/0/0 paris4_afv_central 448 38% 43% 27%>19% 8% 78 2217 81%/7%/6%/7% 16% 147/1/11/46 PHerc0125_3mm_v2 448 59% 41% 28%>0% 28% 0 149 86%/8%/1%/5% 12% 184/6/0/0 gradients: max 41, clipped 0/50 updates Step 15,400 | last 50 updates / 1,600 crops | lr 1.44e-05 points (35/curve, tolerance 2 vox): 46,978 right / 3,059 wrong | 93.9% correct | 5,963 unknown loss 0.9466 | geometry 0.2662 | confidence 1.3608 | mean error 0.666 vox speed: 896 ms/update | data wait 1 ms/update | 35.1 crops/s | 35.1 decisions/s | 1.29 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1492 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 283, from the previous prediction 1209 confidence-labeled crops: terminal 22/1508 (1.5%) | recoverable 114/1508 (7.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 192/1600 (12.0%), 2: 512/1600 (32.0%), 3: 256/1600 (16.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 80 chunks / 160.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 589 rows | mean depth 16 decisions | mean travel 133 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 49% 46% 26%>5% 21% 14 595 89%/7%/1%/4% 11% 161/0/0/0 0175A_5mm_v1 192 45% 48% 30%>7% 23% 12 1044 83%/9%/1%/7% 5% 93/4/0/0 1447_5mm_v1 512 47% 46% 25%>7% 19% 31 1178 88%/6%/2%/4% 5% 182/3/100/55 paris4_afv_central 256 43% 44% 23%>12% 11% 40 2238 82%/9%/3%/6% 13% 61/2/0/52 PHerc0125_3mm_v2 288 57% 43% 27%>0% 27% 0 140 82%/7%/0%/11% 16% 92/0/0/31 gradients: max 54, clipped 0/50 updates Step 15,450 | last 50 updates / 1,600 crops | lr 1.43e-05 points (35/curve, tolerance 2 vox): 46,936 right / 3,422 wrong | 93.2% correct | 5,642 unknown loss 1.0099 | geometry 0.3050 | confidence 1.4098 | mean error 0.698 vox speed: 896 ms/update | data wait 1 ms/update | 35.0 crops/s | 35.0 decisions/s | 1.26 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1498 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 192, from the previous prediction 1306 confidence-labeled crops: terminal 27/1525 (1.8%) | recoverable 125/1525 (8.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 288/1600 (18.0%), 2: 192/1600 (12.0%), 3: 288/1600 (18.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 80 chunks / 160.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 668 rows | mean depth 16 decisions | mean travel 130 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 50% 44% 27%>6% 21% 18 811 88%/8%/0%/4% 11% 158/0/105/23 0175A_5mm_v1 288 51% 42% 28%>7% 22% 17 1153 85%/9%/1%/5% 6% 122/1/10/0 1447_5mm_v1 192 54% 44% 27%>3% 24% 5 500 88%/8%/0%/4% 5% 84/0/0/0 paris4_afv_central 288 42% 41% 25%>17% 8% 43 1692 81%/6%/7%/6% 14% 119/1/0/0 PHerc0125_3mm_v2 416 56% 44% 27%>0% 27% 0 140 87%/7%/1%/5% 11% 185/5/0/0 gradients: max 40, clipped 0/50 updates Step 15,500 | last 50 updates / 1,600 crops | lr 1.42e-05 points (35/curve, tolerance 2 vox): 47,712 right / 2,853 wrong | 94.4% correct | 5,435 unknown loss 0.8913 | geometry 0.2634 | confidence 1.2557 | mean error 0.654 vox speed: 891 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1503 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 247, from the previous prediction 1256 confidence-labeled crops: terminal 25/1527 (1.6%) | recoverable 105/1527 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 320/1600 (20.0%), 2: 352/1600 (22.0%), 3: 320/1600 (20.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 80 chunks / 160.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 639 rows | mean depth 13 decisions | mean travel 107 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 52% 44% 26%>4% 22% 9 771 89%/7%/0%/4% 10% 76/0/23/65 0175A_5mm_v1 320 46% 48% 26%>6% 20% 18 971 86%/6%/1%/7% 5% 153/2/0/0 1447_5mm_v1 352 51% 43% 27%>7% 20% 22 1017 88%/7%/1%/3% 4% 150/0/0/0 paris4_afv_central 320 40% 47% 23%>13% 10% 37 1296 82%/6%/6%/6% 12% 150/3/0/0 PHerc0125_3mm_v2 288 53% 47% 20%>0% 20% 0 138 90%/7%/0%/3% 11% 110/3/9/24 gradients: max 55, clipped 0/50 updates Step 15,550 | last 50 updates / 1,600 crops | lr 1.41e-05 points (35/curve, tolerance 2 vox): 48,680 right / 2,568 wrong | 95.0% correct | 4,752 unknown loss 0.8740 | geometry 0.2499 | confidence 1.2482 | mean error 0.647 vox speed: 892 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1535 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 285, from the previous prediction 1250 confidence-labeled crops: terminal 16/1544 (1.0%) | recoverable 113/1544 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 288/1600 (18.0%), 2: 288/1600 (18.0%), 3: 192/1600 (12.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 80 chunks / 160.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 603 rows | mean depth 17 decisions | mean travel 136 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 57% 41% 24%>2% 21% 7 470 90%/7%/0%/3% 9% 109/1/0/49 0175A_5mm_v1 288 44% 48% 22%>8% 14% 22 1889 87%/7%/3%/4% 7% 77/0/20/61 1447_5mm_v1 288 49% 48% 23%>3% 19% 9 588 91%/7%/0%/1% 1% 137/2/0/0 paris4_afv_central 192 38% 51% 20%>11% 8% 20 1149 83%/8%/4%/5% 11% 97/3/22/1 PHerc0125_3mm_v2 448 59% 41% 28%>0% 28% 0 145 88%/7%/0%/4% 12% 183/4/0/0 gradients: max 53, clipped 0/50 updates Step 15,600 | last 50 updates / 1,600 crops | lr 1.41e-05 points (35/curve, tolerance 2 vox): 48,022 right / 2,810 wrong | 94.5% correct | 5,168 unknown loss 0.8216 | geometry 0.2355 | confidence 1.1722 | mean error 0.627 vox speed: 891 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1512 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 262, from the previous prediction 1250 confidence-labeled crops: terminal 24/1534 (1.6%) | recoverable 116/1534 (7.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 352/1600 (22.0%), 2: 384/1600 (24.0%), 3: 288/1600 (18.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 80 chunks / 160.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 644 rows | mean depth 16 decisions | mean travel 132 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 51% 44% 24%>5% 20% 9 589 87%/8%/0%/4% 8% 113/0/0/0 0175A_5mm_v1 352 48% 44% 25%>8% 17% 23 1451 86%/6%/3%/5% 8% 126/5/0/29 1447_5mm_v1 384 55% 41% 28%>4% 24% 11 711 88%/10%/0%/2% 2% 156/1/59/2 paris4_afv_central 288 39% 50% 22%>11% 11% 31 1582 84%/7%/3%/6% 11% 94/2/34/50 PHerc0125_3mm_v2 320 49% 51% 23%>0% 23% 0 137 89%/6%/1%/5% 13% 155/4/1/8 gradients: max 52, clipped 0/50 updates Step 15,650 | last 50 updates / 1,600 crops | lr 1.40e-05 points (35/curve, tolerance 2 vox): 48,506 right / 2,133 wrong | 95.8% correct | 5,361 unknown loss 0.8058 | geometry 0.2231 | confidence 1.1655 | mean error 0.598 vox speed: 894 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1518 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 230, from the previous prediction 1288 confidence-labeled crops: terminal 17/1539 (1.1%) | recoverable 96/1539 (6.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 288/1600 (18.0%), 2: 320/1600 (20.0%), 3: 384/1600 (24.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 80 chunks / 160.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 682 rows | mean depth 12 decisions | mean travel 101 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 56% 44% 24%>0% 24% 0 164 92%/6%/0%/2% 7% 141/0/0/0 0175A_5mm_v1 288 51% 46% 24%>3% 22% 7 583 89%/8%/1%/2% 5% 126/6/0/6 1447_5mm_v1 320 47% 49% 21%>4% 17% 13 751 91%/6%/0%/3% 3% 134/1/4/22 paris4_afv_central 384 38% 46% 24%>15% 9% 54 1535 85%/6%/3%/6% 11% 176/2/0/2 PHerc0125_3mm_v2 288 55% 45% 22%>0% 22% 0 135 89%/5%/0%/6% 9% 105/1/0/26 gradients: max 44, clipped 0/50 updates Step 15,700 | last 50 updates / 1,600 crops | lr 1.39e-05 points (35/curve, tolerance 2 vox): 47,151 right / 3,156 wrong | 93.7% correct | 5,693 unknown loss 0.9239 | geometry 0.2662 | confidence 1.3156 | mean error 0.659 vox speed: 895 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.24 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1486 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 272, from the previous prediction 1214 confidence-labeled crops: terminal 26/1533 (1.7%) | recoverable 96/1533 (6.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 288/1600 (18.0%), 2: 224/1600 (14.0%), 3: 416/1600 (26.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 80 chunks / 160.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 619 rows | mean depth 15 decisions | mean travel 123 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 52% 45% 23%>3% 19% 12 450 88%/7%/0%/5% 10% 199/1/23/1 0175A_5mm_v1 288 47% 52% 21%>1% 19% 4 338 92%/5%/1%/2% 5% 149/4/0/0 1447_5mm_v1 224 54% 42% 22%>4% 18% 9 720 89%/6%/2%/4% 4% 94/2/0/0 paris4_afv_central 416 40% 44% 25%>16% 9% 69 3186 85%/5%/4%/6% 15% 71/2/0/112 PHerc0125_3mm_v2 224 53% 47% 24%>0% 24% 0 148 88%/7%/1%/4% 13% 106/1/0/0 gradients: max 66, clipped 1/50 updates Step 15,750 | last 50 updates / 1,600 crops | lr 1.38e-05 points (35/curve, tolerance 2 vox): 47,776 right / 2,395 wrong | 95.2% correct | 5,829 unknown loss 0.8026 | geometry 0.2288 | confidence 1.1476 | mean error 0.623 vox speed: 891 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1531 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 222, from the previous prediction 1309 confidence-labeled crops: terminal 17/1523 (1.1%) | recoverable 97/1523 (6.4%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 320/1600 (20.0%), 2: 384/1600 (24.0%), 3: 160/1600 (10.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 80 chunks / 160.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 672 rows | mean depth 16 decisions | mean travel 128 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 47% 47% 22%>6% 16% 13 767 90%/5%/0%/5% 9% 107/1/17/29 0175A_5mm_v1 320 52% 45% 20%>3% 17% 11 790 89%/5%/1%/5% 4% 117/2/0/27 1447_5mm_v1 384 49% 46% 24%>6% 18% 19 896 88%/6%/1%/5% 4% 175/3/0/0 paris4_afv_central 160 44% 44% 20%>12% 8% 17 1335 85%/7%/4%/4% 10% 71/0/0/0 PHerc0125_3mm_v2 448 55% 45% 24%>0% 24% 0 148 88%/7%/0%/5% 12% 202/3/0/0 gradients: max 53, clipped 0/50 updates Step 15,800 | last 50 updates / 1,600 crops | lr 1.37e-05 points (35/curve, tolerance 2 vox): 47,513 right / 2,832 wrong | 94.4% correct | 5,655 unknown loss 0.8065 | geometry 0.2242 | confidence 1.1645 | mean error 0.613 vox speed: 891 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.26 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1486 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 226, from the previous prediction 1260 confidence-labeled crops: terminal 32/1528 (2.1%) | recoverable 106/1528 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 160/1600 (10.0%), 1: 288/1600 (18.0%), 2: 448/1600 (28.0%), 3: 384/1600 (24.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 80 chunks / 160.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 638 rows | mean depth 13 decisions | mean travel 109 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 160 52% 42% 24%>5% 19% 7 644 89%/8%/0%/3% 9% 68/0/0/0 0175A_5mm_v1 288 51% 44% 26%>5% 22% 12 858 89%/6%/1%/4% 5% 127/4/0/0 1447_5mm_v1 448 50% 45% 25%>5% 20% 18 1007 88%/7%/2%/4% 5% 153/2/2/49 paris4_afv_central 384 38% 44% 26%>18% 8% 63 1795 83%/7%/5%/5% 14% 169/1/0/0 PHerc0125_3mm_v2 320 61% 39% 29%>0% 29% 0 148 88%/6%/0%/6% 9% 121/1/6/4 gradients: max 54, clipped 0/50 updates Step 15,850 | last 50 updates / 1,600 crops | lr 1.37e-05 points (35/curve, tolerance 2 vox): 47,481 right / 2,891 wrong | 94.3% correct | 5,628 unknown loss 0.8292 | geometry 0.2438 | confidence 1.1707 | mean error 0.624 vox speed: 891 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1511 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 288, from the previous prediction 1223 confidence-labeled crops: terminal 28/1528 (1.8%) | recoverable 97/1528 (6.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 128/1600 (8.0%), 2: 416/1600 (26.0%), 3: 320/1600 (20.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 80 chunks / 160.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 617 rows | mean depth 15 decisions | mean travel 128 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 50% 49% 21%>2% 19% 6 311 92%/5%/0%/3% 7% 202/0/0/0 0175A_5mm_v1 128 45% 48% 26%>6% 20% 8 1038 84%/8%/2%/7% 5% 62/0/0/0 1447_5mm_v1 416 52% 44% 29%>3% 25% 12 977 88%/8%/1%/3% 3% 97/3/0/87 paris4_afv_central 320 40% 42% 28%>19% 9% 53 1945 82%/5%/6%/8% 16% 131/2/45/2 PHerc0125_3mm_v2 320 52% 48% 25%>0% 25% 0 137 88%/5%/1%/6% 14% 125/4/0/27 gradients: max 46, clipped 0/50 updates Step 15,900 | last 50 updates / 1,600 crops | lr 1.36e-05 points (35/curve, tolerance 2 vox): 47,923 right / 2,160 wrong | 95.7% correct | 5,917 unknown loss 0.8259 | geometry 0.2372 | confidence 1.1773 | mean error 0.618 vox speed: 893 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1555 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 256, from the previous prediction 1299 confidence-labeled crops: terminal 15/1515 (1.0%) | recoverable 99/1515 (6.5%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 320/1600 (20.0%), 2: 416/1600 (26.0%), 3: 160/1600 (10.0%), 4: 480/1600 (30.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 80 chunks / 160.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 634 rows | mean depth 15 decisions | mean travel 124 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 54% 43% 22%>3% 19% 6 423 89%/6%/0%/4% 8% 97/0/1/0 0175A_5mm_v1 320 52% 45% 24%>3% 21% 9 674 86%/8%/1%/5% 5% 144/3/0/0 1447_5mm_v1 416 53% 43% 23%>4% 19% 17 938 90%/4%/1%/5% 3% 141/5/0/38 paris4_afv_central 160 46% 48% 18%>6% 12% 10 814 86%/6%/3%/5% 9% 76/0/0/0 PHerc0125_3mm_v2 480 57% 43% 25%>0% 25% 0 138 86%/7%/1%/6% 11% 176/8/0/31 gradients: max 50, clipped 0/50 updates Step 15,950 | last 50 updates / 1,600 crops | lr 1.35e-05 points (35/curve, tolerance 2 vox): 48,288 right / 2,989 wrong | 94.2% correct | 4,723 unknown loss 0.8974 | geometry 0.2836 | confidence 1.2276 | mean error 0.680 vox speed: 893 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1517 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 281, from the previous prediction 1236 confidence-labeled crops: terminal 23/1532 (1.5%) | recoverable 107/1532 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 480/1600 (30.0%), 1: 256/1600 (16.0%), 2: 192/1600 (12.0%), 3: 320/1600 (20.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 80 chunks / 160.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 594 rows | mean depth 17 decisions | mean travel 137 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 480 48% 48% 25%>5% 20% 17 938 89%/7%/0%/4% 9% 134/0/18/94 0175A_5mm_v1 256 54% 41% 24%>5% 19% 12 1004 88%/7%/2%/4% 6% 106/2/0/0 1447_5mm_v1 192 47% 48% 26%>5% 21% 8 764 88%/8%/2%/2% 4% 93/0/0/0 paris4_afv_central 320 48% 42% 25%>11% 14% 31 1325 83%/8%/4%/5% 11% 134/3/0/0 PHerc0125_3mm_v2 352 55% 45% 23%>0% 23% 0 136 89%/5%/0%/6% 11% 127/5/0/32 gradients: max 71, clipped 1/50 updates Step 16,000 | last 50 updates / 1,600 crops | lr 1.34e-05 points (35/curve, tolerance 2 vox): 47,427 right / 2,911 wrong | 94.2% correct | 5,662 unknown loss 0.8413 | geometry 0.2586 | confidence 1.1654 | mean error 0.637 vox speed: 891 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1498 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 292, from the previous prediction 1206 confidence-labeled crops: terminal 27/1527 (1.8%) | recoverable 89/1527 (5.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 416/1600 (26.0%), 2: 224/1600 (14.0%), 3: 384/1600 (24.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 80 chunks / 160.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 609 rows | mean depth 14 decisions | mean travel 118 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 53% 43% 31%>3% 27% 9 564 88%/9%/0%/3% 12% 104/0/0/21 0175A_5mm_v1 416 46% 47% 21%>7% 14% 26 1761 89%/5%/1%/5% 5% 111/3/28/83 1447_5mm_v1 224 46% 52% 15%>2% 12% 5 453 92%/4%/0%/4% 3% 116/1/0/0 paris4_afv_central 384 41% 45% 27%>14% 12% 49 1830 82%/7%/5%/6% 13% 143/4/28/30 PHerc0125_3mm_v2 288 53% 47% 22%>0% 22% 0 138 92%/3%/1%/5% 11% 135/5/0/0 gradients: max 51, clipped 0/50 updates Step 16,000 | dagger_launched: true Step 16,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 246, "stopped": 1, "mean_example_error": 0.4156117860613198, "geometry_examples": 29, "examples": 32, "inference_and_measurement_seconds": 1.8394761681556702, "total_diagnostic_seconds": 3.6251387037336826} Step 16,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 6/8 (75.0%) | false stops 0 1.5-3: recovered 7/8 (87.5%) | false stops 0 3-4.5: recovered 6/8 (75.0%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 0 entire proposed 8-point prefix correct (distance): 22/28 (78.6%) gate @ 0.50: false stops 0/23 (0.0%) | accepted wrong 5/27 (18.5%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.585 1.567 following 16 464 1.139 1.191 recoverable 12 348 2.181 2.069 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 1.139 1.191 332 32 812 1.585 1.567 improved/worsened/compared states: 3/1/28 Step 16,000 | batch_diagnostics_seconds: 3.627621151506901 | recovery_seconds: 14.188569899648428 Step 16,050 | last 50 updates / 1,600 crops | lr 1.34e-05 points (35/curve, tolerance 2 vox): 47,760 right / 2,495 wrong | 95.0% correct | 5,745 unknown loss 0.7957 | geometry 0.2429 | confidence 1.1056 | mean error 0.636 vox speed: 989 ms/update | data wait 1 ms/update | 23.1 crops/s | 23.1 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1510 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 259, from the previous prediction 1251 confidence-labeled crops: terminal 16/1517 (1.1%) | recoverable 100/1517 (6.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 320/1600 (20.0%), 2: 448/1600 (28.0%), 3: 256/1600 (16.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 80 chunks / 160.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 630 rows | mean depth 16 decisions | mean travel 128 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 55% 42% 24%>3% 21% 10 493 88%/9%/0%/3% 10% 133/0/0/0 0175A_5mm_v1 320 51% 46% 19%>2% 17% 8 652 88%/5%/1%/6% 6% 114/1/0/34 1447_5mm_v1 448 49% 43% 27%>8% 20% 29 1312 87%/7%/1%/5% 4% 172/5/128/20 paris4_afv_central 256 39% 46% 27%>14% 12% 33 1924 83%/5%/4%/8% 11% 85/3/0/33 PHerc0125_3mm_v2 256 51% 49% 20%>0% 20% 0 149 90%/4%/0%/6% 8% 126/3/0/0 gradients: max 58, clipped 0/50 updates Step 16,100 | last 50 updates / 1,600 crops | lr 1.33e-05 points (35/curve, tolerance 2 vox): 48,860 right / 2,231 wrong | 95.6% correct | 4,909 unknown loss 0.7521 | geometry 0.2300 | confidence 1.0442 | mean error 0.617 vox speed: 1449 ms/update | data wait 1 ms/update | 21.8 crops/s | 21.8 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1534 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 286, from the previous prediction 1248 confidence-labeled crops: terminal 14/1533 (0.9%) | recoverable 102/1533 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 448/1600 (28.0%), 2: 352/1600 (22.0%), 3: 224/1600 (14.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 80 chunks / 160.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 608 rows | mean depth 18 decisions | mean travel 147 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 47% 48% 26%>5% 21% 13 634 92%/5%/0%/3% 8% 137/1/0/0 0175A_5mm_v1 448 52% 46% 25%>2% 23% 7 549 90%/7%/0%/4% 4% 145/4/0/61 1447_5mm_v1 352 55% 41% 28%>3% 24% 11 816 90%/7%/0%/3% 2% 101/4/0/45 paris4_afv_central 224 41% 45% 24%>14% 9% 27 1568 82%/6%/5%/6% 12% 101/0/0/0 PHerc0125_3mm_v2 288 57% 43% 25%>0% 25% 0 143 86%/6%/0%/7% 11% 124/5/49/1 gradients: max 61, clipped 1/50 updates Step 16,150 | last 50 updates / 1,600 crops | lr 1.32e-05 points (35/curve, tolerance 2 vox): 47,824 right / 2,924 wrong | 94.2% correct | 5,252 unknown loss 0.9099 | geometry 0.2705 | confidence 1.2787 | mean error 0.657 vox speed: 1447 ms/update | data wait 1 ms/update | 21.8 crops/s | 21.8 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1512 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 229, from the previous prediction 1283 confidence-labeled crops: terminal 17/1532 (1.1%) | recoverable 104/1532 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 192/1600 (12.0%), 2: 320/1600 (20.0%), 3: 352/1600 (22.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 81 chunks / 162.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 646 rows | mean depth 14 decisions | mean travel 112 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 51% 47% 23%>2% 21% 14 739 89%/8%/0%/3% 11% 147/1/83/48 0175A_5mm_v1 192 49% 46% 29%>5% 23% 10 958 90%/6%/1%/4% 7% 88/1/0/0 1447_5mm_v1 320 51% 45% 22%>4% 18% 12 772 92%/4%/1%/2% 3% 130/6/0/13 paris4_afv_central 352 43% 44% 25%>13% 12% 42 1506 82%/7%/4%/8% 12% 154/0/0/0 PHerc0125_3mm_v2 320 56% 44% 26%>0% 26% 0 135 88%/7%/0%/5% 11% 127/2/13/15 gradients: max 44, clipped 0/50 updates Step 16,167 | dagger_states: 2140 | dagger_source_step: 16000 | dagger_caches: 4 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/decisions_016000.npz" | dagger_fibers: 128 | dagger_supply: {"states": 2140, "traces": 128, "fibers": 128, "rows": {"terminal": 18, "premature_stop": 114, "recoverable": 67, "pre_excursion": 95, "ordinary": 1844}, "events": {"terminal": 3, "premature_stop": 114, "recoverable": 17, "pre_excursion": 18, "ordinary": 123}, "supervision": {"following": 2047, "recoverable": 67, "terminal": 18, "unknown": 8}, "departed_traces": 14, "terminal_traces": 3, "stop_reasons": {"after_failure": 2, "confidence": 117, "unannotated": 9}, "traced_voxels": 30715.43919374601, "seconds": 231.23186920955777, "voxels_per_second": 132.83393547240507} | dagger_coverage: {"requested": 128, "seeds": 128, "skipped": [], "epoch": 0} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 0 | dagger_launch_interval: 5000 | dagger_publication_age: 167 | dataset: "paris4" Step 16,200 | last 50 updates / 1,600 crops | lr 1.31e-05 points (35/curve, tolerance 2 vox): 47,863 right / 2,950 wrong | 94.2% correct | 5,187 unknown loss 0.8216 | geometry 0.2581 | confidence 1.1271 | mean error 0.627 vox speed: 1021 ms/update | data wait 1 ms/update | 31.0 crops/s | 31.0 decisions/s | 1.24 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1513 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 184, from the previous prediction 1329 confidence-labeled crops: terminal 33/1522 (2.2%) | recoverable 115/1522 (7.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 416/1600 (26.0%), 2: 352/1600 (22.0%), 3: 256/1600 (16.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 81 chunks / 162.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 676 rows | mean depth 16 decisions | mean travel 127 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 53% 44% 28%>4% 24% 6 527 89%/6%/0%/5% 7% 112/0/0/0 0175A_5mm_v1 416 51% 46% 25%>3% 22% 13 697 86%/7%/2%/5% 6% 191/3/21/0 1447_5mm_v1 352 50% 47% 25%>4% 21% 12 671 87%/8%/1%/4% 3% 158/4/0/6 paris4_afv_central 256 39% 41% 32%>20% 12% 43 2125 75%/10%/7%/7% 17% 105/2/0/0 PHerc0125_3mm_v2 320 57% 43% 23%>0% 23% 0 136 89%/5%/0%/5% 10% 110/1/0/29 gradients: max 53, clipped 0/50 updates Step 16,250 | last 50 updates / 1,600 crops | lr 1.30e-05 points (35/curve, tolerance 2 vox): 47,346 right / 2,691 wrong | 94.6% correct | 5,963 unknown loss 0.9194 | geometry 0.2680 | confidence 1.3028 | mean error 0.658 vox speed: 891 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1525 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 236, from the previous prediction 1289 confidence-labeled crops: terminal 16/1512 (1.1%) | recoverable 107/1512 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 384/1600 (24.0%), 2: 192/1600 (12.0%), 3: 320/1600 (20.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 81 chunks / 162.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 633 rows | mean depth 17 decisions | mean travel 135 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 56% 41% 31%>3% 27% 9 540 89%/8%/0%/3% 9% 117/0/0/0 0175A_5mm_v1 384 53% 45% 21%>2% 19% 8 506 90%/6%/0%/4% 4% 173/3/1/0 1447_5mm_v1 192 52% 45% 22%>3% 19% 5 531 91%/6%/1%/2% 3% 87/3/0/0 paris4_afv_central 320 43% 43% 28%>14% 14% 46 2133 82%/8%/3%/7% 11% 107/1/35/31 PHerc0125_3mm_v2 416 57% 43% 23%>0% 23% 0 138 84%/7%/0%/9% 13% 149/4/0/29 gradients: max 56, clipped 0/50 updates Step 16,300 | last 50 updates / 1,600 crops | lr 1.30e-05 points (35/curve, tolerance 2 vox): 48,149 right / 2,225 wrong | 95.6% correct | 5,626 unknown loss 0.8214 | geometry 0.2329 | confidence 1.1771 | mean error 0.622 vox speed: 889 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1531 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 271, from the previous prediction 1260 confidence-labeled crops: terminal 13/1519 (0.9%) | recoverable 105/1519 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 512/1600 (32.0%), 2: 224/1600 (14.0%), 3: 352/1600 (22.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 81 chunks / 162.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 638 rows | mean depth 15 decisions | mean travel 119 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 46% 48% 25%>5% 20% 11 680 86%/9%/0%/5% 9% 124/0/0/0 0175A_5mm_v1 512 54% 45% 24%>1% 23% 5 419 87%/7%/1%/5% 5% 140/1/0/91 1447_5mm_v1 224 50% 47% 21%>2% 19% 5 463 91%/7%/0%/3% 1% 106/3/0/0 paris4_afv_central 352 42% 46% 22%>12% 10% 41 1579 86%/3%/2%/8% 9% 147/6/0/14 PHerc0125_3mm_v2 256 53% 47% 22%>0% 22% 0 145 89%/7%/0%/4% 11% 121/4/0/0 gradients: max 66, clipped 1/50 updates Step 16,350 | last 50 updates / 1,600 crops | lr 1.29e-05 points (35/curve, tolerance 2 vox): 48,420 right / 2,521 wrong | 95.1% correct | 5,059 unknown loss 0.8527 | geometry 0.2383 | confidence 1.2287 | mean error 0.642 vox speed: 891 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1542 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 215, from the previous prediction 1327 confidence-labeled crops: terminal 14/1522 (0.9%) | recoverable 99/1522 (6.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 352/1600 (22.0%), 2: 224/1600 (14.0%), 3: 256/1600 (16.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 81 chunks / 162.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 676 rows | mean depth 16 decisions | mean travel 127 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 56% 42% 23%>2% 22% 5 363 89%/7%/0%/3% 8% 136/0/0/0 0175A_5mm_v1 352 53% 45% 23%>2% 21% 6 441 90%/5%/0%/5% 4% 157/2/0/3 1447_5mm_v1 224 50% 45% 27%>4% 23% 10 800 86%/8%/1%/4% 4% 101/2/0/0 paris4_afv_central 256 40% 46% 25%>14% 11% 35 1582 84%/6%/4%/7% 11% 118/3/0/0 PHerc0125_3mm_v2 448 54% 46% 22%>0% 22% 0 138 89%/5%/0%/5% 11% 164/6/0/41 gradients: max 47, clipped 0/50 updates Step 16,400 | last 50 updates / 1,600 crops | lr 1.28e-05 points (35/curve, tolerance 2 vox): 46,966 right / 3,181 wrong | 93.7% correct | 5,853 unknown loss 0.9303 | geometry 0.2707 | confidence 1.3193 | mean error 0.664 vox speed: 890 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.24 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1509 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 303, from the previous prediction 1206 confidence-labeled crops: terminal 32/1531 (2.1%) | recoverable 130/1531 (8.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 224/1600 (14.0%), 2: 192/1600 (12.0%), 3: 448/1600 (28.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 81 chunks / 162.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 588 rows | mean depth 17 decisions | mean travel 138 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 54% 44% 25%>2% 23% 6 363 88%/9%/0%/3% 9% 139/0/27/3 0175A_5mm_v1 224 62% 38% 28%>0% 28% 0 138 86%/11%/0%/3% 4% 71/1/0/15 1447_5mm_v1 192 45% 52% 20%>3% 18% 4 522 92%/8%/0%/1% 3% 100/1/0/0 paris4_afv_central 448 39% 45% 27%>17% 10% 62 2209 78%/8%/7%/6% 14% 161/5/0/39 PHerc0125_3mm_v2 416 58% 42% 24%>0% 24% 0 134 88%/6%/0%/6% 13% 117/2/0/59 gradients: max 63, clipped 1/50 updates Step 16,450 | last 50 updates / 1,600 crops | lr 1.27e-05 points (35/curve, tolerance 2 vox): 47,236 right / 2,876 wrong | 94.3% correct | 5,888 unknown loss 0.9514 | geometry 0.2605 | confidence 1.3819 | mean error 0.663 vox speed: 890 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1522 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 239, from the previous prediction 1283 confidence-labeled crops: terminal 23/1515 (1.5%) | recoverable 90/1515 (5.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 288/1600 (18.0%), 2: 288/1600 (18.0%), 3: 288/1600 (18.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 82 chunks / 164.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 635 rows | mean depth 16 decisions | mean travel 126 vox | 6 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 47% 49% 22%>4% 18% 14 782 92%/4%/0%/3% 10% 137/1/18/19 0175A_5mm_v1 288 49% 49% 24%>2% 22% 7 396 89%/7%/1%/3% 6% 140/2/0/0 1447_5mm_v1 288 51% 44% 27%>5% 22% 15 949 90%/4%/0%/6% 2% 126/1/0/0 paris4_afv_central 288 44% 44% 24%>11% 12% 32 1746 79%/6%/6%/8% 13% 108/1/0/20 PHerc0125_3mm_v2 416 56% 44% 23%>0% 23% 0 133 87%/6%/0%/6% 12% 124/3/0/61 gradients: max 43, clipped 0/50 updates Step 16,500 | last 50 updates / 1,600 crops | lr 1.27e-05 points (35/curve, tolerance 2 vox): 48,570 right / 2,501 wrong | 95.1% correct | 4,929 unknown loss 0.8082 | geometry 0.2273 | confidence 1.1619 | mean error 0.613 vox speed: 891 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1521 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 165, from the previous prediction 1356 confidence-labeled crops: terminal 21/1532 (1.4%) | recoverable 115/1532 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 288/1600 (18.0%), 2: 384/1600 (24.0%), 3: 352/1600 (22.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 82 chunks / 164.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 718 rows | mean depth 17 decisions | mean travel 140 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 52% 48% 22%>0% 22% 0 142 89%/9%/0%/2% 8% 93/0/0/0 0175A_5mm_v1 288 52% 44% 30%>5% 26% 12 940 89%/7%/0%/4% 5% 126/3/0/0 1447_5mm_v1 384 52% 44% 27%>4% 23% 15 770 89%/8%/1%/3% 4% 162/3/64/7 paris4_afv_central 352 36% 50% 23%>14% 9% 47 1634 82%/6%/4%/7% 11% 165/3/0/11 PHerc0125_3mm_v2 384 55% 45% 24%>0% 24% 0 145 88%/7%/1%/5% 12% 172/2/0/0 gradients: max 47, clipped 0/50 updates Step 16,550 | last 50 updates / 1,600 crops | lr 1.26e-05 points (35/curve, tolerance 2 vox): 47,772 right / 2,999 wrong | 94.1% correct | 5,229 unknown loss 0.8839 | geometry 0.2492 | confidence 1.2693 | mean error 0.649 vox speed: 889 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1532 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 223, from the previous prediction 1309 confidence-labeled crops: terminal 23/1531 (1.5%) | recoverable 123/1531 (8.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 416/1600 (26.0%), 2: 288/1600 (18.0%), 3: 192/1600 (12.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 82 chunks / 164.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 653 rows | mean depth 15 decisions | mean travel 124 vox | 6 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 53% 42% 26%>6% 21% 17 790 87%/9%/0%/4% 11% 147/0/0/0 0175A_5mm_v1 416 52% 44% 26%>4% 22% 16 732 88%/9%/1%/3% 4% 181/4/0/0 1447_5mm_v1 288 56% 42% 23%>3% 20% 8 789 90%/6%/1%/4% 2% 83/1/0/37 paris4_afv_central 192 43% 45% 22%>12% 10% 22 1479 82%/4%/7%/7% 15% 87/1/0/0 PHerc0125_3mm_v2 352 55% 45% 28%>0% 28% 0 141 85%/9%/1%/5% 15% 155/4/0/2 gradients: max 46, clipped 0/50 updates Step 16,600 | last 50 updates / 1,600 crops | lr 1.25e-05 points (35/curve, tolerance 2 vox): 47,773 right / 2,797 wrong | 94.5% correct | 5,430 unknown loss 0.7744 | geometry 0.2467 | confidence 1.0555 | mean error 0.619 vox speed: 893 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1506 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 275, from the previous prediction 1231 confidence-labeled crops: terminal 26/1528 (1.7%) | recoverable 109/1528 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 384/1600 (24.0%), 2: 416/1600 (26.0%), 3: 384/1600 (24.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 82 chunks / 164.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 610 rows | mean depth 16 decisions | mean travel 130 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 39% 50% 22%>11% 10% 16 1157 89%/5%/0%/6% 13% 96/0/0/0 0175A_5mm_v1 384 55% 41% 26%>4% 22% 15 679 87%/7%/1%/5% 6% 144/2/103/12 1447_5mm_v1 416 54% 44% 25%>2% 23% 10 794 90%/6%/0%/4% 2% 103/3/0/79 paris4_afv_central 384 45% 43% 27%>12% 15% 41 1518 80%/9%/5%/6% 13% 166/2/14/0 PHerc0125_3mm_v2 224 55% 45% 24%>0% 24% 0 157 92%/6%/0%/2% 12% 101/1/0/0 gradients: max 46, clipped 0/50 updates Step 16,650 | last 50 updates / 1,600 crops | lr 1.24e-05 points (35/curve, tolerance 2 vox): 47,904 right / 2,295 wrong | 95.4% correct | 5,801 unknown loss 0.7624 | geometry 0.2267 | confidence 1.0714 | mean error 0.614 vox speed: 892 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1506 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 251, from the previous prediction 1255 confidence-labeled crops: terminal 16/1514 (1.1%) | recoverable 113/1514 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 448/1600 (28.0%), 2: 256/1600 (16.0%), 3: 320/1600 (20.0%), 4: 160/1600 (10.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 83 chunks / 166.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 619 rows | mean depth 18 decisions | mean travel 143 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 52% 40% 25%>8% 17% 20 1026 89%/6%/0%/5% 10% 167/0/0/0 0175A_5mm_v1 448 55% 43% 28%>1% 27% 6 490 89%/8%/0%/3% 3% 136/2/0/58 1447_5mm_v1 256 52% 45% 21%>4% 18% 8 693 86%/8%/0%/6% 3% 115/0/0/0 paris4_afv_central 320 43% 42% 26%>15% 11% 35 1835 80%/5%/4%/11% 12% 134/1/0/0 PHerc0125_3mm_v2 160 58% 42% 31%>0% 31% 0 157 86%/9%/1%/5% 15% 67/1/0/0 gradients: max 42, clipped 0/50 updates Step 16,700 | last 50 updates / 1,600 crops | lr 1.23e-05 points (35/curve, tolerance 2 vox): 48,310 right / 2,728 wrong | 94.7% correct | 4,962 unknown loss 0.7842 | geometry 0.2279 | confidence 1.1125 | mean error 0.610 vox speed: 890 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1530 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 230, from the previous prediction 1300 confidence-labeled crops: terminal 26/1538 (1.7%) | recoverable 111/1538 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 384/1600 (24.0%), 2: 448/1600 (28.0%), 3: 256/1600 (16.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 83 chunks / 166.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 673 rows | mean depth 18 decisions | mean travel 145 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 53% 45% 25%>2% 23% 6 446 89%/8%/0%/3% 9% 118/0/20/26 0175A_5mm_v1 384 55% 42% 26%>2% 23% 9 443 89%/7%/1%/4% 5% 163/5/0/0 1447_5mm_v1 448 46% 49% 21%>5% 17% 20 717 87%/6%/2%/5% 6% 218/1/0/0 paris4_afv_central 256 44% 44% 29%>12% 17% 27 1692 80%/9%/6%/5% 12% 103/3/0/9 PHerc0125_3mm_v2 192 54% 46% 23%>0% 23% 0 189 91%/5%/1%/4% 11% 71/0/83/17 gradients: max 55, clipped 0/50 updates Step 16,750 | last 50 updates / 1,600 crops | lr 1.23e-05 points (35/curve, tolerance 2 vox): 47,784 right / 3,018 wrong | 94.1% correct | 5,198 unknown loss 0.9413 | geometry 0.2845 | confidence 1.3135 | mean error 0.684 vox speed: 893 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1544 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 310, from the previous prediction 1234 confidence-labeled crops: terminal 18/1529 (1.2%) | recoverable 122/1529 (8.0%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 192/1600 (12.0%), 2: 288/1600 (18.0%), 3: 288/1600 (18.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 83 chunks / 166.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 578 rows | mean depth 15 decisions | mean travel 127 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 52% 46% 27%>3% 24% 13 882 91%/7%/0%/3% 10% 97/0/0/107 0175A_5mm_v1 192 53% 47% 26%>0% 26% 0 138 89%/3%/0%/8% 3% 77/2/0/13 1447_5mm_v1 288 51% 47% 23%>2% 21% 7 519 85%/11%/0%/3% 2% 134/3/0/0 paris4_afv_central 288 47% 42% 25%>12% 14% 32 1613 81%/7%/6%/6% 12% 120/6/0/0 PHerc0125_3mm_v2 384 57% 43% 23%>0% 23% 0 144 86%/8%/0%/5% 15% 150/0/32/16 gradients: max 77, clipped 1/50 updates Step 16,800 | last 50 updates / 1,600 crops | lr 1.22e-05 points (35/curve, tolerance 2 vox): 48,358 right / 2,346 wrong | 95.4% correct | 5,296 unknown loss 0.7312 | geometry 0.2293 | confidence 1.0039 | mean error 0.600 vox speed: 891 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1509 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 260, from the previous prediction 1249 confidence-labeled crops: terminal 25/1527 (1.6%) | recoverable 107/1527 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 480/1600 (30.0%), 2: 352/1600 (22.0%), 3: 288/1600 (18.0%), 4: 160/1600 (10.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 83 chunks / 166.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 621 rows | mean depth 19 decisions | mean travel 153 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 45% 50% 24%>5% 19% 15 651 93%/5%/0%/2% 8% 160/0/0/0 0175A_5mm_v1 480 51% 43% 27%>6% 21% 24 906 88%/7%/2%/4% 6% 206/6/0/0 1447_5mm_v1 352 58% 39% 28%>3% 25% 8 802 85%/9%/0%/6% 2% 97/1/13/41 paris4_afv_central 288 45% 41% 26%>14% 12% 35 1785 81%/7%/5%/7% 10% 119/2/0/0 PHerc0125_3mm_v2 160 56% 44% 24%>0% 24% 0 152 88%/6%/0%/6% 15% 39/3/0/32 gradients: max 55, clipped 0/50 updates Step 16,850 | last 50 updates / 1,600 crops | lr 1.21e-05 points (35/curve, tolerance 2 vox): 46,609 right / 3,485 wrong | 93.0% correct | 5,906 unknown loss 0.9417 | geometry 0.2691 | confidence 1.3451 | mean error 0.681 vox speed: 889 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.27 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1516 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 282, from the previous prediction 1234 confidence-labeled crops: terminal 33/1502 (2.2%) | recoverable 97/1502 (6.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 384/1600 (24.0%), 2: 160/1600 (10.0%), 3: 320/1600 (20.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 83 chunks / 166.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 616 rows | mean depth 16 decisions | mean travel 127 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 45% 47% 24%>8% 16% 20 964 89%/6%/0%/5% 13% 136/0/0/0 0175A_5mm_v1 384 49% 46% 25%>5% 20% 17 865 88%/5%/3%/5% 7% 152/3/55/23 1447_5mm_v1 160 58% 41% 23%>1% 22% 2 390 91%/6%/0%/4% 4% 65/2/0/0 paris4_afv_central 320 44% 44% 26%>12% 14% 33 1578 78%/7%/6%/10% 12% 140/2/0/0 PHerc0125_3mm_v2 448 54% 46% 21%>0% 21% 0 131 85%/7%/1%/7% 13% 123/1/0/82 gradients: max 62, clipped 1/50 updates Step 16,900 | last 50 updates / 1,600 crops | lr 1.20e-05 points (35/curve, tolerance 2 vox): 47,092 right / 2,984 wrong | 94.0% correct | 5,924 unknown loss 0.8295 | geometry 0.2664 | confidence 1.1263 | mean error 0.628 vox speed: 890 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.24 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1502 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 278, from the previous prediction 1224 confidence-labeled crops: terminal 31/1513 (2.0%) | recoverable 116/1513 (7.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 320/1600 (20.0%), 2: 256/1600 (16.0%), 3: 448/1600 (28.0%), 4: 160/1600 (10.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 84 chunks / 168.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 588 rows | mean depth 15 decisions | mean travel 121 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 53% 41% 30%>6% 24% 18 827 87%/8%/0%/5% 11% 171/1/22/1 0175A_5mm_v1 320 52% 46% 26%>2% 24% 7 634 89%/8%/0%/3% 5% 84/1/0/64 1447_5mm_v1 256 50% 48% 23%>2% 21% 3 457 89%/7%/0%/4% 3% 109/2/18/14 paris4_afv_central 448 45% 42% 27%>14% 14% 52 2098 79%/7%/6%/8% 13% 151/4/0/35 PHerc0125_3mm_v2 160 54% 46% 22%>0% 22% 0 152 84%/6%/1%/9% 13% 73/2/0/0 gradients: max 53, clipped 0/50 updates Step 16,950 | last 50 updates / 1,600 crops | lr 1.20e-05 points (35/curve, tolerance 2 vox): 48,447 right / 2,533 wrong | 95.0% correct | 5,020 unknown loss 0.8713 | geometry 0.2535 | confidence 1.2357 | mean error 0.619 vox speed: 892 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1504 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 174, from the previous prediction 1330 confidence-labeled crops: terminal 22/1534 (1.4%) | recoverable 122/1534 (8.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 448/1600 (28.0%), 2: 192/1600 (12.0%), 3: 320/1600 (20.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 84 chunks / 168.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 689 rows | mean depth 16 decisions | mean travel 130 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 43% 47% 25%>10% 15% 18 1130 88%/6%/0%/5% 9% 121/0/2/0 0175A_5mm_v1 448 47% 45% 24%>8% 16% 24 957 86%/9%/2%/4% 8% 194/0/0/9 1447_5mm_v1 192 55% 43% 27%>2% 25% 3 449 89%/9%/0%/3% 2% 83/1/5/0 paris4_afv_central 320 47% 42% 26%>11% 15% 30 1596 84%/8%/4%/5% 9% 124/3/0/12 PHerc0125_3mm_v2 384 54% 46% 24%>0% 24% 0 141 88%/7%/0%/5% 10% 167/4/0/10 gradients: max 41, clipped 0/50 updates Step 17,000 | last 50 updates / 1,600 crops | lr 1.19e-05 points (35/curve, tolerance 2 vox): 48,392 right / 3,090 wrong | 94.0% correct | 4,518 unknown loss 1.0194 | geometry 0.2861 | confidence 1.4666 | mean error 0.681 vox speed: 892 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1515 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 249, from the previous prediction 1266 confidence-labeled crops: terminal 15/1535 (1.0%) | recoverable 93/1535 (6.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 512/1600 (32.0%), 1: 288/1600 (18.0%), 2: 192/1600 (12.0%), 3: 288/1600 (18.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 84 chunks / 168.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 634 rows | mean depth 16 decisions | mean travel 128 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 512 52% 43% 27%>5% 23% 19 794 90%/7%/0%/3% 11% 189/2/3/32 0175A_5mm_v1 288 52% 41% 25%>7% 18% 17 1154 87%/5%/1%/7% 6% 119/5/0/0 1447_5mm_v1 192 53% 41% 26%>7% 19% 10 1299 90%/5%/0%/5% 5% 77/2/29/1 paris4_afv_central 288 41% 50% 22%>10% 12% 25 1283 86%/6%/3%/5% 8% 139/6/0/4 PHerc0125_3mm_v2 320 58% 42% 28%>0% 28% 0 136 89%/5%/0%/5% 15% 110/3/0/24 gradients: max 42, clipped 0/50 updates Step 17,000 | dagger_launched: true Step 17,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 256, "stopped": 0, "mean_example_error": 0.44303751587867735, "geometry_examples": 30, "examples": 32, "inference_and_measurement_seconds": 1.8134752251207829, "total_diagnostic_seconds": 3.6720812506973743} Step 17,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 6/8 (75.0%) | false stops 0 1.5-3: recovered 6/8 (75.0%) | false stops 0 3-4.5: recovered 6/8 (75.0%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 0 entire proposed 8-point prefix correct (distance): 22/28 (78.6%) gate @ 0.50: false stops 0/23 (0.0%) | accepted wrong 4/26 (15.4%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.606 1.586 following 16 464 1.178 1.224 recoverable 12 348 2.177 2.069 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 1.178 1.224 332 32 812 1.606 1.586 improved/worsened/compared states: 3/1/28 Step 17,000 | batch_diagnostics_seconds: 3.678754333406687 | recovery_seconds: 15.000713400542736 Step 17,050 | last 50 updates / 1,600 crops | lr 1.18e-05 points (35/curve, tolerance 2 vox): 48,019 right / 2,479 wrong | 95.1% correct | 5,502 unknown loss 0.8024 | geometry 0.2353 | confidence 1.1342 | mean error 0.621 vox speed: 1342 ms/update | data wait 1 ms/update | 18.2 crops/s | 18.2 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1536 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 219, from the previous prediction 1317 confidence-labeled crops: terminal 18/1525 (1.2%) | recoverable 96/1525 (6.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 320/1600 (20.0%), 2: 320/1600 (20.0%), 3: 256/1600 (16.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 84 chunks / 168.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 653 rows | mean depth 14 decisions | mean travel 117 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 54% 41% 28%>5% 23% 14 743 89%/8%/0%/3% 9% 140/2/0/6 0175A_5mm_v1 320 51% 43% 26%>5% 21% 15 986 88%/6%/2%/4% 4% 139/4/0/0 1447_5mm_v1 320 53% 43% 24%>4% 20% 13 826 90%/5%/1%/4% 3% 138/2/0/0 paris4_afv_central 256 52% 41% 22%>7% 15% 17 1157 81%/7%/4%/8% 9% 104/0/0/0 PHerc0125_3mm_v2 352 56% 44% 25%>0% 25% 0 134 89%/4%/1%/6% 11% 132/1/0/23 gradients: max 55, clipped 0/50 updates Step 17,100 | last 50 updates / 1,600 crops | lr 1.17e-05 points (35/curve, tolerance 2 vox): 47,695 right / 2,294 wrong | 95.4% correct | 6,011 unknown loss 0.8422 | geometry 0.2413 | confidence 1.2019 | mean error 0.634 vox speed: 1381 ms/update | data wait 1 ms/update | 23.0 crops/s | 23.0 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1537 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 345, from the previous prediction 1192 confidence-labeled crops: terminal 13/1503 (0.9%) | recoverable 94/1503 (6.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 320/1600 (20.0%), 2: 352/1600 (22.0%), 3: 288/1600 (18.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 84 chunks / 168.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 558 rows | mean depth 19 decisions | mean travel 151 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 57% 41% 28%>2% 26% 8 818 89%/8%/0%/3% 9% 93/0/0/25 0175A_5mm_v1 320 42% 50% 21%>8% 13% 20 905 88%/5%/2%/5% 7% 161/1/0/0 1447_5mm_v1 352 53% 46% 24%>1% 23% 2 314 89%/6%/0%/5% 2% 87/1/22/75 paris4_afv_central 288 43% 48% 19%>8% 11% 22 1174 89%/5%/2%/5% 6% 139/2/0/0 PHerc0125_3mm_v2 352 60% 40% 26%>0% 26% 0 142 82%/6%/0%/13% 13% 78/2/0/63 gradients: max 70, clipped 2/50 updates Step 17,150 | last 50 updates / 1,600 crops | lr 1.17e-05 points (35/curve, tolerance 2 vox): 48,407 right / 2,201 wrong | 95.7% correct | 5,392 unknown loss 0.8081 | geometry 0.2419 | confidence 1.1324 | mean error 0.611 vox speed: 1369 ms/update | data wait 1 ms/update | 23.2 crops/s | 23.2 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1560 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 322, from the previous prediction 1238 confidence-labeled crops: terminal 11/1529 (0.7%) | recoverable 122/1529 (8.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 352/1600 (22.0%), 2: 512/1600 (32.0%), 3: 128/1600 (8.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 84 chunks / 168.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 592 rows | mean depth 16 decisions | mean travel 130 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 51% 47% 22%>2% 20% 4 402 89%/9%/0%/2% 9% 90/1/0/0 0175A_5mm_v1 352 53% 43% 22%>4% 18% 13 673 89%/6%/1%/5% 6% 150/0/0/0 1447_5mm_v1 512 52% 45% 25%>3% 22% 13 824 86%/9%/0%/5% 4% 113/0/0/119 paris4_afv_central 128 48% 46% 27%>6% 20% 7 1000 86%/5%/4%/5% 8% 59/0/0/0 PHerc0125_3mm_v2 416 56% 44% 27%>0% 27% 0 153 88%/8%/0%/5% 13% 180/4/7/4 gradients: max 43, clipped 0/50 updates Step 17,200 | last 50 updates / 1,600 crops | lr 1.16e-05 points (35/curve, tolerance 2 vox): 48,393 right / 2,493 wrong | 95.1% correct | 5,114 unknown loss 0.8622 | geometry 0.2605 | confidence 1.2035 | mean error 0.653 vox speed: 1379 ms/update | data wait 1 ms/update | 23.0 crops/s | 23.0 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1540 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 302, from the previous prediction 1238 confidence-labeled crops: terminal 17/1537 (1.1%) | recoverable 96/1537 (6.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 480/1600 (30.0%), 2: 192/1600 (12.0%), 3: 224/1600 (14.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 84 chunks / 168.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 612 rows | mean depth 12 decisions | mean travel 102 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 55% 43% 26%>3% 24% 9 510 91%/7%/0%/3% 7% 151/2/0/0 0175A_5mm_v1 480 48% 48% 20%>4% 16% 16 948 90%/5%/1%/3% 5% 131/2/41/100 1447_5mm_v1 192 51% 48% 23%>1% 22% 2 325 90%/7%/0%/3% 2% 85/0/0/7 paris4_afv_central 224 41% 48% 23%>11% 12% 24 1501 83%/5%/4%/7% 12% 107/1/0/0 PHerc0125_3mm_v2 352 54% 46% 24%>0% 24% 0 137 88%/6%/0%/5% 12% 138/5/0/24 gradients: max 57, clipped 0/50 updates Step 17,250 | last 50 updates / 1,600 crops | lr 1.15e-05 points (35/curve, tolerance 2 vox): 48,764 right / 2,449 wrong | 95.2% correct | 4,787 unknown loss 0.8835 | geometry 0.2706 | confidence 1.2259 | mean error 0.649 vox speed: 1363 ms/update | data wait 1 ms/update | 23.3 crops/s | 23.3 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1528 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 177, from the previous prediction 1351 confidence-labeled crops: terminal 12/1534 (0.8%) | recoverable 112/1534 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 384/1600 (24.0%), 2: 128/1600 (8.0%), 3: 352/1600 (22.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 84 chunks / 168.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 739 rows | mean depth 19 decisions | mean travel 156 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 49% 48% 24%>4% 21% 12 605 91%/6%/0%/3% 10% 211/2/36/2 0175A_5mm_v1 384 47% 46% 21%>8% 14% 22 896 89%/7%/1%/3% 7% 176/4/0/0 1447_5mm_v1 128 52% 48% 24%>0% 24% 0 153 92%/5%/1%/2% 2% 62/0/0/0 paris4_afv_central 352 45% 48% 23%>7% 15% 25 1104 84%/7%/2%/7% 5% 168/1/0/0 PHerc0125_3mm_v2 288 57% 43% 27%>0% 27% 0 141 84%/10%/0%/6% 11% 122/7/0/3 gradients: max 60, clipped 0/50 updates Step 17,300 | last 50 updates / 1,600 crops | lr 1.14e-05 points (35/curve, tolerance 2 vox): 48,237 right / 2,477 wrong | 95.1% correct | 5,286 unknown loss 0.8786 | geometry 0.2462 | confidence 1.2647 | mean error 0.626 vox speed: 1363 ms/update | data wait 1 ms/update | 23.2 crops/s | 23.2 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1526 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 252, from the previous prediction 1274 confidence-labeled crops: terminal 17/1525 (1.1%) | recoverable 115/1525 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 288/1600 (18.0%), 2: 416/1600 (26.0%), 3: 224/1600 (14.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 84 chunks / 168.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 664 rows | mean depth 16 decisions | mean travel 129 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 52% 44% 25%>4% 21% 11 1015 89%/7%/0%/4% 9% 78/1/0/64 0175A_5mm_v1 288 37% 50% 26%>14% 12% 19 1242 84%/8%/3%/5% 11% 143/6/0/0 1447_5mm_v1 416 50% 48% 26%>2% 25% 6 314 88%/7%/0%/5% 3% 195/3/105/5 paris4_afv_central 224 46% 47% 23%>6% 17% 14 1053 90%/5%/2%/3% 8% 103/0/25/3 PHerc0125_3mm_v2 352 51% 49% 22%>0% 22% 0 136 84%/8%/0%/7% 14% 145/2/0/27 gradients: max 41, clipped 0/50 updates Step 17,350 | last 50 updates / 1,600 crops | lr 1.13e-05 points (35/curve, tolerance 2 vox): 48,820 right / 2,410 wrong | 95.3% correct | 4,770 unknown loss 0.8081 | geometry 0.2364 | confidence 1.1433 | mean error 0.621 vox speed: 1377 ms/update | data wait 1 ms/update | 23.0 crops/s | 23.0 decisions/s | 1.14 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1520 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 274, from the previous prediction 1246 confidence-labeled crops: terminal 25/1544 (1.6%) | recoverable 105/1544 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 288/1600 (18.0%), 2: 416/1600 (26.0%), 3: 352/1600 (22.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 84 chunks / 168.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 599 rows | mean depth 13 decisions | mean travel 104 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 53% 44% 28%>3% 25% 8 558 88%/9%/0%/4% 7% 112/0/0/0 0175A_5mm_v1 288 48% 48% 25%>5% 20% 11 759 92%/7%/0%/1% 5% 137/3/0/0 1447_5mm_v1 416 53% 44% 24%>3% 21% 14 892 90%/6%/1%/3% 5% 98/2/0/86 paris4_afv_central 352 42% 47% 24%>11% 13% 32 1855 84%/5%/5%/6% 11% 127/2/0/38 PHerc0125_3mm_v2 288 57% 43% 28%>0% 28% 0 147 89%/7%/0%/4% 12% 125/3/0/0 gradients: max 47, clipped 0/50 updates Step 17,400 | last 50 updates / 1,600 crops | lr 1.13e-05 points (35/curve, tolerance 2 vox): 48,030 right / 2,560 wrong | 94.9% correct | 5,410 unknown loss 0.8287 | geometry 0.2461 | confidence 1.1652 | mean error 0.633 vox speed: 1392 ms/update | data wait 1 ms/update | 22.8 crops/s | 22.8 decisions/s | 1.25 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1534 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 323, from the previous prediction 1211 confidence-labeled crops: terminal 15/1519 (1.0%) | recoverable 102/1519 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 224/1600 (14.0%), 2: 384/1600 (24.0%), 3: 256/1600 (16.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 84 chunks / 168.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 574 rows | mean depth 20 decisions | mean travel 166 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 52% 45% 23%>3% 20% 10 594 89%/7%/0%/4% 10% 173/0/0/0 0175A_5mm_v1 224 43% 51% 21%>6% 15% 11 782 87%/7%/3%/4% 9% 114/0/0/0 1447_5mm_v1 384 55% 42% 25%>2% 23% 9 1532 87%/7%/0%/6% 2% 44/2/0/119 paris4_afv_central 256 46% 43% 25%>10% 14% 30 2038 85%/5%/3%/7% 10% 87/1/0/24 PHerc0125_3mm_v2 352 56% 44% 24%>0% 24% 0 143 89%/6%/0%/5% 9% 156/7/0/0 gradients: max 52, clipped 0/50 updates Step 17,450 | last 50 updates / 1,600 crops | lr 1.12e-05 points (35/curve, tolerance 2 vox): 48,539 right / 2,779 wrong | 94.6% correct | 4,682 unknown loss 0.8971 | geometry 0.2447 | confidence 1.3049 | mean error 0.627 vox speed: 1382 ms/update | data wait 1 ms/update | 23.0 crops/s | 23.0 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1533 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 202, from the previous prediction 1331 confidence-labeled crops: terminal 24/1545 (1.6%) | recoverable 116/1545 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 256/1600 (16.0%), 2: 352/1600 (22.0%), 3: 448/1600 (28.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 84 chunks / 168.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 690 rows | mean depth 16 decisions | mean travel 131 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 57% 41% 24%>2% 22% 4 389 89%/8%/0%/3% 10% 105/1/0/0 0175A_5mm_v1 256 48% 45% 22%>7% 16% 16 845 89%/5%/2%/3% 7% 115/1/0/0 1447_5mm_v1 352 55% 44% 22%>1% 21% 4 302 88%/9%/1%/2% 3% 154/0/0/0 paris4_afv_central 448 44% 47% 24%>9% 15% 38 1427 85%/6%/3%/5% 9% 191/1/0/18 PHerc0125_3mm_v2 288 56% 44% 27%>0% 27% 0 143 88%/7%/0%/5% 10% 125/2/27/1 gradients: max 47, clipped 0/50 updates Step 17,458 | dagger_states: 6435 | dagger_source_step: 17000 | dagger_caches: 4 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/0175A_5mm_v1/decisions_017000.npz" | dagger_fibers: 384 | dagger_supply: {"states": 6435, "traces": 384, "fibers": 384, "rows": {"terminal": 145, "premature_stop": 118, "recoverable": 33, "pre_excursion": 156, "ordinary": 5970}, "events": {"terminal": 21, "premature_stop": 118, "recoverable": 8, "pre_excursion": 27, "ordinary": 381}, "supervision": {"following": 6243, "recoverable": 23, "terminal": 145, "unknown": 24}, "departed_traces": 27, "terminal_traces": 21, "stop_reasons": {"after_failure": 17, "confidence": 122, "unannotated": 245}, "traced_voxels": 97723.8340029882, "seconds": 644.4714206084609, "voxels_per_second": 151.6340847367984} | dagger_coverage: {"requested": 384, "seeds": 384, "skipped": [], "epoch": 0} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 0 | dagger_launch_interval: 5000 | dagger_publication_age: 458 | dataset: "0175A_5mm_v1" Step 17,500 | last 50 updates / 1,600 crops | lr 1.11e-05 points (35/curve, tolerance 2 vox): 47,611 right / 2,549 wrong | 94.9% correct | 5,840 unknown loss 0.7982 | geometry 0.2501 | confidence 1.0963 | mean error 0.616 vox speed: 925 ms/update | data wait 1 ms/update | 34.2 crops/s | 34.2 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1542 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 232, from the previous prediction 1310 confidence-labeled crops: terminal 20/1527 (1.3%) | recoverable 103/1527 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 352/1600 (22.0%), 2: 352/1600 (22.0%), 3: 256/1600 (16.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 84 chunks / 168.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 691 rows | mean depth 13 decisions | mean travel 106 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 48% 48% 23%>4% 20% 9 657 91%/5%/0%/4% 8% 114/0/61/8 0175A_5mm_v1 352 51% 49% 25%>0% 25% 0 163 89%/7%/1%/3% 4% 172/3/0/0 1447_5mm_v1 352 52% 47% 26%>1% 25% 4 350 90%/5%/0%/4% 4% 164/3/0/0 paris4_afv_central 256 43% 43% 23%>14% 9% 38 2644 81%/6%/6%/6% 11% 85/3/0/26 PHerc0125_3mm_v2 384 54% 46% 27%>0% 27% 0 137 86%/8%/0%/6% 14% 156/4/2/20 gradients: max 40, clipped 0/50 updates Step 17,550 | last 50 updates / 1,600 crops | lr 1.10e-05 points (35/curve, tolerance 2 vox): 48,898 right / 2,692 wrong | 94.8% correct | 4,410 unknown loss 0.8752 | geometry 0.2450 | confidence 1.2604 | mean error 0.643 vox speed: 892 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1541 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 302, from the previous prediction 1239 confidence-labeled crops: terminal 19/1542 (1.2%) | recoverable 113/1542 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 544/1600 (34.0%), 1: 352/1600 (22.0%), 2: 192/1600 (12.0%), 3: 288/1600 (18.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 84 chunks / 168.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 556 rows | mean depth 21 decisions | mean travel 170 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 544 56% 43% 27%>1% 26% 6 481 91%/7%/0%/2% 9% 158/1/0/77 0175A_5mm_v1 352 59% 39% 26%>2% 24% 8 600 89%/7%/1%/3% 4% 113/0/67/24 1447_5mm_v1 192 51% 45% 28%>4% 23% 6 518 89%/8%/1%/3% 6% 87/3/0/0 paris4_afv_central 288 42% 46% 25%>12% 13% 29 1647 82%/7%/5%/6% 10% 133/3/0/0 PHerc0125_3mm_v2 224 59% 41% 29%>0% 29% 0 149 86%/7%/0%/7% 12% 65/0/0/26 gradients: max 47, clipped 0/50 updates Step 17,600 | last 50 updates / 1,600 crops | lr 1.10e-05 points (35/curve, tolerance 2 vox): 48,145 right / 2,742 wrong | 94.6% correct | 5,113 unknown loss 0.8448 | geometry 0.2191 | confidence 1.2514 | mean error 0.610 vox speed: 892 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1547 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 198, from the previous prediction 1349 confidence-labeled crops: terminal 29/1533 (1.9%) | recoverable 112/1533 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 128/1600 (8.0%), 1: 224/1600 (14.0%), 2: 448/1600 (28.0%), 3: 416/1600 (26.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 85 chunks / 170.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 683 rows | mean depth 15 decisions | mean travel 124 vox | 8 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 128 65% 34% 23%>2% 22% 2 437 90%/9%/0%/1% 11% 43/0/0/0 0175A_5mm_v1 224 49% 46% 25%>5% 20% 9 825 89%/5%/2%/4% 6% 103/3/0/0 1447_5mm_v1 448 51% 47% 26%>2% 24% 10 464 88%/8%/0%/4% 3% 176/1/107/34 paris4_afv_central 416 47% 46% 24%>7% 17% 29 1156 82%/8%/5%/5% 10% 189/3/20/1 PHerc0125_3mm_v2 384 54% 46% 23%>0% 23% 0 147 88%/6%/1%/5% 12% 172/5/0/6 gradients: max 49, clipped 0/50 updates Step 17,650 | last 50 updates / 1,600 crops | lr 1.09e-05 points (35/curve, tolerance 2 vox): 48,122 right / 2,751 wrong | 94.6% correct | 5,127 unknown loss 0.8095 | geometry 0.2494 | confidence 1.1202 | mean error 0.618 vox speed: 893 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1529 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 301, from the previous prediction 1228 confidence-labeled crops: terminal 34/1539 (2.2%) | recoverable 106/1539 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 192/1600 (12.0%), 2: 288/1600 (18.0%), 3: 480/1600 (30.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 85 chunks / 170.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 615 rows | mean depth 17 decisions | mean travel 136 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 49% 50% 23%>1% 22% 5 301 92%/6%/0%/2% 8% 207/3/0/0 0175A_5mm_v1 192 50% 42% 26%>8% 18% 15 1253 87%/7%/3%/4% 6% 80/0/0/0 1447_5mm_v1 288 51% 47% 24%>2% 22% 6 520 91%/5%/0%/4% 2% 127/1/0/8 paris4_afv_central 480 43% 48% 25%>9% 16% 34 2407 82%/8%/6%/5% 11% 101/2/0/131 PHerc0125_3mm_v2 224 55% 45% 25%>0% 25% 0 142 87%/8%/1%/4% 12% 100/6/0/0 gradients: max 66, clipped 2/50 updates Step 17,700 | last 50 updates / 1,600 crops | lr 1.08e-05 points (35/curve, tolerance 2 vox): 47,632 right / 3,183 wrong | 93.7% correct | 5,185 unknown loss 0.9079 | geometry 0.2660 | confidence 1.2837 | mean error 0.656 vox speed: 899 ms/update | data wait 1 ms/update | 35.1 crops/s | 35.1 decisions/s | 1.25 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1502 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 233, from the previous prediction 1269 confidence-labeled crops: terminal 30/1523 (2.0%) | recoverable 112/1523 (7.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 320/1600 (20.0%), 2: 320/1600 (20.0%), 3: 384/1600 (24.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 85 chunks / 170.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 625 rows | mean depth 20 decisions | mean travel 163 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 55% 44% 24%>0% 24% 1 202 91%/7%/0%/2% 11% 113/0/8/0 0175A_5mm_v1 320 46% 40% 28%>14% 14% 23 1610 82%/8%/4%/6% 10% 127/0/4/0 1447_5mm_v1 320 57% 41% 25%>2% 23% 6 609 90%/8%/0%/2% 3% 102/1/0/30 paris4_afv_central 384 47% 41% 27%>12% 15% 40 2049 80%/7%/4%/9% 12% 135/4/0/24 PHerc0125_3mm_v2 320 54% 46% 25%>0% 25% 0 144 90%/6%/0%/4% 10% 148/3/0/0 gradients: max 42, clipped 0/50 updates Step 17,750 | last 50 updates / 1,600 crops | lr 1.07e-05 points (35/curve, tolerance 2 vox): 47,820 right / 2,544 wrong | 94.9% correct | 5,636 unknown loss 0.8670 | geometry 0.2351 | confidence 1.2638 | mean error 0.626 vox speed: 894 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1514 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 219, from the previous prediction 1295 confidence-labeled crops: terminal 23/1514 (1.5%) | recoverable 104/1514 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 384/1600 (24.0%), 2: 224/1600 (14.0%), 3: 352/1600 (22.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 85 chunks / 170.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 638 rows | mean depth 14 decisions | mean travel 114 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 57% 42% 26%>1% 25% 4 309 90%/8%/0%/2% 9% 163/1/0/0 0175A_5mm_v1 384 52% 39% 26%>9% 16% 23 1466 87%/4%/3%/6% 6% 116/1/23/32 1447_5mm_v1 224 46% 47% 20%>7% 13% 9 720 87%/8%/0%/5% 4% 105/2/0/0 paris4_afv_central 352 48% 43% 28%>9% 19% 25 1428 81%/7%/3%/9% 11% 151/0/0/0 PHerc0125_3mm_v2 256 60% 40% 28%>0% 28% 0 144 88%/6%/0%/5% 11% 103/4/0/0 gradients: max 54, clipped 0/50 updates Step 17,800 | last 50 updates / 1,600 crops | lr 1.07e-05 points (35/curve, tolerance 2 vox): 48,542 right / 2,399 wrong | 95.3% correct | 5,059 unknown loss 0.7543 | geometry 0.2074 | confidence 1.0938 | mean error 0.596 vox speed: 892 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.14 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1532 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 241, from the previous prediction 1291 confidence-labeled crops: terminal 24/1526 (1.6%) | recoverable 115/1526 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 352/1600 (22.0%), 2: 288/1600 (18.0%), 3: 512/1600 (32.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 85 chunks / 170.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 667 rows | mean depth 14 decisions | mean travel 113 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 62% 38% 28%>0% 28% 0 151 87%/9%/0%/4% 9% 84/1/0/0 0175A_5mm_v1 352 47% 45% 23%>8% 16% 23 1804 86%/7%/2%/5% 7% 86/3/0/72 1447_5mm_v1 288 49% 47% 23%>3% 20% 9 639 91%/6%/1%/3% 3% 136/2/0/0 paris4_afv_central 512 42% 52% 23%>6% 17% 25 943 84%/8%/3%/5% 7% 266/6/1/2 PHerc0125_3mm_v2 224 58% 42% 20%>0% 20% 0 147 88%/6%/0%/5% 10% 95/3/0/0 gradients: max 45, clipped 0/50 updates Step 17,850 | last 50 updates / 1,600 crops | lr 1.06e-05 points (35/curve, tolerance 2 vox): 48,381 right / 2,921 wrong | 94.3% correct | 4,698 unknown loss 0.9109 | geometry 0.2836 | confidence 1.2546 | mean error 0.671 vox speed: 891 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1533 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 233, from the previous prediction 1300 confidence-labeled crops: terminal 14/1531 (0.9%) | recoverable 100/1531 (6.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 288/1600 (18.0%), 2: 288/1600 (18.0%), 3: 320/1600 (20.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 85 chunks / 170.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 662 rows | mean depth 20 decisions | mean travel 166 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 46% 45% 27%>9% 19% 28 1295 88%/7%/0%/5% 11% 174/0/7/15 0175A_5mm_v1 288 48% 49% 25%>3% 22% 8 584 91%/6%/1%/2% 5% 128/2/0/13 1447_5mm_v1 288 57% 42% 27%>1% 27% 2 293 88%/7%/0%/5% 3% 121/2/0/0 paris4_afv_central 320 47% 47% 24%>6% 18% 19 1251 84%/7%/3%/6% 9% 112/2/0/39 PHerc0125_3mm_v2 288 56% 44% 25%>0% 25% 0 150 90%/5%/0%/5% 12% 127/3/0/0 gradients: max 50, clipped 0/50 updates Step 17,900 | last 50 updates / 1,600 crops | lr 1.05e-05 points (35/curve, tolerance 2 vox): 48,541 right / 2,528 wrong | 95.0% correct | 4,931 unknown loss 0.8759 | geometry 0.2446 | confidence 1.2626 | mean error 0.622 vox speed: 892 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1536 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 208, from the previous prediction 1328 confidence-labeled crops: terminal 20/1543 (1.3%) | recoverable 114/1543 (7.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 352/1600 (22.0%), 2: 288/1600 (18.0%), 3: 320/1600 (20.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 85 chunks / 170.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 1747 chunk references / 1 windows live chains: 683 rows | mean depth 14 decisions | mean travel 111 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 51% 45% 26%>4% 22% 10 694 88%/9%/0%/2% 12% 157/0/0/0 0175A_5mm_v1 352 49% 47% 22%>5% 17% 14 729 89%/5%/2%/3% 4% 154/2/0/10 1447_5mm_v1 288 52% 45% 26%>3% 24% 8 551 90%/6%/1%/3% 2% 96/0/48/33 paris4_afv_central 320 48% 44% 20%>8% 12% 19 1313 86%/6%/3%/6% 9% 140/5/0/0 PHerc0125_3mm_v2 288 53% 47% 28%>0% 28% 0 160 86%/9%/0%/5% 11% 136/1/0/0 gradients: max 54, clipped 0/50 updates Step 17,950 | last 50 updates / 1,600 crops | lr 1.04e-05 points (35/curve, tolerance 2 vox): 47,103 right / 3,250 wrong | 93.5% correct | 5,647 unknown loss 0.8136 | geometry 0.2547 | confidence 1.1178 | mean error 0.636 vox speed: 893 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1501 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 326, from the previous prediction 1175 confidence-labeled crops: terminal 44/1520 (2.9%) | recoverable 106/1520 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 480/1600 (30.0%), 2: 256/1600 (16.0%), 3: 288/1600 (18.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 85 chunks / 170.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 542 rows | mean depth 16 decisions | mean travel 127 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 52% 48% 21%>0% 21% 0 135 93%/5%/0%/2% 5% 86/0/0/53 0175A_5mm_v1 480 43% 46% 26%>11% 15% 27 1768 84%/6%/6%/4% 11% 148/2/0/73 1447_5mm_v1 256 57% 41% 30%>2% 28% 5 452 88%/7%/2%/4% 3% 104/2/0/0 paris4_afv_central 288 45% 42% 26%>12% 14% 30 1887 80%/8%/4%/8% 12% 122/2/0/0 PHerc0125_3mm_v2 288 53% 47% 25%>0% 25% 0 145 84%/8%/0%/8% 14% 82/2/28/54 gradients: max 62, clipped 1/50 updates Step 18,000 | last 50 updates / 1,600 crops | lr 1.04e-05 points (35/curve, tolerance 2 vox): 48,067 right / 3,218 wrong | 93.7% correct | 4,715 unknown loss 0.9003 | geometry 0.2464 | confidence 1.3077 | mean error 0.640 vox speed: 891 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1508 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 204, from the previous prediction 1304 confidence-labeled crops: terminal 36/1545 (2.3%) | recoverable 120/1545 (7.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 320/1600 (20.0%), 2: 224/1600 (14.0%), 3: 384/1600 (24.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 88 chunks / 176.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 680 rows | mean depth 18 decisions | mean travel 148 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 55% 43% 27%>2% 25% 8 423 88%/10%/0%/2% 10% 186/3/0/8 0175A_5mm_v1 320 48% 43% 25%>9% 16% 25 1749 86%/7%/3%/4% 7% 121/3/0/17 1447_5mm_v1 224 51% 44% 27%>5% 21% 10 699 87%/8%/2%/4% 4% 98/0/0/0 paris4_afv_central 384 43% 46% 23%>11% 12% 37 1689 86%/5%/5%/4% 12% 177/2/0/0 PHerc0125_3mm_v2 224 56% 44% 29%>0% 29% 0 138 87%/8%/0%/4% 10% 98/1/0/0 gradients: max 68, clipped 1/50 updates Step 18,000 | dagger_launched: true Step 18,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 240, "stopped": 2, "mean_example_error": 0.5141678314123835, "geometry_examples": 28, "examples": 32, "inference_and_measurement_seconds": 2.241976011544466, "total_diagnostic_seconds": 4.3569180481135845} Step 18,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 6/8 (75.0%) | false stops 0 1.5-3: recovered 7/8 (87.5%) | false stops 0 3-4.5: recovered 6/8 (75.0%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 0 entire proposed 8-point prefix correct (distance): 22/28 (78.6%) gate @ 0.50: false stops 0/23 (0.0%) | accepted wrong 4/26 (15.4%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.617 1.584 following 16 464 1.180 1.212 recoverable 12 348 2.199 2.079 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 1.180 1.212 332 32 812 1.617 1.584 improved/worsened/compared states: 2/2/28 Step 18,000 | batch_diagnostics_seconds: 4.3592722453176975 | recovery_seconds: 14.615048374980688 Step 18,050 | last 50 updates / 1,600 crops | lr 1.03e-05 points (35/curve, tolerance 2 vox): 47,403 right / 3,273 wrong | 93.5% correct | 5,324 unknown loss 0.9452 | geometry 0.2462 | confidence 1.3979 | mean error 0.639 vox speed: 1324 ms/update | data wait 1 ms/update | 18.4 crops/s | 18.4 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1507 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 185, from the previous prediction 1322 confidence-labeled crops: terminal 41/1541 (2.7%) | recoverable 96/1541 (6.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 352/1600 (22.0%), 2: 256/1600 (16.0%), 3: 384/1600 (24.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 88 chunks / 176.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 690 rows | mean depth 17 decisions | mean travel 140 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 55% 42% 28%>4% 24% 11 714 89%/9%/0%/2% 11% 120/4/0/0 0175A_5mm_v1 352 43% 45% 25%>11% 14% 26 1418 87%/6%/5%/2% 9% 160/2/0/0 1447_5mm_v1 256 51% 47% 24%>2% 22% 4 401 89%/6%/1%/4% 3% 121/0/10/0 paris4_afv_central 384 48% 42% 27%>10% 17% 29 1725 85%/5%/5%/6% 12% 151/3/37/11 PHerc0125_3mm_v2 320 57% 43% 27%>0% 27% 0 148 90%/6%/0%/4% 12% 138/2/0/0 gradients: max 49, clipped 0/50 updates Step 18,100 | last 50 updates / 1,600 crops | lr 1.02e-05 points (35/curve, tolerance 2 vox): 48,693 right / 2,904 wrong | 94.4% correct | 4,403 unknown loss 0.8813 | geometry 0.2752 | confidence 1.2120 | mean error 0.659 vox speed: 1363 ms/update | data wait 1 ms/update | 23.3 crops/s | 23.3 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1533 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 274, from the previous prediction 1259 confidence-labeled crops: terminal 20/1544 (1.3%) | recoverable 101/1544 (6.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 224/1600 (14.0%), 2: 320/1600 (20.0%), 3: 224/1600 (14.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 88 chunks / 176.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 607 rows | mean depth 15 decisions | mean travel 121 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 49% 46% 24%>5% 19% 20 849 91%/7%/0%/3% 9% 176/1/0/0 0175A_5mm_v1 224 51% 40% 31%>9% 22% 15 1289 85%/8%/2%/4% 7% 89/2/0/0 1447_5mm_v1 320 48% 50% 21%>2% 18% 7 422 93%/3%/1%/3% 3% 121/3/32/40 paris4_afv_central 224 46% 46% 23%>8% 14% 17 1887 82%/8%/6%/4% 11% 67/0/0/35 PHerc0125_3mm_v2 448 58% 42% 28%>0% 28% 0 139 89%/6%/0%/4% 13% 154/2/0/33 gradients: max 49, clipped 0/50 updates Step 18,150 | last 50 updates / 1,600 crops | lr 1.01e-05 points (35/curve, tolerance 2 vox): 48,832 right / 2,699 wrong | 94.8% correct | 4,469 unknown loss 0.8529 | geometry 0.2414 | confidence 1.2230 | mean error 0.615 vox speed: 1375 ms/update | data wait 1 ms/update | 23.1 crops/s | 23.1 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1534 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 253, from the previous prediction 1281 confidence-labeled crops: terminal 21/1545 (1.4%) | recoverable 102/1545 (6.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 384/1600 (24.0%), 2: 320/1600 (20.0%), 3: 288/1600 (18.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 88 chunks / 176.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 662 rows | mean depth 18 decisions | mean travel 148 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 53% 45% 25%>2% 23% 5 423 92%/6%/0%/2% 9% 143/0/0/0 0175A_5mm_v1 384 47% 48% 25%>5% 20% 15 736 88%/6%/3%/3% 6% 182/4/75/3 1447_5mm_v1 320 46% 49% 24%>4% 20% 9 721 90%/7%/0%/3% 3% 113/2/0/45 paris4_afv_central 288 42% 49% 27%>9% 18% 24 1173 85%/6%/3%/6% 11% 141/2/0/0 PHerc0125_3mm_v2 288 56% 44% 29%>0% 29% 0 145 89%/7%/1%/4% 13% 83/4/0/45 gradients: max 52, clipped 0/50 updates Step 18,200 | last 50 updates / 1,600 crops | lr 1.01e-05 points (35/curve, tolerance 2 vox): 48,739 right / 2,324 wrong | 95.4% correct | 4,937 unknown loss 0.8154 | geometry 0.2472 | confidence 1.1364 | mean error 0.628 vox speed: 1361 ms/update | data wait 1 ms/update | 23.3 crops/s | 23.3 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1533 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 234, from the previous prediction 1299 confidence-labeled crops: terminal 20/1542 (1.3%) | recoverable 120/1542 (7.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 352/1600 (22.0%), 2: 352/1600 (22.0%), 3: 288/1600 (18.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 88 chunks / 176.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 645 rows | mean depth 17 decisions | mean travel 134 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 52% 46% 25%>1% 23% 3 351 91%/6%/0%/4% 7% 104/0/0/0 0175A_5mm_v1 352 47% 45% 25%>7% 18% 19 1042 84%/10%/2%/3% 7% 153/3/4/7 1447_5mm_v1 352 48% 44% 24%>8% 16% 25 1323 89%/6%/1%/3% 5% 118/0/0/38 paris4_afv_central 288 53% 44% 24%>2% 22% 7 577 88%/7%/2%/2% 7% 128/3/0/0 PHerc0125_3mm_v2 384 56% 44% 26%>0% 26% 0 132 87%/8%/0%/5% 11% 142/0/0/28 gradients: max 63, clipped 1/50 updates Step 18,250 | last 50 updates / 1,600 crops | lr 9.99e-06 points (35/curve, tolerance 2 vox): 49,056 right / 2,014 wrong | 96.1% correct | 4,930 unknown loss 0.7844 | geometry 0.2426 | confidence 1.0837 | mean error 0.624 vox speed: 1358 ms/update | data wait 1 ms/update | 23.4 crops/s | 23.4 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1561 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 318, from the previous prediction 1243 confidence-labeled crops: terminal 7/1538 (0.5%) | recoverable 122/1538 (7.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 160/1600 (10.0%), 2: 448/1600 (28.0%), 3: 256/1600 (16.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 88 chunks / 176.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 562 rows | mean depth 14 decisions | mean travel 117 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 52% 41% 30%>7% 23% 18 1170 91%/6%/0%/3% 8% 119/0/0/0 0175A_5mm_v1 160 54% 43% 22%>2% 20% 4 772 89%/9%/0%/2% 3% 47/0/0/22 1447_5mm_v1 448 52% 46% 22%>2% 20% 11 582 88%/7%/1%/4% 3% 134/7/0/71 paris4_afv_central 256 49% 50% 27%>1% 25% 3 347 89%/8%/0%/3% 5% 128/3/0/0 PHerc0125_3mm_v2 448 57% 43% 27%>0% 27% 0 140 85%/9%/1%/6% 12% 134/4/0/58 gradients: max 50, clipped 0/50 updates Step 18,300 | last 50 updates / 1,600 crops | lr 9.92e-06 points (35/curve, tolerance 2 vox): 48,736 right / 2,752 wrong | 94.7% correct | 4,512 unknown loss 0.9145 | geometry 0.2666 | confidence 1.2959 | mean error 0.639 vox speed: 1393 ms/update | data wait 1 ms/update | 22.8 crops/s | 22.8 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1540 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 174, from the previous prediction 1366 confidence-labeled crops: terminal 19/1547 (1.2%) | recoverable 110/1547 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 384/1600 (24.0%), 2: 192/1600 (12.0%), 3: 288/1600 (18.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 92 chunks / 184.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 699 rows | mean depth 20 decisions | mean travel 161 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 53% 44% 25%>3% 22% 8 554 90%/7%/0%/3% 11% 185/0/0/0 0175A_5mm_v1 384 49% 44% 24%>7% 17% 20 1008 89%/5%/1%/4% 7% 168/2/0/0 1447_5mm_v1 192 51% 46% 23%>3% 20% 6 549 88%/7%/3%/2% 5% 89/1/0/0 paris4_afv_central 288 51% 43% 28%>6% 23% 16 1077 85%/9%/3%/3% 7% 124/0/0/0 PHerc0125_3mm_v2 320 57% 43% 29%>0% 29% 0 142 89%/6%/0%/5% 12% 133/1/0/4 gradients: max 50, clipped 0/50 updates Step 18,350 | last 50 updates / 1,600 crops | lr 9.85e-06 points (35/curve, tolerance 2 vox): 48,187 right / 2,543 wrong | 95.0% correct | 5,270 unknown loss 0.7206 | geometry 0.1992 | confidence 1.0428 | mean error 0.588 vox speed: 1365 ms/update | data wait 1 ms/update | 23.3 crops/s | 23.3 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1522 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 179, from the previous prediction 1343 confidence-labeled crops: terminal 30/1526 (2.0%) | recoverable 105/1526 (6.9%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 384/1600 (24.0%), 2: 320/1600 (20.0%), 3: 288/1600 (18.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 92 chunks / 184.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 705 rows | mean depth 16 decisions | mean travel 132 vox | 6 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 53% 46% 21%>1% 20% 3 307 90%/7%/0%/3% 5% 129/2/23/3 0175A_5mm_v1 384 51% 42% 27%>7% 21% 18 1164 87%/7%/3%/4% 7% 160/6/2/1 1447_5mm_v1 320 48% 43% 27%>9% 18% 24 1075 87%/5%/2%/6% 6% 137/1/23/0 paris4_afv_central 288 49% 44% 25%>7% 18% 19 1221 84%/7%/3%/6% 9% 127/4/0/0 PHerc0125_3mm_v2 320 52% 48% 24%>0% 24% 0 137 86%/7%/1%/6% 10% 152/1/16/1 gradients: max 61, clipped 1/50 updates Step 18,400 | last 50 updates / 1,600 crops | lr 9.77e-06 points (35/curve, tolerance 2 vox): 47,984 right / 2,641 wrong | 94.8% correct | 5,375 unknown loss 0.8092 | geometry 0.2616 | confidence 1.0951 | mean error 0.636 vox speed: 1369 ms/update | data wait 1 ms/update | 23.0 crops/s | 23.0 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1518 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 269, from the previous prediction 1249 confidence-labeled crops: terminal 27/1532 (1.8%) | recoverable 105/1532 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 320/1600 (20.0%), 2: 288/1600 (18.0%), 3: 384/1600 (24.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 92 chunks / 184.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 628 rows | mean depth 15 decisions | mean travel 125 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 49% 50% 23%>1% 22% 3 300 92%/6%/0%/2% 7% 134/0/0/9 0175A_5mm_v1 320 46% 46% 22%>9% 13% 27 1822 88%/4%/1%/6% 7% 95/3/0/51 1447_5mm_v1 288 52% 42% 26%>7% 20% 15 1010 86%/8%/2%/3% 4% 120/0/0/0 paris4_afv_central 384 49% 43% 29%>8% 21% 27 1535 81%/10%/4%/5% 10% 147/3/39/20 PHerc0125_3mm_v2 320 53% 47% 20%>0% 20% 0 144 90%/4%/1%/5% 12% 132/6/38/19 gradients: max 46, clipped 0/50 updates Step 18,450 | last 50 updates / 1,600 crops | lr 9.70e-06 points (35/curve, tolerance 2 vox): 49,244 right / 2,379 wrong | 95.4% correct | 4,377 unknown loss 0.8019 | geometry 0.2449 | confidence 1.1140 | mean error 0.613 vox speed: 1376 ms/update | data wait 1 ms/update | 23.1 crops/s | 23.1 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1549 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 217, from the previous prediction 1332 confidence-labeled crops: terminal 19/1547 (1.2%) | recoverable 109/1547 (7.0%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 288/1600 (18.0%), 2: 384/1600 (24.0%), 3: 256/1600 (16.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 93 chunks / 186.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 694 rows | mean depth 21 decisions | mean travel 170 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 55% 45% 27%>0% 27% 0 144 90%/7%/0%/3% 7% 159/3/0/0 0175A_5mm_v1 288 45% 49% 23%>6% 17% 15 915 90%/8%/1%/1% 5% 126/0/0/15 1447_5mm_v1 384 54% 43% 26%>3% 23% 10 614 89%/7%/2%/3% 4% 164/2/32/2 paris4_afv_central 256 46% 45% 23%>9% 14% 23 1299 85%/5%/3%/6% 10% 114/4/0/0 PHerc0125_3mm_v2 320 51% 49% 22%>0% 22% 0 132 89%/6%/1%/4% 12% 131/4/0/27 gradients: max 60, clipped 1/50 updates Step 18,500 | last 50 updates / 1,600 crops | lr 9.63e-06 points (35/curve, tolerance 2 vox): 48,772 right / 2,387 wrong | 95.3% correct | 4,841 unknown loss 0.7493 | geometry 0.2153 | confidence 1.0680 | mean error 0.599 vox speed: 1331 ms/update | data wait 1 ms/update | 23.9 crops/s | 23.9 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1541 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 311, from the previous prediction 1230 confidence-labeled crops: terminal 20/1536 (1.3%) | recoverable 108/1536 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 256/1600 (16.0%), 2: 416/1600 (26.0%), 3: 320/1600 (20.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 93 chunks / 186.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 573 rows | mean depth 16 decisions | mean travel 129 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 56% 40% 31%>4% 27% 8 818 86%/9%/0%/4% 7% 90/0/0/0 0175A_5mm_v1 256 49% 46% 23%>5% 18% 9 785 91%/5%/2%/2% 5% 118/3/0/0 1447_5mm_v1 416 50% 46% 22%>4% 18% 16 979 88%/7%/1%/3% 5% 127/2/78/65 paris4_afv_central 320 52% 44% 23%>4% 19% 13 1133 86%/7%/2%/5% 8% 103/0/0/38 PHerc0125_3mm_v2 384 53% 47% 25%>0% 25% 0 138 88%/6%/1%/6% 10% 135/4/0/45 gradients: max 61, clipped 1/50 updates Step 18,501 | dagger_states: 6791 | dagger_source_step: 18000 | dagger_caches: 4 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/1447_5mm_v1/decisions_018000.npz" | dagger_fibers: 384 | dagger_supply: {"states": 6791, "traces": 384, "fibers": 384, "rows": {"terminal": 151, "premature_stop": 126, "recoverable": 26, "pre_excursion": 160, "ordinary": 6318}, "events": {"terminal": 22, "premature_stop": 126, "recoverable": 9, "pre_excursion": 28, "ordinary": 378}, "supervision": {"following": 6604, "recoverable": 22, "terminal": 151, "unknown": 14}, "departed_traces": 24, "terminal_traces": 22, "stop_reasons": {"after_failure": 17, "confidence": 131, "unannotated": 236}, "traced_voxels": 102786.27247571103, "seconds": 700.9849819801748, "voxels_per_second": 146.63120483031693} | dagger_coverage: {"requested": 384, "seeds": 384, "skipped": [], "epoch": 0} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 0 | dagger_launch_interval: 5000 | dagger_publication_age: 501 | dataset: "1447_5mm_v1" Step 18,550 | last 50 updates / 1,600 crops | lr 9.55e-06 points (35/curve, tolerance 2 vox): 48,838 right / 2,497 wrong | 95.1% correct | 4,665 unknown loss 0.8392 | geometry 0.2613 | confidence 1.1558 | mean error 0.636 vox speed: 890 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1530 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 221, from the previous prediction 1309 confidence-labeled crops: terminal 15/1535 (1.0%) | recoverable 94/1535 (6.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 384/1600 (24.0%), 2: 224/1600 (14.0%), 3: 256/1600 (16.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 93 chunks / 186.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 668 rows | mean depth 16 decisions | mean travel 128 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 55% 43% 26%>2% 23% 7 547 94%/5%/0%/1% 7% 124/2/0/0 0175A_5mm_v1 384 48% 45% 28%>7% 21% 20 962 90%/5%/2%/4% 6% 173/3/0/0 1447_5mm_v1 224 50% 46% 23%>4% 19% 9 760 91%/6%/1%/3% 4% 69/1/0/34 paris4_afv_central 256 37% 52% 21%>11% 9% 27 1382 85%/6%/2%/7% 8% 133/8/0/0 PHerc0125_3mm_v2 448 56% 44% 23%>0% 23% 0 134 86%/8%/0%/6% 11% 169/4/0/26 gradients: max 60, clipped 1/50 updates Step 18,600 | last 50 updates / 1,600 crops | lr 9.48e-06 points (35/curve, tolerance 2 vox): 49,631 right / 2,362 wrong | 95.5% correct | 4,007 unknown loss 0.7891 | geometry 0.2289 | confidence 1.1203 | mean error 0.611 vox speed: 892 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1538 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 220, from the previous prediction 1318 confidence-labeled crops: terminal 19/1550 (1.2%) | recoverable 115/1550 (7.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 352/1600 (22.0%), 2: 288/1600 (18.0%), 3: 320/1600 (20.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 93 chunks / 186.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 661 rows | mean depth 20 decisions | mean travel 165 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 50% 46% 26%>4% 22% 14 697 87%/10%/0%/3% 9% 191/0/0/0 0175A_5mm_v1 352 46% 45% 22%>9% 13% 24 1323 90%/4%/3%/3% 7% 142/2/45/17 1447_5mm_v1 288 57% 42% 28%>1% 27% 2 289 89%/8%/1%/2% 2% 108/1/0/12 paris4_afv_central 320 45% 52% 24%>4% 20% 11 816 88%/7%/2%/3% 6% 138/3/17/27 PHerc0125_3mm_v2 224 60% 40% 27%>0% 27% 0 142 88%/7%/0%/5% 13% 82/2/0/7 gradients: max 60, clipped 0/50 updates Step 18,650 | last 50 updates / 1,600 crops | lr 9.41e-06 points (35/curve, tolerance 2 vox): 48,572 right / 1,997 wrong | 96.1% correct | 5,431 unknown loss 0.7835 | geometry 0.2338 | confidence 1.0993 | mean error 0.595 vox speed: 891 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1567 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 314, from the previous prediction 1253 confidence-labeled crops: terminal 11/1527 (0.7%) | recoverable 98/1527 (6.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 224/1600 (14.0%), 2: 320/1600 (20.0%), 3: 448/1600 (28.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 93 chunks / 186.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 547 rows | mean depth 20 decisions | mean travel 159 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 54% 45% 27%>1% 26% 4 357 90%/7%/0%/3% 6% 133/0/11/41 0175A_5mm_v1 224 48% 48% 21%>4% 17% 8 707 88%/6%/2%/4% 5% 93/3/0/15 1447_5mm_v1 320 63% 36% 33%>1% 32% 4 382 88%/7%/0%/4% 2% 115/1/0/0 paris4_afv_central 448 52% 45% 23%>4% 19% 16 1021 88%/5%/1%/6% 8% 110/3/0/91 PHerc0125_3mm_v2 224 57% 43% 25%>0% 25% 0 164 87%/6%/0%/7% 12% 96/3/0/0 gradients: max 47, clipped 0/50 updates Step 18,700 | last 50 updates / 1,600 crops | lr 9.33e-06 points (35/curve, tolerance 2 vox): 48,866 right / 2,408 wrong | 95.3% correct | 4,726 unknown loss 0.8084 | geometry 0.2389 | confidence 1.1389 | mean error 0.621 vox speed: 889 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1541 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 244, from the previous prediction 1297 confidence-labeled crops: terminal 20/1538 (1.3%) | recoverable 113/1538 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 224/1600 (14.0%), 2: 416/1600 (26.0%), 3: 320/1600 (20.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 93 chunks / 186.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 619 rows | mean depth 15 decisions | mean travel 126 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 51% 47% 24%>2% 22% 5 506 91%/7%/0%/2% 10% 115/0/0/67 0175A_5mm_v1 224 46% 47% 27%>6% 21% 12 913 88%/8%/1%/3% 5% 106/2/0/0 1447_5mm_v1 416 53% 45% 25%>2% 24% 7 379 89%/7%/1%/3% 3% 181/2/64/8 paris4_afv_central 320 54% 36% 31%>10% 21% 25 1780 82%/7%/4%/7% 9% 105/2/0/10 PHerc0125_3mm_v2 256 56% 44% 26%>0% 26% 0 176 87%/7%/0%/5% 10% 112/0/0/0 gradients: max 58, clipped 0/50 updates Step 18,750 | last 50 updates / 1,600 crops | lr 9.26e-06 points (35/curve, tolerance 2 vox): 48,429 right / 2,634 wrong | 94.8% correct | 4,937 unknown loss 0.8653 | geometry 0.2425 | confidence 1.2455 | mean error 0.635 vox speed: 890 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1535 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 187, from the previous prediction 1348 confidence-labeled crops: terminal 23/1535 (1.5%) | recoverable 91/1535 (5.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 352/1600 (22.0%), 2: 224/1600 (14.0%), 3: 352/1600 (22.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 93 chunks / 186.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 714 rows | mean depth 21 decisions | mean travel 173 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 53% 45% 23%>2% 22% 6 429 92%/5%/0%/3% 8% 151/1/0/22 0175A_5mm_v1 352 47% 45% 25%>7% 18% 19 1067 84%/7%/2%/7% 9% 160/1/0/0 1447_5mm_v1 224 50% 50% 17%>0% 17% 0 141 92%/5%/0%/3% 2% 108/0/12/4 paris4_afv_central 352 45% 45% 25%>10% 16% 33 1306 85%/5%/4%/6% 10% 159/3/0/0 PHerc0125_3mm_v2 288 53% 47% 25%>0% 25% 0 159 92%/6%/1%/2% 10% 136/5/0/0 gradients: max 49, clipped 0/50 updates Step 18,800 | last 50 updates / 1,600 crops | lr 9.19e-06 points (35/curve, tolerance 2 vox): 48,224 right / 2,393 wrong | 95.3% correct | 5,383 unknown loss 0.8100 | geometry 0.2361 | confidence 1.1477 | mean error 0.607 vox speed: 893 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1542 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 354, from the previous prediction 1188 confidence-labeled crops: terminal 21/1527 (1.4%) | recoverable 96/1527 (6.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 480/1600 (30.0%), 2: 256/1600 (16.0%), 3: 224/1600 (14.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 93 chunks / 186.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 584 rows | mean depth 16 decisions | mean travel 132 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 46% 51% 21%>3% 18% 5 507 91%/6%/0%/3% 6% 98/0/0/0 0175A_5mm_v1 480 47% 46% 27%>7% 20% 21 1190 86%/6%/2%/5% 7% 182/7/2/39 1447_5mm_v1 256 46% 48% 21%>5% 16% 12 724 88%/7%/2%/2% 6% 112/3/0/12 paris4_afv_central 224 50% 47% 25%>3% 22% 6 613 85%/8%/1%/5% 5% 105/1/0/0 PHerc0125_3mm_v2 448 54% 46% 21%>0% 21% 0 133 90%/4%/0%/6% 10% 87/0/84/119 gradients: max 67, clipped 1/50 updates Step 18,850 | last 50 updates / 1,600 crops | lr 9.12e-06 points (35/curve, tolerance 2 vox): 49,182 right / 2,441 wrong | 95.3% correct | 4,377 unknown loss 0.8384 | geometry 0.2546 | confidence 1.1675 | mean error 0.636 vox speed: 890 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.12 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1539 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 225, from the previous prediction 1314 confidence-labeled crops: terminal 16/1539 (1.0%) | recoverable 101/1539 (6.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 224/1600 (14.0%), 2: 384/1600 (24.0%), 3: 256/1600 (16.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 93 chunks / 186.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 682 rows | mean depth 18 decisions | mean travel 142 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 49% 47% 25%>4% 21% 16 755 93%/5%/0%/2% 9% 211/2/31/1 0175A_5mm_v1 224 48% 48% 22%>4% 17% 6 1231 86%/4%/1%/8% 8% 51/2/0/56 1447_5mm_v1 384 47% 45% 30%>7% 23% 20 934 86%/8%/3%/3% 5% 174/5/0/0 paris4_afv_central 256 51% 47% 23%>2% 21% 4 487 88%/7%/0%/4% 4% 121/0/5/0 PHerc0125_3mm_v2 288 54% 46% 26%>0% 26% 0 136 89%/6%/1%/4% 12% 125/6/0/8 gradients: max 54, clipped 0/50 updates Step 18,900 | last 50 updates / 1,600 crops | lr 9.04e-06 points (35/curve, tolerance 2 vox): 48,553 right / 2,363 wrong | 95.4% correct | 5,084 unknown loss 0.8788 | geometry 0.2728 | confidence 1.2119 | mean error 0.662 vox speed: 891 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1546 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 261, from the previous prediction 1285 confidence-labeled crops: terminal 10/1536 (0.7%) | recoverable 110/1536 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 256/1600 (16.0%), 2: 224/1600 (14.0%), 3: 320/1600 (20.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 93 chunks / 186.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 624 rows | mean depth 21 decisions | mean travel 172 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 52% 46% 24%>2% 22% 8 476 91%/6%/0%/2% 9% 178/1/33/29 0175A_5mm_v1 256 53% 45% 23%>2% 20% 6 489 89%/6%/0%/5% 4% 114/2/0/0 1447_5mm_v1 224 54% 42% 29%>5% 24% 8 746 86%/9%/1%/4% 5% 93/1/0/0 paris4_afv_central 320 47% 44% 27%>9% 18% 28 1429 88%/8%/2%/2% 8% 140/7/0/0 PHerc0125_3mm_v2 352 54% 46% 25%>0% 25% 0 135 87%/6%/0%/7% 11% 99/3/0/64 gradients: max 61, clipped 1/50 updates Step 18,950 | last 50 updates / 1,600 crops | lr 8.97e-06 points (35/curve, tolerance 2 vox): 49,629 right / 2,356 wrong | 95.5% correct | 4,015 unknown loss 0.8008 | geometry 0.2746 | confidence 1.0524 | mean error 0.631 vox speed: 890 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1530 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 316, from the previous prediction 1214 confidence-labeled crops: terminal 14/1551 (0.9%) | recoverable 111/1551 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 384/1600 (24.0%), 2: 256/1600 (16.0%), 3: 320/1600 (20.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 580 rows | mean depth 18 decisions | mean travel 148 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 52% 46% 21%>2% 20% 11 1076 92%/7%/0%/2% 9% 80/0/0/126 0175A_5mm_v1 384 51% 41% 27%>8% 18% 28 1280 89%/6%/2%/3% 7% 158/0/0/0 1447_5mm_v1 256 52% 43% 30%>4% 26% 8 727 86%/9%/1%/4% 4% 111/0/6/0 paris4_afv_central 320 47% 48% 19%>5% 13% 17 905 89%/5%/2%/4% 7% 148/4/95/6 PHerc0125_3mm_v2 192 57% 43% 29%>0% 29% 0 137 86%/9%/0%/4% 10% 83/2/0/0 gradients: max 53, clipped 0/50 updates Step 19,000 | last 50 updates / 1,600 crops | lr 8.90e-06 points (35/curve, tolerance 2 vox): 49,125 right / 2,314 wrong | 95.5% correct | 4,561 unknown loss 0.8624 | geometry 0.2526 | confidence 1.2196 | mean error 0.626 vox speed: 891 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1553 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 258, from the previous prediction 1295 confidence-labeled crops: terminal 17/1544 (1.1%) | recoverable 112/1544 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 384/1600 (24.0%), 2: 256/1600 (16.0%), 3: 416/1600 (26.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 635 rows | mean depth 17 decisions | mean travel 136 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 54% 44% 26%>2% 23% 6 549 94%/4%/0%/2% 7% 113/0/0/0 0175A_5mm_v1 384 55% 39% 27%>5% 21% 15 985 85%/7%/2%/6% 6% 151/3/0/0 1447_5mm_v1 256 44% 54% 24%>2% 21% 5 481 92%/5%/0%/2% 3% 138/2/0/0 paris4_afv_central 416 49% 49% 23%>3% 20% 14 990 87%/9%/1%/4% 6% 98/2/0/104 PHerc0125_3mm_v2 288 53% 47% 25%>0% 25% 0 146 88%/8%/1%/3% 14% 135/3/0/0 gradients: max 43, clipped 0/50 updates Step 19,000 | dagger_launched: true Step 19,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 256, "stopped": 0, "mean_example_error": 0.4035499352030456, "geometry_examples": 32, "examples": 32, "inference_and_measurement_seconds": 1.926613736897707, "total_diagnostic_seconds": 3.722803831100464} Step 19,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 6/8 (75.0%) | false stops 0 1.5-3: recovered 6/8 (75.0%) | false stops 0 3-4.5: recovered 6/8 (75.0%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 0 entire proposed 8-point prefix correct (distance): 22/28 (78.6%) gate @ 0.50: false stops 0/23 (0.0%) | accepted wrong 4/26 (15.4%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.636 1.593 following 16 464 1.172 1.200 recoverable 12 348 2.255 2.117 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 1.172 1.200 332 32 812 1.636 1.593 improved/worsened/compared states: 2/2/28 Step 19,000 | batch_diagnostics_seconds: 3.728037714958191 | recovery_seconds: 16.835036490112543 Step 19,050 | last 50 updates / 1,600 crops | lr 8.83e-06 points (35/curve, tolerance 2 vox): 49,709 right / 2,045 wrong | 96.0% correct | 4,246 unknown loss 0.8059 | geometry 0.2212 | confidence 1.1693 | mean error 0.598 vox speed: 1381 ms/update | data wait 1 ms/update | 17.4 crops/s | 17.4 decisions/s | 1.14 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1572 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 251, from the previous prediction 1321 confidence-labeled crops: terminal 11/1541 (0.7%) | recoverable 115/1541 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 288/1600 (18.0%), 2: 352/1600 (22.0%), 3: 384/1600 (24.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 659 rows | mean depth 21 decisions | mean travel 171 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 50% 49% 23%>1% 22% 3 310 92%/7%/0%/1% 8% 156/1/0/0 0175A_5mm_v1 288 46% 50% 25%>4% 21% 10 881 87%/7%/1%/5% 4% 100/0/37/44 1447_5mm_v1 352 57% 42% 24%>1% 23% 3 302 88%/7%/0%/5% 3% 147/4/0/0 paris4_afv_central 384 52% 46% 26%>2% 23% 9 678 88%/7%/2%/4% 7% 138/4/0/38 PHerc0125_3mm_v2 256 54% 46% 24%>0% 24% 0 142 88%/8%/0%/4% 13% 118/7/0/0 gradients: max 51, clipped 0/50 updates Step 19,100 | last 50 updates / 1,600 crops | lr 8.76e-06 points (35/curve, tolerance 2 vox): 49,350 right / 1,874 wrong | 96.3% correct | 4,776 unknown loss 0.7055 | geometry 0.2317 | confidence 0.9476 | mean error 0.591 vox speed: 1403 ms/update | data wait 1 ms/update | 22.6 crops/s | 22.6 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1542 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 294, from the previous prediction 1248 confidence-labeled crops: terminal 8/1532 (0.5%) | recoverable 120/1532 (7.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 448/1600 (28.0%), 2: 384/1600 (24.0%), 3: 320/1600 (20.0%), 4: 128/1600 (8.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 582 rows | mean depth 17 decisions | mean travel 141 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 57% 42% 23%>1% 22% 3 353 93%/6%/0%/1% 8% 133/0/0/0 0175A_5mm_v1 448 54% 43% 25%>3% 21% 12 639 86%/8%/1%/6% 4% 179/4/0/12 1447_5mm_v1 384 51% 46% 25%>3% 22% 11 743 85%/9%/1%/5% 3% 96/1/9/80 paris4_afv_central 320 51% 44% 26%>5% 21% 17 918 88%/8%/0%/4% 6% 111/1/0/31 PHerc0125_3mm_v2 128 44% 49% 25%>7% 18% 9 643 87%/5%/1%/8% 14% 63/1/0/0 gradients: max 57, clipped 0/50 updates Step 19,150 | last 50 updates / 1,600 crops | lr 8.69e-06 points (35/curve, tolerance 2 vox): 47,570 right / 2,784 wrong | 94.5% correct | 5,646 unknown loss 0.8741 | geometry 0.2578 | confidence 1.2327 | mean error 0.637 vox speed: 1414 ms/update | data wait 1 ms/update | 22.4 crops/s | 22.4 decisions/s | 1.24 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1516 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 313, from the previous prediction 1203 confidence-labeled crops: terminal 24/1524 (1.6%) | recoverable 118/1524 (7.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 160/1600 (10.0%), 1: 320/1600 (20.0%), 2: 384/1600 (24.0%), 3: 288/1600 (18.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 566 rows | mean depth 18 decisions | mean travel 146 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 160 57% 40% 29%>2% 26% 4 631 88%/10%/0%/2% 9% 64/1/0/0 0175A_5mm_v1 320 47% 47% 21%>7% 14% 17 1305 87%/5%/2%/6% 7% 110/2/0/39 1447_5mm_v1 384 44% 46% 27%>11% 16% 28 1931 86%/7%/2%/5% 8% 119/0/0/56 paris4_afv_central 288 55% 43% 26%>2% 23% 7 628 88%/6%/2%/4% 6% 123/2/0/0 PHerc0125_3mm_v2 448 56% 44% 27%>0% 27% 0 136 85%/9%/1%/6% 14% 150/3/128/49 gradients: max 41, clipped 0/50 updates Step 19,200 | last 50 updates / 1,600 crops | lr 8.61e-06 points (35/curve, tolerance 2 vox): 49,073 right / 1,999 wrong | 96.1% correct | 4,928 unknown loss 0.7592 | geometry 0.2173 | confidence 1.0839 | mean error 0.600 vox speed: 1405 ms/update | data wait 1 ms/update | 22.6 crops/s | 22.6 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1560 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 262, from the previous prediction 1298 confidence-labeled crops: terminal 10/1533 (0.7%) | recoverable 100/1533 (6.5%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 288/1600 (18.0%), 2: 384/1600 (24.0%), 3: 256/1600 (16.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 636 rows | mean depth 22 decisions | mean travel 175 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 52% 45% 24%>2% 22% 6 542 92%/6%/0%/2% 6% 145/0/0/0 0175A_5mm_v1 288 53% 43% 30%>4% 26% 10 705 86%/9%/1%/5% 5% 125/0/0/0 1447_5mm_v1 384 47% 49% 23%>4% 19% 12 630 89%/7%/1%/4% 2% 188/0/0/0 paris4_afv_central 256 56% 41% 25%>3% 23% 7 711 89%/4%/2%/4% 8% 106/1/0/0 PHerc0125_3mm_v2 352 55% 45% 23%>0% 23% 0 132 88%/6%/0%/7% 12% 72/1/0/85 gradients: max 53, clipped 0/50 updates Step 19,250 | last 50 updates / 1,600 crops | lr 8.54e-06 points (35/curve, tolerance 2 vox): 49,240 right / 1,981 wrong | 96.1% correct | 4,779 unknown loss 0.7957 | geometry 0.2249 | confidence 1.1415 | mean error 0.606 vox speed: 1421 ms/update | data wait 1 ms/update | 22.3 crops/s | 22.3 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1544 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 191, from the previous prediction 1353 confidence-labeled crops: terminal 13/1536 (0.8%) | recoverable 132/1536 (8.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 480/1600 (30.0%), 1: 416/1600 (26.0%), 2: 192/1600 (12.0%), 3: 416/1600 (26.0%), 4: 96/1600 (6.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 668 rows | mean depth 17 decisions | mean travel 136 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 480 53% 45% 24%>2% 22% 8 468 89%/9%/0%/2% 8% 191/1/128/27 0175A_5mm_v1 416 52% 46% 27%>1% 25% 6 351 87%/8%/0%/5% 3% 192/2/15/0 1447_5mm_v1 192 51% 44% 27%>5% 22% 7 738 88%/7%/1%/5% 3% 85/2/0/0 paris4_afv_central 416 53% 39% 29%>8% 21% 30 1353 83%/8%/3%/5% 9% 159/4/64/4 PHerc0125_3mm_v2 96 57% 43% 27%>0% 27% 0 160 88%/6%/0%/6% 11% 41/1/0/0 gradients: max 70, clipped 1/50 updates Step 19,300 | last 50 updates / 1,600 crops | lr 8.47e-06 points (35/curve, tolerance 2 vox): 49,757 right / 1,814 wrong | 96.5% correct | 4,429 unknown loss 0.7120 | geometry 0.2450 | confidence 0.9339 | mean error 0.599 vox speed: 1414 ms/update | data wait 1 ms/update | 22.4 crops/s | 22.4 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1578 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 239, from the previous prediction 1339 confidence-labeled crops: terminal 8/1552 (0.5%) | recoverable 129/1552 (8.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 384/1600 (24.0%), 2: 384/1600 (24.0%), 3: 288/1600 (18.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 626 rows | mean depth 20 decisions | mean travel 165 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 55% 43% 29%>2% 27% 6 603 92%/8%/0%/1% 8% 98/1/0/26 0175A_5mm_v1 384 54% 45% 24%>1% 23% 5 374 88%/8%/0%/4% 4% 143/2/7/28 1447_5mm_v1 384 56% 44% 25%>0% 25% 0 144 88%/10%/1%/2% 3% 166/0/47/2 paris4_afv_central 288 51% 45% 27%>4% 23% 11 957 86%/7%/2%/5% 7% 112/0/60/17 PHerc0125_3mm_v2 256 58% 42% 27%>0% 27% 0 168 88%/7%/0%/4% 8% 107/1/0/0 gradients: max 66, clipped 1/50 updates Step 19,350 | last 50 updates / 1,600 crops | lr 8.40e-06 points (35/curve, tolerance 2 vox): 48,376 right / 2,717 wrong | 94.7% correct | 4,907 unknown loss 0.8765 | geometry 0.2592 | confidence 1.2346 | mean error 0.647 vox speed: 1407 ms/update | data wait 1 ms/update | 22.6 crops/s | 22.6 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1536 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 207, from the previous prediction 1329 confidence-labeled crops: terminal 21/1533 (1.4%) | recoverable 99/1533 (6.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 128/1600 (8.0%), 2: 320/1600 (20.0%), 3: 416/1600 (26.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 656 rows | mean depth 17 decisions | mean travel 138 vox | 6 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 57% 41% 30%>3% 27% 13 1184 90%/8%/0%/2% 10% 109/0/0/21 0175A_5mm_v1 128 50% 48% 20%>2% 17% 3 487 91%/5%/1%/4% 3% 61/3/0/0 1447_5mm_v1 320 51% 44% 22%>5% 18% 13 767 89%/6%/2%/2% 4% 141/3/0/0 paris4_afv_central 416 50% 44% 26%>6% 20% 23 1251 86%/7%/2%/5% 7% 152/3/0/32 PHerc0125_3mm_v2 416 53% 46% 20%>1% 19% 3 232 88%/4%/1%/7% 11% 193/3/15/0 gradients: max 66, clipped 1/50 updates Step 19,357 | dagger_states: 5586 | dagger_source_step: 19000 | dagger_caches: 4 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/paris4_afv_central/decisions_019000.npz" | dagger_fibers: 384 | dagger_supply: {"states": 5586, "traces": 384, "fibers": 384, "rows": {"terminal": 290, "premature_stop": 185, "recoverable": 28, "pre_excursion": 274, "ordinary": 4806}, "events": {"terminal": 39, "premature_stop": 185, "recoverable": 13, "pre_excursion": 50, "ordinary": 368}, "supervision": {"following": 5261, "recoverable": 28, "terminal": 290, "unknown": 7}, "departed_traces": 48, "terminal_traces": 39, "stop_reasons": {"after_failure": 33, "confidence": 192, "unannotated": 159}, "traced_voxels": 80677.53125464602, "seconds": 519.4632880687714, "voxels_per_second": 155.30939935059507} | dagger_coverage: {"requested": 384, "seeds": 384, "skipped": [], "epoch": 0} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 0 | dagger_launch_interval: 5000 | dagger_publication_age: 357 | dataset: "paris4_afv_central" Step 19,400 | last 50 updates / 1,600 crops | lr 8.33e-06 points (35/curve, tolerance 2 vox): 48,998 right / 1,958 wrong | 96.2% correct | 5,044 unknown loss 0.7947 | geometry 0.2376 | confidence 1.1143 | mean error 0.608 vox speed: 923 ms/update | data wait 1 ms/update | 34.2 crops/s | 34.2 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1547 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 317, from the previous prediction 1230 confidence-labeled crops: terminal 11/1524 (0.7%) | recoverable 124/1524 (8.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 256/1600 (16.0%), 2: 320/1600 (20.0%), 3: 352/1600 (22.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 571 rows | mean depth 17 decisions | mean travel 134 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 58% 42% 21%>0% 21% 0 160 88%/9%/0%/3% 5% 107/2/0/0 0175A_5mm_v1 256 54% 46% 25%>0% 25% 1 236 89%/7%/0%/4% 2% 118/1/0/0 1447_5mm_v1 320 53% 44% 28%>3% 25% 9 650 85%/10%/2%/4% 5% 140/1/0/0 paris4_afv_central 352 42% 49% 22%>9% 13% 34 1569 86%/7%/2%/6% 10% 114/1/0/58 PHerc0125_3mm_v2 416 55% 45% 25%>0% 25% 0 141 87%/6%/0%/6% 10% 92/4/0/97 gradients: max 51, clipped 0/50 updates Step 19,450 | last 50 updates / 1,600 crops | lr 8.26e-06 points (35/curve, tolerance 2 vox): 47,852 right / 2,821 wrong | 94.4% correct | 5,327 unknown loss 0.7821 | geometry 0.2286 | confidence 1.1071 | mean error 0.608 vox speed: 893 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1530 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 236, from the previous prediction 1294 confidence-labeled crops: terminal 29/1537 (1.9%) | recoverable 105/1537 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 416/1600 (26.0%), 2: 352/1600 (22.0%), 3: 224/1600 (14.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 648 rows | mean depth 21 decisions | mean travel 174 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 59% 41% 23%>0% 23% 0 171 92%/6%/0%/2% 8% 78/0/0/0 0175A_5mm_v1 416 52% 41% 26%>7% 19% 19 1241 86%/7%/3%/5% 5% 171/4/0/0 1447_5mm_v1 352 42% 47% 28%>11% 17% 26 1369 83%/8%/5%/4% 8% 165/5/0/0 paris4_afv_central 224 60% 40% 29%>0% 29% 0 143 89%/8%/0%/3% 4% 89/3/0/0 PHerc0125_3mm_v2 416 56% 44% 25%>0% 25% 0 134 90%/5%/0%/5% 12% 145/3/0/39 gradients: max 69, clipped 2/50 updates Step 19,500 | last 50 updates / 1,600 crops | lr 8.19e-06 points (35/curve, tolerance 2 vox): 49,143 right / 2,115 wrong | 95.9% correct | 4,742 unknown loss 0.8217 | geometry 0.2243 | confidence 1.1948 | mean error 0.615 vox speed: 893 ms/update | data wait 1 ms/update | 35.1 crops/s | 35.1 decisions/s | 1.14 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1566 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 263, from the previous prediction 1303 confidence-labeled crops: terminal 9/1527 (0.6%) | recoverable 97/1527 (6.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 288/1600 (18.0%), 2: 288/1600 (18.0%), 3: 256/1600 (16.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 627 rows | mean depth 19 decisions | mean travel 157 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 51% 49% 24%>0% 24% 1 193 94%/4%/0%/2% 6% 218/1/0/0 0175A_5mm_v1 288 57% 42% 25%>1% 24% 3 422 87%/6%/0%/7% 3% 77/1/27/45 1447_5mm_v1 288 51% 46% 23%>3% 20% 9 682 90%/8%/0%/2% 3% 106/4/6/26 paris4_afv_central 256 50% 42% 29%>8% 21% 16 1080 82%/7%/3%/8% 7% 108/2/0/0 PHerc0125_3mm_v2 320 53% 47% 26%>0% 26% 0 141 88%/7%/0%/6% 10% 118/2/0/33 gradients: max 44, clipped 0/50 updates Step 19,550 | last 50 updates / 1,600 crops | lr 8.12e-06 points (35/curve, tolerance 2 vox): 49,092 right / 2,342 wrong | 95.4% correct | 4,566 unknown loss 0.8165 | geometry 0.2414 | confidence 1.1502 | mean error 0.623 vox speed: 891 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.14 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1556 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 287, from the previous prediction 1269 confidence-labeled crops: terminal 15/1537 (1.0%) | recoverable 106/1537 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 384/1600 (24.0%), 2: 128/1600 (8.0%), 3: 320/1600 (20.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 600 rows | mean depth 17 decisions | mean travel 137 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 59% 40% 27%>0% 26% 1 233 90%/7%/0%/3% 8% 83/0/73/98 0175A_5mm_v1 384 54% 43% 24%>3% 21% 8 593 87%/8%/1%/4% 5% 167/2/0/0 1447_5mm_v1 128 45% 41% 28%>13% 15% 14 1695 88%/7%/3%/2% 8% 53/1/0/0 paris4_afv_central 320 46% 49% 26%>5% 22% 13 987 86%/6%/2%/6% 6% 154/1/48/3 PHerc0125_3mm_v2 320 55% 45% 24%>0% 24% 0 152 91%/5%/0%/4% 12% 143/2/0/0 gradients: max 47, clipped 0/50 updates Step 19,600 | last 50 updates / 1,600 crops | lr 8.05e-06 points (35/curve, tolerance 2 vox): 47,824 right / 2,513 wrong | 95.0% correct | 5,663 unknown loss 0.8665 | geometry 0.2702 | confidence 1.1925 | mean error 0.638 vox speed: 891 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1565 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 287, from the previous prediction 1278 confidence-labeled crops: terminal 16/1522 (1.1%) | recoverable 106/1522 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 320/1600 (20.0%), 2: 192/1600 (12.0%), 3: 448/1600 (28.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 602 rows | mean depth 20 decisions | mean travel 165 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 54% 45% 23%>1% 22% 3 393 91%/7%/0%/2% 10% 97/0/0/47 0175A_5mm_v1 320 55% 42% 27%>2% 24% 7 566 88%/7%/1%/4% 5% 133/3/0/2 1447_5mm_v1 192 51% 42% 23%>7% 16% 9 1125 81%/8%/5%/6% 7% 80/5/0/0 paris4_afv_central 448 54% 44% 25%>2% 23% 10 524 85%/7%/1%/7% 5% 134/4/4/62 PHerc0125_3mm_v2 320 50% 50% 22%>0% 22% 0 138 90%/5%/0%/5% 12% 158/7/60/3 gradients: max 54, clipped 0/50 updates Step 19,650 | last 50 updates / 1,600 crops | lr 7.98e-06 points (35/curve, tolerance 2 vox): 48,586 right / 2,297 wrong | 95.5% correct | 5,117 unknown loss 0.7747 | geometry 0.2284 | confidence 1.0926 | mean error 0.601 vox speed: 892 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.13 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1536 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 326, from the previous prediction 1210 confidence-labeled crops: terminal 19/1539 (1.2%) | recoverable 97/1539 (6.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 576/1600 (36.0%), 2: 288/1600 (18.0%), 3: 288/1600 (18.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 590 rows | mean depth 17 decisions | mean travel 139 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 52% 48% 23%>1% 22% 1 237 93%/4%/0%/3% 6% 92/0/0/0 0175A_5mm_v1 576 51% 47% 25%>2% 23% 9 625 89%/6%/1%/5% 5% 136/0/0/137 1447_5mm_v1 288 56% 40% 27%>5% 22% 12 865 88%/8%/1%/4% 4% 115/3/41/0 paris4_afv_central 288 48% 47% 24%>6% 18% 17 812 89%/6%/1%/4% 7% 121/1/0/13 PHerc0125_3mm_v2 256 38% 54% 22%>7% 14% 24 862 87%/5%/3%/4% 16% 126/2/0/13 gradients: max 48, clipped 0/50 updates Step 19,700 | last 50 updates / 1,600 crops | lr 7.91e-06 points (35/curve, tolerance 2 vox): 47,723 right / 3,085 wrong | 93.9% correct | 5,192 unknown loss 0.9215 | geometry 0.2772 | confidence 1.2886 | mean error 0.674 vox speed: 892 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1524 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 185, from the previous prediction 1339 confidence-labeled crops: terminal 26/1522 (1.7%) | recoverable 102/1522 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 544/1600 (34.0%), 1: 256/1600 (16.0%), 2: 160/1600 (10.0%), 3: 384/1600 (24.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 719 rows | mean depth 16 decisions | mean travel 132 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 544 49% 47% 27%>3% 24% 13 753 90%/7%/0%/3% 9% 256/2/0/0 0175A_5mm_v1 256 53% 43% 26%>4% 22% 8 830 84%/6%/3%/7% 6% 109/0/0/0 1447_5mm_v1 160 44% 46% 23%>9% 14% 11 1318 85%/6%/3%/6% 9% 69/2/1/5 paris4_afv_central 384 51% 42% 24%>7% 17% 25 1055 86%/5%/3%/5% 10% 162/5/0/0 PHerc0125_3mm_v2 256 50% 48% 23%>2% 21% 5 315 86%/7%/0%/6% 14% 123/3/0/0 gradients: max 52, clipped 0/50 updates Step 19,750 | last 50 updates / 1,600 crops | lr 7.84e-06 points (35/curve, tolerance 2 vox): 49,116 right / 2,115 wrong | 95.9% correct | 4,769 unknown loss 0.6950 | geometry 0.2173 | confidence 0.9554 | mean error 0.593 vox speed: 893 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1554 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 243, from the previous prediction 1311 confidence-labeled crops: terminal 15/1537 (1.0%) | recoverable 107/1537 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 320/1600 (20.0%), 2: 384/1600 (24.0%), 3: 384/1600 (24.0%), 4: 160/1600 (10.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 646 rows | mean depth 20 decisions | mean travel 159 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 51% 47% 25%>2% 23% 5 294 91%/6%/1%/2% 9% 164/0/4/0 0175A_5mm_v1 320 49% 45% 26%>6% 20% 17 1068 87%/8%/2%/3% 6% 143/4/0/0 1447_5mm_v1 384 52% 46% 25%>3% 22% 9 716 90%/7%/0%/3% 2% 105/1/0/71 paris4_afv_central 384 55% 43% 26%>2% 24% 9 508 86%/7%/1%/6% 6% 160/1/0/5 PHerc0125_3mm_v2 160 54% 46% 26%>0% 26% 0 160 87%/6%/1%/7% 9% 74/1/0/0 gradients: max 55, clipped 0/50 updates Step 19,800 | last 50 updates / 1,600 crops | lr 7.78e-06 points (35/curve, tolerance 2 vox): 47,611 right / 2,791 wrong | 94.5% correct | 5,598 unknown loss 0.8233 | geometry 0.2389 | confidence 1.1688 | mean error 0.629 vox speed: 889 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1520 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 223, from the previous prediction 1297 confidence-labeled crops: terminal 26/1526 (1.7%) | recoverable 103/1526 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 352/1600 (22.0%), 2: 384/1600 (24.0%), 3: 288/1600 (18.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 648 rows | mean depth 16 decisions | mean travel 132 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 60% 40% 24%>0% 24% 0 155 91%/6%/0%/3% 7% 61/0/0/15 0175A_5mm_v1 352 54% 44% 24%>2% 22% 7 471 90%/6%/1%/3% 3% 156/2/0/0 1447_5mm_v1 384 44% 47% 27%>10% 18% 23 1835 86%/5%/4%/4% 6% 125/3/0/54 paris4_afv_central 288 50% 46% 24%>4% 20% 11 715 86%/7%/2%/6% 7% 124/3/0/9 PHerc0125_3mm_v2 384 48% 47% 27%>5% 22% 17 598 84%/8%/1%/7% 17% 182/5/0/0 gradients: max 36, clipped 0/50 updates Step 19,850 | last 50 updates / 1,600 crops | lr 7.71e-06 points (35/curve, tolerance 2 vox): 47,774 right / 2,857 wrong | 94.4% correct | 5,369 unknown loss 0.8984 | geometry 0.2692 | confidence 1.2584 | mean error 0.648 vox speed: 891 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1526 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 317, from the previous prediction 1209 confidence-labeled crops: terminal 23/1523 (1.5%) | recoverable 119/1523 (7.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 224/1600 (14.0%), 2: 416/1600 (26.0%), 3: 320/1600 (20.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 568 rows | mean depth 16 decisions | mean travel 136 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 54% 45% 27%>0% 26% 1 229 88%/11%/0%/1% 6% 116/0/0/0 0175A_5mm_v1 224 53% 46% 29%>1% 28% 1 356 89%/7%/1%/3% 5% 102/2/0/0 1447_5mm_v1 416 46% 47% 24%>8% 17% 23 2269 84%/9%/2%/5% 7% 92/2/0/102 paris4_afv_central 320 49% 45% 22%>5% 17% 16 861 87%/7%/2%/4% 10% 145/3/0/0 PHerc0125_3mm_v2 384 53% 45% 25%>2% 23% 10 521 85%/4%/1%/9% 11% 113/3/1/61 gradients: max 64, clipped 2/50 updates Step 19,900 | last 50 updates / 1,600 crops | lr 7.64e-06 points (35/curve, tolerance 2 vox): 48,410 right / 2,549 wrong | 95.0% correct | 5,041 unknown loss 0.9170 | geometry 0.2616 | confidence 1.3108 | mean error 0.651 vox speed: 894 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1546 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 364, from the previous prediction 1182 confidence-labeled crops: terminal 13/1532 (0.8%) | recoverable 120/1532 (7.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 448/1600 (28.0%), 2: 224/1600 (14.0%), 3: 192/1600 (12.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 524 rows | mean depth 21 decisions | mean travel 172 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 57% 43% 26%>0% 26% 0 181 89%/9%/0%/2% 9% 123/0/4/0 0175A_5mm_v1 448 51% 46% 28%>3% 25% 11 784 88%/7%/1%/4% 5% 145/1/128/62 1447_5mm_v1 224 56% 42% 27%>2% 25% 5 525 91%/7%/0%/2% 2% 93/3/0/0 paris4_afv_central 192 51% 45% 28%>4% 24% 7 566 88%/7%/2%/3% 7% 87/1/0/0 PHerc0125_3mm_v2 448 51% 44% 27%>5% 22% 29 1444 83%/7%/1%/9% 13% 76/4/0/121 gradients: max 58, clipped 0/50 updates Step 19,950 | last 50 updates / 1,600 crops | lr 7.57e-06 points (35/curve, tolerance 2 vox): 48,736 right / 2,676 wrong | 94.8% correct | 4,588 unknown loss 0.8740 | geometry 0.2700 | confidence 1.2079 | mean error 0.656 vox speed: 892 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.14 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1527 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 231, from the previous prediction 1296 confidence-labeled crops: terminal 18/1536 (1.2%) | recoverable 98/1536 (6.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 512/1600 (32.0%), 1: 128/1600 (8.0%), 2: 352/1600 (22.0%), 3: 352/1600 (22.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 621 rows | mean depth 18 decisions | mean travel 143 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 512 52% 46% 24%>2% 22% 11 1039 91%/5%/0%/3% 9% 169/1/0/68 0175A_5mm_v1 128 56% 39% 26%>5% 21% 5 1150 88%/6%/0%/6% 2% 41/1/0/9 1447_5mm_v1 352 48% 45% 26%>7% 18% 19 1073 89%/6%/2%/3% 4% 158/4/30/1 paris4_afv_central 352 53% 40% 25%>7% 18% 22 1040 86%/8%/2%/4% 8% 142/5/0/0 PHerc0125_3mm_v2 256 57% 43% 24%>0% 24% 0 146 86%/6%/1%/7% 13% 111/6/0/0 gradients: max 59, clipped 0/50 updates Step 20,000 | last 50 updates / 1,600 crops | lr 7.50e-06 points (35/curve, tolerance 2 vox): 48,617 right / 2,392 wrong | 95.3% correct | 4,991 unknown loss 0.8049 | geometry 0.2588 | confidence 1.0923 | mean error 0.622 vox speed: 891 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1512 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 189, from the previous prediction 1323 confidence-labeled crops: terminal 14/1537 (0.9%) | recoverable 121/1537 (7.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 352/1600 (22.0%), 2: 384/1600 (24.0%), 3: 352/1600 (22.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 668 rows | mean depth 15 decisions | mean travel 126 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 62% 38% 30%>0% 30% 0 135 89%/9%/0%/3% 7% 80/0/0/17 0175A_5mm_v1 352 53% 47% 20%>1% 20% 2 264 89%/7%/0%/4% 3% 164/3/0/0 1447_5mm_v1 384 47% 45% 27%>8% 18% 24 1188 88%/9%/1%/2% 7% 174/2/1/0 paris4_afv_central 352 49% 42% 29%>9% 20% 26 1184 87%/7%/2%/5% 8% 149/5/0/0 PHerc0125_3mm_v2 256 51% 39% 25%>10% 16% 24 1114 84%/7%/2%/7% 16% 101/3/0/0 gradients: max 47, clipped 0/50 updates Step 20,000 | dagger_launched: true Step 20,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 256, "stopped": 0, "mean_example_error": 0.5306433506550328, "geometry_examples": 31, "examples": 32, "inference_and_measurement_seconds": 1.9948398433625698, "total_diagnostic_seconds": 3.9225452691316605} Step 20,000 | monitor rollout @ 0.50 coverage 75.4% | precision 82.4% | diverged 21.9% Step 20,000 | monitor rollout @ 0.50 coverage 92.9% | precision 100.0% | diverged 0.0% Step 20,000 | monitor rollout @ 0.50 coverage 95.3% | precision 99.8% | diverged 3.1% Step 20,000 | monitor rollout @ 0.50 coverage 92.0% | precision 97.0% | diverged 3.1% Step 20,000 | monitor rollout @ 0.50 coverage 82.4% | precision 73.6% | diverged 18.8% Step 20,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 6/8 (75.0%) | false stops 0 1.5-3: recovered 7/8 (87.5%) | false stops 0 3-4.5: recovered 6/8 (75.0%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 0 entire proposed 8-point prefix correct (distance): 22/28 (78.6%) gate @ 0.50: false stops 0/23 (0.0%) | accepted wrong 4/26 (15.4%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.643 1.605 following 16 464 1.156 1.182 recoverable 12 348 2.292 2.169 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 1.156 1.182 332 32 812 1.643 1.605 improved/worsened/compared states: 1/1/28 Step 20,000 | batch_diagnostics_seconds: 3.934725422412157 | diagnostics_seconds: 718.0496779046953 | recovery_seconds: 19.2936856970191 Step 20,001 | dagger_states: 4923 | dagger_source_step: 20000 | dagger_caches: 3 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/PHerc0125_3mm_v2/decisions_020000.npz" | dagger_fibers: 384 | dagger_supply: {"states": 4923, "traces": 384, "fibers": 384, "rows": {"terminal": 633, "premature_stop": 124, "recoverable": 76, "pre_excursion": 602, "ordinary": 3457}, "events": {"terminal": 89, "premature_stop": 124, "recoverable": 34, "pre_excursion": 114, "ordinary": 351}, "supervision": {"following": 4178, "recoverable": 63, "terminal": 633, "unknown": 49}, "departed_traces": 93, "terminal_traces": 89, "stop_reasons": {"after_failure": 73, "confidence": 138, "unannotated": 173}, "traced_voxels": 65260.634052693866, "seconds": 335.7459265962243, "voxels_per_second": 194.37505828976978} | dagger_coverage: {"requested": 384, "seeds": 384, "skipped": [], "epoch": 0} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 0 | dagger_launch_interval: 5000 | dagger_publication_age: 1 | dataset: "PHerc0125_3mm_v2" Step 20,050 | last 50 updates / 1,600 crops | lr 7.43e-06 points (35/curve, tolerance 2 vox): 48,187 right / 2,875 wrong | 94.4% correct | 4,938 unknown loss 0.8895 | geometry 0.2755 | confidence 1.2280 | mean error 0.665 vox speed: 892 ms/update | data wait 1 ms/update | 2.0 crops/s | 2.0 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1520 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 305, from the previous prediction 1215 confidence-labeled crops: terminal 18/1531 (1.2%) | recoverable 98/1531 (6.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 160/1600 (10.0%), 2: 480/1600 (30.0%), 3: 288/1600 (18.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 1902 chunk references / 1 windows live chains: 582 rows | mean depth 23 decisions | mean travel 184 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 52% 48% 21%>0% 21% 0 139 90%/8%/0%/2% 10% 169/2/0/0 0175A_5mm_v1 160 57% 42% 23%>0% 23% 0 157 92%/5%/0%/3% 3% 68/2/0/0 1447_5mm_v1 480 50% 44% 26%>6% 20% 25 1951 87%/6%/1%/6% 4% 112/4/0/99 paris4_afv_central 288 45% 51% 21%>4% 17% 14 1097 88%/6%/2%/4% 8% 94/1/14/53 PHerc0125_3mm_v2 320 50% 43% 23%>7% 16% 22 821 86%/4%/2%/7% 14% 139/3/0/0 gradients: max 46, clipped 0/50 updates Step 20,100 | last 50 updates / 1,600 crops | lr 7.37e-06 points (35/curve, tolerance 2 vox): 47,895 right / 3,175 wrong | 93.8% correct | 4,930 unknown loss 0.8453 | geometry 0.2509 | confidence 1.1889 | mean error 0.610 vox speed: 897 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1503 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 263, from the previous prediction 1240 confidence-labeled crops: terminal 31/1541 (2.0%) | recoverable 101/1541 (6.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 224/1600 (14.0%), 2: 512/1600 (32.0%), 3: 256/1600 (16.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 657 rows | mean depth 16 decisions | mean travel 133 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 51% 47% 25%>2% 23% 5 497 91%/8%/0%/1% 8% 121/0/0/0 0175A_5mm_v1 224 52% 48% 24%>0% 24% 0 177 92%/7%/0%/1% 3% 107/4/1/0 1447_5mm_v1 512 42% 51% 23%>8% 15% 28 1493 88%/4%/3%/5% 6% 187/1/0/72 paris4_afv_central 256 51% 42% 27%>7% 20% 15 1002 85%/9%/1%/5% 9% 107/2/0/0 PHerc0125_3mm_v2 352 42% 50% 26%>8% 18% 30 1082 85%/6%/3%/6% 19% 135/6/120/42 gradients: max 52, clipped 0/50 updates Step 20,150 | last 50 updates / 1,600 crops | lr 7.30e-06 points (35/curve, tolerance 2 vox): 48,168 right / 2,786 wrong | 94.5% correct | 5,046 unknown loss 0.8448 | geometry 0.2628 | confidence 1.1642 | mean error 0.645 vox speed: 896 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1505 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 287, from the previous prediction 1218 confidence-labeled crops: terminal 22/1517 (1.5%) | recoverable 109/1517 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 256/1600 (16.0%), 2: 192/1600 (12.0%), 3: 544/1600 (34.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 603 rows | mean depth 18 decisions | mean travel 152 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 47% 51% 19%>3% 17% 9 429 90%/6%/0%/4% 9% 168/1/20/26 0175A_5mm_v1 256 51% 45% 25%>4% 20% 11 857 84%/11%/2%/4% 4% 114/1/0/0 1447_5mm_v1 192 51% 42% 28%>8% 20% 12 1245 83%/6%/3%/7% 5% 80/0/0/0 paris4_afv_central 544 48% 48% 21%>4% 17% 27 1088 89%/6%/1%/5% 6% 185/6/0/75 PHerc0125_3mm_v2 224 49% 44% 28%>7% 21% 22 1638 80%/7%/3%/10% 16% 56/2/0/43 gradients: max 52, clipped 0/50 updates Step 20,200 | last 50 updates / 1,600 crops | lr 7.23e-06 points (35/curve, tolerance 2 vox): 48,346 right / 2,878 wrong | 94.4% correct | 4,776 unknown loss 0.9284 | geometry 0.2958 | confidence 1.2653 | mean error 0.655 vox speed: 907 ms/update | data wait 1 ms/update | 34.6 crops/s | 34.6 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1521 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 295, from the previous prediction 1226 confidence-labeled crops: terminal 16/1540 (1.0%) | recoverable 118/1540 (7.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 256/1600 (16.0%), 2: 224/1600 (14.0%), 3: 544/1600 (34.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 586 rows | mean depth 17 decisions | mean travel 142 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 50% 50% 21%>0% 21% 0 141 90%/9%/0%/1% 10% 109/0/0/3 0175A_5mm_v1 256 56% 43% 27%>1% 26% 3 382 89%/7%/0%/4% 3% 110/1/0/0 1447_5mm_v1 224 53% 42% 24%>4% 19% 9 830 91%/7%/1%/1% 4% 95/2/0/0 paris4_afv_central 544 47% 48% 23%>5% 17% 30 1520 85%/7%/1%/7% 9% 121/3/0/141 PHerc0125_3mm_v2 352 49% 43% 29%>9% 20% 27 976 86%/7%/2%/5% 15% 151/6/0/0 gradients: max 74, clipped 1/50 updates Step 20,250 | last 50 updates / 1,600 crops | lr 7.16e-06 points (35/curve, tolerance 2 vox): 48,582 right / 2,834 wrong | 94.5% correct | 4,584 unknown loss 0.9273 | geometry 0.2701 | confidence 1.3143 | mean error 0.647 vox speed: 900 ms/update | data wait 1 ms/update | 34.8 crops/s | 34.8 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1547 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 292, from the previous prediction 1255 confidence-labeled crops: terminal 17/1546 (1.1%) | recoverable 92/1546 (6.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 512/1600 (32.0%), 1: 288/1600 (18.0%), 2: 224/1600 (14.0%), 3: 288/1600 (18.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 591 rows | mean depth 20 decisions | mean travel 165 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 512 58% 42% 27%>0% 27% 0 141 91%/6%/0%/3% 8% 136/1/0/77 0175A_5mm_v1 288 55% 41% 30%>5% 26% 12 964 89%/7%/2%/2% 5% 117/1/0/0 1447_5mm_v1 224 52% 44% 25%>4% 21% 6 772 88%/5%/3%/4% 5% 98/2/33/1 paris4_afv_central 288 52% 43% 27%>5% 22% 14 1013 88%/6%/1%/5% 8% 101/0/0/22 PHerc0125_3mm_v2 288 45% 50% 21%>5% 16% 12 610 89%/5%/1%/6% 13% 139/6/0/4 gradients: max 58, clipped 0/50 updates Step 20,300 | last 50 updates / 1,600 crops | lr 7.10e-06 points (35/curve, tolerance 2 vox): 48,941 right / 2,925 wrong | 94.4% correct | 4,134 unknown loss 0.8639 | geometry 0.2577 | confidence 1.2124 | mean error 0.639 vox speed: 901 ms/update | data wait 1 ms/update | 35.0 crops/s | 35.0 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1510 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 247, from the previous prediction 1263 confidence-labeled crops: terminal 25/1551 (1.6%) | recoverable 125/1551 (8.1%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 256/1600 (16.0%), 2: 352/1600 (22.0%), 3: 288/1600 (18.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 658 rows | mean depth 16 decisions | mean travel 134 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 47% 53% 21%>0% 21% 1 187 91%/7%/0%/2% 7% 205/2/0/14 0175A_5mm_v1 256 52% 46% 21%>2% 20% 4 409 93%/5%/0%/2% 2% 119/5/0/0 1447_5mm_v1 352 50% 44% 29%>6% 23% 11 967 82%/11%/3%/4% 6% 156/5/0/0 paris4_afv_central 288 51% 41% 28%>8% 20% 21 1186 85%/10%/2%/3% 8% 118/1/0/0 PHerc0125_3mm_v2 288 45% 44% 24%>10% 14% 35 2219 86%/5%/3%/6% 21% 60/1/0/68 gradients: max 63, clipped 1/50 updates Step 20,350 | last 50 updates / 1,600 crops | lr 7.03e-06 points (35/curve, tolerance 2 vox): 48,956 right / 2,847 wrong | 94.5% correct | 4,197 unknown loss 0.8663 | geometry 0.2509 | confidence 1.2308 | mean error 0.622 vox speed: 902 ms/update | data wait 1 ms/update | 34.9 crops/s | 34.9 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1504 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 308, from the previous prediction 1196 confidence-labeled crops: terminal 25/1548 (1.6%) | recoverable 107/1548 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 288/1600 (18.0%), 2: 416/1600 (26.0%), 3: 288/1600 (18.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 611 rows | mean depth 21 decisions | mean travel 168 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 50% 45% 28%>5% 23% 18 866 88%/10%/0%/3% 12% 120/2/0/40 0175A_5mm_v1 288 55% 45% 26%>0% 26% 0 142 91%/7%/0%/2% 2% 124/1/49/5 1447_5mm_v1 416 46% 49% 25%>6% 19% 20 1608 88%/6%/2%/4% 5% 96/0/128/106 paris4_afv_central 288 42% 50% 24%>9% 15% 23 1076 87%/7%/3%/3% 11% 143/3/0/0 PHerc0125_3mm_v2 256 40% 50% 21%>10% 11% 26 1029 90%/3%/2%/4% 11% 128/2/0/0 gradients: max 65, clipped 2/50 updates Step 20,400 | last 50 updates / 1,600 crops | lr 6.96e-06 points (35/curve, tolerance 2 vox): 48,497 right / 2,626 wrong | 94.9% correct | 4,877 unknown loss 0.8357 | geometry 0.2631 | confidence 1.1453 | mean error 0.647 vox speed: 899 ms/update | data wait 1 ms/update | 34.9 crops/s | 34.9 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1535 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 216, from the previous prediction 1319 confidence-labeled crops: terminal 18/1534 (1.2%) | recoverable 118/1534 (7.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 320/1600 (20.0%), 2: 320/1600 (20.0%), 3: 224/1600 (14.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 675 rows | mean depth 19 decisions | mean travel 152 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 47% 52% 24%>1% 23% 5 363 90%/7%/0%/3% 7% 206/2/0/25 0175A_5mm_v1 320 55% 45% 24%>0% 24% 1 206 88%/8%/0%/5% 4% 116/1/20/27 1447_5mm_v1 320 50% 44% 29%>6% 23% 16 1326 85%/8%/3%/4% 5% 120/2/0/22 paris4_afv_central 224 43% 47% 20%>10% 10% 19 1224 85%/8%/3%/4% 10% 106/4/0/0 PHerc0125_3mm_v2 288 50% 44% 26%>6% 20% 16 730 87%/6%/1%/6% 13% 127/3/0/0 gradients: max 64, clipped 1/50 updates Step 20,450 | last 50 updates / 1,600 crops | lr 6.90e-06 points (35/curve, tolerance 2 vox): 47,052 right / 3,607 wrong | 92.9% correct | 5,341 unknown loss 1.0118 | geometry 0.2824 | confidence 1.4587 | mean error 0.663 vox speed: 905 ms/update | data wait 1 ms/update | 34.8 crops/s | 34.8 decisions/s | 1.24 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1472 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 210, from the previous prediction 1262 confidence-labeled crops: terminal 35/1529 (2.3%) | recoverable 114/1529 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 128/1600 (8.0%), 2: 352/1600 (22.0%), 3: 320/1600 (20.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 645 rows | mean depth 16 decisions | mean travel 131 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 49% 46% 24%>5% 20% 16 668 89%/7%/1%/3% 10% 155/1/0/23 0175A_5mm_v1 128 56% 41% 23%>2% 21% 3 597 93%/6%/0%/1% 3% 53/1/0/0 1447_5mm_v1 352 45% 42% 26%>13% 14% 29 2128 85%/6%/4%/5% 8% 136/1/0/11 paris4_afv_central 320 46% 45% 28%>9% 19% 25 1245 84%/9%/2%/5% 9% 141/1/32/2 PHerc0125_3mm_v2 416 51% 43% 27%>6% 21% 25 885 83%/7%/2%/7% 16% 160/2/0/18 gradients: max 51, clipped 0/50 updates Step 20,500 | last 50 updates / 1,600 crops | lr 6.83e-06 points (35/curve, tolerance 2 vox): 48,212 right / 3,481 wrong | 93.3% correct | 4,307 unknown loss 0.9773 | geometry 0.3146 | confidence 1.3254 | mean error 0.709 vox speed: 899 ms/update | data wait 2 ms/update | 34.7 crops/s | 34.7 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1494 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 248, from the previous prediction 1246 confidence-labeled crops: terminal 21/1541 (1.4%) | recoverable 122/1541 (7.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 384/1600 (24.0%), 2: 192/1600 (12.0%), 3: 352/1600 (22.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 644 rows | mean depth 21 decisions | mean travel 173 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 50% 48% 22%>2% 20% 5 320 91%/7%/0%/2% 10% 122/1/0/0 0175A_5mm_v1 384 55% 43% 27%>2% 25% 6 286 88%/7%/1%/4% 4% 165/6/0/0 1447_5mm_v1 192 45% 51% 27%>4% 23% 7 691 87%/8%/2%/3% 4% 98/1/0/0 paris4_afv_central 352 42% 46% 26%>12% 14% 42 1753 82%/12%/2%/5% 13% 140/3/96/22 PHerc0125_3mm_v2 416 45% 48% 25%>6% 19% 37 1426 89%/5%/2%/5% 15% 119/1/0/82 gradients: max 44, clipped 0/50 updates Step 20,550 | last 50 updates / 1,600 crops | lr 6.77e-06 points (35/curve, tolerance 2 vox): 46,425 right / 3,697 wrong | 92.6% correct | 5,878 unknown loss 0.9303 | geometry 0.3105 | confidence 1.2397 | mean error 0.704 vox speed: 902 ms/update | data wait 1 ms/update | 34.6 crops/s | 34.6 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1462 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 311, from the previous prediction 1151 confidence-labeled crops: terminal 29/1515 (1.9%) | recoverable 121/1515 (8.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 224/1600 (14.0%), 2: 256/1600 (16.0%), 3: 576/1600 (36.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 558 rows | mean depth 19 decisions | mean travel 156 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 54% 45% 25%>1% 24% 2 240 90%/8%/0%/2% 9% 144/0/0/0 0175A_5mm_v1 224 53% 46% 29%>2% 27% 2 461 84%/10%/1%/4% 5% 102/3/0/0 1447_5mm_v1 256 49% 44% 25%>7% 18% 13 1153 87%/7%/3%/3% 4% 113/2/0/0 paris4_afv_central 576 44% 44% 28%>11% 17% 59 3051 83%/7%/3%/7% 13% 103/3/0/153 PHerc0125_3mm_v2 224 43% 43% 27%>14% 13% 28 1458 80%/7%/1%/12% 19% 96/2/0/0 gradients: max 52, clipped 0/50 updates Step 20,600 | last 50 updates / 1,600 crops | lr 6.70e-06 points (35/curve, tolerance 2 vox): 48,687 right / 2,572 wrong | 95.0% correct | 4,741 unknown loss 0.7899 | geometry 0.2422 | confidence 1.0955 | mean error 0.604 vox speed: 894 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1520 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 280, from the previous prediction 1240 confidence-labeled crops: terminal 22/1533 (1.4%) | recoverable 124/1533 (8.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 448/1600 (28.0%), 2: 384/1600 (24.0%), 3: 256/1600 (16.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 580 rows | mean depth 16 decisions | mean travel 129 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 60% 39% 25%>0% 24% 1 288 89%/7%/0%/4% 8% 88/0/0/0 0175A_5mm_v1 448 54% 42% 27%>4% 23% 18 1034 86%/8%/0%/5% 5% 107/0/25/79 1447_5mm_v1 384 50% 41% 29%>9% 20% 25 1531 84%/8%/3%/4% 7% 148/3/82/11 paris4_afv_central 256 50% 45% 22%>5% 16% 14 967 92%/5%/1%/2% 5% 114/2/0/0 PHerc0125_3mm_v2 288 53% 43% 29%>4% 25% 9 596 82%/9%/2%/8% 18% 123/2/0/0 gradients: max 48, clipped 0/50 updates Step 20,650 | last 50 updates / 1,600 crops | lr 6.64e-06 points (35/curve, tolerance 2 vox): 48,589 right / 2,856 wrong | 94.4% correct | 4,555 unknown loss 0.8550 | geometry 0.2617 | confidence 1.1865 | mean error 0.639 vox speed: 892 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1519 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 233, from the previous prediction 1286 confidence-labeled crops: terminal 20/1536 (1.3%) | recoverable 92/1536 (6.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 320/1600 (20.0%), 2: 320/1600 (20.0%), 3: 288/1600 (18.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 658 rows | mean depth 23 decisions | mean travel 184 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 45% 48% 24%>6% 17% 15 678 90%/6%/0%/3% 10% 139/0/7/0 0175A_5mm_v1 320 57% 42% 24%>1% 23% 4 392 92%/4%/0%/4% 3% 129/0/0/5 1447_5mm_v1 320 53% 42% 26%>5% 21% 15 1318 90%/5%/2%/3% 4% 96/1/17/40 paris4_afv_central 288 41% 50% 24%>9% 15% 26 1371 85%/8%/3%/4% 10% 133/2/1/10 PHerc0125_3mm_v2 384 55% 42% 26%>3% 23% 9 468 86%/6%/1%/7% 12% 161/0/28/1 gradients: max 69, clipped 3/50 updates Step 20,700 | last 50 updates / 1,600 crops | lr 6.57e-06 points (35/curve, tolerance 2 vox): 47,572 right / 3,450 wrong | 93.2% correct | 4,978 unknown loss 0.9406 | geometry 0.3165 | confidence 1.2483 | mean error 0.678 vox speed: 894 ms/update | data wait 1 ms/update | 35.1 crops/s | 35.1 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1470 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 349, from the previous prediction 1121 confidence-labeled crops: terminal 21/1539 (1.4%) | recoverable 126/1539 (8.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 128/1600 (8.0%), 2: 576/1600 (36.0%), 3: 224/1600 (14.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 532 rows | mean depth 17 decisions | mean travel 141 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 52% 44% 28%>4% 24% 13 540 88%/8%/1%/3% 8% 156/0/0/0 0175A_5mm_v1 128 44% 49% 17%>7% 10% 9 979 92%/3%/2%/3% 7% 59/2/0/4 1447_5mm_v1 576 51% 44% 24%>5% 19% 25 1651 86%/8%/2%/5% 5% 127/1/0/125 paris4_afv_central 224 40% 44% 26%>17% 9% 26 1935 84%/11%/0%/4% 15% 98/1/0/0 PHerc0125_3mm_v2 320 44% 47% 23%>9% 13% 35 1761 85%/8%/2%/5% 18% 92/1/38/58 gradients: max 65, clipped 1/50 updates Step 20,750 | last 50 updates / 1,600 crops | lr 6.51e-06 points (35/curve, tolerance 2 vox): 49,572 right / 1,965 wrong | 96.2% correct | 4,463 unknown loss 0.6959 | geometry 0.2181 | confidence 0.9555 | mean error 0.578 vox speed: 896 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1539 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 230, from the previous prediction 1309 confidence-labeled crops: terminal 15/1548 (1.0%) | recoverable 93/1548 (6.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 352/1600 (22.0%), 2: 416/1600 (26.0%), 3: 224/1600 (14.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 672 rows | mean depth 19 decisions | mean travel 159 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 55% 45% 24%>0% 24% 0 145 89%/8%/0%/2% 7% 144/0/0/0 0175A_5mm_v1 352 49% 49% 23%>2% 21% 7 332 90%/6%/1%/3% 3% 171/2/0/0 1447_5mm_v1 416 56% 39% 26%>5% 21% 18 1087 92%/4%/1%/3% 4% 133/0/0/30 paris4_afv_central 224 48% 45% 22%>8% 15% 16 1111 90%/6%/2%/1% 9% 100/0/0/0 PHerc0125_3mm_v2 288 47% 48% 22%>5% 17% 17 813 87%/5%/1%/7% 12% 124/6/0/14 gradients: max 54, clipped 0/50 updates Step 20,800 | last 50 updates / 1,600 crops | lr 6.44e-06 points (35/curve, tolerance 2 vox): 47,943 right / 3,353 wrong | 93.5% correct | 4,704 unknown loss 0.8862 | geometry 0.2595 | confidence 1.2534 | mean error 0.634 vox speed: 893 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1490 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 206, from the previous prediction 1284 confidence-labeled crops: terminal 33/1542 (2.1%) | recoverable 119/1542 (7.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 384/1600 (24.0%), 2: 448/1600 (28.0%), 3: 160/1600 (10.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 648 rows | mean depth 19 decisions | mean travel 154 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 53% 43% 26%>4% 22% 12 677 90%/7%/0%/3% 9% 134/0/92/16 0175A_5mm_v1 384 53% 41% 26%>6% 20% 17 954 85%/9%/3%/3% 8% 157/6/22/0 1447_5mm_v1 448 51% 43% 27%>6% 21% 17 1051 88%/7%/3%/3% 5% 192/1/0/0 paris4_afv_central 160 43% 43% 21%>14% 8% 22 1766 83%/9%/3%/5% 14% 69/0/0/0 PHerc0125_3mm_v2 256 46% 43% 27%>10% 16% 25 1339 84%/7%/2%/8% 15% 96/1/0/15 gradients: max 56, clipped 0/50 updates Step 20,850 | last 50 updates / 1,600 crops | lr 6.38e-06 points (35/curve, tolerance 2 vox): 48,677 right / 2,409 wrong | 95.3% correct | 4,914 unknown loss 0.8229 | geometry 0.2703 | confidence 1.1052 | mean error 0.626 vox speed: 893 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1543 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 328, from the previous prediction 1215 confidence-labeled crops: terminal 19/1540 (1.2%) | recoverable 106/1540 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 448/1600 (28.0%), 2: 320/1600 (20.0%), 3: 224/1600 (14.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 559 rows | mean depth 16 decisions | mean travel 134 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 56% 44% 22%>0% 22% 0 137 93%/5%/0%/2% 7% 116/1/0/12 0175A_5mm_v1 448 51% 46% 24%>3% 21% 10 736 87%/6%/1%/5% 6% 79/0/0/129 1447_5mm_v1 320 52% 43% 24%>5% 19% 14 935 88%/8%/2%/2% 5% 138/0/0/0 paris4_afv_central 224 56% 41% 25%>3% 21% 7 656 90%/6%/1%/3% 7% 91/3/0/0 PHerc0125_3mm_v2 320 51% 42% 24%>7% 17% 20 950 85%/8%/2%/6% 15% 135/4/1/0 gradients: max 67, clipped 1/50 updates Step 20,900 | last 50 updates / 1,600 crops | lr 6.31e-06 points (35/curve, tolerance 2 vox): 49,659 right / 2,260 wrong | 95.6% correct | 4,081 unknown loss 0.7190 | geometry 0.2192 | confidence 0.9995 | mean error 0.588 vox speed: 894 ms/update | data wait 1 ms/update | 35.0 crops/s | 35.0 decisions/s | 1.13 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1535 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 188, from the previous prediction 1347 confidence-labeled crops: terminal 21/1547 (1.4%) | recoverable 111/1547 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 352/1600 (22.0%), 2: 416/1600 (26.0%), 3: 192/1600 (12.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 683 rows | mean depth 21 decisions | mean travel 168 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 56% 42% 25%>2% 23% 5 414 90%/7%/1%/3% 8% 135/1/0/0 0175A_5mm_v1 352 57% 42% 25%>1% 24% 4 380 90%/7%/1%/3% 4% 142/4/0/6 1447_5mm_v1 416 50% 44% 25%>6% 18% 19 1108 86%/8%/2%/4% 5% 183/0/37/0 paris4_afv_central 192 44% 49% 22%>7% 16% 13 995 89%/5%/2%/4% 5% 94/2/0/0 PHerc0125_3mm_v2 320 52% 44% 26%>5% 21% 15 741 86%/7%/2%/6% 12% 129/2/34/11 gradients: max 58, clipped 0/50 updates Step 20,950 | last 50 updates / 1,600 crops | lr 6.25e-06 points (35/curve, tolerance 2 vox): 47,708 right / 3,503 wrong | 93.2% correct | 4,789 unknown loss 1.0124 | geometry 0.3092 | confidence 1.4064 | mean error 0.697 vox speed: 893 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1486 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 242, from the previous prediction 1244 confidence-labeled crops: terminal 25/1543 (1.6%) | recoverable 102/1543 (6.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 320/1600 (20.0%), 2: 192/1600 (12.0%), 3: 352/1600 (22.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 980 chunk references / 1 windows live chains: 652 rows | mean depth 19 decisions | mean travel 158 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 46% 50% 25%>4% 20% 22 768 90%/7%/0%/3% 11% 189/1/0/33 0175A_5mm_v1 320 55% 44% 27%>1% 26% 3 396 88%/8%/1%/3% 4% 141/1/0/0 1447_5mm_v1 192 54% 45% 19%>1% 18% 2 314 91%/5%/1%/3% 2% 87/0/0/0 paris4_afv_central 352 41% 49% 20%>10% 10% 33 1445 89%/5%/3%/3% 9% 155/2/70/18 PHerc0125_3mm_v2 288 44% 43% 28%>12% 16% 42 2276 83%/6%/3%/8% 20% 80/6/11/44 gradients: max 39, clipped 0/50 updates Step 21,000 | last 50 updates / 1,600 crops | lr 6.18e-06 points (35/curve, tolerance 2 vox): 47,277 right / 2,953 wrong | 94.1% correct | 5,770 unknown loss 0.9249 | geometry 0.2936 | confidence 1.2625 | mean error 0.661 vox speed: 894 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1486 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 283, from the previous prediction 1203 confidence-labeled crops: terminal 16/1525 (1.0%) | recoverable 110/1525 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 192/1600 (12.0%), 2: 384/1600 (24.0%), 3: 384/1600 (24.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 607 rows | mean depth 15 decisions | mean travel 126 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 45% 45% 25%>9% 16% 31 1674 87%/8%/1%/5% 14% 104/0/0/70 0175A_5mm_v1 192 46% 53% 23%>1% 22% 2 317 92%/6%/0%/3% 4% 102/3/0/0 1447_5mm_v1 384 49% 49% 27%>3% 25% 9 551 88%/8%/0%/4% 3% 187/2/0/0 paris4_afv_central 384 45% 48% 24%>7% 17% 33 2141 84%/8%/3%/5% 10% 99/1/69/85 PHerc0125_3mm_v2 256 44% 50% 22%>5% 16% 14 773 87%/4%/1%/7% 14% 115/2/0/14 gradients: max 39, clipped 0/50 updates Step 21,000 | dagger_launched: true Step 21,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 235, "stopped": 1, "mean_example_error": 0.4972187466919422, "geometry_examples": 28, "examples": 32, "inference_and_measurement_seconds": 2.6089258305728436, "total_diagnostic_seconds": 5.32966797798872} Step 21,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 6/8 (75.0%) | false stops 1 1.5-3: recovered 7/8 (87.5%) | false stops 0 3-4.5: recovered 6/8 (75.0%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 0 entire proposed 8-point prefix correct (distance): 22/28 (78.6%) gate @ 0.50: false stops 0/23 (0.0%) | accepted wrong 5/27 (18.5%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.645 1.618 following 16 464 1.151 1.173 recoverable 12 348 2.304 2.210 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 1.151 1.173 332 32 812 1.645 1.618 improved/worsened/compared states: 1/2/28 Step 21,000 | batch_diagnostics_seconds: 5.333457548171282 | recovery_seconds: 14.761819809675217 Step 21,050 | last 50 updates / 1,600 crops | lr 6.12e-06 points (35/curve, tolerance 2 vox): 46,473 right / 3,726 wrong | 92.6% correct | 5,801 unknown loss 0.9501 | geometry 0.3117 | confidence 1.2769 | mean error 0.679 vox speed: 943 ms/update | data wait 1 ms/update | 23.1 crops/s | 23.1 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1436 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 291, from the previous prediction 1145 confidence-labeled crops: terminal 37/1521 (2.4%) | recoverable 111/1521 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 256/1600 (16.0%), 2: 352/1600 (22.0%), 3: 384/1600 (24.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 544 rows | mean depth 19 decisions | mean travel 154 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 57% 42% 23%>0% 23% 1 263 91%/6%/0%/3% 10% 95/0/0/0 0175A_5mm_v1 256 54% 39% 26%>7% 20% 16 1013 88%/7%/2%/3% 6% 101/1/0/0 1447_5mm_v1 352 56% 40% 24%>4% 20% 12 863 89%/7%/2%/3% 4% 141/2/0/0 paris4_afv_central 384 42% 49% 25%>9% 16% 42 3501 80%/9%/3%/9% 14% 63/1/0/127 PHerc0125_3mm_v2 384 38% 44% 28%>17% 10% 58 2101 82%/5%/4%/9% 18% 144/6/0/26 gradients: max 80, clipped 1/50 updates Step 21,100 | last 50 updates / 1,600 crops | lr 6.06e-06 points (35/curve, tolerance 2 vox): 47,214 right / 2,762 wrong | 94.5% correct | 6,024 unknown loss 0.8613 | geometry 0.2572 | confidence 1.2084 | mean error 0.635 vox speed: 1353 ms/update | data wait 1 ms/update | 23.4 crops/s | 23.4 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1512 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 220, from the previous prediction 1292 confidence-labeled crops: terminal 18/1527 (1.2%) | recoverable 100/1527 (6.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 320/1600 (20.0%), 2: 352/1600 (22.0%), 3: 256/1600 (16.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 653 rows | mean depth 21 decisions | mean travel 170 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 52% 48% 22%>0% 22% 0 180 90%/8%/0%/2% 9% 123/0/0/0 0175A_5mm_v1 320 56% 43% 26%>1% 25% 3 340 89%/5%/1%/5% 5% 137/1/0/0 1447_5mm_v1 352 53% 44% 25%>3% 22% 9 581 90%/6%/1%/4% 3% 155/4/0/0 paris4_afv_central 256 50% 40% 30%>9% 20% 21 1448 84%/7%/1%/8% 7% 103/3/0/0 PHerc0125_3mm_v2 416 42% 47% 25%>11% 14% 42 1796 84%/6%/3%/7% 18% 135/3/0/62 gradients: max 60, clipped 1/50 updates Step 21,150 | last 50 updates / 1,600 crops | lr 5.99e-06 points (35/curve, tolerance 2 vox): 48,837 right / 2,551 wrong | 95.0% correct | 4,612 unknown loss 0.7751 | geometry 0.2422 | confidence 1.0657 | mean error 0.622 vox speed: 1377 ms/update | data wait 1 ms/update | 23.0 crops/s | 23.0 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1509 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 195, from the previous prediction 1314 confidence-labeled crops: terminal 20/1537 (1.3%) | recoverable 111/1537 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 416/1600 (26.0%), 2: 288/1600 (18.0%), 3: 288/1600 (18.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 658 rows | mean depth 16 decisions | mean travel 134 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 52% 43% 29%>6% 23% 13 738 87%/9%/0%/4% 11% 123/2/0/0 0175A_5mm_v1 416 55% 44% 25%>1% 24% 4 341 90%/6%/1%/2% 3% 182/4/48/3 1447_5mm_v1 288 50% 43% 27%>7% 20% 13 1253 86%/6%/2%/5% 4% 124/2/0/0 paris4_afv_central 288 50% 43% 23%>7% 16% 18 1169 87%/6%/2%/5% 7% 124/4/0/0 PHerc0125_3mm_v2 320 46% 47% 24%>8% 16% 28 1462 86%/7%/1%/6% 14% 105/2/0/44 gradients: max 45, clipped 0/50 updates Step 21,200 | last 50 updates / 1,600 crops | lr 5.93e-06 points (35/curve, tolerance 2 vox): 47,528 right / 3,110 wrong | 93.9% correct | 5,362 unknown loss 0.9187 | geometry 0.2730 | confidence 1.2914 | mean error 0.662 vox speed: 1389 ms/update | data wait 1 ms/update | 22.9 crops/s | 22.9 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1492 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 291, from the previous prediction 1201 confidence-labeled crops: terminal 24/1532 (1.6%) | recoverable 103/1532 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 384/1600 (24.0%), 2: 160/1600 (10.0%), 3: 416/1600 (26.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 630 rows | mean depth 18 decisions | mean travel 147 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 47% 48% 22%>5% 18% 11 739 90%/7%/0%/4% 9% 96/0/94/27 0175A_5mm_v1 384 52% 47% 23%>1% 22% 3 372 89%/7%/1%/3% 5% 110/3/32/70 1447_5mm_v1 160 48% 51% 22%>1% 21% 1 257 88%/8%/1%/4% 2% 82/0/0/0 paris4_afv_central 416 43% 48% 24%>9% 14% 37 1539 87%/6%/3%/4% 11% 183/4/47/16 PHerc0125_3mm_v2 384 45% 44% 27%>11% 16% 40 1494 84%/6%/3%/7% 15% 159/7/0/10 gradients: max 57, clipped 0/50 updates Step 21,250 | last 50 updates / 1,600 crops | lr 5.87e-06 points (35/curve, tolerance 2 vox): 46,307 right / 3,759 wrong | 92.5% correct | 5,934 unknown loss 0.9815 | geometry 0.3028 | confidence 1.3575 | mean error 0.677 vox speed: 1388 ms/update | data wait 1 ms/update | 22.9 crops/s | 22.9 decisions/s | 1.25 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1481 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 248, from the previous prediction 1233 confidence-labeled crops: terminal 38/1525 (2.5%) | recoverable 92/1525 (6.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 192/1600 (12.0%), 2: 352/1600 (22.0%), 3: 256/1600 (16.0%), 4: 480/1600 (30.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 659 rows | mean depth 18 decisions | mean travel 151 vox | 11 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 51% 43% 23%>6% 17% 14 922 88%/6%/0%/5% 10% 116/1/0/21 0175A_5mm_v1 192 56% 40% 31%>5% 27% 7 1111 85%/5%/5%/5% 9% 76/0/0/0 1447_5mm_v1 352 50% 49% 24%>1% 23% 5 419 90%/7%/1%/2% 2% 172/2/0/0 paris4_afv_central 256 40% 48% 21%>11% 10% 28 1841 86%/6%/4%/5% 12% 98/1/42/26 PHerc0125_3mm_v2 480 44% 47% 22%>9% 12% 48 1441 85%/5%/4%/7% 18% 197/4/0/29 gradients: max 55, clipped 0/50 updates Step 21,300 | last 50 updates / 1,600 crops | lr 5.81e-06 points (35/curve, tolerance 2 vox): 48,466 right / 2,586 wrong | 94.9% correct | 4,948 unknown loss 0.8262 | geometry 0.2467 | confidence 1.1589 | mean error 0.622 vox speed: 1403 ms/update | data wait 1 ms/update | 22.6 crops/s | 22.6 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1515 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 229, from the previous prediction 1286 confidence-labeled crops: terminal 20/1544 (1.3%) | recoverable 106/1544 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 288/1600 (18.0%), 2: 320/1600 (20.0%), 3: 320/1600 (20.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 673 rows | mean depth 17 decisions | mean travel 139 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 54% 45% 22%>1% 21% 2 231 90%/8%/0%/2% 7% 116/0/0/0 0175A_5mm_v1 288 54% 46% 24%>0% 24% 0 156 87%/9%/0%/4% 3% 133/4/0/0 1447_5mm_v1 320 51% 46% 23%>3% 20% 10 672 90%/8%/1%/2% 3% 146/1/0/0 paris4_afv_central 320 48% 43% 24%>8% 15% 24 1393 88%/4%/2%/5% 8% 131/2/0/7 PHerc0125_3mm_v2 416 45% 45% 23%>11% 12% 40 1576 88%/5%/2%/5% 16% 147/0/0/39 gradients: max 44, clipped 0/50 updates Step 21,345 | dagger_states: 3738 | dagger_source_step: 21000 | dagger_caches: 4 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/decisions_021000.npz" | dagger_fibers: 128 | dagger_supply: {"states": 3738, "traces": 128, "fibers": 128, "rows": {"terminal": 12, "premature_stop": 105, "recoverable": 47, "pre_excursion": 100, "ordinary": 3471}, "events": {"terminal": 3, "premature_stop": 105, "recoverable": 17, "pre_excursion": 19, "ordinary": 123}, "supervision": {"following": 3674, "recoverable": 41, "terminal": 12, "unknown": 11}, "departed_traces": 18, "terminal_traces": 3, "stop_reasons": {"after_failure": 2, "confidence": 107, "unannotated": 19}, "traced_voxels": 56786.49606187242, "seconds": 469.1714581735432, "voxels_per_second": 121.03570043015596} | dagger_coverage: {"requested": 128, "seeds": 128, "skipped": [], "epoch": 1} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 0 | dagger_launch_interval: 5000 | dagger_publication_age: 345 | dataset: "paris4" Step 21,350 | last 50 updates / 1,600 crops | lr 5.75e-06 points (35/curve, tolerance 2 vox): 47,281 right / 3,516 wrong | 93.1% correct | 5,203 unknown loss 1.0160 | geometry 0.3368 | confidence 1.3584 | mean error 0.715 vox speed: 1248 ms/update | data wait 1 ms/update | 25.4 crops/s | 25.4 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1476 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 259, from the previous prediction 1217 confidence-labeled crops: terminal 21/1542 (1.4%) | recoverable 135/1542 (8.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 288/1600 (18.0%), 2: 320/1600 (20.0%), 3: 256/1600 (16.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 605 rows | mean depth 18 decisions | mean travel 147 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 49% 48% 23%>3% 20% 12 744 88%/9%/0%/3% 10% 124/1/0/44 0175A_5mm_v1 288 51% 47% 28%>3% 25% 7 681 86%/10%/1%/2% 4% 134/3/0/0 1447_5mm_v1 320 50% 48% 23%>2% 21% 5 469 88%/9%/1%/3% 3% 142/3/25/12 paris4_afv_central 256 39% 45% 26%>17% 9% 38 2114 86%/9%/4%/2% 13% 114/1/0/0 PHerc0125_3mm_v2 384 47% 40% 26%>13% 13% 44 2414 83%/6%/2%/9% 20% 91/0/0/63 gradients: max 47, clipped 0/50 updates Step 21,400 | last 50 updates / 1,600 crops | lr 5.68e-06 points (35/curve, tolerance 2 vox): 48,732 right / 2,788 wrong | 94.6% correct | 4,480 unknown loss 0.7613 | geometry 0.2239 | confidence 1.0748 | mean error 0.606 vox speed: 893 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1524 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 252, from the previous prediction 1272 confidence-labeled crops: terminal 20/1546 (1.3%) | recoverable 106/1546 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 352/1600 (22.0%), 2: 416/1600 (26.0%), 3: 320/1600 (20.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 650 rows | mean depth 19 decisions | mean travel 157 vox | 7 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 50% 46% 24%>4% 21% 9 565 88%/9%/0%/3% 10% 114/1/0/4 0175A_5mm_v1 352 52% 47% 26%>1% 24% 5 572 91%/6%/0%/3% 4% 102/2/71/62 1447_5mm_v1 416 51% 47% 25%>2% 23% 6 464 90%/7%/0%/3% 2% 169/5/0/26 paris4_afv_central 320 46% 46% 26%>8% 17% 24 1272 88%/6%/2%/4% 7% 146/2/15/0 PHerc0125_3mm_v2 256 43% 46% 23%>11% 12% 25 1348 82%/5%/5%/8% 19% 119/4/0/0 gradients: max 41, clipped 0/50 updates Step 21,450 | last 50 updates / 1,600 crops | lr 5.62e-06 points (35/curve, tolerance 2 vox): 46,179 right / 4,060 wrong | 91.9% correct | 5,761 unknown loss 0.9360 | geometry 0.3273 | confidence 1.2174 | mean error 0.695 vox speed: 891 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1427 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 218, from the previous prediction 1209 confidence-labeled crops: terminal 39/1526 (2.6%) | recoverable 99/1526 (6.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 224/1600 (14.0%), 2: 320/1600 (20.0%), 3: 352/1600 (22.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 640 rows | mean depth 18 decisions | mean travel 143 vox | 10 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 43% 43% 25%>13% 12% 31 1464 88%/6%/2%/4% 13% 144/0/0/9 0175A_5mm_v1 224 53% 44% 23%>3% 20% 5 780 92%/3%/2%/3% 5% 94/2/0/5 1447_5mm_v1 320 51% 43% 28%>6% 22% 18 1323 88%/7%/1%/4% 4% 114/3/0/23 paris4_afv_central 352 40% 40% 27%>20% 7% 52 2884 82%/8%/4%/7% 14% 124/4/0/18 PHerc0125_3mm_v2 352 47% 47% 21%>7% 14% 22 959 86%/6%/3%/5% 15% 164/5/0/0 gradients: max 49, clipped 0/50 updates Step 21,500 | last 50 updates / 1,600 crops | lr 5.56e-06 points (35/curve, tolerance 2 vox): 47,434 right / 3,302 wrong | 93.5% correct | 5,264 unknown loss 0.9295 | geometry 0.2985 | confidence 1.2620 | mean error 0.657 vox speed: 892 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1503 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 183, from the previous prediction 1320 confidence-labeled crops: terminal 23/1535 (1.5%) | recoverable 119/1535 (7.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 288/1600 (18.0%), 2: 352/1600 (22.0%), 3: 352/1600 (22.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 655 rows | mean depth 19 decisions | mean travel 155 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 57% 40% 27%>2% 24% 7 434 88%/9%/0%/3% 10% 155/0/0/0 0175A_5mm_v1 288 56% 44% 27%>0% 26% 1 226 89%/7%/0%/5% 3% 127/3/0/0 1447_5mm_v1 352 53% 45% 26%>2% 25% 6 458 89%/7%/0%/3% 3% 151/2/0/8 paris4_afv_central 352 46% 42% 25%>12% 13% 34 1905 86%/6%/2%/6% 12% 132/1/0/17 PHerc0125_3mm_v2 224 42% 40% 27%>17% 10% 33 2060 79%/7%/6%/7% 25% 90/4/0/0 gradients: max 67, clipped 1/50 updates Step 21,550 | last 50 updates / 1,600 crops | lr 5.50e-06 points (35/curve, tolerance 2 vox): 47,733 right / 3,099 wrong | 93.9% correct | 5,168 unknown loss 0.7946 | geometry 0.2467 | confidence 1.0959 | mean error 0.619 vox speed: 893 ms/update | data wait 1 ms/update | 35.0 crops/s | 35.0 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1495 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 248, from the previous prediction 1247 confidence-labeled crops: terminal 29/1527 (1.9%) | recoverable 91/1527 (6.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 352/1600 (22.0%), 2: 416/1600 (26.0%), 3: 256/1600 (16.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 637 rows | mean depth 18 decisions | mean travel 146 vox | 8 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 52% 42% 24%>6% 18% 12 923 91%/6%/1%/2% 9% 100/1/11/22 0175A_5mm_v1 352 54% 46% 23%>0% 23% 0 156 91%/5%/0%/3% 4% 163/3/0/0 1447_5mm_v1 416 54% 43% 29%>4% 25% 12 986 87%/7%/1%/6% 2% 134/1/0/44 paris4_afv_central 256 42% 46% 22%>12% 11% 26 1644 84%/6%/4%/6% 13% 118/2/0/0 PHerc0125_3mm_v2 288 34% 51% 21%>15% 6% 39 1801 85%/5%/3%/7% 18% 122/4/122/26 gradients: max 60, clipped 1/50 updates Step 21,600 | last 50 updates / 1,600 crops | lr 5.44e-06 points (35/curve, tolerance 2 vox): 47,398 right / 3,221 wrong | 93.6% correct | 5,381 unknown loss 0.8342 | geometry 0.2896 | confidence 1.0892 | mean error 0.657 vox speed: 893 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.24 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1447 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 216, from the previous prediction 1231 confidence-labeled crops: terminal 21/1534 (1.4%) | recoverable 105/1534 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 416/1600 (26.0%), 2: 320/1600 (20.0%), 3: 256/1600 (16.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 644 rows | mean depth 18 decisions | mean travel 149 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 43% 47% 25%>10% 15% 22 1110 90%/6%/0%/4% 10% 121/0/0/0 0175A_5mm_v1 416 54% 45% 25%>1% 24% 5 416 89%/7%/0%/3% 3% 152/2/25/36 1447_5mm_v1 320 52% 47% 24%>1% 22% 4 362 88%/8%/1%/3% 2% 151/3/0/0 paris4_afv_central 256 42% 44% 28%>14% 14% 35 1990 87%/6%/3%/4% 10% 112/4/0/0 PHerc0125_3mm_v2 352 38% 42% 26%>20% 6% 70 2914 81%/5%/3%/10% 23% 108/2/0/41 gradients: max 50, clipped 0/50 updates Step 21,650 | last 50 updates / 1,600 crops | lr 5.38e-06 points (35/curve, tolerance 2 vox): 48,976 right / 2,599 wrong | 95.0% correct | 4,425 unknown loss 0.8546 | geometry 0.2752 | confidence 1.1588 | mean error 0.642 vox speed: 890 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1522 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 238, from the previous prediction 1284 confidence-labeled crops: terminal 15/1539 (1.0%) | recoverable 104/1539 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 128/1600 (8.0%), 2: 448/1600 (28.0%), 3: 320/1600 (20.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 655 rows | mean depth 19 decisions | mean travel 152 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 49% 45% 27%>6% 21% 14 840 90%/7%/1%/2% 11% 131/2/0/0 0175A_5mm_v1 128 42% 52% 25%>5% 20% 6 559 91%/5%/1%/3% 5% 67/1/0/0 1447_5mm_v1 448 52% 44% 27%>3% 24% 13 708 89%/8%/1%/3% 2% 194/5/38/5 paris4_afv_central 320 52% 46% 23%>2% 21% 6 444 90%/5%/0%/4% 4% 147/3/0/0 PHerc0125_3mm_v2 416 50% 43% 24%>6% 17% 28 1607 85%/6%/2%/7% 15% 116/2/0/64 gradients: max 56, clipped 0/50 updates Step 21,700 | last 50 updates / 1,600 crops | lr 5.32e-06 points (35/curve, tolerance 2 vox): 47,350 right / 3,025 wrong | 94.0% correct | 5,625 unknown loss 0.9519 | geometry 0.3060 | confidence 1.2918 | mean error 0.675 vox speed: 891 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1509 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 258, from the previous prediction 1251 confidence-labeled crops: terminal 15/1525 (1.0%) | recoverable 94/1525 (6.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 192/1600 (12.0%), 2: 320/1600 (20.0%), 3: 224/1600 (14.0%), 4: 480/1600 (30.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 627 rows | mean depth 18 decisions | mean travel 145 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 49% 47% 25%>4% 21% 13 677 89%/7%/1%/3% 9% 171/1/9/10 0175A_5mm_v1 192 56% 43% 28%>1% 27% 2 391 91%/7%/0%/2% 3% 82/2/0/0 1447_5mm_v1 320 53% 47% 20%>0% 20% 1 203 92%/4%/0%/4% 3% 125/3/0/25 paris4_afv_central 224 46% 44% 22%>10% 12% 22 1888 87%/6%/3%/4% 11% 86/1/14/12 PHerc0125_3mm_v2 480 48% 46% 26%>6% 20% 41 1524 85%/5%/1%/8% 16% 163/7/0/57 gradients: max 55, clipped 0/50 updates Step 21,750 | last 50 updates / 1,600 crops | lr 5.26e-06 points (35/curve, tolerance 2 vox): 45,361 right / 4,771 wrong | 90.5% correct | 5,868 unknown loss 1.1227 | geometry 0.4427 | confidence 1.3600 | mean error 0.823 vox speed: 891 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.31 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1397 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 287, from the previous prediction 1110 confidence-labeled crops: terminal 38/1518 (2.5%) | recoverable 106/1518 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 192/1600 (12.0%), 2: 288/1600 (18.0%), 3: 288/1600 (18.0%), 4: 512/1600 (32.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 575 rows | mean depth 20 decisions | mean travel 165 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 48% 45% 24%>7% 17% 21 970 89%/8%/0%/3% 12% 131/0/0/12 0175A_5mm_v1 192 49% 44% 20%>7% 14% 11 939 89%/6%/2%/3% 8% 85/1/0/0 1447_5mm_v1 288 57% 42% 25%>1% 24% 3 418 92%/5%/1%/2% 3% 121/2/0/0 paris4_afv_central 288 44% 43% 28%>13% 15% 31 1902 84%/8%/3%/5% 10% 123/0/0/0 PHerc0125_3mm_v2 512 34% 46% 25%>20% 5% 94 3577 79%/6%/4%/11% 24% 115/2/0/121 gradients: max 47, clipped 0/50 updates Step 21,800 | last 50 updates / 1,600 crops | lr 5.20e-06 points (35/curve, tolerance 2 vox): 47,072 right / 4,164 wrong | 91.9% correct | 4,764 unknown loss 0.9709 | geometry 0.3611 | confidence 1.2197 | mean error 0.737 vox speed: 890 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.30 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1442 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 284, from the previous prediction 1158 confidence-labeled crops: terminal 30/1529 (2.0%) | recoverable 128/1529 (8.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 608/1600 (38.0%), 1: 288/1600 (18.0%), 2: 256/1600 (16.0%), 3: 352/1600 (22.0%), 4: 96/1600 (6.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 576 rows | mean depth 22 decisions | mean travel 175 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 608 43% 47% 25%>10% 16% 37 2054 86%/8%/1%/5% 16% 148/2/0/136 0175A_5mm_v1 288 52% 45% 27%>2% 24% 7 435 87%/8%/0%/4% 5% 131/2/2/0 1447_5mm_v1 256 52% 47% 24%>1% 23% 3 360 90%/8%/0%/2% 1% 121/1/0/0 paris4_afv_central 352 42% 45% 27%>14% 14% 44 2241 80%/9%/5%/6% 12% 137/1/0/20 PHerc0125_3mm_v2 96 39% 41% 27%>21% 6% 19 2385 82%/8%/5%/4% 22% 39/2/0/0 gradients: max 57, clipped 0/50 updates Step 21,850 | last 50 updates / 1,600 crops | lr 5.14e-06 points (35/curve, tolerance 2 vox): 46,717 right / 3,923 wrong | 92.3% correct | 5,360 unknown loss 0.9479 | geometry 0.3275 | confidence 1.2408 | mean error 0.691 vox speed: 889 ms/update | data wait 1 ms/update | 35.6 crops/s | 35.6 decisions/s | 1.29 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1439 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 341, from the previous prediction 1098 confidence-labeled crops: terminal 36/1536 (2.3%) | recoverable 102/1536 (6.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 352/1600 (22.0%), 2: 256/1600 (16.0%), 3: 448/1600 (28.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 565 rows | mean depth 19 decisions | mean travel 155 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 47% 46% 23%>7% 16% 19 1303 89%/8%/0%/2% 11% 97/0/0/51 0175A_5mm_v1 352 54% 46% 23%>0% 23% 0 135 89%/7%/0%/5% 4% 94/0/33/68 1447_5mm_v1 256 48% 50% 22%>2% 20% 6 556 94%/5%/1%/0% 3% 128/2/0/0 paris4_afv_central 448 32% 48% 23%>20% 3% 56 3113 82%/7%/5%/7% 17% 154/5/0/62 PHerc0125_3mm_v2 224 44% 41% 21%>15% 6% 30 1929 85%/4%/5%/6% 18% 92/2/0/0 gradients: max 67, clipped 2/50 updates Step 21,900 | last 50 updates / 1,600 crops | lr 5.08e-06 points (35/curve, tolerance 2 vox): 47,834 right / 3,351 wrong | 93.5% correct | 4,815 unknown loss 0.9203 | geometry 0.3212 | confidence 1.1981 | mean error 0.701 vox speed: 892 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.25 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1484 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 237, from the previous prediction 1247 confidence-labeled crops: terminal 17/1533 (1.1%) | recoverable 109/1533 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 224/1600 (14.0%), 2: 384/1600 (24.0%), 3: 256/1600 (16.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 629 rows | mean depth 19 decisions | mean travel 154 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 51% 40% 29%>9% 19% 28 1239 86%/9%/0%/4% 13% 178/1/0/0 0175A_5mm_v1 224 50% 46% 23%>4% 19% 9 657 91%/4%/1%/4% 6% 80/3/0/23 1447_5mm_v1 384 53% 45% 26%>2% 24% 8 566 92%/5%/0%/3% 3% 173/4/0/0 paris4_afv_central 256 55% 42% 24%>3% 21% 11 1125 85%/9%/2%/5% 5% 75/1/0/33 PHerc0125_3mm_v2 288 42% 43% 28%>16% 12% 40 1963 83%/6%/3%/8% 20% 123/4/0/0 gradients: max 98, clipped 1/50 updates Step 21,950 | last 50 updates / 1,600 crops | lr 5.02e-06 points (35/curve, tolerance 2 vox): 48,117 right / 2,896 wrong | 94.3% correct | 4,987 unknown loss 0.8215 | geometry 0.2575 | confidence 1.1280 | mean error 0.636 vox speed: 896 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1500 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 267, from the previous prediction 1233 confidence-labeled crops: terminal 22/1537 (1.4%) | recoverable 99/1537 (6.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 352/1600 (22.0%), 2: 288/1600 (18.0%), 3: 352/1600 (22.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 630 rows | mean depth 19 decisions | mean travel 158 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 48% 45% 25%>8% 17% 20 967 91%/7%/0%/2% 12% 144/1/12/0 0175A_5mm_v1 352 51% 46% 25%>3% 22% 10 534 89%/7%/1%/4% 5% 126/2/0/35 1447_5mm_v1 288 46% 54% 23%>0% 23% 0 143 91%/5%/0%/4% 2% 128/0/27/27 paris4_afv_central 352 45% 47% 27%>8% 19% 26 1480 87%/8%/3%/3% 7% 149/2/0/15 PHerc0125_3mm_v2 288 47% 42% 26%>11% 15% 34 2174 84%/5%/3%/9% 16% 83/0/128/37 gradients: max 38, clipped 0/50 updates Step 22,000 | last 50 updates / 1,600 crops | lr 4.96e-06 points (35/curve, tolerance 2 vox): 46,302 right / 3,355 wrong | 93.2% correct | 6,343 unknown loss 0.8658 | geometry 0.2880 | confidence 1.1556 | mean error 0.660 vox speed: 893 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.27 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1440 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 226, from the previous prediction 1214 confidence-labeled crops: terminal 27/1506 (1.8%) | recoverable 123/1506 (8.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 352/1600 (22.0%), 2: 352/1600 (22.0%), 3: 224/1600 (14.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 94 chunks / 188.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 633 rows | mean depth 17 decisions | mean travel 145 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 41% 43% 24%>16% 8% 29 1792 85%/9%/1%/6% 16% 133/1/79/4 0175A_5mm_v1 352 53% 45% 28%>2% 27% 6 346 84%/8%/0%/8% 3% 157/4/0/1 1447_5mm_v1 352 49% 48% 20%>4% 16% 11 914 88%/7%/1%/3% 4% 135/4/0/33 paris4_afv_central 224 50% 44% 29%>7% 22% 13 1187 85%/9%/2%/4% 4% 98/2/0/0 PHerc0125_3mm_v2 352 35% 46% 27%>19% 8% 66 2886 80%/6%/5%/9% 22% 110/4/0/52 gradients: max 53, clipped 0/50 updates Step 22,000 | dagger_launched: true Step 22,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 250, "stopped": 0, "mean_example_error": 0.5609543142574174, "geometry_examples": 28, "examples": 32, "inference_and_measurement_seconds": 1.873542431741953, "total_diagnostic_seconds": 3.6666650138795376} Step 22,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 6/8 (75.0%) | false stops 1 1.5-3: recovered 6/8 (75.0%) | false stops 1 3-4.5: recovered 6/8 (75.0%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 0 entire proposed 8-point prefix correct (distance): 22/28 (78.6%) gate @ 0.50: false stops 0/23 (0.0%) | accepted wrong 4/26 (15.4%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.682 1.652 following 16 464 1.160 1.186 recoverable 12 348 2.379 2.273 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 1.160 1.186 332 32 812 1.682 1.652 improved/worsened/compared states: 2/1/28 Step 22,000 | batch_diagnostics_seconds: 3.6692148223519325 | recovery_seconds: 15.152668613940477 Step 22,050 | last 50 updates / 1,600 crops | lr 4.91e-06 points (35/curve, tolerance 2 vox): 47,537 right / 3,593 wrong | 93.0% correct | 4,870 unknown loss 0.8828 | geometry 0.3154 | confidence 1.1349 | mean error 0.690 vox speed: 1296 ms/update | data wait 1 ms/update | 18.7 crops/s | 18.7 decisions/s | 1.25 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1481 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 193, from the previous prediction 1288 confidence-labeled crops: terminal 25/1532 (1.6%) | recoverable 129/1532 (8.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 320/1600 (20.0%), 2: 320/1600 (20.0%), 3: 256/1600 (16.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 97 chunks / 194.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 689 rows | mean depth 19 decisions | mean travel 153 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 49% 44% 25%>7% 18% 19 1006 90%/7%/0%/3% 10% 179/0/9/5 0175A_5mm_v1 320 57% 41% 27%>2% 25% 5 327 84%/8%/1%/7% 4% 126/2/0/5 1447_5mm_v1 320 51% 48% 25%>2% 23% 6 477 89%/8%/1%/2% 3% 152/2/0/0 paris4_afv_central 256 38% 51% 24%>11% 13% 22 1578 83%/7%/4%/6% 12% 131/3/0/0 PHerc0125_3mm_v2 288 41% 43% 26%>16% 10% 43 2361 79%/10%/3%/8% 22% 101/0/0/24 gradients: max 42, clipped 0/50 updates Step 22,100 | last 50 updates / 1,600 crops | lr 4.85e-06 points (35/curve, tolerance 2 vox): 47,527 right / 3,165 wrong | 93.8% correct | 5,308 unknown loss 0.9006 | geometry 0.2767 | confidence 1.2478 | mean error 0.659 vox speed: 1372 ms/update | data wait 1 ms/update | 23.1 crops/s | 23.1 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1496 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 246, from the previous prediction 1250 confidence-labeled crops: terminal 20/1535 (1.3%) | recoverable 93/1535 (6.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 352/1600 (22.0%), 2: 288/1600 (18.0%), 3: 256/1600 (16.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 97 chunks / 194.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 645 rows | mean depth 20 decisions | mean travel 160 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 47% 47% 20%>6% 13% 23 1132 91%/7%/0%/2% 13% 125/0/0/26 0175A_5mm_v1 352 51% 49% 23%>0% 23% 0 134 91%/5%/0%/4% 3% 150/2/0/22 1447_5mm_v1 288 48% 48% 23%>4% 19% 10 870 91%/5%/0%/3% 3% 124/3/0/14 paris4_afv_central 256 46% 49% 23%>5% 18% 13 953 88%/8%/2%/2% 7% 125/4/0/0 PHerc0125_3mm_v2 384 46% 43% 27%>11% 16% 44 2274 82%/5%/3%/10% 18% 121/4/0/43 gradients: max 58, clipped 0/50 updates Step 22,150 | last 50 updates / 1,600 crops | lr 4.79e-06 points (35/curve, tolerance 2 vox): 48,767 right / 3,050 wrong | 94.1% correct | 4,183 unknown loss 0.8537 | geometry 0.3131 | confidence 1.0813 | mean error 0.679 vox speed: 1342 ms/update | data wait 1 ms/update | 23.6 crops/s | 23.6 decisions/s | 1.26 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1496 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 236, from the previous prediction 1260 confidence-labeled crops: terminal 12/1551 (0.8%) | recoverable 135/1551 (8.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 480/1600 (30.0%), 1: 192/1600 (12.0%), 2: 416/1600 (26.0%), 3: 288/1600 (18.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 97 chunks / 194.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 634 rows | mean depth 18 decisions | mean travel 145 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 480 47% 45% 25%>8% 17% 29 1532 88%/9%/0%/3% 14% 151/0/0/63 0175A_5mm_v1 192 59% 41% 29%>0% 29% 0 146 89%/8%/0%/3% 2% 78/1/0/0 1447_5mm_v1 416 53% 46% 26%>1% 25% 4 329 92%/6%/0%/2% 1% 192/4/0/0 paris4_afv_central 288 49% 45% 26%>5% 20% 15 966 85%/10%/1%/3% 8% 131/2/0/0 PHerc0125_3mm_v2 224 46% 37% 28%>17% 11% 34 2390 82%/8%/4%/5% 21% 82/1/0/0 gradients: max 75, clipped 2/50 updates Step 22,200 | last 50 updates / 1,600 crops | lr 4.73e-06 points (35/curve, tolerance 2 vox): 47,601 right / 2,861 wrong | 94.3% correct | 5,538 unknown loss 0.8534 | geometry 0.2794 | confidence 1.1481 | mean error 0.632 vox speed: 1346 ms/update | data wait 1 ms/update | 23.6 crops/s | 23.6 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1482 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 237, from the previous prediction 1245 confidence-labeled crops: terminal 24/1528 (1.6%) | recoverable 115/1528 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 512/1600 (32.0%), 2: 192/1600 (12.0%), 3: 352/1600 (22.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 97 chunks / 194.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 619 rows | mean depth 19 decisions | mean travel 155 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 52% 44% 25%>4% 21% 9 653 90%/9%/0%/0% 7% 99/0/0/0 0175A_5mm_v1 512 49% 47% 24%>4% 20% 21 950 89%/7%/1%/4% 5% 193/8/22/47 1447_5mm_v1 192 46% 48% 24%>6% 18% 10 615 89%/5%/2%/4% 7% 92/1/0/0 paris4_afv_central 352 50% 41% 26%>9% 16% 29 2197 84%/8%/2%/7% 8% 99/1/44/46 PHerc0125_3mm_v2 320 45% 43% 28%>12% 16% 35 1706 82%/7%/3%/8% 17% 136/4/15/2 gradients: max 43, clipped 0/50 updates Step 22,250 | last 50 updates / 1,600 crops | lr 4.68e-06 points (35/curve, tolerance 2 vox): 46,015 right / 3,931 wrong | 92.1% correct | 6,054 unknown loss 0.9335 | geometry 0.2821 | confidence 1.3030 | mean error 0.650 vox speed: 1387 ms/update | data wait 1 ms/update | 22.9 crops/s | 22.9 decisions/s | 1.28 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1444 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 240, from the previous prediction 1204 confidence-labeled crops: terminal 48/1522 (3.2%) | recoverable 103/1522 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 96/1600 (6.0%), 1: 416/1600 (26.0%), 2: 320/1600 (20.0%), 3: 352/1600 (22.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 97 chunks / 194.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 609 rows | mean depth 17 decisions | mean travel 143 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 96 47% 53% 17%>0% 17% 0 164 90%/10%/0%/0% 13% 51/2/0/0 0175A_5mm_v1 416 55% 42% 25%>4% 22% 15 669 87%/8%/0%/4% 5% 129/3/0/44 1447_5mm_v1 320 50% 47% 24%>3% 21% 9 724 89%/5%/1%/4% 4% 150/3/0/0 paris4_afv_central 352 43% 43% 30%>14% 16% 42 2149 83%/8%/5%/5% 12% 152/6/0/0 PHerc0125_3mm_v2 416 44% 38% 26%>18% 8% 63 2894 81%/4%/6%/9% 23% 127/2/0/32 gradients: max 45, clipped 0/50 updates Step 22,300 | last 50 updates / 1,600 crops | lr 4.62e-06 points (35/curve, tolerance 2 vox): 49,303 right / 2,172 wrong | 95.8% correct | 4,525 unknown loss 0.6745 | geometry 0.2319 | confidence 0.8852 | mean error 0.578 vox speed: 1351 ms/update | data wait 1 ms/update | 23.5 crops/s | 23.5 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1527 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 307, from the previous prediction 1220 confidence-labeled crops: terminal 18/1538 (1.2%) | recoverable 125/1538 (8.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 448/1600 (28.0%), 2: 512/1600 (32.0%), 3: 352/1600 (22.0%), 4: 96/1600 (6.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 97 chunks / 194.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 568 rows | mean depth 19 decisions | mean travel 155 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 52% 42% 28%>7% 21% 12 1035 87%/8%/0%/5% 11% 80/1/45/0 0175A_5mm_v1 448 56% 42% 28%>2% 26% 12 678 86%/10%/1%/3% 6% 101/3/0/86 1447_5mm_v1 512 57% 41% 25%>2% 23% 8 423 90%/6%/1%/4% 2% 188/4/40/23 paris4_afv_central 352 45% 45% 28%>9% 19% 29 1615 85%/9%/3%/4% 10% 149/3/0/11 PHerc0125_3mm_v2 96 43% 52% 11%>5% 6% 5 813 86%/7%/0%/6% 12% 50/0/0/0 gradients: max 51, clipped 0/50 updates Step 22,350 | last 50 updates / 1,600 crops | lr 4.56e-06 points (35/curve, tolerance 2 vox): 47,209 right / 3,685 wrong | 92.8% correct | 5,106 unknown loss 0.9512 | geometry 0.3241 | confidence 1.2542 | mean error 0.721 vox speed: 1353 ms/update | data wait 1 ms/update | 23.4 crops/s | 23.4 decisions/s | 1.25 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1474 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 236, from the previous prediction 1238 confidence-labeled crops: terminal 23/1536 (1.5%) | recoverable 104/1536 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 352/1600 (22.0%), 2: 320/1600 (20.0%), 3: 224/1600 (14.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 97 chunks / 194.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 626 rows | mean depth 19 decisions | mean travel 155 vox | 7 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 50% 43% 29%>7% 22% 20 1045 89%/9%/0%/2% 11% 137/0/9/0 0175A_5mm_v1 352 49% 44% 23%>7% 16% 18 888 91%/3%/3%/4% 7% 156/8/0/0 1447_5mm_v1 320 56% 44% 29%>0% 28% 1 248 86%/9%/0%/5% 2% 81/2/0/60 paris4_afv_central 224 44% 49% 22%>7% 15% 13 1264 87%/8%/1%/4% 7% 94/3/0/16 PHerc0125_3mm_v2 384 42% 41% 27%>17% 10% 56 2219 85%/5%/3%/8% 21% 158/5/0/0 gradients: max 67, clipped 1/50 updates Step 22,400 | last 50 updates / 1,600 crops | lr 4.51e-06 points (35/curve, tolerance 2 vox): 46,177 right / 3,713 wrong | 92.6% correct | 6,110 unknown loss 0.9852 | geometry 0.3333 | confidence 1.3038 | mean error 0.718 vox speed: 1350 ms/update | data wait 1 ms/update | 23.5 crops/s | 23.5 decisions/s | 1.32 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1435 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 339, from the previous prediction 1096 confidence-labeled crops: terminal 26/1511 (1.7%) | recoverable 93/1511 (6.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 320/1600 (20.0%), 2: 320/1600 (20.0%), 3: 192/1600 (12.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 97 chunks / 194.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 521 rows | mean depth 18 decisions | mean travel 144 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 50% 40% 26%>9% 17% 25 2469 88%/8%/1%/3% 14% 62/0/72/67 0175A_5mm_v1 320 52% 46% 27%>2% 25% 5 320 90%/5%/1%/4% 5% 147/1/0/0 1447_5mm_v1 320 50% 47% 21%>3% 18% 11 651 88%/6%/1%/5% 3% 118/0/0/31 paris4_afv_central 192 42% 44% 27%>15% 12% 26 2216 87%/5%/4%/4% 8% 83/1/0/1 PHerc0125_3mm_v2 448 40% 47% 20%>13% 7% 71 3271 81%/4%/3%/11% 20% 111/2/0/98 gradients: max 44, clipped 0/50 updates Step 22,450 | last 50 updates / 1,600 crops | lr 4.45e-06 points (35/curve, tolerance 2 vox): 47,662 right / 3,850 wrong | 92.5% correct | 4,488 unknown loss 0.9032 | geometry 0.2892 | confidence 1.2281 | mean error 0.676 vox speed: 1333 ms/update | data wait 1 ms/update | 23.7 crops/s | 23.7 decisions/s | 1.28 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1461 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 247, from the previous prediction 1214 confidence-labeled crops: terminal 37/1539 (2.4%) | recoverable 100/1539 (6.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 160/1600 (10.0%), 1: 320/1600 (20.0%), 2: 416/1600 (26.0%), 3: 384/1600 (24.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 97 chunks / 194.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 662 rows | mean depth 18 decisions | mean travel 149 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 160 41% 45% 26%>14% 12% 22 2292 85%/12%/0%/3% 16% 52/0/0/20 0175A_5mm_v1 320 48% 44% 26%>8% 18% 24 1180 87%/4%/4%/5% 6% 141/3/12/0 1447_5mm_v1 416 46% 50% 21%>3% 18% 13 421 89%/7%/1%/2% 4% 209/2/0/0 paris4_afv_central 384 51% 45% 24%>4% 20% 10 770 90%/6%/2%/3% 6% 173/4/0/0 PHerc0125_3mm_v2 320 36% 49% 22%>14% 8% 52 3022 82%/4%/4%/10% 23% 87/2/0/71 gradients: max 62, clipped 1/50 updates Step 22,500 | last 50 updates / 1,600 crops | lr 4.39e-06 points (35/curve, tolerance 2 vox): 45,594 right / 4,420 wrong | 91.2% correct | 5,986 unknown loss 1.1225 | geometry 0.3980 | confidence 1.4491 | mean error 0.769 vox speed: 1354 ms/update | data wait 1 ms/update | 23.3 crops/s | 23.3 decisions/s | 1.37 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1417 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 318, from the previous prediction 1099 confidence-labeled crops: terminal 29/1518 (1.9%) | recoverable 110/1518 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 128/1600 (8.0%), 2: 320/1600 (20.0%), 3: 288/1600 (18.0%), 4: 576/1600 (36.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 97 chunks / 194.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 557 rows | mean depth 20 decisions | mean travel 164 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 48% 41% 24%>11% 12% 26 1546 86%/9%/0%/4% 13% 117/0/0/0 0175A_5mm_v1 128 48% 52% 26%>0% 26% 0 164 90%/5%/0%/5% 2% 66/2/0/0 1447_5mm_v1 320 53% 47% 22%>0% 22% 1 203 91%/5%/0%/4% 1% 149/4/19/1 paris4_afv_central 288 44% 42% 27%>13% 14% 31 2243 85%/9%/3%/3% 12% 116/2/8/6 PHerc0125_3mm_v2 576 38% 46% 28%>16% 12% 87 3822 82%/6%/3%/9% 24% 109/1/0/155 gradients: max 57, clipped 0/50 updates Step 22,550 | last 50 updates / 1,600 crops | lr 4.34e-06 points (35/curve, tolerance 2 vox): 47,896 right / 2,923 wrong | 94.2% correct | 5,181 unknown loss 0.8261 | geometry 0.2557 | confidence 1.1410 | mean error 0.637 vox speed: 1363 ms/update | data wait 1 ms/update | 23.3 crops/s | 23.3 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1513 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 277, from the previous prediction 1236 confidence-labeled crops: terminal 24/1530 (1.6%) | recoverable 99/1530 (6.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 288/1600 (18.0%), 2: 352/1600 (22.0%), 3: 416/1600 (26.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 97 chunks / 194.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 647 rows | mean depth 15 decisions | mean travel 121 vox | 6 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 41% 52% 22%>6% 16% 17 846 91%/7%/0%/2% 11% 164/0/0/4 0175A_5mm_v1 288 45% 50% 21%>5% 16% 13 653 88%/7%/1%/4% 6% 144/5/4/0 1447_5mm_v1 352 50% 47% 22%>3% 20% 9 505 90%/7%/0%/3% 2% 129/4/7/38 paris4_afv_central 416 49% 47% 27%>5% 22% 23 1750 85%/5%/3%/7% 8% 109/0/0/85 PHerc0125_3mm_v2 224 46% 45% 23%>8% 15% 18 1317 84%/4%/3%/8% 15% 101/1/0/0 gradients: max 52, clipped 0/50 updates Step 22,596 | dagger_states: 7394 | dagger_source_step: 22000 | dagger_caches: 4 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/0175A_5mm_v1/decisions_022000.npz" | dagger_fibers: 384 | dagger_supply: {"states": 7394, "traces": 384, "fibers": 384, "rows": {"terminal": 207, "premature_stop": 93, "recoverable": 17, "pre_excursion": 204, "ordinary": 6862}, "events": {"terminal": 28, "premature_stop": 93, "recoverable": 9, "pre_excursion": 36, "ordinary": 381}, "supervision": {"following": 7159, "recoverable": 15, "terminal": 207, "unknown": 13}, "departed_traces": 35, "terminal_traces": 28, "stop_reasons": {"after_failure": 26, "confidence": 96, "unannotated": 262}, "traced_voxels": 112655.3290258026, "seconds": 822.5135770849884, "voxels_per_second": 136.96470449163442} | dagger_coverage: {"requested": 384, "seeds": 384, "skipped": [], "epoch": 0} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 0 | dagger_launch_interval: 5000 | dagger_publication_age: 596 | dataset: "0175A_5mm_v1" Step 22,600 | last 50 updates / 1,600 crops | lr 4.28e-06 points (35/curve, tolerance 2 vox): 46,877 right / 3,847 wrong | 92.4% correct | 5,276 unknown loss 1.0045 | geometry 0.3386 | confidence 1.3318 | mean error 0.720 vox speed: 1269 ms/update | data wait 1 ms/update | 24.9 crops/s | 24.9 decisions/s | 1.32 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1459 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 289, from the previous prediction 1170 confidence-labeled crops: terminal 14/1531 (0.9%) | recoverable 128/1531 (8.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 288/1600 (18.0%), 2: 256/1600 (16.0%), 3: 256/1600 (16.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 97 chunks / 194.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 573 rows | mean depth 17 decisions | mean travel 138 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 48% 42% 27%>10% 17% 37 3127 88%/10%/0%/2% 16% 77/0/0/113 0175A_5mm_v1 288 48% 50% 22%>3% 19% 7 450 91%/6%/0%/3% 4% 143/1/0/0 1447_5mm_v1 256 50% 43% 30%>7% 23% 16 1004 85%/9%/1%/4% 5% 110/0/0/0 paris4_afv_central 256 50% 41% 28%>9% 19% 20 1679 85%/9%/2%/5% 10% 104/0/0/0 PHerc0125_3mm_v2 352 49% 44% 22%>7% 15% 22 1282 83%/6%/2%/9% 15% 139/3/58/16 gradients: max 67, clipped 1/50 updates Step 22,650 | last 50 updates / 1,600 crops | lr 4.23e-06 points (35/curve, tolerance 2 vox): 47,874 right / 2,698 wrong | 94.7% correct | 5,428 unknown loss 0.7668 | geometry 0.2460 | confidence 1.0416 | mean error 0.599 vox speed: 894 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1483 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 289, from the previous prediction 1194 confidence-labeled crops: terminal 21/1529 (1.4%) | recoverable 96/1529 (6.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 512/1600 (32.0%), 2: 352/1600 (22.0%), 3: 224/1600 (14.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 97 chunks / 194.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 597 rows | mean depth 19 decisions | mean travel 158 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 50% 46% 23%>5% 19% 11 772 90%/8%/0%/2% 10% 111/0/0/6 0175A_5mm_v1 512 48% 46% 23%>5% 18% 31 1724 90%/5%/1%/5% 5% 103/1/71/135 1447_5mm_v1 352 51% 45% 27%>3% 24% 10 596 88%/8%/0%/4% 3% 159/1/0/0 paris4_afv_central 224 53% 45% 23%>3% 21% 6 636 90%/4%/1%/5% 7% 100/1/0/0 PHerc0125_3mm_v2 256 34% 49% 25%>18% 8% 42 2141 80%/6%/5%/9% 22% 124/3/17/1 gradients: max 46, clipped 0/50 updates Step 22,700 | last 50 updates / 1,600 crops | lr 4.17e-06 points (35/curve, tolerance 2 vox): 47,146 right / 4,077 wrong | 92.0% correct | 4,777 unknown loss 0.9971 | geometry 0.3115 | confidence 1.3711 | mean error 0.715 vox speed: 895 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.26 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1471 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 212, from the previous prediction 1259 confidence-labeled crops: terminal 33/1549 (2.1%) | recoverable 115/1549 (7.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 256/1600 (16.0%), 2: 224/1600 (14.0%), 3: 320/1600 (20.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 97 chunks / 194.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 636 rows | mean depth 14 decisions | mean travel 116 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 54% 42% 26%>4% 22% 15 784 90%/8%/0%/2% 10% 164/0/0/10 0175A_5mm_v1 256 54% 42% 27%>4% 23% 11 753 87%/7%/2%/4% 7% 89/3/0/19 1447_5mm_v1 224 49% 47% 26%>4% 22% 9 527 92%/5%/1%/2% 4% 105/3/0/0 paris4_afv_central 320 52% 41% 31%>7% 24% 20 1395 84%/8%/2%/5% 7% 130/2/0/0 PHerc0125_3mm_v2 384 38% 46% 24%>15% 9% 60 2536 84%/6%/5%/6% 23% 148/2/0/30 gradients: max 55, clipped 0/50 updates Step 22,750 | last 50 updates / 1,600 crops | lr 4.12e-06 points (35/curve, tolerance 2 vox): 48,216 right / 2,874 wrong | 94.4% correct | 4,910 unknown loss 0.8899 | geometry 0.3091 | confidence 1.1617 | mean error 0.671 vox speed: 896 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.25 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1513 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 236, from the previous prediction 1277 confidence-labeled crops: terminal 12/1534 (0.8%) | recoverable 115/1534 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 384/1600 (24.0%), 2: 288/1600 (18.0%), 3: 224/1600 (14.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 97 chunks / 194.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 700 rows | mean depth 16 decisions | mean travel 136 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 48% 47% 24%>5% 19% 16 898 90%/8%/0%/2% 13% 139/1/0/27 0175A_5mm_v1 384 46% 48% 23%>6% 17% 21 784 88%/7%/1%/5% 5% 184/7/0/0 1447_5mm_v1 288 54% 44% 22%>2% 19% 7 392 90%/6%/0%/4% 2% 126/1/0/0 paris4_afv_central 224 54% 46% 21%>0% 21% 1 236 89%/8%/0%/4% 4% 102/4/0/0 PHerc0125_3mm_v2 352 37% 52% 22%>11% 11% 34 1713 82%/8%/2%/8% 18% 149/1/0/33 gradients: max 58, clipped 0/50 updates Step 22,800 | last 50 updates / 1,600 crops | lr 4.07e-06 points (35/curve, tolerance 2 vox): 47,212 right / 3,372 wrong | 93.3% correct | 5,416 unknown loss 0.9018 | geometry 0.2811 | confidence 1.2415 | mean error 0.663 vox speed: 894 ms/update | data wait 1 ms/update | 35.1 crops/s | 35.1 decisions/s | 1.26 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1473 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 270, from the previous prediction 1203 confidence-labeled crops: terminal 27/1527 (1.8%) | recoverable 109/1527 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 512/1600 (32.0%), 1: 416/1600 (26.0%), 2: 224/1600 (14.0%), 3: 160/1600 (10.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 97 chunks / 194.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 616 rows | mean depth 18 decisions | mean travel 147 vox | 8 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 512 46% 47% 23%>7% 16% 30 1292 89%/7%/0%/4% 9% 184/0/0/55 0175A_5mm_v1 416 47% 45% 27%>8% 19% 29 1123 85%/7%/3%/5% 9% 188/5/0/0 1447_5mm_v1 224 53% 46% 30%>1% 29% 2 343 88%/9%/0%/3% 2% 103/3/0/0 paris4_afv_central 160 50% 50% 26%>0% 26% 0 156 89%/4%/1%/6% 4% 80/0/0/0 PHerc0125_3mm_v2 288 43% 43% 29%>14% 15% 45 3649 81%/6%/3%/9% 21% 61/2/0/63 gradients: max 49, clipped 0/50 updates Step 22,850 | last 50 updates / 1,600 crops | lr 4.01e-06 points (35/curve, tolerance 2 vox): 48,469 right / 2,845 wrong | 94.5% correct | 4,686 unknown loss 0.8004 | geometry 0.2581 | confidence 1.0846 | mean error 0.627 vox speed: 894 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1516 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 301, from the previous prediction 1215 confidence-labeled crops: terminal 20/1539 (1.3%) | recoverable 122/1539 (7.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 224/1600 (14.0%), 2: 288/1600 (18.0%), 3: 608/1600 (38.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 97 chunks / 194.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 584 rows | mean depth 15 decisions | mean travel 125 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 50% 40% 26%>10% 16% 20 1563 90%/8%/0%/2% 14% 107/2/0/9 0175A_5mm_v1 224 46% 46% 22%>8% 14% 17 999 88%/4%/2%/5% 7% 102/1/0/0 1447_5mm_v1 288 50% 48% 26%>2% 24% 7 631 85%/9%/2%/4% 3% 134/0/38/4 paris4_afv_central 608 54% 44% 29%>2% 27% 13 867 86%/9%/1%/4% 4% 153/4/78/112 PHerc0125_3mm_v2 192 46% 46% 19%>8% 10% 15 1348 85%/4%/3%/8% 19% 88/5/0/0 gradients: max 44, clipped 0/50 updates Step 22,900 | last 50 updates / 1,600 crops | lr 3.96e-06 points (35/curve, tolerance 2 vox): 47,158 right / 3,462 wrong | 93.2% correct | 5,380 unknown loss 0.8478 | geometry 0.2737 | confidence 1.1481 | mean error 0.652 vox speed: 893 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.25 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1472 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 261, from the previous prediction 1211 confidence-labeled crops: terminal 34/1527 (2.2%) | recoverable 106/1527 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 160/1600 (10.0%), 1: 384/1600 (24.0%), 2: 192/1600 (12.0%), 3: 480/1600 (30.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 97 chunks / 194.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 609 rows | mean depth 18 decisions | mean travel 149 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 160 52% 37% 25%>11% 14% 14 1744 91%/8%/0%/1% 14% 59/0/0/0 0175A_5mm_v1 384 49% 47% 28%>4% 24% 14 590 87%/7%/2%/5% 5% 179/7/0/0 1447_5mm_v1 192 53% 47% 24%>0% 24% 0 161 91%/7%/0%/2% 1% 91/0/0/0 paris4_afv_central 480 48% 44% 26%>9% 17% 47 3029 84%/7%/3%/6% 9% 113/1/0/97 PHerc0125_3mm_v2 384 46% 43% 25%>10% 15% 35 1614 85%/5%/4%/7% 17% 167/2/0/0 gradients: max 73, clipped 1/50 updates Step 22,950 | last 50 updates / 1,600 crops | lr 3.91e-06 points (35/curve, tolerance 2 vox): 47,529 right / 2,762 wrong | 94.5% correct | 5,709 unknown loss 0.8364 | geometry 0.2521 | confidence 1.1687 | mean error 0.632 vox speed: 893 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.24 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1530 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 194, from the previous prediction 1336 confidence-labeled crops: terminal 20/1526 (1.3%) | recoverable 95/1526 (6.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 320/1600 (20.0%), 2: 288/1600 (18.0%), 3: 320/1600 (20.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 97 chunks / 194.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 692 rows | mean depth 15 decisions | mean travel 124 vox | 6 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 45% 48% 25%>6% 19% 14 946 90%/8%/0%/2% 12% 155/0/0/0 0175A_5mm_v1 320 52% 46% 28%>2% 26% 6 573 86%/6%/2%/6% 5% 146/4/0/0 1447_5mm_v1 288 50% 45% 22%>5% 17% 12 696 90%/5%/1%/4% 3% 130/0/0/0 paris4_afv_central 320 49% 48% 22%>3% 19% 12 1001 89%/5%/1%/5% 6% 119/2/0/34 PHerc0125_3mm_v2 352 54% 41% 28%>5% 23% 18 1024 86%/5%/2%/7% 14% 142/2/36/2 gradients: max 48, clipped 0/50 updates Step 23,000 | last 50 updates / 1,600 crops | lr 3.85e-06 points (35/curve, tolerance 2 vox): 48,296 right / 2,994 wrong | 94.2% correct | 4,710 unknown loss 0.8623 | geometry 0.2855 | confidence 1.1535 | mean error 0.648 vox speed: 894 ms/update | data wait 1 ms/update | 34.9 crops/s | 34.9 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1505 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 230, from the previous prediction 1275 confidence-labeled crops: terminal 19/1545 (1.2%) | recoverable 114/1545 (7.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 320/1600 (20.0%), 2: 352/1600 (22.0%), 3: 288/1600 (18.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 97 chunks / 194.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 655 rows | mean depth 18 decisions | mean travel 145 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 47% 45% 26%>8% 18% 19 1172 89%/7%/0%/4% 15% 115/1/0/0 0175A_5mm_v1 320 50% 43% 25%>7% 18% 17 940 88%/6%/2%/4% 6% 139/3/0/0 1447_5mm_v1 352 50% 47% 25%>3% 22% 12 483 88%/9%/1%/2% 3% 165/2/0/0 paris4_afv_central 288 54% 46% 21%>0% 21% 0 140 91%/7%/0%/3% 3% 125/0/0/8 PHerc0125_3mm_v2 384 46% 45% 22%>9% 12% 39 2250 84%/7%/3%/7% 20% 111/3/0/62 gradients: max 60, clipped 0/50 updates Step 23,000 | dagger_launched: true Step 23,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 243, "stopped": 0, "mean_example_error": 0.6080392588067938, "geometry_examples": 27, "examples": 32, "inference_and_measurement_seconds": 2.5325449854135513, "total_diagnostic_seconds": 5.64941905811429} Step 23,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 6/8 (75.0%) | false stops 1 1.5-3: recovered 6/8 (75.0%) | false stops 0 3-4.5: recovered 6/8 (75.0%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 0 entire proposed 8-point prefix correct (distance): 22/28 (78.6%) gate @ 0.50: false stops 0/23 (0.0%) | accepted wrong 4/26 (15.4%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.660 1.632 following 16 464 1.160 1.182 recoverable 12 348 2.327 2.233 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 1.160 1.182 332 32 812 1.660 1.632 improved/worsened/compared states: 2/1/28 Step 23,000 | batch_diagnostics_seconds: 5.66001745313406 | recovery_seconds: 18.654080625623465 Step 23,050 | last 50 updates / 1,600 crops | lr 3.80e-06 points (35/curve, tolerance 2 vox): 48,716 right / 2,599 wrong | 94.9% correct | 4,685 unknown loss 0.8010 | geometry 0.2456 | confidence 1.1108 | mean error 0.615 vox speed: 1358 ms/update | data wait 1 ms/update | 17.0 crops/s | 17.0 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1510 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 276, from the previous prediction 1234 confidence-labeled crops: terminal 12/1538 (0.8%) | recoverable 115/1538 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 480/1600 (30.0%), 2: 256/1600 (16.0%), 3: 352/1600 (22.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 97 chunks / 194.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 603 rows | mean depth 17 decisions | mean travel 138 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 45% 46% 27%>10% 18% 26 2611 85%/9%/0%/5% 15% 79/1/86/67 0175A_5mm_v1 480 53% 44% 22%>3% 19% 16 691 91%/5%/1%/3% 4% 171/2/21/41 1447_5mm_v1 256 57% 43% 23%>0% 23% 0 143 87%/9%/0%/4% 2% 110/5/0/0 paris4_afv_central 352 49% 45% 28%>7% 21% 18 1298 87%/7%/1%/4% 6% 155/2/57/3 PHerc0125_3mm_v2 192 51% 46% 23%>4% 19% 7 728 89%/7%/1%/4% 13% 88/3/0/0 gradients: max 64, clipped 2/50 updates Step 23,100 | last 50 updates / 1,600 crops | lr 3.75e-06 points (35/curve, tolerance 2 vox): 48,024 right / 3,436 wrong | 93.3% correct | 4,540 unknown loss 0.8853 | geometry 0.2580 | confidence 1.2546 | mean error 0.645 vox speed: 1344 ms/update | data wait 1 ms/update | 23.6 crops/s | 23.6 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1504 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 294, from the previous prediction 1210 confidence-labeled crops: terminal 23/1545 (1.5%) | recoverable 93/1545 (6.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 288/1600 (18.0%), 2: 416/1600 (26.0%), 3: 256/1600 (16.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 97 chunks / 194.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 622 rows | mean depth 16 decisions | mean travel 135 vox | 7 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 49% 43% 24%>8% 16% 22 1305 91%/7%/0%/2% 13% 160/0/0/5 0175A_5mm_v1 288 45% 53% 24%>2% 22% 9 672 90%/5%/1%/4% 8% 111/4/0/42 1447_5mm_v1 416 47% 50% 24%>3% 21% 20 1529 89%/5%/2%/4% 4% 125/2/4/82 paris4_afv_central 256 55% 39% 31%>6% 25% 14 832 87%/7%/2%/5% 7% 100/2/0/0 PHerc0125_3mm_v2 256 46% 49% 21%>5% 16% 13 900 86%/5%/3%/6% 14% 126/6/0/0 gradients: max 46, clipped 0/50 updates Step 23,150 | last 50 updates / 1,600 crops | lr 3.70e-06 points (35/curve, tolerance 2 vox): 47,725 right / 3,174 wrong | 93.8% correct | 5,101 unknown loss 0.9488 | geometry 0.2856 | confidence 1.3265 | mean error 0.688 vox speed: 1356 ms/update | data wait 1 ms/update | 23.4 crops/s | 23.4 decisions/s | 1.27 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1542 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 332, from the previous prediction 1210 confidence-labeled crops: terminal 18/1521 (1.2%) | recoverable 101/1521 (6.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 480/1600 (30.0%), 1: 160/1600 (10.0%), 2: 384/1600 (24.0%), 3: 96/1600 (6.0%), 4: 480/1600 (30.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 97 chunks / 194.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 548 rows | mean depth 15 decisions | mean travel 124 vox | 6 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 480 53% 44% 24%>3% 21% 13 1089 91%/6%/0%/3% 11% 114/1/0/99 0175A_5mm_v1 160 54% 46% 25%>0% 25% 0 145 91%/6%/0%/3% 4% 74/3/0/0 1447_5mm_v1 384 57% 41% 25%>2% 23% 7 445 89%/7%/0%/4% 3% 141/1/0/18 paris4_afv_central 96 52% 48% 23%>0% 23% 0 172 85%/5%/0%/9% 2% 46/1/0/0 PHerc0125_3mm_v2 480 48% 45% 25%>7% 18% 31 1424 82%/7%/4%/8% 16% 173/3/2/43 gradients: max 37, clipped 0/50 updates Step 23,200 | last 50 updates / 1,600 crops | lr 3.65e-06 points (35/curve, tolerance 2 vox): 48,625 right / 2,521 wrong | 95.1% correct | 4,854 unknown loss 0.7668 | geometry 0.2383 | confidence 1.0569 | mean error 0.604 vox speed: 1342 ms/update | data wait 1 ms/update | 23.6 crops/s | 23.6 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1517 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 237, from the previous prediction 1280 confidence-labeled crops: terminal 16/1542 (1.0%) | recoverable 103/1542 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 320/1600 (20.0%), 2: 448/1600 (28.0%), 3: 320/1600 (20.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 97 chunks / 194.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 689 rows | mean depth 19 decisions | mean travel 158 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 48% 49% 25%>3% 22% 10 618 91%/7%/0%/2% 11% 140/1/0/0 0175A_5mm_v1 320 46% 47% 22%>7% 15% 19 897 89%/4%/2%/5% 6% 151/0/0/0 1447_5mm_v1 448 53% 44% 22%>3% 19% 17 820 90%/6%/1%/3% 4% 127/1/0/70 paris4_afv_central 320 47% 48% 25%>5% 21% 14 972 90%/8%/0%/2% 6% 154/5/0/0 PHerc0125_3mm_v2 224 40% 52% 21%>8% 13% 14 1160 82%/7%/1%/9% 15% 117/0/0/0 gradients: max 60, clipped 0/50 updates Step 23,250 | last 50 updates / 1,600 crops | lr 3.59e-06 points (35/curve, tolerance 2 vox): 47,333 right / 3,776 wrong | 92.6% correct | 4,891 unknown loss 1.0027 | geometry 0.3335 | confidence 1.3385 | mean error 0.723 vox speed: 1383 ms/update | data wait 1 ms/update | 22.9 crops/s | 22.9 decisions/s | 1.30 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1482 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 242, from the previous prediction 1240 confidence-labeled crops: terminal 25/1534 (1.6%) | recoverable 116/1534 (7.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 224/1600 (14.0%), 2: 224/1600 (14.0%), 3: 352/1600 (22.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 97 chunks / 194.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 616 rows | mean depth 18 decisions | mean travel 150 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 45% 46% 26%>8% 17% 22 1216 90%/8%/0%/2% 12% 178/0/0/0 0175A_5mm_v1 224 46% 45% 27%>8% 19% 17 1124 87%/6%/3%/5% 7% 101/3/0/0 1447_5mm_v1 224 61% 39% 29%>0% 29% 0 142 90%/8%/0%/2% 1% 87/3/0/0 paris4_afv_central 352 52% 43% 28%>4% 23% 13 985 86%/9%/1%/4% 6% 151/1/32/2 PHerc0125_3mm_v2 416 44% 45% 23%>11% 12% 43 2919 82%/5%/4%/8% 21% 99/7/0/90 gradients: max 52, clipped 0/50 updates Step 23,300 | last 50 updates / 1,600 crops | lr 3.54e-06 points (35/curve, tolerance 2 vox): 48,018 right / 3,473 wrong | 93.3% correct | 4,509 unknown loss 0.9212 | geometry 0.2860 | confidence 1.2704 | mean error 0.663 vox speed: 1375 ms/update | data wait 1 ms/update | 22.9 crops/s | 22.9 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1496 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 208, from the previous prediction 1288 confidence-labeled crops: terminal 21/1535 (1.4%) | recoverable 115/1535 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 384/1600 (24.0%), 2: 320/1600 (20.0%), 3: 256/1600 (16.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 97 chunks / 194.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 675 rows | mean depth 15 decisions | mean travel 124 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 45% 47% 22%>8% 14% 24 1232 89%/8%/0%/3% 13% 166/3/0/14 0175A_5mm_v1 384 51% 43% 27%>7% 20% 21 998 86%/9%/1%/5% 7% 163/3/32/2 1447_5mm_v1 320 52% 43% 21%>5% 16% 15 676 90%/5%/2%/4% 4% 139/2/0/0 paris4_afv_central 256 51% 43% 25%>6% 19% 13 1422 86%/7%/2%/4% 9% 92/0/11/19 PHerc0125_3mm_v2 256 48% 45% 24%>7% 18% 16 1191 83%/7%/2%/8% 18% 115/3/0/0 gradients: max 46, clipped 0/50 updates Step 23,350 | last 50 updates / 1,600 crops | lr 3.49e-06 points (35/curve, tolerance 2 vox): 47,984 right / 2,861 wrong | 94.4% correct | 5,155 unknown loss 0.7801 | geometry 0.2347 | confidence 1.0906 | mean error 0.605 vox speed: 1337 ms/update | data wait 1 ms/update | 23.7 crops/s | 23.7 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1501 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 189, from the previous prediction 1312 confidence-labeled crops: terminal 26/1536 (1.7%) | recoverable 87/1536 (5.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 320/1600 (20.0%), 2: 384/1600 (24.0%), 3: 288/1600 (18.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 97 chunks / 194.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 722 rows | mean depth 16 decisions | mean travel 130 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 49% 47% 23%>4% 19% 10 714 88%/8%/0%/4% 8% 115/4/0/6 0175A_5mm_v1 320 50% 41% 26%>9% 17% 26 1266 88%/5%/2%/5% 8% 131/2/0/0 1447_5mm_v1 384 47% 50% 20%>3% 17% 11 460 91%/4%/1%/4% 3% 186/1/77/6 paris4_afv_central 288 50% 46% 24%>4% 20% 11 592 89%/6%/1%/4% 5% 132/4/0/0 PHerc0125_3mm_v2 352 44% 46% 24%>10% 14% 32 1664 85%/4%/4%/7% 16% 158/3/9/4 gradients: max 41, clipped 0/50 updates Step 23,400 | last 50 updates / 1,600 crops | lr 3.44e-06 points (35/curve, tolerance 2 vox): 47,462 right / 3,360 wrong | 93.4% correct | 5,178 unknown loss 0.8744 | geometry 0.2951 | confidence 1.1586 | mean error 0.668 vox speed: 1362 ms/update | data wait 1 ms/update | 23.2 crops/s | 23.2 decisions/s | 1.26 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1486 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 335, from the previous prediction 1151 confidence-labeled crops: terminal 21/1532 (1.4%) | recoverable 110/1532 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 352/1600 (22.0%), 2: 352/1600 (22.0%), 3: 384/1600 (24.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 99 chunks / 198.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 559 rows | mean depth 17 decisions | mean travel 143 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 39% 44% 25%>17% 8% 20 2126 86%/9%/1%/5% 17% 85/1/0/0 0175A_5mm_v1 352 51% 41% 26%>8% 18% 27 1672 86%/6%/3%/6% 10% 98/4/46/45 1447_5mm_v1 352 53% 45% 25%>2% 23% 7 422 90%/6%/1%/3% 2% 122/3/2/38 paris4_afv_central 384 53% 45% 26%>2% 23% 8 806 88%/8%/0%/4% 6% 115/1/0/56 PHerc0125_3mm_v2 320 40% 50% 22%>10% 12% 33 1526 85%/7%/2%/6% 17% 139/3/0/20 gradients: max 48, clipped 0/50 updates Step 23,450 | last 50 updates / 1,600 crops | lr 3.39e-06 points (35/curve, tolerance 2 vox): 48,525 right / 3,079 wrong | 94.0% correct | 4,396 unknown loss 0.9382 | geometry 0.2960 | confidence 1.2845 | mean error 0.667 vox speed: 1348 ms/update | data wait 1 ms/update | 23.3 crops/s | 23.3 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1485 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 231, from the previous prediction 1254 confidence-labeled crops: terminal 21/1540 (1.4%) | recoverable 104/1540 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 480/1600 (30.0%), 1: 224/1600 (14.0%), 2: 320/1600 (20.0%), 3: 256/1600 (16.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 99 chunks / 198.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 639 rows | mean depth 19 decisions | mean travel 158 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 480 49% 45% 25%>6% 19% 25 992 90%/8%/0%/2% 10% 217/3/0/0 0175A_5mm_v1 224 54% 40% 29%>7% 22% 13 1061 87%/8%/2%/3% 8% 89/1/0/0 1447_5mm_v1 320 49% 42% 25%>10% 16% 27 1277 88%/5%/2%/5% 5% 113/1/0/20 paris4_afv_central 256 50% 45% 25%>5% 21% 13 906 88%/7%/1%/4% 6% 92/0/0/23 PHerc0125_3mm_v2 320 48% 44% 22%>8% 14% 27 1629 87%/5%/2%/6% 15% 128/3/0/13 gradients: max 57, clipped 0/50 updates Step 23,500 | last 50 updates / 1,600 crops | lr 3.34e-06 points (35/curve, tolerance 2 vox): 48,482 right / 2,829 wrong | 94.5% correct | 4,689 unknown loss 0.8914 | geometry 0.2866 | confidence 1.2096 | mean error 0.664 vox speed: 1354 ms/update | data wait 1 ms/update | 23.4 crops/s | 23.4 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1489 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 246, from the previous prediction 1243 confidence-labeled crops: terminal 15/1529 (1.0%) | recoverable 114/1529 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 512/1600 (32.0%), 1: 320/1600 (20.0%), 2: 224/1600 (14.0%), 3: 224/1600 (14.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 99 chunks / 198.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 624 rows | mean depth 16 decisions | mean travel 131 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 512 45% 45% 26%>10% 16% 38 2403 86%/9%/1%/4% 13% 132/1/14/99 0175A_5mm_v1 320 50% 44% 27%>6% 21% 18 875 88%/8%/0%/4% 3% 142/3/0/0 1447_5mm_v1 224 53% 42% 24%>5% 18% 11 749 87%/4%/1%/8% 3% 94/1/0/0 paris4_afv_central 224 51% 49% 21%>0% 21% 0 153 91%/5%/0%/4% 2% 109/4/0/0 PHerc0125_3mm_v2 320 47% 46% 24%>7% 17% 22 1021 86%/7%/2%/5% 14% 147/3/0/0 gradients: max 55, clipped 0/50 updates Step 23,550 | last 50 updates / 1,600 crops | lr 3.29e-06 points (35/curve, tolerance 2 vox): 49,619 right / 2,144 wrong | 95.9% correct | 4,237 unknown loss 0.7316 | geometry 0.2291 | confidence 1.0050 | mean error 0.592 vox speed: 1350 ms/update | data wait 1 ms/update | 23.5 crops/s | 23.5 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1558 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 220, from the previous prediction 1338 confidence-labeled crops: terminal 13/1547 (0.8%) | recoverable 102/1547 (6.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 224/1600 (14.0%), 2: 416/1600 (26.0%), 3: 416/1600 (26.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 99 chunks / 198.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 681 rows | mean depth 18 decisions | mean travel 144 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 52% 47% 25%>1% 24% 3 404 90%/8%/0%/2% 6% 80/1/0/41 0175A_5mm_v1 224 47% 49% 21%>4% 17% 10 697 90%/5%/3%/2% 5% 109/3/0/0 1447_5mm_v1 416 55% 44% 25%>1% 24% 5 301 89%/7%/0%/3% 2% 183/6/0/0 paris4_afv_central 416 51% 49% 27%>0% 27% 0 149 89%/7%/0%/4% 5% 202/5/0/0 PHerc0125_3mm_v2 288 51% 40% 27%>8% 19% 24 1671 86%/4%/1%/8% 18% 107/0/0/9 gradients: max 59, clipped 0/50 updates Step 23,599 | dagger_states: 7477 | dagger_source_step: 23000 | dagger_caches: 4 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/1447_5mm_v1/decisions_023000.npz" | dagger_fibers: 384 | dagger_supply: {"states": 7477, "traces": 384, "fibers": 384, "rows": {"terminal": 198, "premature_stop": 121, "recoverable": 17, "pre_excursion": 187, "ordinary": 6943}, "events": {"terminal": 27, "premature_stop": 121, "recoverable": 9, "pre_excursion": 33, "ordinary": 379}, "supervision": {"following": 7251, "recoverable": 11, "terminal": 198, "unknown": 17}, "departed_traces": 31, "terminal_traces": 27, "stop_reasons": {"after_failure": 24, "confidence": 124, "unannotated": 236}, "traced_voxels": 113451.66331140256, "seconds": 834.1407309807837, "voxels_per_second": 136.01021877687947} | dagger_coverage: {"requested": 384, "seeds": 384, "skipped": [], "epoch": 0} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 0 | dagger_launch_interval: 5000 | dagger_publication_age: 599 | dataset: "1447_5mm_v1" Step 23,600 | last 50 updates / 1,600 crops | lr 3.25e-06 points (35/curve, tolerance 2 vox): 48,092 right / 2,844 wrong | 94.4% correct | 5,064 unknown loss 0.8657 | geometry 0.2662 | confidence 1.1989 | mean error 0.625 vox speed: 1272 ms/update | data wait 1 ms/update | 24.7 crops/s | 24.7 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1503 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 233, from the previous prediction 1270 confidence-labeled crops: terminal 20/1538 (1.3%) | recoverable 107/1538 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 192/1600 (12.0%), 2: 416/1600 (26.0%), 3: 384/1600 (24.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 99 chunks / 198.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 670 rows | mean depth 18 decisions | mean travel 149 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 49% 47% 24%>4% 20% 11 741 90%/8%/0%/2% 10% 150/0/0/0 0175A_5mm_v1 192 46% 51% 24%>4% 20% 7 576 88%/7%/1%/5% 4% 97/0/0/0 1447_5mm_v1 416 48% 46% 25%>6% 18% 25 824 89%/7%/1%/3% 5% 189/2/19/3 paris4_afv_central 384 52% 45% 26%>3% 23% 14 1347 89%/7%/1%/4% 8% 109/2/0/63 PHerc0125_3mm_v2 288 45% 43% 26%>12% 14% 31 1955 82%/6%/3%/9% 17% 125/2/0/0 gradients: max 60, clipped 0/50 updates Step 23,650 | last 50 updates / 1,600 crops | lr 3.20e-06 points (35/curve, tolerance 2 vox): 47,379 right / 2,977 wrong | 94.1% correct | 5,644 unknown loss 0.8356 | geometry 0.2748 | confidence 1.1216 | mean error 0.645 vox speed: 900 ms/update | data wait 1 ms/update | 34.9 crops/s | 34.9 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1493 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 351, from the previous prediction 1142 confidence-labeled crops: terminal 17/1520 (1.1%) | recoverable 110/1520 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 160/1600 (10.0%), 1: 192/1600 (12.0%), 2: 384/1600 (24.0%), 3: 544/1600 (34.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 99 chunks / 198.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 495 rows | mean depth 14 decisions | mean travel 118 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 160 49% 41% 27%>9% 18% 14 1553 94%/5%/0%/1% 11% 66/0/0/0 0175A_5mm_v1 192 42% 39% 36%>19% 17% 28 2291 81%/8%/4%/7% 14% 75/0/0/0 1447_5mm_v1 384 58% 39% 25%>2% 23% 9 925 85%/9%/0%/5% 3% 77/0/0/74 paris4_afv_central 544 52% 44% 26%>4% 22% 26 1821 88%/6%/1%/6% 8% 133/1/0/105 PHerc0125_3mm_v2 320 52% 45% 26%>3% 23% 10 703 87%/7%/0%/6% 13% 144/6/0/0 gradients: max 48, clipped 0/50 updates Step 23,700 | last 50 updates / 1,600 crops | lr 3.15e-06 points (35/curve, tolerance 2 vox): 48,803 right / 2,843 wrong | 94.5% correct | 4,354 unknown loss 0.8402 | geometry 0.2565 | confidence 1.1674 | mean error 0.627 vox speed: 900 ms/update | data wait 1 ms/update | 35.0 crops/s | 35.0 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1520 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 254, from the previous prediction 1266 confidence-labeled crops: terminal 18/1551 (1.2%) | recoverable 104/1551 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 224/1600 (14.0%), 2: 416/1600 (26.0%), 3: 256/1600 (16.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 99 chunks / 198.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 597 rows | mean depth 22 decisions | mean travel 175 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 53% 44% 24%>3% 21% 11 633 90%/9%/0%/1% 9% 183/3/0/0 0175A_5mm_v1 224 54% 39% 24%>7% 17% 12 1153 88%/5%/1%/6% 6% 86/1/20/1 1447_5mm_v1 416 52% 45% 21%>3% 18% 13 710 91%/5%/1%/3% 4% 138/5/0/49 paris4_afv_central 256 55% 43% 28%>2% 26% 10 1055 90%/6%/1%/3% 6% 94/4/0/15 PHerc0125_3mm_v2 288 49% 42% 30%>9% 20% 27 2020 84%/6%/4%/6% 19% 96/0/12/25 gradients: max 49, clipped 0/50 updates Step 23,750 | last 50 updates / 1,600 crops | lr 3.10e-06 points (35/curve, tolerance 2 vox): 48,887 right / 3,141 wrong | 94.0% correct | 3,972 unknown loss 0.8908 | geometry 0.2721 | confidence 1.2374 | mean error 0.663 vox speed: 898 ms/update | data wait 1 ms/update | 35.0 crops/s | 35.0 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1512 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 252, from the previous prediction 1260 confidence-labeled crops: terminal 20/1556 (1.3%) | recoverable 123/1556 (7.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 480/1600 (30.0%), 1: 320/1600 (20.0%), 2: 224/1600 (14.0%), 3: 224/1600 (14.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 99 chunks / 198.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 631 rows | mean depth 15 decisions | mean travel 122 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 480 55% 40% 26%>5% 21% 23 1353 89%/9%/0%/2% 10% 142/2/128/49 0175A_5mm_v1 320 49% 48% 23%>3% 21% 7 518 88%/8%/2%/2% 6% 155/3/0/0 1447_5mm_v1 224 46% 51% 19%>2% 17% 5 454 91%/5%/0%/4% 3% 106/2/0/9 paris4_afv_central 224 56% 37% 29%>7% 22% 14 1641 84%/9%/4%/2% 9% 83/2/12/0 PHerc0125_3mm_v2 352 45% 47% 23%>9% 14% 28 1651 88%/7%/1%/5% 15% 145/2/0/20 gradients: max 45, clipped 0/50 updates Step 23,800 | last 50 updates / 1,600 crops | lr 3.05e-06 points (35/curve, tolerance 2 vox): 47,618 right / 3,137 wrong | 93.8% correct | 5,245 unknown loss 0.8876 | geometry 0.2777 | confidence 1.2197 | mean error 0.671 vox speed: 899 ms/update | data wait 1 ms/update | 34.7 crops/s | 34.7 decisions/s | 1.25 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1498 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 367, from the previous prediction 1131 confidence-labeled crops: terminal 19/1519 (1.3%) | recoverable 96/1519 (6.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 320/1600 (20.0%), 2: 256/1600 (16.0%), 3: 320/1600 (20.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 99 chunks / 198.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 507 rows | mean depth 18 decisions | mean travel 148 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 53% 41% 26%>6% 20% 20 2060 89%/8%/0%/3% 11% 77/0/0/69 0175A_5mm_v1 320 47% 48% 24%>5% 18% 20 2389 88%/4%/2%/6% 6% 49/1/29/104 1447_5mm_v1 256 54% 43% 24%>3% 21% 7 505 88%/7%/1%/5% 3% 111/3/0/0 paris4_afv_central 320 52% 47% 24%>2% 22% 4 483 90%/5%/0%/4% 3% 150/3/0/0 PHerc0125_3mm_v2 352 45% 46% 26%>9% 17% 39 2293 84%/6%/3%/7% 19% 120/2/0/42 gradients: max 52, clipped 0/50 updates Step 23,850 | last 50 updates / 1,600 crops | lr 3.01e-06 points (35/curve, tolerance 2 vox): 47,431 right / 3,220 wrong | 93.6% correct | 5,349 unknown loss 0.9719 | geometry 0.2828 | confidence 1.3782 | mean error 0.685 vox speed: 895 ms/update | data wait 1 ms/update | 35.0 crops/s | 35.0 decisions/s | 1.28 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1492 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 379, from the previous prediction 1113 confidence-labeled crops: terminal 21/1522 (1.4%) | recoverable 100/1522 (6.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 416/1600 (26.0%), 2: 128/1600 (8.0%), 3: 160/1600 (10.0%), 4: 512/1600 (32.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 99 chunks / 198.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 518 rows | mean depth 20 decisions | mean travel 167 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 50% 46% 26%>4% 22% 13 732 90%/8%/1%/2% 8% 178/0/0/0 0175A_5mm_v1 416 51% 42% 27%>7% 20% 26 1598 86%/7%/2%/5% 7% 114/2/0/60 1447_5mm_v1 128 39% 44% 23%>17% 5% 23 2588 88%/7%/2%/3% 10% 35/0/74/21 paris4_afv_central 160 51% 45% 27%>4% 22% 6 1020 86%/9%/1%/4% 6% 72/0/0/0 PHerc0125_3mm_v2 512 46% 49% 24%>5% 20% 28 1484 87%/4%/1%/8% 15% 119/4/0/131 gradients: max 47, clipped 0/50 updates Step 23,900 | last 50 updates / 1,600 crops | lr 2.96e-06 points (35/curve, tolerance 2 vox): 49,123 right / 2,254 wrong | 95.6% correct | 4,623 unknown loss 0.7805 | geometry 0.2461 | confidence 1.0688 | mean error 0.610 vox speed: 897 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1546 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 200, from the previous prediction 1346 confidence-labeled crops: terminal 15/1540 (1.0%) | recoverable 106/1540 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 320/1600 (20.0%), 2: 448/1600 (28.0%), 3: 288/1600 (18.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 99 chunks / 198.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 718 rows | mean depth 19 decisions | mean travel 151 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 47% 53% 22%>0% 22% 0 148 89%/8%/0%/3% 10% 101/2/0/0 0175A_5mm_v1 320 43% 48% 24%>8% 16% 22 1145 87%/5%/2%/5% 6% 155/1/0/0 1447_5mm_v1 448 50% 44% 24%>6% 18% 23 876 90%/6%/1%/3% 5% 185/2/113/14 paris4_afv_central 288 51% 49% 21%>0% 21% 0 153 89%/8%/0%/3% 6% 141/0/0/0 PHerc0125_3mm_v2 352 54% 46% 25%>0% 25% 0 136 89%/7%/1%/4% 11% 136/2/0/26 gradients: max 71, clipped 1/50 updates Step 23,950 | last 50 updates / 1,600 crops | lr 2.91e-06 points (35/curve, tolerance 2 vox): 46,140 right / 3,283 wrong | 93.4% correct | 6,577 unknown loss 0.8790 | geometry 0.2719 | confidence 1.2142 | mean error 0.661 vox speed: 898 ms/update | data wait 1 ms/update | 35.1 crops/s | 35.1 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1473 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 324, from the previous prediction 1149 confidence-labeled crops: terminal 26/1511 (1.7%) | recoverable 109/1511 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 128/1600 (8.0%), 1: 352/1600 (22.0%), 2: 384/1600 (24.0%), 3: 320/1600 (20.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 99 chunks / 198.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 539 rows | mean depth 12 decisions | mean travel 100 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 128 48% 50% 23%>2% 21% 2 394 90%/6%/0%/4% 9% 64/0/0/0 0175A_5mm_v1 352 46% 43% 28%>11% 17% 36 1752 84%/6%/3%/6% 11% 138/1/0/13 1447_5mm_v1 384 43% 47% 22%>10% 12% 35 2477 85%/7%/2%/6% 10% 75/2/0/106 paris4_afv_central 320 53% 44% 27%>2% 25% 11 948 86%/9%/1%/5% 6% 122/0/73/20 PHerc0125_3mm_v2 416 50% 45% 25%>4% 20% 19 952 85%/6%/1%/8% 14% 140/2/0/49 gradients: max 40, clipped 0/50 updates Step 24,000 | last 50 updates / 1,600 crops | lr 2.87e-06 points (35/curve, tolerance 2 vox): 48,321 right / 2,675 wrong | 94.8% correct | 5,004 unknown loss 0.8877 | geometry 0.2846 | confidence 1.2062 | mean error 0.643 vox speed: 894 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1523 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 320, from the previous prediction 1203 confidence-labeled crops: terminal 12/1527 (0.8%) | recoverable 99/1527 (6.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 480/1600 (30.0%), 1: 224/1600 (14.0%), 2: 320/1600 (20.0%), 3: 352/1600 (22.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 99 chunks / 198.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 1079 chunk references / 1 windows live chains: 581 rows | mean depth 20 decisions | mean travel 162 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 480 49% 47% 24%>4% 20% 16 806 89%/8%/0%/3% 10% 208/2/107/16 0175A_5mm_v1 224 53% 45% 25%>2% 23% 4 401 85%/9%/2%/4% 3% 101/0/0/0 1447_5mm_v1 320 53% 46% 24%>1% 23% 3 345 90%/5%/0%/5% 2% 79/2/0/69 paris4_afv_central 352 49% 46% 27%>6% 21% 26 2166 87%/4%/1%/8% 9% 83/2/55/78 PHerc0125_3mm_v2 224 40% 49% 27%>11% 16% 24 1775 87%/5%/2%/6% 19% 110/2/0/0 gradients: max 75, clipped 1/50 updates Step 24,000 | dagger_launched: true Step 24,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 248, "stopped": 1, "mean_example_error": 1.0108891201430354, "geometry_examples": 29, "examples": 32, "inference_and_measurement_seconds": 2.009846493601799, "total_diagnostic_seconds": 4.138042479753494} Step 24,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 6/8 (75.0%) | false stops 1 1.5-3: recovered 6/8 (75.0%) | false stops 1 3-4.5: recovered 6/8 (75.0%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 0 entire proposed 8-point prefix correct (distance): 22/28 (78.6%) gate @ 0.50: false stops 0/23 (0.0%) | accepted wrong 5/27 (18.5%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.605 1.595 following 16 464 1.157 1.184 recoverable 12 348 2.204 2.144 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 1.157 1.184 332 32 812 1.605 1.595 improved/worsened/compared states: 1/2/28 Step 24,000 | batch_diagnostics_seconds: 4.1409726440906525 | recovery_seconds: 14.7688562348485 Step 24,050 | last 50 updates / 1,600 crops | lr 2.82e-06 points (35/curve, tolerance 2 vox): 47,558 right / 3,334 wrong | 93.4% correct | 5,108 unknown loss 0.9250 | geometry 0.3202 | confidence 1.2097 | mean error 0.700 vox speed: 1332 ms/update | data wait 1 ms/update | 18.2 crops/s | 18.2 decisions/s | 1.27 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1495 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 310, from the previous prediction 1185 confidence-labeled crops: terminal 20/1539 (1.3%) | recoverable 118/1539 (7.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 256/1600 (16.0%), 2: 192/1600 (12.0%), 3: 448/1600 (28.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 99 chunks / 198.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 548 rows | mean depth 19 decisions | mean travel 155 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 56% 41% 27%>3% 24% 24 2339 88%/9%/0%/3% 11% 80/0/0/104 0175A_5mm_v1 256 46% 43% 29%>11% 18% 22 1560 86%/7%/3%/4% 8% 109/1/0/0 1447_5mm_v1 192 50% 46% 21%>4% 17% 7 639 91%/8%/0%/2% 4% 88/0/0/0 paris4_afv_central 448 51% 46% 22%>3% 19% 14 548 87%/7%/1%/5% 6% 151/3/0/53 PHerc0125_3mm_v2 256 44% 47% 23%>9% 13% 22 1623 85%/6%/4%/5% 18% 120/1/0/0 gradients: max 50, clipped 0/50 updates Step 24,100 | last 50 updates / 1,600 crops | lr 2.77e-06 points (35/curve, tolerance 2 vox): 47,178 right / 3,213 wrong | 93.6% correct | 5,609 unknown loss 0.8905 | geometry 0.2722 | confidence 1.2367 | mean error 0.645 vox speed: 1375 ms/update | data wait 1 ms/update | 23.0 crops/s | 23.0 decisions/s | 1.24 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1474 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 285, from the previous prediction 1189 confidence-labeled crops: terminal 30/1528 (2.0%) | recoverable 120/1528 (7.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 256/1600 (16.0%), 2: 352/1600 (22.0%), 3: 416/1600 (26.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 99 chunks / 198.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 587 rows | mean depth 15 decisions | mean travel 127 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 51% 47% 24%>2% 22% 8 1100 89%/8%/0%/3% 10% 72/1/0/79 0175A_5mm_v1 256 52% 41% 26%>8% 18% 17 1289 86%/7%/2%/5% 7% 104/4/0/0 1447_5mm_v1 352 43% 46% 25%>11% 14% 32 1339 84%/7%/3%/6% 9% 161/6/27/1 paris4_afv_central 416 49% 46% 25%>6% 20% 28 1074 87%/7%/1%/5% 7% 146/1/0/44 PHerc0125_3mm_v2 256 48% 41% 28%>11% 16% 27 1828 82%/9%/2%/6% 16% 104/0/0/0 gradients: max 47, clipped 0/50 updates Step 24,150 | last 50 updates / 1,600 crops | lr 2.73e-06 points (35/curve, tolerance 2 vox): 48,912 right / 2,297 wrong | 95.5% correct | 4,791 unknown loss 0.7393 | geometry 0.2422 | confidence 0.9941 | mean error 0.609 vox speed: 1390 ms/update | data wait 1 ms/update | 22.8 crops/s | 22.8 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1543 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 265, from the previous prediction 1278 confidence-labeled crops: terminal 18/1540 (1.2%) | recoverable 89/1540 (5.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 352/1600 (22.0%), 2: 192/1600 (12.0%), 3: 256/1600 (16.0%), 4: 512/1600 (32.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 100 chunks / 200.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 650 rows | mean depth 18 decisions | mean travel 146 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 55% 43% 26%>2% 24% 5 454 91%/8%/0%/1% 5% 124/0/0/0 0175A_5mm_v1 352 45% 52% 21%>4% 17% 12 650 92%/5%/1%/3% 4% 172/2/43/10 1447_5mm_v1 192 48% 48% 24%>3% 21% 6 509 89%/6%/1%/4% 2% 93/1/0/0 paris4_afv_central 256 54% 42% 27%>4% 23% 10 602 90%/5%/2%/3% 7% 108/2/0/0 PHerc0125_3mm_v2 512 49% 48% 23%>3% 20% 21 1266 87%/5%/2%/6% 11% 153/4/128/93 gradients: max 54, clipped 0/50 updates Step 24,200 | last 50 updates / 1,600 crops | lr 2.68e-06 points (35/curve, tolerance 2 vox): 48,098 right / 2,757 wrong | 94.6% correct | 5,145 unknown loss 0.8928 | geometry 0.2547 | confidence 1.2761 | mean error 0.628 vox speed: 1387 ms/update | data wait 1 ms/update | 22.9 crops/s | 22.9 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1483 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 261, from the previous prediction 1222 confidence-labeled crops: terminal 23/1525 (1.5%) | recoverable 121/1525 (7.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 320/1600 (20.0%), 2: 288/1600 (18.0%), 3: 352/1600 (22.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 100 chunks / 200.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 586 rows | mean depth 20 decisions | mean travel 161 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 51% 43% 27%>6% 21% 16 1078 89%/8%/0%/2% 11% 139/1/0/0 0175A_5mm_v1 320 46% 42% 31%>12% 19% 35 2830 81%/9%/3%/7% 8% 79/1/0/55 1447_5mm_v1 288 48% 41% 25%>11% 14% 29 1739 86%/6%/3%/5% 7% 97/3/61/22 paris4_afv_central 352 57% 43% 26%>0% 26% 0 142 88%/8%/0%/4% 3% 152/2/0/0 PHerc0125_3mm_v2 320 53% 42% 23%>4% 19% 14 1147 87%/6%/2%/6% 14% 119/1/0/17 gradients: max 57, clipped 0/50 updates Step 24,250 | last 50 updates / 1,600 crops | lr 2.64e-06 points (35/curve, tolerance 2 vox): 47,814 right / 2,508 wrong | 95.0% correct | 5,678 unknown loss 0.8114 | geometry 0.2347 | confidence 1.1535 | mean error 0.611 vox speed: 1378 ms/update | data wait 1 ms/update | 22.9 crops/s | 22.9 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1522 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 311, from the previous prediction 1211 confidence-labeled crops: terminal 19/1526 (1.2%) | recoverable 85/1526 (5.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 320/1600 (20.0%), 2: 384/1600 (24.0%), 3: 352/1600 (22.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 100 chunks / 200.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 583 rows | mean depth 13 decisions | mean travel 112 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 51% 47% 26%>2% 23% 4 487 88%/8%/0%/5% 7% 90/0/0/0 0175A_5mm_v1 320 49% 45% 28%>5% 23% 17 1629 89%/4%/1%/6% 6% 64/0/0/81 1447_5mm_v1 384 48% 47% 22%>5% 16% 18 804 89%/6%/1%/4% 5% 172/5/0/9 paris4_afv_central 352 47% 48% 26%>4% 21% 15 562 89%/5%/3%/3% 9% 168/3/60/2 PHerc0125_3mm_v2 352 51% 44% 21%>5% 16% 19 1936 88%/4%/1%/7% 14% 89/2/0/65 gradients: max 56, clipped 0/50 updates Step 24,300 | last 50 updates / 1,600 crops | lr 2.59e-06 points (35/curve, tolerance 2 vox): 47,755 right / 3,049 wrong | 94.0% correct | 5,196 unknown loss 0.9257 | geometry 0.3058 | confidence 1.2398 | mean error 0.666 vox speed: 1367 ms/update | data wait 1 ms/update | 23.1 crops/s | 23.1 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1500 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 282, from the previous prediction 1218 confidence-labeled crops: terminal 15/1529 (1.0%) | recoverable 118/1529 (7.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 288/1600 (18.0%), 2: 224/1600 (14.0%), 3: 480/1600 (30.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 100 chunks / 200.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 611 rows | mean depth 20 decisions | mean travel 162 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 54% 44% 24%>2% 22% 6 574 91%/7%/0%/2% 10% 126/0/0/0 0175A_5mm_v1 288 48% 47% 24%>6% 18% 15 953 89%/7%/1%/3% 8% 134/4/0/0 1447_5mm_v1 224 48% 48% 24%>4% 20% 10 1031 90%/6%/0%/4% 4% 70/0/0/37 paris4_afv_central 480 48% 44% 27%>8% 19% 41 1735 83%/9%/2%/6% 10% 126/2/0/87 PHerc0125_3mm_v2 320 46% 48% 28%>5% 22% 15 1043 86%/6%/1%/8% 13% 155/4/0/0 gradients: max 53, clipped 0/50 updates Step 24,350 | last 50 updates / 1,600 crops | lr 2.55e-06 points (35/curve, tolerance 2 vox): 49,324 right / 2,592 wrong | 95.0% correct | 4,084 unknown loss 0.7995 | geometry 0.2310 | confidence 1.1370 | mean error 0.612 vox speed: 1377 ms/update | data wait 1 ms/update | 23.0 crops/s | 23.0 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1521 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 243, from the previous prediction 1278 confidence-labeled crops: terminal 25/1553 (1.6%) | recoverable 115/1553 (7.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 320/1600 (20.0%), 2: 288/1600 (18.0%), 3: 320/1600 (20.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 101 chunks / 202.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 627 rows | mean depth 20 decisions | mean travel 161 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 53% 46% 25%>1% 24% 5 428 89%/8%/0%/2% 6% 135/0/128/69 0175A_5mm_v1 320 57% 42% 30%>1% 28% 4 368 90%/7%/1%/2% 4% 134/1/0/0 1447_5mm_v1 288 45% 47% 23%>8% 15% 21 1118 86%/8%/3%/3% 8% 134/0/0/1 paris4_afv_central 320 48% 44% 30%>8% 22% 25 1629 85%/8%/2%/5% 7% 125/2/0/15 PHerc0125_3mm_v2 224 46% 45% 24%>8% 16% 19 1284 88%/3%/3%/5% 14% 99/2/46/2 gradients: max 49, clipped 0/50 updates Step 24,400 | last 50 updates / 1,600 crops | lr 2.51e-06 points (35/curve, tolerance 2 vox): 48,045 right / 2,970 wrong | 94.2% correct | 4,985 unknown loss 0.9087 | geometry 0.2595 | confidence 1.2985 | mean error 0.626 vox speed: 1346 ms/update | data wait 1 ms/update | 23.5 crops/s | 23.5 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1489 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 185, from the previous prediction 1304 confidence-labeled crops: terminal 24/1547 (1.6%) | recoverable 89/1547 (5.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 288/1600 (18.0%), 2: 256/1600 (16.0%), 3: 384/1600 (24.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 101 chunks / 202.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 679 rows | mean depth 12 decisions | mean travel 104 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 56% 42% 29%>2% 27% 6 641 91%/7%/0%/2% 8% 92/0/0/30 0175A_5mm_v1 288 46% 46% 25%>8% 17% 20 1245 90%/5%/3%/3% 6% 129/1/48/3 1447_5mm_v1 256 44% 46% 23%>10% 14% 21 1227 89%/5%/2%/4% 8% 118/3/0/0 paris4_afv_central 384 47% 44% 24%>9% 15% 34 1062 89%/6%/1%/4% 7% 164/5/0/4 PHerc0125_3mm_v2 384 48% 46% 25%>6% 20% 21 815 89%/5%/2%/4% 15% 176/6/10/0 gradients: max 39, clipped 0/50 updates Step 24,424 | dagger_states: 5656 | dagger_source_step: 24000 | dagger_caches: 4 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/paris4_afv_central/decisions_024000.npz" | dagger_fibers: 384 | dagger_supply: {"states": 5656, "traces": 384, "fibers": 384, "rows": {"terminal": 248, "premature_stop": 170, "recoverable": 50, "pre_excursion": 235, "ordinary": 4943}, "events": {"terminal": 36, "premature_stop": 170, "recoverable": 16, "pre_excursion": 45, "ordinary": 366}, "supervision": {"following": 5347, "recoverable": 44, "terminal": 248, "unknown": 17}, "departed_traces": 42, "terminal_traces": 36, "stop_reasons": {"after_failure": 30, "confidence": 176, "unannotated": 178}, "traced_voxels": 82574.24701987025, "seconds": 597.5727325156331, "voxels_per_second": 138.1827558835443} | dagger_coverage: {"requested": 384, "seeds": 384, "skipped": [], "epoch": 0} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 0 | dagger_launch_interval: 5000 | dagger_publication_age: 424 | dataset: "paris4_afv_central" Step 24,450 | last 50 updates / 1,600 crops | lr 2.46e-06 points (35/curve, tolerance 2 vox): 47,320 right / 3,044 wrong | 94.0% correct | 5,636 unknown loss 0.8653 | geometry 0.2619 | confidence 1.2068 | mean error 0.639 vox speed: 1080 ms/update | data wait 1 ms/update | 29.1 crops/s | 29.1 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1458 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 333, from the previous prediction 1125 confidence-labeled crops: terminal 26/1517 (1.7%) | recoverable 113/1517 (7.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 448/1600 (28.0%), 2: 448/1600 (28.0%), 3: 192/1600 (12.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 101 chunks / 202.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 565 rows | mean depth 21 decisions | mean travel 171 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 49% 47% 23%>4% 19% 11 1360 89%/7%/0%/3% 12% 92/0/0/28 0175A_5mm_v1 448 46% 47% 24%>8% 17% 35 1859 85%/8%/2%/5% 8% 139/2/42/71 1447_5mm_v1 448 41% 51% 21%>8% 13% 35 1566 87%/7%/1%/5% 6% 154/2/0/75 paris4_afv_central 192 53% 40% 24%>8% 16% 13 1461 87%/6%/1%/6% 9% 76/2/0/0 PHerc0125_3mm_v2 256 46% 45% 28%>9% 19% 21 1232 82%/6%/3%/9% 16% 104/1/12/11 gradients: max 40, clipped 0/50 updates Step 24,500 | last 50 updates / 1,600 crops | lr 2.42e-06 points (35/curve, tolerance 2 vox): 48,495 right / 2,658 wrong | 94.8% correct | 4,847 unknown loss 0.8027 | geometry 0.2248 | confidence 1.1558 | mean error 0.605 vox speed: 896 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1507 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 296, from the previous prediction 1211 confidence-labeled crops: terminal 24/1534 (1.6%) | recoverable 115/1534 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 416/1600 (26.0%), 2: 192/1600 (12.0%), 3: 416/1600 (26.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 101 chunks / 202.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 575 rows | mean depth 20 decisions | mean travel 166 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 55% 45% 23%>0% 23% 0 158 90%/8%/0%/2% 9% 131/0/0/0 0175A_5mm_v1 416 48% 42% 29%>10% 19% 30 2480 85%/8%/4%/3% 8% 98/1/0/76 1447_5mm_v1 192 48% 40% 28%>11% 17% 20 1560 87%/7%/2%/5% 4% 77/1/0/0 paris4_afv_central 416 50% 45% 23%>4% 19% 15 610 88%/6%/1%/5% 4% 189/1/0/0 PHerc0125_3mm_v2 288 53% 45% 25%>2% 23% 7 481 86%/7%/1%/6% 9% 80/2/0/50 gradients: max 43, clipped 0/50 updates Step 24,550 | last 50 updates / 1,600 crops | lr 2.38e-06 points (35/curve, tolerance 2 vox): 49,291 right / 2,318 wrong | 95.5% correct | 4,391 unknown loss 0.7854 | geometry 0.2507 | confidence 1.0695 | mean error 0.607 vox speed: 897 ms/update | data wait 1 ms/update | 35.1 crops/s | 35.1 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1550 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 193, from the previous prediction 1357 confidence-labeled crops: terminal 16/1552 (1.0%) | recoverable 106/1552 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 192/1600 (12.0%), 2: 416/1600 (26.0%), 3: 352/1600 (22.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 101 chunks / 202.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 711 rows | mean depth 16 decisions | mean travel 134 vox | 8 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 52% 47% 26%>1% 25% 3 345 89%/9%/0%/2% 7% 121/0/0/0 0175A_5mm_v1 192 55% 45% 26%>0% 26% 0 140 89%/6%/1%/4% 3% 81/3/0/6 1447_5mm_v1 416 46% 48% 21%>6% 16% 20 839 89%/6%/2%/2% 6% 199/1/0/2 paris4_afv_central 352 54% 42% 26%>4% 22% 13 633 87%/8%/1%/4% 7% 138/5/96/10 PHerc0125_3mm_v2 384 53% 45% 22%>2% 20% 7 359 90%/4%/1%/5% 11% 172/2/0/0 gradients: max 34, clipped 0/50 updates Step 24,600 | last 50 updates / 1,600 crops | lr 2.34e-06 points (35/curve, tolerance 2 vox): 48,647 right / 2,252 wrong | 95.6% correct | 5,101 unknown loss 0.8040 | geometry 0.2453 | confidence 1.1174 | mean error 0.600 vox speed: 895 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1541 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 294, from the previous prediction 1247 confidence-labeled crops: terminal 14/1537 (0.9%) | recoverable 117/1537 (7.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 192/1600 (12.0%), 2: 448/1600 (28.0%), 3: 256/1600 (16.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 103 chunks / 206.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 593 rows | mean depth 17 decisions | mean travel 140 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 53% 46% 24%>1% 23% 4 376 88%/9%/0%/3% 7% 141/0/81/51 0175A_5mm_v1 192 54% 45% 25%>2% 23% 2 418 85%/9%/1%/4% 6% 86/1/0/0 1447_5mm_v1 448 52% 41% 25%>8% 17% 28 1148 88%/7%/1%/4% 6% 183/2/0/0 paris4_afv_central 256 54% 42% 29%>4% 26% 9 592 87%/7%/2%/5% 9% 99/0/0/9 PHerc0125_3mm_v2 288 53% 44% 26%>3% 23% 9 562 87%/6%/2%/6% 12% 84/1/0/42 gradients: max 49, clipped 0/50 updates Step 24,650 | last 50 updates / 1,600 crops | lr 2.29e-06 points (35/curve, tolerance 2 vox): 47,682 right / 2,885 wrong | 94.3% correct | 5,433 unknown loss 0.8792 | geometry 0.2663 | confidence 1.2257 | mean error 0.633 vox speed: 901 ms/update | data wait 1 ms/update | 34.8 crops/s | 34.8 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1499 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 280, from the previous prediction 1219 confidence-labeled crops: terminal 21/1515 (1.4%) | recoverable 114/1515 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 256/1600 (16.0%), 2: 288/1600 (18.0%), 3: 320/1600 (20.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 103 chunks / 206.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 612 rows | mean depth 19 decisions | mean travel 152 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 51% 45% 25%>3% 21% 10 761 88%/9%/0%/2% 9% 126/2/0/4 0175A_5mm_v1 256 48% 44% 27%>7% 19% 18 1256 85%/5%/2%/7% 8% 113/3/0/0 1447_5mm_v1 288 51% 45% 25%>4% 21% 14 1547 85%/9%/1%/5% 6% 61/0/0/68 paris4_afv_central 320 46% 49% 20%>5% 15% 17 854 87%/7%/1%/5% 6% 156/1/0/0 PHerc0125_3mm_v2 448 48% 45% 25%>7% 18% 37 1379 86%/6%/2%/7% 13% 156/0/0/47 gradients: max 50, clipped 0/50 updates Step 24,700 | last 50 updates / 1,600 crops | lr 2.25e-06 points (35/curve, tolerance 2 vox): 47,928 right / 2,735 wrong | 94.6% correct | 5,337 unknown loss 0.8193 | geometry 0.2497 | confidence 1.1393 | mean error 0.627 vox speed: 893 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1523 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 283, from the previous prediction 1240 confidence-labeled crops: terminal 20/1530 (1.3%) | recoverable 117/1530 (7.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 192/1600 (12.0%), 2: 256/1600 (16.0%), 3: 480/1600 (30.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 103 chunks / 206.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 614 rows | mean depth 19 decisions | mean travel 156 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 54% 46% 23%>0% 23% 0 133 88%/10%/0%/3% 10% 149/0/0/42 0175A_5mm_v1 192 51% 46% 24%>3% 22% 4 569 90%/8%/0%/3% 3% 89/1/0/0 1447_5mm_v1 256 43% 48% 25%>9% 16% 21 1237 85%/7%/4%/4% 8% 120/2/0/2 paris4_afv_central 480 49% 44% 23%>7% 16% 29 1265 86%/6%/1%/6% 7% 133/4/0/77 PHerc0125_3mm_v2 256 46% 48% 23%>5% 18% 14 868 86%/6%/1%/7% 11% 123/4/0/0 gradients: max 55, clipped 0/50 updates Step 24,750 | last 50 updates / 1,600 crops | lr 2.21e-06 points (35/curve, tolerance 2 vox): 46,746 right / 3,439 wrong | 93.1% correct | 5,815 unknown loss 0.8352 | geometry 0.2803 | confidence 1.1100 | mean error 0.635 vox speed: 893 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1456 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 283, from the previous prediction 1173 confidence-labeled crops: terminal 38/1521 (2.5%) | recoverable 106/1521 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 128/1600 (8.0%), 1: 192/1600 (12.0%), 2: 608/1600 (38.0%), 3: 224/1600 (14.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 103 chunks / 206.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 587 rows | mean depth 13 decisions | mean travel 110 vox | 6 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 128 57% 43% 28%>0% 28% 0 140 91%/6%/0%/3% 8% 31/0/0/24 0175A_5mm_v1 192 43% 42% 28%>16% 12% 26 2235 81%/7%/3%/9% 11% 80/1/0/0 1447_5mm_v1 608 43% 45% 28%>12% 17% 47 1816 84%/8%/4%/5% 9% 236/6/2/38 paris4_afv_central 224 51% 41% 25%>8% 18% 14 1166 85%/5%/3%/7% 11% 76/2/0/16 PHerc0125_3mm_v2 448 49% 48% 25%>4% 21% 17 733 90%/6%/1%/4% 11% 164/3/0/49 gradients: max 37, clipped 0/50 updates Step 24,800 | last 50 updates / 1,600 crops | lr 2.17e-06 points (35/curve, tolerance 2 vox): 48,123 right / 2,642 wrong | 94.8% correct | 5,235 unknown loss 0.8195 | geometry 0.2348 | confidence 1.1694 | mean error 0.619 vox speed: 897 ms/update | data wait 1 ms/update | 35.0 crops/s | 35.0 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1500 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 309, from the previous prediction 1191 confidence-labeled crops: terminal 22/1531 (1.4%) | recoverable 121/1531 (7.9%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 224/1600 (14.0%), 2: 480/1600 (30.0%), 3: 160/1600 (10.0%), 4: 512/1600 (32.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 103 chunks / 206.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 555 rows | mean depth 17 decisions | mean travel 143 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 48% 45% 24%>8% 17% 16 1425 92%/5%/0%/2% 12% 100/0/0/0 0175A_5mm_v1 224 40% 53% 21%>7% 14% 15 1064 85%/8%/2%/5% 6% 118/1/0/0 1447_5mm_v1 480 51% 42% 31%>6% 24% 30 2003 85%/8%/2%/4% 5% 102/1/0/102 paris4_afv_central 160 49% 48% 24%>2% 22% 4 447 88%/8%/1%/4% 3% 77/1/0/0 PHerc0125_3mm_v2 512 52% 44% 26%>4% 22% 23 999 85%/8%/1%/6% 11% 158/1/0/67 gradients: max 42, clipped 0/50 updates Step 24,850 | last 50 updates / 1,600 crops | lr 2.13e-06 points (35/curve, tolerance 2 vox): 47,951 right / 3,194 wrong | 93.8% correct | 4,855 unknown loss 0.8605 | geometry 0.2650 | confidence 1.1909 | mean error 0.638 vox speed: 897 ms/update | data wait 1 ms/update | 35.1 crops/s | 35.1 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1481 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 190, from the previous prediction 1291 confidence-labeled crops: terminal 27/1527 (1.8%) | recoverable 116/1527 (7.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 320/1600 (20.0%), 2: 192/1600 (12.0%), 3: 448/1600 (28.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 103 chunks / 206.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 684 rows | mean depth 20 decisions | mean travel 166 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 53% 46% 21%>1% 20% 4 323 91%/7%/0%/2% 7% 207/2/0/0 0175A_5mm_v1 320 42% 42% 31%>17% 14% 39 2636 81%/8%/2%/8% 11% 111/0/78/22 1447_5mm_v1 192 40% 46% 22%>14% 8% 23 1654 85%/6%/4%/5% 11% 89/0/0/0 paris4_afv_central 448 44% 48% 26%>8% 18% 31 979 85%/8%/2%/4% 9% 213/5/5/0 PHerc0125_3mm_v2 192 63% 37% 34%>0% 34% 0 136 88%/5%/1%/6% 12% 64/2/0/7 gradients: max 49, clipped 0/50 updates Step 24,900 | last 50 updates / 1,600 crops | lr 2.09e-06 points (35/curve, tolerance 2 vox): 47,526 right / 3,232 wrong | 93.6% correct | 5,242 unknown loss 0.9082 | geometry 0.3010 | confidence 1.2143 | mean error 0.650 vox speed: 896 ms/update | data wait 1 ms/update | 35.1 crops/s | 35.1 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1485 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 235, from the previous prediction 1250 confidence-labeled crops: terminal 25/1541 (1.6%) | recoverable 120/1541 (7.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 320/1600 (20.0%), 2: 320/1600 (20.0%), 3: 320/1600 (20.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 103 chunks / 206.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 619 rows | mean depth 14 decisions | mean travel 117 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 51% 44% 24%>6% 19% 10 1218 93%/6%/0%/1% 8% 79/0/88/5 0175A_5mm_v1 320 54% 40% 26%>6% 21% 19 1568 87%/7%/2%/4% 8% 97/0/50/32 1447_5mm_v1 320 41% 48% 23%>11% 12% 28 1385 84%/9%/2%/4% 9% 155/3/0/0 paris4_afv_central 320 48% 46% 23%>6% 17% 19 1354 84%/8%/2%/5% 12% 101/3/0/47 PHerc0125_3mm_v2 448 54% 42% 27%>4% 22% 19 613 87%/7%/1%/5% 12% 187/7/0/0 gradients: max 70, clipped 1/50 updates Step 24,950 | last 50 updates / 1,600 crops | lr 2.05e-06 points (35/curve, tolerance 2 vox): 48,612 right / 2,835 wrong | 94.5% correct | 4,553 unknown loss 0.8414 | geometry 0.2525 | confidence 1.1777 | mean error 0.625 vox speed: 896 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1515 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 281, from the previous prediction 1234 confidence-labeled crops: terminal 19/1537 (1.2%) | recoverable 95/1537 (6.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 320/1600 (20.0%), 2: 288/1600 (18.0%), 3: 320/1600 (20.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 103 chunks / 206.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 588 rows | mean depth 19 decisions | mean travel 157 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 51% 49% 20%>0% 20% 1 199 93%/5%/0%/3% 9% 129/0/1/73 0175A_5mm_v1 320 50% 45% 22%>5% 18% 14 983 90%/4%/2%/4% 5% 132/2/0/13 1447_5mm_v1 288 45% 44% 25%>10% 15% 25 1414 86%/7%/2%/5% 7% 128/0/0/0 paris4_afv_central 320 50% 40% 30%>10% 21% 33 2000 82%/9%/2%/7% 10% 82/3/0/46 PHerc0125_3mm_v2 256 54% 46% 26%>0% 26% 0 146 93%/5%/0%/2% 10% 117/3/0/0 gradients: max 69, clipped 1/50 updates Step 25,000 | last 50 updates / 1,600 crops | lr 2.01e-06 points (35/curve, tolerance 2 vox): 47,903 right / 2,915 wrong | 94.3% correct | 5,182 unknown loss 0.8586 | geometry 0.2544 | confidence 1.2085 | mean error 0.634 vox speed: 894 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1476 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 182, from the previous prediction 1294 confidence-labeled crops: terminal 29/1528 (1.9%) | recoverable 102/1528 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 320/1600 (20.0%), 2: 256/1600 (16.0%), 3: 416/1600 (26.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 103 chunks / 206.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 1584 chunk references / 1 windows live chains: 679 rows | mean depth 19 decisions | mean travel 158 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 53% 45% 22%>2% 20% 5 487 94%/5%/0%/1% 7% 124/0/0/7 0175A_5mm_v1 320 45% 48% 26%>7% 19% 20 1469 86%/8%/2%/5% 6% 117/2/0/38 1447_5mm_v1 256 45% 43% 25%>12% 13% 26 1664 86%/7%/3%/5% 6% 109/2/0/0 paris4_afv_central 416 41% 46% 26%>13% 13% 47 1574 84%/7%/3%/6% 9% 192/5/0/0 PHerc0125_3mm_v2 320 52% 45% 26%>3% 22% 11 506 88%/5%/1%/6% 14% 137/3/93/6 gradients: max 47, clipped 0/50 updates Step 25,000 | dagger_launched: true Step 25,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 256, "stopped": 0, "mean_example_error": 0.7138415677794094, "geometry_examples": 29, "examples": 32, "inference_and_measurement_seconds": 2.129333633929491, "total_diagnostic_seconds": 4.087032720446587} Step 25,000 | monitor rollout @ 0.50 coverage 71.8% | precision 85.9% | diverged 18.8% Step 25,000 | monitor rollout @ 0.50 coverage 90.5% | precision 100.0% | diverged 0.0% Step 25,000 | monitor rollout @ 0.50 coverage 94.3% | precision 100.0% | diverged 0.0% Step 25,000 | monitor rollout @ 0.50 coverage 88.5% | precision 97.2% | diverged 6.2% Step 25,000 | monitor rollout @ 0.50 coverage 70.6% | precision 84.3% | diverged 9.4% Step 25,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 6/8 (75.0%) | false stops 1 1.5-3: recovered 6/8 (75.0%) | false stops 1 3-4.5: recovered 6/8 (75.0%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 0 entire proposed 8-point prefix correct (distance): 22/28 (78.6%) gate @ 0.50: false stops 0/23 (0.0%) | accepted wrong 5/27 (18.5%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.587 1.562 following 16 464 1.148 1.173 recoverable 12 348 2.173 2.080 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 1.148 1.173 332 32 812 1.587 1.562 improved/worsened/compared states: 3/2/28 Step 25,000 | batch_diagnostics_seconds: 4.095002267509699 | diagnostics_seconds: 569.3118962943554 | recovery_seconds: 13.83505753800273 Step 25,001 | dagger_states: 3217 | dagger_source_step: 25000 | dagger_caches: 4 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/PHerc0125_3mm_v2/decisions_025000.npz" | dagger_fibers: 384 | dagger_supply: {"states": 3217, "traces": 384, "fibers": 384, "rows": {"terminal": 310, "premature_stop": 263, "recoverable": 43, "pre_excursion": 377, "ordinary": 2205}, "events": {"terminal": 50, "premature_stop": 263, "recoverable": 25, "pre_excursion": 78, "ordinary": 352}, "supervision": {"following": 2839, "recoverable": 44, "terminal": 310, "unknown": 24}, "departed_traces": 66, "terminal_traces": 50, "stop_reasons": {"after_failure": 27, "confidence": 291, "unannotated": 66}, "traced_voxels": 40171.62067957683, "seconds": 227.68981247022748, "voxels_per_second": 176.43134861305944} | dagger_coverage: {"requested": 384, "seeds": 384, "skipped": [], "epoch": 0} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 0 | dagger_launch_interval: 5000 | dagger_publication_age: 1 | dataset: "PHerc0125_3mm_v2" Step 25,050 | last 50 updates / 1,600 crops | lr 1.97e-06 points (35/curve, tolerance 2 vox): 47,422 right / 3,605 wrong | 92.9% correct | 4,973 unknown loss 0.9596 | geometry 0.2669 | confidence 1.3852 | mean error 0.647 vox speed: 889 ms/update | data wait 1 ms/update | 2.5 crops/s | 2.5 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1470 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 252, from the previous prediction 1218 confidence-labeled crops: terminal 40/1551 (2.6%) | recoverable 115/1551 (7.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 320/1600 (20.0%), 2: 256/1600 (16.0%), 3: 352/1600 (22.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 103 chunks / 206.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 640 rows | mean depth 13 decisions | mean travel 109 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 55% 43% 28%>2% 26% 13 1043 91%/8%/0%/1% 8% 124/0/0/13 0175A_5mm_v1 320 42% 48% 22%>11% 12% 28 2169 86%/6%/4%/4% 11% 122/4/0/30 1447_5mm_v1 256 42% 43% 28%>16% 12% 29 1994 84%/7%/5%/3% 11% 109/2/0/0 paris4_afv_central 352 43% 50% 24%>6% 18% 21 1064 88%/7%/2%/3% 8% 150/4/0/27 PHerc0125_3mm_v2 352 49% 48% 26%>3% 22% 12 543 85%/8%/2%/5% 13% 135/2/0/34 gradients: max 41, clipped 0/50 updates Step 25,100 | last 50 updates / 1,600 crops | lr 1.93e-06 points (35/curve, tolerance 2 vox): 46,932 right / 3,339 wrong | 93.4% correct | 5,729 unknown loss 0.9007 | geometry 0.2878 | confidence 1.2260 | mean error 0.627 vox speed: 906 ms/update | data wait 1 ms/update | 34.8 crops/s | 34.8 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1447 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 218, from the previous prediction 1229 confidence-labeled crops: terminal 36/1526 (2.4%) | recoverable 117/1526 (7.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 320/1600 (20.0%), 2: 224/1600 (14.0%), 3: 416/1600 (26.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 103 chunks / 206.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 652 rows | mean depth 17 decisions | mean travel 142 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 54% 46% 21%>0% 21% 0 141 92%/6%/0%/2% 9% 102/1/0/0 0175A_5mm_v1 320 44% 45% 25%>12% 13% 27 1793 84%/5%/4%/7% 7% 140/6/0/3 1447_5mm_v1 224 46% 46% 30%>8% 22% 16 1211 85%/10%/2%/4% 6% 102/1/0/0 paris4_afv_central 416 37% 47% 26%>16% 10% 49 2177 84%/8%/4%/3% 16% 141/0/0/56 PHerc0125_3mm_v2 416 50% 44% 25%>6% 20% 21 763 84%/7%/1%/8% 12% 167/1/0/15 gradients: max 41, clipped 0/50 updates Step 25,150 | last 50 updates / 1,600 crops | lr 1.89e-06 points (35/curve, tolerance 2 vox): 46,856 right / 3,169 wrong | 93.7% correct | 5,975 unknown loss 0.8445 | geometry 0.2598 | confidence 1.1693 | mean error 0.612 vox speed: 896 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1446 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 275, from the previous prediction 1171 confidence-labeled crops: terminal 35/1517 (2.3%) | recoverable 100/1517 (6.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 448/1600 (28.0%), 2: 352/1600 (22.0%), 3: 160/1600 (10.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 103 chunks / 206.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 578 rows | mean depth 19 decisions | mean travel 158 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 55% 44% 28%>1% 27% 3 383 90%/7%/0%/3% 6% 113/1/0/0 0175A_5mm_v1 448 48% 43% 27%>9% 17% 29 2105 86%/7%/2%/5% 6% 132/1/0/62 1447_5mm_v1 352 44% 44% 21%>12% 9% 35 1653 86%/5%/4%/6% 8% 154/4/0/0 paris4_afv_central 160 40% 44% 27%>16% 11% 23 1703 88%/4%/3%/4% 9% 71/1/0/0 PHerc0125_3mm_v2 384 47% 44% 23%>10% 13% 40 1709 83%/8%/3%/7% 18% 108/2/0/60 gradients: max 52, clipped 0/50 updates Step 25,200 | last 50 updates / 1,600 crops | lr 1.86e-06 points (35/curve, tolerance 2 vox): 47,612 right / 3,278 wrong | 93.6% correct | 5,110 unknown loss 0.9388 | geometry 0.2691 | confidence 1.3393 | mean error 0.654 vox speed: 891 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.24 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1477 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 221, from the previous prediction 1256 confidence-labeled crops: terminal 34/1523 (2.2%) | recoverable 114/1523 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 384/1600 (24.0%), 2: 160/1600 (10.0%), 3: 224/1600 (14.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 103 chunks / 206.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 693 rows | mean depth 15 decisions | mean travel 125 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 54% 43% 25%>3% 22% 10 403 87%/10%/0%/3% 8% 191/2/55/3 0175A_5mm_v1 384 42% 47% 21%>10% 11% 30 1702 86%/6%/4%/4% 10% 165/0/0/17 1447_5mm_v1 160 42% 47% 22%>11% 11% 16 1519 86%/4%/3%/6% 7% 75/0/0/0 paris4_afv_central 224 42% 50% 20%>7% 12% 14 921 88%/6%/1%/4% 7% 113/3/0/0 PHerc0125_3mm_v2 384 47% 44% 27%>9% 18% 34 1116 81%/8%/3%/8% 14% 149/3/0/20 gradients: max 56, clipped 0/50 updates Step 25,250 | last 50 updates / 1,600 crops | lr 1.82e-06 points (35/curve, tolerance 2 vox): 48,087 right / 2,972 wrong | 94.2% correct | 4,941 unknown loss 0.8395 | geometry 0.2527 | confidence 1.1737 | mean error 0.627 vox speed: 891 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1481 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 277, from the previous prediction 1204 confidence-labeled crops: terminal 31/1530 (2.0%) | recoverable 100/1530 (6.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 192/1600 (12.0%), 2: 352/1600 (22.0%), 3: 288/1600 (18.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 103 chunks / 206.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 598 chunk references / 1 windows live chains: 625 rows | mean depth 16 decisions | mean travel 129 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 51% 46% 23%>3% 20% 14 1108 89%/7%/0%/4% 10% 86/0/0/61 0175A_5mm_v1 192 50% 44% 24%>6% 19% 10 1074 85%/9%/2%/4% 5% 85/1/0/0 1447_5mm_v1 352 46% 44% 23%>10% 13% 29 1446 87%/6%/3%/4% 6% 156/2/0/0 paris4_afv_central 288 49% 42% 25%>9% 16% 21 1212 87%/6%/2%/5% 8% 121/2/0/0 PHerc0125_3mm_v2 448 43% 50% 21%>7% 14% 29 952 87%/5%/2%/6% 12% 177/2/0/49 gradients: max 51, clipped 0/50 updates Step 25,300 | last 50 updates / 1,600 crops | lr 1.78e-06 points (35/curve, tolerance 2 vox): 47,350 right / 2,941 wrong | 94.2% correct | 5,709 unknown loss 0.8668 | geometry 0.2288 | confidence 1.2759 | mean error 0.615 vox speed: 892 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1492 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 252, from the previous prediction 1240 confidence-labeled crops: terminal 29/1514 (1.9%) | recoverable 107/1514 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 288/1600 (18.0%), 2: 384/1600 (24.0%), 3: 160/1600 (10.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 103 chunks / 206.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 634 rows | mean depth 19 decisions | mean travel 157 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 57% 43% 25%>1% 25% 2 300 86%/11%/0%/3% 6% 111/0/0/26 0175A_5mm_v1 288 44% 43% 28%>14% 14% 30 2115 84%/6%/5%/6% 10% 123/2/0/0 1447_5mm_v1 384 46% 46% 22%>8% 14% 30 1457 86%/6%/3%/5% 7% 159/1/56/16 paris4_afv_central 160 51% 46% 30%>3% 27% 5 555 89%/7%/1%/3% 6% 73/1/0/0 PHerc0125_3mm_v2 448 47% 47% 26%>5% 20% 25 1032 85%/5%/1%/9% 14% 168/3/0/44 gradients: max 31, clipped 0/50 updates Step 25,350 | last 50 updates / 1,600 crops | lr 1.74e-06 points (35/curve, tolerance 2 vox): 48,167 right / 2,575 wrong | 94.9% correct | 5,258 unknown loss 0.8843 | geometry 0.2895 | confidence 1.1895 | mean error 0.626 vox speed: 892 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1501 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 260, from the previous prediction 1241 confidence-labeled crops: terminal 22/1537 (1.4%) | recoverable 124/1537 (8.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 320/1600 (20.0%), 2: 288/1600 (18.0%), 3: 384/1600 (24.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 103 chunks / 206.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 632 rows | mean depth 14 decisions | mean travel 114 vox | 6 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 54% 46% 25%>0% 25% 1 224 93%/5%/0%/2% 5% 117/0/0/0 0175A_5mm_v1 320 39% 44% 28%>16% 11% 38 2348 83%/9%/3%/5% 11% 142/3/0/0 1447_5mm_v1 288 47% 43% 25%>10% 16% 24 1901 86%/8%/2%/3% 6% 89/0/72/36 paris4_afv_central 384 49% 49% 24%>2% 22% 6 379 86%/9%/1%/5% 5% 181/3/49/8 PHerc0125_3mm_v2 352 51% 46% 24%>3% 22% 9 1043 88%/7%/1%/5% 14% 103/1/0/59 gradients: max 43, clipped 0/50 updates Step 25,400 | last 50 updates / 1,600 crops | lr 1.71e-06 points (35/curve, tolerance 2 vox): 47,001 right / 3,221 wrong | 93.6% correct | 5,778 unknown loss 0.9224 | geometry 0.2698 | confidence 1.3052 | mean error 0.640 vox speed: 895 ms/update | data wait 1 ms/update | 35.1 crops/s | 35.1 decisions/s | 1.25 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1443 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 241, from the previous prediction 1202 confidence-labeled crops: terminal 25/1514 (1.7%) | recoverable 93/1514 (6.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 160/1600 (10.0%), 2: 288/1600 (18.0%), 3: 480/1600 (30.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 104 chunks / 208.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 652 rows | mean depth 20 decisions | mean travel 163 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 47% 51% 21%>2% 19% 6 519 93%/5%/0%/2% 7% 146/0/0/0 0175A_5mm_v1 160 56% 44% 23%>0% 23% 0 171 89%/6%/0%/6% 4% 70/0/0/0 1447_5mm_v1 288 43% 50% 22%>6% 15% 18 1117 89%/5%/1%/5% 4% 134/2/0/11 paris4_afv_central 480 43% 44% 27%>13% 14% 57 2610 81%/10%/2%/7% 13% 126/2/29/85 PHerc0125_3mm_v2 384 42% 46% 24%>12% 12% 44 1451 87%/2%/3%/8% 15% 176/3/0/0 gradients: max 43, clipped 0/50 updates Step 25,450 | last 50 updates / 1,600 crops | lr 1.67e-06 points (35/curve, tolerance 2 vox): 48,034 right / 2,696 wrong | 94.7% correct | 5,270 unknown loss 0.8272 | geometry 0.2467 | confidence 1.1612 | mean error 0.621 vox speed: 894 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1492 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 250, from the previous prediction 1242 confidence-labeled crops: terminal 26/1523 (1.7%) | recoverable 112/1523 (7.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 288/1600 (18.0%), 2: 288/1600 (18.0%), 3: 320/1600 (20.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 104 chunks / 208.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 605 rows | mean depth 18 decisions | mean travel 146 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 51% 45% 25%>4% 21% 15 642 91%/6%/0%/3% 8% 160/2/31/26 0175A_5mm_v1 288 50% 42% 27%>8% 19% 19 1476 84%/8%/2%/6% 6% 122/2/0/0 1447_5mm_v1 288 55% 44% 25%>1% 24% 3 318 89%/7%/1%/3% 3% 127/1/0/0 paris4_afv_central 320 46% 45% 31%>8% 22% 32 2157 82%/9%/3%/7% 11% 76/4/0/69 PHerc0125_3mm_v2 288 48% 43% 22%>9% 13% 29 1341 85%/5%/3%/7% 13% 120/7/17/5 gradients: max 40, clipped 0/50 updates Step 25,500 | last 50 updates / 1,600 crops | lr 1.64e-06 points (35/curve, tolerance 2 vox): 47,042 right / 3,091 wrong | 93.8% correct | 5,867 unknown loss 0.8463 | geometry 0.2519 | confidence 1.1887 | mean error 0.626 vox speed: 893 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1470 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 196, from the previous prediction 1274 confidence-labeled crops: terminal 31/1522 (2.0%) | recoverable 116/1522 (7.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 352/1600 (22.0%), 2: 352/1600 (22.0%), 3: 288/1600 (18.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 105 chunks / 210.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 690 rows | mean depth 13 decisions | mean travel 111 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 42% 47% 28%>11% 17% 17 968 85%/8%/2%/5% 12% 132/2/0/4 0175A_5mm_v1 352 48% 45% 25%>7% 17% 23 1343 86%/8%/2%/3% 8% 157/6/0/0 1447_5mm_v1 352 57% 39% 28%>3% 24% 12 730 85%/7%/2%/6% 4% 138/2/0/0 paris4_afv_central 288 47% 46% 27%>7% 20% 20 1216 87%/8%/1%/4% 8% 122/2/0/11 PHerc0125_3mm_v2 320 42% 48% 23%>10% 14% 34 1226 85%/6%/2%/8% 12% 141/2/71/12 gradients: max 51, clipped 0/50 updates Step 25,550 | last 50 updates / 1,600 crops | lr 1.60e-06 points (35/curve, tolerance 2 vox): 46,693 right / 3,678 wrong | 92.7% correct | 5,629 unknown loss 0.9575 | geometry 0.2888 | confidence 1.3373 | mean error 0.663 vox speed: 899 ms/update | data wait 1 ms/update | 35.1 crops/s | 35.1 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1401 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 274, from the previous prediction 1127 confidence-labeled crops: terminal 36/1520 (2.4%) | recoverable 106/1520 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 416/1600 (26.0%), 2: 256/1600 (16.0%), 3: 288/1600 (18.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 105 chunks / 210.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 612 rows | mean depth 22 decisions | mean travel 177 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 49% 47% 22%>4% 18% 8 1232 90%/7%/0%/3% 10% 91/0/0/15 0175A_5mm_v1 416 50% 44% 24%>6% 18% 32 2333 87%/8%/2%/4% 6% 113/0/89/71 1447_5mm_v1 256 43% 45% 21%>12% 9% 26 1722 87%/5%/4%/4% 7% 115/3/0/0 paris4_afv_central 288 41% 45% 28%>14% 15% 35 1688 84%/9%/2%/5% 12% 122/0/32/9 PHerc0125_3mm_v2 416 37% 47% 22%>17% 5% 65 1949 84%/4%/3%/9% 16% 171/2/13/23 gradients: max 41, clipped 0/50 updates Step 25,600 | last 50 updates / 1,600 crops | lr 1.57e-06 points (35/curve, tolerance 2 vox): 46,993 right / 3,252 wrong | 93.5% correct | 5,755 unknown loss 0.8801 | geometry 0.2800 | confidence 1.2002 | mean error 0.647 vox speed: 892 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1456 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 225, from the previous prediction 1231 confidence-labeled crops: terminal 29/1528 (1.9%) | recoverable 113/1528 (7.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 320/1600 (20.0%), 2: 320/1600 (20.0%), 3: 256/1600 (16.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 642 rows | mean depth 17 decisions | mean travel 142 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 54% 45% 22%>1% 21% 4 264 89%/8%/0%/2% 6% 158/0/0/0 0175A_5mm_v1 320 43% 49% 23%>8% 15% 24 1720 87%/5%/2%/6% 7% 139/1/0/19 1447_5mm_v1 320 48% 42% 30%>10% 20% 25 1541 86%/9%/2%/4% 5% 134/1/0/0 paris4_afv_central 256 41% 45% 22%>14% 8% 29 1742 84%/7%/4%/5% 16% 112/5/41/2 PHerc0125_3mm_v2 352 45% 45% 24%>10% 14% 35 1699 85%/7%/1%/7% 16% 99/2/0/60 gradients: max 48, clipped 0/50 updates Step 25,650 | last 50 updates / 1,600 crops | lr 1.53e-06 points (35/curve, tolerance 2 vox): 48,632 right / 2,597 wrong | 94.9% correct | 4,771 unknown loss 0.8297 | geometry 0.2561 | confidence 1.1471 | mean error 0.623 vox speed: 891 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1503 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 247, from the previous prediction 1256 confidence-labeled crops: terminal 16/1532 (1.0%) | recoverable 108/1532 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 352/1600 (22.0%), 2: 192/1600 (12.0%), 3: 448/1600 (28.0%), 4: 160/1600 (10.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 606 rows | mean depth 15 decisions | mean travel 128 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 53% 43% 28%>4% 24% 16 623 88%/9%/0%/3% 9% 192/0/7/0 0175A_5mm_v1 352 58% 41% 25%>1% 24% 4 408 89%/5%/1%/5% 5% 121/1/0/24 1447_5mm_v1 192 51% 45% 26%>4% 22% 8 787 89%/6%/1%/4% 3% 86/2/0/0 paris4_afv_central 448 44% 48% 21%>8% 14% 40 1981 88%/6%/1%/6% 9% 144/4/35/70 PHerc0125_3mm_v2 160 54% 39% 28%>6% 21% 10 901 85%/8%/2%/4% 14% 63/3/0/0 gradients: max 72, clipped 1/50 updates Step 25,700 | last 50 updates / 1,600 crops | lr 1.50e-06 points (35/curve, tolerance 2 vox): 49,786 right / 1,942 wrong | 96.2% correct | 4,272 unknown loss 0.6491 | geometry 0.1942 | confidence 0.9098 | mean error 0.551 vox speed: 890 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1539 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 226, from the previous prediction 1313 confidence-labeled crops: terminal 17/1549 (1.1%) | recoverable 112/1549 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 288/1600 (18.0%), 2: 416/1600 (26.0%), 3: 352/1600 (22.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 660 rows | mean depth 24 decisions | mean travel 193 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 57% 41% 28%>2% 26% 5 408 89%/8%/0%/2% 6% 109/0/0/21 0175A_5mm_v1 288 55% 45% 26%>0% 26% 0 140 92%/6%/0%/2% 3% 129/1/0/0 1447_5mm_v1 416 51% 46% 27%>3% 24% 13 628 88%/9%/1%/2% 2% 190/3/0/0 paris4_afv_central 352 52% 45% 23%>4% 19% 15 872 85%/7%/2%/7% 7% 118/0/0/39 PHerc0125_3mm_v2 224 38% 51% 20%>12% 8% 24 1196 90%/3%/3%/4% 14% 114/2/0/0 gradients: max 46, clipped 0/50 updates Step 25,750 | last 50 updates / 1,600 crops | lr 1.46e-06 points (35/curve, tolerance 2 vox): 47,900 right / 2,825 wrong | 94.4% correct | 5,275 unknown loss 0.8721 | geometry 0.2337 | confidence 1.2767 | mean error 0.607 vox speed: 892 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1511 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 251, from the previous prediction 1260 confidence-labeled crops: terminal 31/1543 (2.0%) | recoverable 98/1543 (6.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 352/1600 (22.0%), 2: 256/1600 (16.0%), 3: 352/1600 (22.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 677 rows | mean depth 14 decisions | mean travel 120 vox | 7 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 51% 49% 24%>0% 24% 1 211 89%/8%/0%/2% 7% 142/0/0/15 0175A_5mm_v1 352 50% 42% 27%>7% 20% 23 1424 88%/6%/3%/3% 7% 149/2/10/0 1447_5mm_v1 256 47% 50% 23%>4% 20% 9 663 89%/8%/1%/2% 4% 122/2/30/5 paris4_afv_central 352 50% 45% 23%>5% 18% 16 1077 89%/5%/2%/5% 7% 108/5/0/49 PHerc0125_3mm_v2 320 40% 49% 24%>11% 13% 33 1342 86%/4%/4%/6% 16% 156/1/0/0 gradients: max 48, clipped 0/50 updates Step 25,800 | last 50 updates / 1,600 crops | lr 1.43e-06 points (35/curve, tolerance 2 vox): 47,698 right / 2,605 wrong | 94.8% correct | 5,697 unknown loss 0.7950 | geometry 0.2235 | confidence 1.1430 | mean error 0.597 vox speed: 891 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1472 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 275, from the previous prediction 1197 confidence-labeled crops: terminal 29/1523 (1.9%) | recoverable 97/1523 (6.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 352/1600 (22.0%), 2: 352/1600 (22.0%), 3: 224/1600 (14.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 587 rows | mean depth 15 decisions | mean travel 127 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 57% 39% 34%>4% 30% 10 598 90%/7%/0%/3% 5% 112/2/0/0 0175A_5mm_v1 352 46% 45% 24%>9% 15% 26 1556 86%/5%/3%/6% 8% 158/1/27/1 1447_5mm_v1 352 49% 47% 23%>4% 20% 17 1535 88%/8%/1%/3% 4% 89/0/0/77 paris4_afv_central 224 44% 44% 28%>12% 16% 26 1618 84%/5%/3%/8% 9% 98/2/0/0 PHerc0125_3mm_v2 384 44% 48% 23%>8% 15% 33 1634 86%/5%/1%/7% 13% 130/1/43/55 gradients: max 50, clipped 0/50 updates Step 25,850 | last 50 updates / 1,600 crops | lr 1.40e-06 points (35/curve, tolerance 2 vox): 48,150 right / 2,678 wrong | 94.7% correct | 5,172 unknown loss 0.7821 | geometry 0.2435 | confidence 1.0772 | mean error 0.610 vox speed: 892 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1468 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 240, from the previous prediction 1228 confidence-labeled crops: terminal 24/1525 (1.6%) | recoverable 99/1525 (6.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 192/1600 (12.0%), 2: 480/1600 (30.0%), 3: 192/1600 (12.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 627 rows | mean depth 24 decisions | mean travel 200 vox | 0 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 48% 47% 24%>5% 20% 14 603 90%/6%/1%/3% 8% 177/0/0/3 0175A_5mm_v1 192 48% 46% 24%>6% 19% 11 1126 93%/4%/1%/3% 5% 88/1/0/0 1447_5mm_v1 480 50% 43% 25%>7% 18% 34 1710 85%/7%/1%/6% 4% 167/3/0/38 paris4_afv_central 192 46% 44% 23%>10% 13% 19 1389 86%/7%/3%/4% 8% 84/0/0/0 PHerc0125_3mm_v2 352 48% 41% 24%>11% 14% 36 1672 86%/6%/2%/5% 13% 111/2/0/34 gradients: max 51, clipped 0/50 updates Step 25,900 | last 50 updates / 1,600 crops | lr 1.36e-06 points (35/curve, tolerance 2 vox): 46,045 right / 4,041 wrong | 91.9% correct | 5,914 unknown loss 0.9067 | geometry 0.2640 | confidence 1.2854 | mean error 0.619 vox speed: 891 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1434 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 218, from the previous prediction 1216 confidence-labeled crops: terminal 55/1529 (3.6%) | recoverable 100/1529 (6.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 288/1600 (18.0%), 2: 416/1600 (26.0%), 3: 352/1600 (22.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 627 rows | mean depth 15 decisions | mean travel 130 vox | 8 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 54% 46% 24%>0% 24% 0 141 90%/5%/0%/4% 5% 115/0/0/2 0175A_5mm_v1 288 57% 40% 28%>3% 25% 9 800 89%/7%/2%/3% 6% 115/1/0/0 1447_5mm_v1 416 44% 43% 28%>13% 15% 50 3048 81%/7%/6%/6% 12% 111/2/0/67 paris4_afv_central 352 41% 45% 24%>14% 10% 38 1799 86%/7%/4%/3% 14% 158/6/0/0 PHerc0125_3mm_v2 288 40% 48% 22%>12% 10% 36 1474 84%/5%/4%/7% 17% 128/5/0/9 gradients: max 47, clipped 0/50 updates Step 25,950 | last 50 updates / 1,600 crops | lr 1.33e-06 points (35/curve, tolerance 2 vox): 47,597 right / 3,159 wrong | 93.8% correct | 5,244 unknown loss 0.8607 | geometry 0.2464 | confidence 1.2287 | mean error 0.633 vox speed: 893 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1476 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 231, from the previous prediction 1245 confidence-labeled crops: terminal 29/1520 (1.9%) | recoverable 109/1520 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 352/1600 (22.0%), 2: 320/1600 (20.0%), 3: 192/1600 (12.0%), 4: 480/1600 (30.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 666 rows | mean depth 15 decisions | mean travel 126 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 54% 46% 22%>0% 22% 0 159 88%/10%/0%/2% 7% 118/2/0/0 0175A_5mm_v1 352 48% 45% 24%>7% 17% 19 1260 88%/6%/2%/5% 7% 160/3/0/0 1447_5mm_v1 320 48% 47% 24%>5% 19% 16 1030 88%/6%/1%/5% 2% 123/2/0/28 paris4_afv_central 192 45% 46% 27%>9% 18% 16 1317 85%/7%/2%/7% 6% 87/0/13/1 PHerc0125_3mm_v2 480 43% 43% 27%>14% 13% 58 1718 82%/6%/3%/8% 17% 178/3/0/28 gradients: max 41, clipped 0/50 updates Step 26,000 | last 50 updates / 1,600 crops | lr 1.30e-06 points (35/curve, tolerance 2 vox): 48,760 right / 2,277 wrong | 95.5% correct | 4,963 unknown loss 0.8080 | geometry 0.2496 | confidence 1.1167 | mean error 0.597 vox speed: 891 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1527 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 213, from the previous prediction 1314 confidence-labeled crops: terminal 16/1528 (1.0%) | recoverable 105/1528 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 480/1600 (30.0%), 1: 288/1600 (18.0%), 2: 352/1600 (22.0%), 3: 288/1600 (18.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 698 rows | mean depth 25 decisions | mean travel 205 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 480 54% 45% 24%>1% 23% 5 254 90%/8%/0%/3% 8% 217/0/0/0 0175A_5mm_v1 288 47% 46% 26%>7% 19% 16 1458 87%/5%/3%/5% 7% 114/0/13/18 1447_5mm_v1 352 51% 47% 23%>2% 21% 8 508 89%/6%/0%/4% 3% 165/2/0/0 paris4_afv_central 288 42% 49% 25%>9% 17% 23 1447 85%/7%/1%/7% 7% 114/1/115/27 PHerc0125_3mm_v2 192 47% 46% 24%>7% 18% 13 893 85%/7%/1%/7% 15% 88/2/0/0 gradients: max 46, clipped 0/50 updates Step 26,000 | dagger_launched: true Step 26,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 256, "stopped": 0, "mean_example_error": 0.45053778076544404, "geometry_examples": 32, "examples": 32, "inference_and_measurement_seconds": 1.887630868703127, "total_diagnostic_seconds": 3.6802063807845116} Step 26,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 6/8 (75.0%) | false stops 1 1.5-3: recovered 6/8 (75.0%) | false stops 1 3-4.5: recovered 6/8 (75.0%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 0 entire proposed 8-point prefix correct (distance): 22/28 (78.6%) gate @ 0.50: false stops 0/23 (0.0%) | accepted wrong 5/27 (18.5%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.596 1.564 following 16 464 1.151 1.174 recoverable 12 348 2.190 2.084 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 1.151 1.174 332 32 812 1.596 1.564 improved/worsened/compared states: 3/2/28 Step 26,000 | batch_diagnostics_seconds: 3.6835751198232174 | recovery_seconds: 14.084121622145176 Step 26,050 | last 50 updates / 1,600 crops | lr 1.27e-06 points (35/curve, tolerance 2 vox): 47,359 right / 3,474 wrong | 93.2% correct | 5,167 unknown loss 0.9849 | geometry 0.3122 | confidence 1.3453 | mean error 0.677 vox speed: 962 ms/update | data wait 1 ms/update | 23.6 crops/s | 23.6 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1454 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 270, from the previous prediction 1184 confidence-labeled crops: terminal 29/1544 (1.9%) | recoverable 99/1544 (6.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 256/1600 (16.0%), 2: 256/1600 (16.0%), 3: 288/1600 (18.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 567 rows | mean depth 13 decisions | mean travel 113 vox | 6 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 53% 41% 26%>6% 21% 21 1334 92%/5%/1%/3% 11% 95/2/0/51 0175A_5mm_v1 256 52% 48% 21%>0% 21% 0 136 90%/7%/0%/2% 4% 67/3/0/57 1447_5mm_v1 256 48% 43% 27%>9% 18% 22 1501 86%/6%/2%/5% 7% 110/5/0/0 paris4_afv_central 288 43% 42% 22%>15% 7% 37 1952 85%/7%/4%/5% 13% 121/5/0/0 PHerc0125_3mm_v2 448 47% 41% 26%>11% 15% 51 1776 88%/6%/2%/4% 15% 174/6/24/11 gradients: max 50, clipped 0/50 updates Step 26,100 | last 50 updates / 1,600 crops | lr 1.23e-06 points (35/curve, tolerance 2 vox): 48,505 right / 2,651 wrong | 94.8% correct | 4,844 unknown loss 0.8141 | geometry 0.2258 | confidence 1.1765 | mean error 0.592 vox speed: 1352 ms/update | data wait 1 ms/update | 23.4 crops/s | 23.4 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1492 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 248, from the previous prediction 1244 confidence-labeled crops: terminal 28/1538 (1.8%) | recoverable 109/1538 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 416/1600 (26.0%), 2: 256/1600 (16.0%), 3: 288/1600 (18.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 625 rows | mean depth 19 decisions | mean travel 154 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 53% 45% 24%>3% 21% 8 441 90%/7%/0%/3% 7% 152/0/0/19 0175A_5mm_v1 416 54% 41% 25%>5% 19% 20 1131 89%/6%/1%/3% 5% 171/1/0/0 1447_5mm_v1 256 42% 52% 23%>6% 17% 20 1429 85%/7%/3%/5% 6% 103/2/66/30 paris4_afv_central 288 47% 43% 27%>9% 17% 26 1899 86%/7%/2%/5% 7% 97/1/0/28 PHerc0125_3mm_v2 256 45% 48% 21%>7% 14% 23 1261 84%/6%/4%/6% 18% 102/0/0/21 gradients: max 59, clipped 0/50 updates Step 26,150 | last 50 updates / 1,600 crops | lr 1.20e-06 points (35/curve, tolerance 2 vox): 48,007 right / 2,631 wrong | 94.8% correct | 5,362 unknown loss 0.8194 | geometry 0.2538 | confidence 1.1312 | mean error 0.611 vox speed: 1393 ms/update | data wait 1 ms/update | 22.7 crops/s | 22.7 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1495 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 218, from the previous prediction 1277 confidence-labeled crops: terminal 21/1530 (1.4%) | recoverable 87/1530 (5.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 480/1600 (30.0%), 2: 320/1600 (20.0%), 3: 192/1600 (12.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 650 rows | mean depth 22 decisions | mean travel 179 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 55% 41% 23%>4% 20% 7 546 89%/7%/0%/4% 8% 106/1/0/0 0175A_5mm_v1 480 54% 43% 22%>3% 20% 11 673 89%/6%/1%/4% 5% 207/3/0/0 1447_5mm_v1 320 44% 52% 24%>4% 19% 17 1349 91%/4%/2%/3% 4% 118/1/62/48 paris4_afv_central 192 44% 43% 28%>13% 15% 23 1973 86%/7%/3%/4% 11% 76/3/0/6 PHerc0125_3mm_v2 352 44% 48% 19%>8% 11% 33 1314 88%/4%/1%/8% 14% 143/2/0/25 gradients: max 40, clipped 0/50 updates Step 26,200 | last 50 updates / 1,600 crops | lr 1.17e-06 points (35/curve, tolerance 2 vox): 47,450 right / 2,938 wrong | 94.2% correct | 5,612 unknown loss 0.8684 | geometry 0.2527 | confidence 1.2315 | mean error 0.616 vox speed: 1279 ms/update | data wait 1 ms/update | 24.8 crops/s | 24.8 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1469 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 254, from the previous prediction 1215 confidence-labeled crops: terminal 25/1532 (1.6%) | recoverable 97/1532 (6.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 320/1600 (20.0%), 2: 384/1600 (24.0%), 3: 320/1600 (20.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 669 rows | mean depth 15 decisions | mean travel 129 vox | 7 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 54% 46% 28%>0% 28% 0 181 93%/5%/0%/2% 7% 102/0/11/0 0175A_5mm_v1 320 47% 48% 17%>5% 12% 20 2113 90%/5%/2%/3% 6% 75/2/0/79 1447_5mm_v1 384 49% 46% 21%>5% 16% 17 904 89%/7%/2%/2% 5% 172/0/0/4 paris4_afv_central 320 43% 47% 24%>9% 15% 29 1512 86%/7%/2%/6% 11% 141/4/0/10 PHerc0125_3mm_v2 352 33% 51% 22%>16% 6% 49 1615 83%/6%/2%/9% 19% 179/2/0/0 gradients: max 59, clipped 0/50 updates Step 26,250 | last 50 updates / 1,600 crops | lr 1.14e-06 points (35/curve, tolerance 2 vox): 46,301 right / 3,513 wrong | 92.9% correct | 6,186 unknown loss 1.0147 | geometry 0.3094 | confidence 1.4107 | mean error 0.680 vox speed: 1385 ms/update | data wait 1 ms/update | 22.8 crops/s | 22.8 decisions/s | 1.25 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1447 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 266, from the previous prediction 1181 confidence-labeled crops: terminal 28/1519 (1.8%) | recoverable 102/1519 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 288/1600 (18.0%), 2: 320/1600 (20.0%), 3: 224/1600 (14.0%), 4: 480/1600 (30.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 613 rows | mean depth 18 decisions | mean travel 155 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 53% 47% 26%>0% 26% 1 202 88%/9%/0%/3% 7% 134/1/0/0 0175A_5mm_v1 288 56% 42% 27%>2% 25% 7 639 91%/6%/1%/3% 4% 121/2/0/0 1447_5mm_v1 320 51% 44% 26%>5% 21% 14 1150 88%/6%/1%/5% 4% 105/1/0/37 paris4_afv_central 224 29% 50% 26%>21% 5% 38 2249 84%/5%/4%/6% 15% 111/0/0/0 PHerc0125_3mm_v2 480 41% 46% 23%>14% 9% 75 2383 82%/6%/2%/9% 21% 142/2/0/77 gradients: max 45, clipped 0/50 updates Step 26,300 | last 50 updates / 1,600 crops | lr 1.11e-06 points (35/curve, tolerance 2 vox): 46,211 right / 3,436 wrong | 93.1% correct | 6,353 unknown loss 0.9437 | geometry 0.2931 | confidence 1.3012 | mean error 0.651 vox speed: 1363 ms/update | data wait 1 ms/update | 23.3 crops/s | 23.3 decisions/s | 1.28 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1454 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 277, from the previous prediction 1177 confidence-labeled crops: terminal 27/1515 (1.8%) | recoverable 104/1515 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 160/1600 (10.0%), 2: 416/1600 (26.0%), 3: 192/1600 (12.0%), 4: 512/1600 (32.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 588 rows | mean depth 24 decisions | mean travel 198 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 53% 47% 25%>0% 25% 0 166 91%/7%/0%/3% 7% 149/0/0/0 0175A_5mm_v1 160 50% 46% 24%>4% 19% 7 962 84%/8%/1%/8% 6% 73/3/0/0 1447_5mm_v1 416 50% 45% 21%>5% 16% 24 1267 89%/5%/2%/4% 6% 174/6/0/14 paris4_afv_central 192 42% 43% 33%>15% 19% 26 1938 83%/10%/2%/4% 10% 83/0/0/0 PHerc0125_3mm_v2 512 42% 46% 23%>13% 11% 68 2789 82%/6%/3%/9% 19% 109/4/0/125 gradients: max 58, clipped 0/50 updates Step 26,350 | last 50 updates / 1,600 crops | lr 1.08e-06 points (35/curve, tolerance 2 vox): 47,798 right / 3,192 wrong | 93.7% correct | 5,010 unknown loss 0.8636 | geometry 0.2792 | confidence 1.1688 | mean error 0.621 vox speed: 1367 ms/update | data wait 1 ms/update | 23.1 crops/s | 23.1 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1488 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 213, from the previous prediction 1275 confidence-labeled crops: terminal 31/1544 (2.0%) | recoverable 111/1544 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 288/1600 (18.0%), 2: 320/1600 (20.0%), 3: 480/1600 (30.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 665 rows | mean depth 16 decisions | mean travel 132 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 54% 43% 23%>3% 20% 7 455 93%/6%/0%/1% 9% 125/1/0/0 0175A_5mm_v1 288 56% 44% 28%>0% 28% 0 176 88%/9%/0%/3% 4% 128/2/0/0 1447_5mm_v1 320 52% 46% 23%>2% 21% 5 420 89%/7%/1%/3% 3% 142/2/0/5 paris4_afv_central 480 44% 46% 25%>10% 15% 42 2170 84%/8%/4%/4% 13% 165/4/22/55 PHerc0125_3mm_v2 224 38% 47% 25%>15% 10% 34 1990 83%/3%/4%/9% 16% 105/3/0/0 gradients: max 48, clipped 0/50 updates Step 26,400 | last 50 updates / 1,600 crops | lr 1.05e-06 points (35/curve, tolerance 2 vox): 48,216 right / 2,472 wrong | 95.1% correct | 5,312 unknown loss 0.7807 | geometry 0.2630 | confidence 1.0354 | mean error 0.625 vox speed: 1327 ms/update | data wait 1 ms/update | 23.9 crops/s | 23.9 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1509 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 252, from the previous prediction 1257 confidence-labeled crops: terminal 12/1528 (0.8%) | recoverable 122/1528 (8.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 288/1600 (18.0%), 2: 576/1600 (36.0%), 3: 192/1600 (12.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 618 rows | mean depth 20 decisions | mean travel 164 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 49% 48% 27%>3% 24% 16 711 88%/9%/0%/3% 9% 131/0/66/37 0175A_5mm_v1 288 47% 47% 22%>6% 16% 15 1183 86%/8%/1%/5% 6% 136/0/0/0 1447_5mm_v1 576 56% 43% 25%>1% 24% 8 519 89%/7%/0%/4% 3% 184/2/0/62 paris4_afv_central 192 44% 46% 28%>9% 19% 16 1354 84%/8%/1%/7% 9% 89/0/0/0 PHerc0125_3mm_v2 192 43% 41% 24%>16% 8% 29 1930 83%/6%/3%/8% 22% 78/1/0/0 gradients: max 55, clipped 0/50 updates Step 26,450 | last 50 updates / 1,600 crops | lr 1.03e-06 points (35/curve, tolerance 2 vox): 48,905 right / 2,400 wrong | 95.3% correct | 4,695 unknown loss 0.8107 | geometry 0.2624 | confidence 1.0964 | mean error 0.608 vox speed: 1272 ms/update | data wait 1 ms/update | 24.9 crops/s | 24.9 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1514 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 291, from the previous prediction 1223 confidence-labeled crops: terminal 16/1545 (1.0%) | recoverable 110/1545 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 352/1600 (22.0%), 2: 352/1600 (22.0%), 3: 224/1600 (14.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 595 rows | mean depth 26 decisions | mean travel 214 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 49% 50% 24%>2% 22% 7 405 90%/8%/0%/2% 7% 151/0/0/71 0175A_5mm_v1 352 51% 48% 23%>2% 21% 6 568 91%/5%/0%/5% 3% 125/4/77/43 1447_5mm_v1 352 53% 45% 25%>2% 23% 8 615 89%/6%/1%/4% 6% 131/0/0/26 paris4_afv_central 224 46% 45% 22%>9% 13% 18 1403 87%/8%/1%/4% 10% 100/2/0/0 PHerc0125_3mm_v2 224 42% 39% 32%>19% 13% 41 2236 83%/9%/2%/6% 20% 88/2/0/0 gradients: max 43, clipped 0/50 updates Step 26,459 | dagger_states: 3855 | dagger_source_step: 26000 | dagger_caches: 4 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/decisions_026000.npz" | dagger_fibers: 128 | dagger_supply: {"states": 3855, "traces": 128, "fibers": 128, "rows": {"terminal": 8, "premature_stop": 113, "recoverable": 32, "pre_excursion": 80, "ordinary": 3622}, "events": {"terminal": 1, "premature_stop": 113, "recoverable": 13, "pre_excursion": 15, "ordinary": 123}, "supervision": {"following": 3811, "recoverable": 35, "terminal": 8, "unknown": 1}, "departed_traces": 14, "terminal_traces": 1, "stop_reasons": {"after_failure": 1, "confidence": 113, "unannotated": 14}, "traced_voxels": 61968.38585622282, "seconds": 613.9395554661751, "voxels_per_second": 100.93564635881644} | dagger_coverage: {"requested": 128, "seeds": 128, "skipped": [], "epoch": 1} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 0 | dagger_launch_interval: 5000 | dagger_publication_age: 459 | dataset: "paris4" Step 26,500 | last 50 updates / 1,600 crops | lr 9.97e-07 points (35/curve, tolerance 2 vox): 47,953 right / 3,622 wrong | 93.0% correct | 4,425 unknown loss 0.9902 | geometry 0.2937 | confidence 1.3929 | mean error 0.684 vox speed: 937 ms/update | data wait 1 ms/update | 33.7 crops/s | 33.7 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1477 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 208, from the previous prediction 1269 confidence-labeled crops: terminal 29/1542 (1.9%) | recoverable 104/1542 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 160/1600 (10.0%), 2: 352/1600 (22.0%), 3: 384/1600 (24.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 660 rows | mean depth 15 decisions | mean travel 128 vox | 6 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 49% 44% 27%>7% 20% 23 1373 90%/8%/0%/1% 13% 103/0/0/51 0175A_5mm_v1 160 51% 49% 21%>0% 21% 0 154 89%/6%/1%/3% 4% 79/1/0/0 1447_5mm_v1 352 50% 44% 23%>6% 17% 16 1098 86%/7%/2%/5% 5% 155/5/0/0 paris4_afv_central 384 50% 42% 26%>8% 18% 24 1288 88%/6%/3%/4% 8% 163/3/0/0 PHerc0125_3mm_v2 352 43% 45% 24%>12% 12% 39 1468 87%/5%/3%/6% 17% 160/3/10/0 gradients: max 40, clipped 0/50 updates Step 26,550 | last 50 updates / 1,600 crops | lr 9.69e-07 points (35/curve, tolerance 2 vox): 48,048 right / 2,853 wrong | 94.4% correct | 5,099 unknown loss 0.9335 | geometry 0.3351 | confidence 1.1969 | mean error 0.677 vox speed: 897 ms/update | data wait 1 ms/update | 35.1 crops/s | 35.1 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1522 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 310, from the previous prediction 1212 confidence-labeled crops: terminal 15/1538 (1.0%) | recoverable 103/1538 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 352/1600 (22.0%), 2: 160/1600 (10.0%), 3: 512/1600 (32.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 598 rows | mean depth 19 decisions | mean travel 166 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 51% 47% 23%>2% 21% 8 418 93%/5%/0%/2% 7% 163/0/0/18 0175A_5mm_v1 352 49% 48% 25%>4% 21% 11 849 89%/7%/1%/3% 6% 161/1/0/7 1447_5mm_v1 160 54% 46% 19%>0% 19% 0 153 92%/4%/0%/4% 5% 73/2/0/0 paris4_afv_central 512 49% 46% 22%>5% 17% 25 1352 86%/7%/1%/6% 9% 138/3/4/98 PHerc0125_3mm_v2 192 41% 44% 25%>15% 10% 25 2162 82%/9%/4%/5% 21% 63/1/75/21 gradients: max 47, clipped 0/50 updates Step 26,600 | last 50 updates / 1,600 crops | lr 9.41e-07 points (35/curve, tolerance 2 vox): 47,593 right / 2,634 wrong | 94.8% correct | 5,773 unknown loss 0.7745 | geometry 0.2573 | confidence 1.0344 | mean error 0.597 vox speed: 892 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1470 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 266, from the previous prediction 1204 confidence-labeled crops: terminal 23/1525 (1.5%) | recoverable 110/1525 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 160/1600 (10.0%), 1: 352/1600 (22.0%), 2: 384/1600 (24.0%), 3: 448/1600 (28.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 614 rows | mean depth 22 decisions | mean travel 184 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 160 50% 46% 21%>4% 17% 7 611 91%/6%/0%/3% 6% 73/0/0/0 0175A_5mm_v1 352 52% 48% 25%>0% 25% 0 138 89%/6%/0%/5% 2% 136/4/0/33 1447_5mm_v1 384 45% 48% 24%>7% 17% 25 1246 86%/7%/2%/5% 8% 183/5/0/0 paris4_afv_central 448 46% 42% 26%>12% 13% 51 2925 85%/7%/2%/5% 12% 110/3/0/78 PHerc0125_3mm_v2 256 43% 45% 21%>12% 10% 28 1499 85%/8%/2%/5% 16% 112/2/44/4 gradients: max 76, clipped 1/50 updates Step 26,650 | last 50 updates / 1,600 crops | lr 9.14e-07 points (35/curve, tolerance 2 vox): 47,603 right / 3,453 wrong | 93.2% correct | 4,944 unknown loss 0.9795 | geometry 0.2671 | confidence 1.4249 | mean error 0.652 vox speed: 893 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1451 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 266, from the previous prediction 1185 confidence-labeled crops: terminal 36/1534 (2.3%) | recoverable 112/1534 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 288/1600 (18.0%), 2: 256/1600 (16.0%), 3: 416/1600 (26.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 609 rows | mean depth 18 decisions | mean travel 149 vox | 7 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 48% 47% 24%>6% 18% 16 743 93%/6%/1%/1% 10% 164/0/0/0 0175A_5mm_v1 288 52% 48% 26%>0% 26% 0 136 86%/9%/0%/5% 3% 103/1/0/36 1447_5mm_v1 256 49% 42% 25%>9% 16% 21 1648 89%/5%/3%/3% 6% 107/1/0/0 paris4_afv_central 416 43% 45% 26%>12% 15% 52 2629 81%/7%/4%/7% 12% 122/3/0/67 PHerc0125_3mm_v2 288 41% 47% 22%>12% 9% 42 1989 84%/7%/4%/5% 20% 113/5/9/21 gradients: max 44, clipped 0/50 updates Step 26,700 | last 50 updates / 1,600 crops | lr 8.87e-07 points (35/curve, tolerance 2 vox): 48,959 right / 2,580 wrong | 95.0% correct | 4,461 unknown loss 0.8152 | geometry 0.2390 | confidence 1.1523 | mean error 0.612 vox speed: 892 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1546 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 257, from the previous prediction 1289 confidence-labeled crops: terminal 17/1541 (1.1%) | recoverable 96/1541 (6.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 352/1600 (22.0%), 2: 352/1600 (22.0%), 3: 288/1600 (18.0%), 4: 160/1600 (10.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 648 rows | mean depth 24 decisions | mean travel 196 vox | 8 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 51% 44% 23%>4% 18% 21 777 91%/7%/0%/3% 10% 176/1/128/22 0175A_5mm_v1 352 56% 44% 26%>0% 26% 0 145 91%/7%/0%/3% 3% 136/1/0/19 1447_5mm_v1 352 51% 44% 22%>5% 18% 13 1128 86%/7%/2%/5% 6% 123/2/88/32 paris4_afv_central 288 46% 49% 23%>5% 17% 14 873 88%/4%/3%/5% 9% 140/1/0/0 PHerc0125_3mm_v2 160 54% 46% 25%>0% 25% 0 141 88%/5%/2%/5% 12% 73/2/0/0 gradients: max 43, clipped 0/50 updates Step 26,750 | last 50 updates / 1,600 crops | lr 8.61e-07 points (35/curve, tolerance 2 vox): 47,782 right / 3,415 wrong | 93.3% correct | 4,803 unknown loss 0.9208 | geometry 0.2669 | confidence 1.3078 | mean error 0.635 vox speed: 895 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1441 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 278, from the previous prediction 1163 confidence-labeled crops: terminal 30/1537 (2.0%) | recoverable 105/1537 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 288/1600 (18.0%), 2: 512/1600 (32.0%), 3: 224/1600 (14.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 583 rows | mean depth 18 decisions | mean travel 151 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 43% 50% 26%>7% 18% 22 1790 91%/6%/0%/2% 13% 71/0/0/57 0175A_5mm_v1 288 44% 48% 24%>8% 16% 21 1514 89%/6%/1%/4% 4% 138/2/0/0 1447_5mm_v1 512 52% 43% 25%>5% 20% 30 1817 87%/7%/2%/4% 5% 161/1/0/59 paris4_afv_central 224 51% 41% 28%>8% 19% 16 1453 84%/8%/3%/4% 10% 91/0/0/0 PHerc0125_3mm_v2 320 46% 39% 25%>15% 10% 43 1997 85%/6%/3%/6% 19% 122/3/0/3 gradients: max 53, clipped 0/50 updates Step 26,800 | last 50 updates / 1,600 crops | lr 8.35e-07 points (35/curve, tolerance 2 vox): 47,528 right / 3,179 wrong | 93.7% correct | 5,293 unknown loss 0.9364 | geometry 0.2667 | confidence 1.3395 | mean error 0.649 vox speed: 892 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1466 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 229, from the previous prediction 1237 confidence-labeled crops: terminal 23/1525 (1.5%) | recoverable 107/1525 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 256/1600 (16.0%), 2: 192/1600 (12.0%), 3: 480/1600 (30.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 665 rows | mean depth 18 decisions | mean travel 154 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 46% 48% 26%>7% 19% 23 935 90%/8%/0%/3% 8% 170/0/0/28 0175A_5mm_v1 256 42% 48% 23%>10% 12% 22 1842 86%/6%/4%/4% 8% 122/4/14/0 1447_5mm_v1 192 48% 45% 26%>7% 19% 10 1481 86%/10%/2%/3% 6% 76/2/0/10 paris4_afv_central 480 49% 45% 24%>6% 18% 24 1272 86%/5%/1%/8% 8% 184/2/117/31 PHerc0125_3mm_v2 256 43% 44% 27%>13% 14% 32 1683 85%/6%/2%/6% 17% 113/1/0/0 gradients: max 66, clipped 1/50 updates Step 26,850 | last 50 updates / 1,600 crops | lr 8.09e-07 points (35/curve, tolerance 2 vox): 46,700 right / 4,100 wrong | 91.9% correct | 5,200 unknown loss 1.0481 | geometry 0.3159 | confidence 1.4644 | mean error 0.696 vox speed: 891 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1432 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 214, from the previous prediction 1218 confidence-labeled crops: terminal 32/1538 (2.1%) | recoverable 126/1538 (8.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 416/1600 (26.0%), 2: 224/1600 (14.0%), 3: 160/1600 (10.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 1449 chunk references / 1 windows live chains: 654 rows | mean depth 22 decisions | mean travel 190 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 45% 44% 28%>10% 17% 32 1214 86%/11%/1%/2% 14% 185/0/0/0 0175A_5mm_v1 416 47% 47% 26%>6% 20% 19 1240 87%/7%/2%/5% 5% 195/2/0/1 1447_5mm_v1 224 51% 40% 29%>8% 21% 16 1665 87%/7%/3%/4% 8% 90/1/0/0 paris4_afv_central 160 49% 47% 20%>4% 16% 6 711 90%/4%/1%/6% 6% 75/2/0/0 PHerc0125_3mm_v2 384 41% 42% 24%>16% 8% 61 2864 83%/8%/4%/6% 24% 109/2/0/54 gradients: max 50, clipped 0/50 updates Step 26,900 | last 50 updates / 1,600 crops | lr 7.84e-07 points (35/curve, tolerance 2 vox): 47,978 right / 2,902 wrong | 94.3% correct | 5,120 unknown loss 0.8274 | geometry 0.2555 | confidence 1.1438 | mean error 0.623 vox speed: 893 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1494 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 294, from the previous prediction 1200 confidence-labeled crops: terminal 27/1534 (1.8%) | recoverable 112/1534 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 448/1600 (28.0%), 2: 256/1600 (16.0%), 3: 352/1600 (22.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 575 rows | mean depth 17 decisions | mean travel 141 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 51% 44% 26%>5% 21% 11 939 88%/10%/0%/1% 11% 86/0/0/13 0175A_5mm_v1 448 54% 44% 27%>2% 25% 13 1037 88%/8%/0%/3% 4% 137/2/0/60 1447_5mm_v1 256 51% 46% 27%>3% 24% 4 649 88%/6%/1%/5% 4% 119/4/0/0 paris4_afv_central 352 47% 47% 20%>6% 14% 19 1083 88%/6%/2%/4% 8% 144/4/0/22 PHerc0125_3mm_v2 320 39% 47% 24%>14% 11% 47 2592 82%/5%/5%/9% 20% 89/1/0/62 gradients: max 49, clipped 0/50 updates Step 26,950 | last 50 updates / 1,600 crops | lr 7.59e-07 points (35/curve, tolerance 2 vox): 48,221 right / 2,867 wrong | 94.4% correct | 4,912 unknown loss 0.8223 | geometry 0.2607 | confidence 1.1232 | mean error 0.622 vox speed: 894 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1496 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 199, from the previous prediction 1297 confidence-labeled crops: terminal 20/1534 (1.3%) | recoverable 111/1534 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 352/1600 (22.0%), 2: 384/1600 (24.0%), 3: 192/1600 (12.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 108 chunks / 216.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 656 rows | mean depth 19 decisions | mean travel 158 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 51% 43% 27%>7% 21% 23 930 90%/8%/0%/3% 10% 164/0/0/0 0175A_5mm_v1 352 54% 46% 26%>0% 26% 0 136 89%/7%/0%/4% 3% 124/2/0/37 1447_5mm_v1 384 51% 46% 23%>3% 20% 11 712 86%/8%/1%/5% 4% 178/1/0/0 paris4_afv_central 192 43% 46% 24%>11% 12% 21 1703 86%/5%/5%/4% 12% 88/2/0/0 PHerc0125_3mm_v2 288 46% 39% 26%>15% 11% 39 2187 84%/7%/2%/7% 18% 102/1/0/9 gradients: max 58, clipped 0/50 updates Step 27,000 | last 50 updates / 1,600 crops | lr 7.35e-07 points (35/curve, tolerance 2 vox): 45,206 right / 4,367 wrong | 91.2% correct | 6,427 unknown loss 1.0717 | geometry 0.3583 | confidence 1.4268 | mean error 0.730 vox speed: 893 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.28 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1412 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 294, from the previous prediction 1118 confidence-labeled crops: terminal 35/1508 (2.3%) | recoverable 124/1508 (8.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 256/1600 (16.0%), 2: 288/1600 (18.0%), 3: 160/1600 (10.0%), 4: 608/1600 (38.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 117 chunks / 234.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 530 rows | mean depth 24 decisions | mean travel 200 vox | 6 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 41% 45% 30%>14% 16% 28 1492 87%/9%/1%/3% 14% 131/0/0/0 0175A_5mm_v1 256 44% 48% 27%>8% 20% 17 1512 86%/7%/3%/4% 7% 124/3/0/0 1447_5mm_v1 288 60% 39% 26%>1% 25% 5 851 84%/8%/1%/7% 3% 57/1/89/55 paris4_afv_central 160 51% 42% 21%>8% 13% 12 1346 87%/4%/2%/8% 8% 67/1/0/0 PHerc0125_3mm_v2 608 41% 44% 26%>15% 12% 79 3050 80%/9%/3%/8% 22% 151/3/0/118 gradients: max 51, clipped 0/50 updates Step 27,000 | dagger_launched: true Step 27,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 256, "stopped": 0, "mean_example_error": 0.472078085907044, "geometry_examples": 31, "examples": 32, "inference_and_measurement_seconds": 1.951594352722168, "total_diagnostic_seconds": 3.817750196903944} Step 27,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 6/8 (75.0%) | false stops 1 1.5-3: recovered 6/8 (75.0%) | false stops 1 3-4.5: recovered 6/8 (75.0%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 0 entire proposed 8-point prefix correct (distance): 22/28 (78.6%) gate @ 0.50: false stops 0/23 (0.0%) | accepted wrong 5/27 (18.5%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.608 1.567 following 16 464 1.157 1.172 recoverable 12 348 2.209 2.094 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 1.157 1.172 332 32 812 1.608 1.567 improved/worsened/compared states: 3/2/28 Step 27,000 | batch_diagnostics_seconds: 3.820063002407551 | recovery_seconds: 14.780859522521496 Step 27,050 | last 50 updates / 1,600 crops | lr 7.11e-07 points (35/curve, tolerance 2 vox): 47,774 right / 3,209 wrong | 93.7% correct | 5,017 unknown loss 0.8554 | geometry 0.2623 | confidence 1.1861 | mean error 0.645 vox speed: 1286 ms/update | data wait 1 ms/update | 18.8 crops/s | 18.8 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1492 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 220, from the previous prediction 1272 confidence-labeled crops: terminal 25/1521 (1.6%) | recoverable 125/1521 (8.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 352/1600 (22.0%), 2: 288/1600 (18.0%), 3: 448/1600 (28.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 117 chunks / 234.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 645 rows | mean depth 16 decisions | mean travel 138 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 53% 41% 23%>7% 16% 16 994 88%/10%/0%/2% 14% 104/0/0/0 0175A_5mm_v1 352 57% 42% 26%>1% 26% 3 347 86%/9%/0%/6% 2% 149/1/0/0 1447_5mm_v1 288 49% 44% 26%>7% 20% 15 1565 83%/7%/3%/6% 8% 99/2/0/29 paris4_afv_central 448 54% 40% 30%>6% 24% 26 1192 85%/8%/2%/5% 7% 179/3/0/0 PHerc0125_3mm_v2 256 39% 45% 22%>16% 5% 39 1997 88%/4%/3%/5% 17% 114/2/0/0 gradients: max 47, clipped 0/50 updates Step 27,100 | last 50 updates / 1,600 crops | lr 6.87e-07 points (35/curve, tolerance 2 vox): 47,095 right / 3,281 wrong | 93.5% correct | 5,624 unknown loss 0.9411 | geometry 0.3242 | confidence 1.2337 | mean error 0.678 vox speed: 1347 ms/update | data wait 1 ms/update | 23.5 crops/s | 23.5 decisions/s | 1.27 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1447 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 272, from the previous prediction 1175 confidence-labeled crops: terminal 25/1526 (1.6%) | recoverable 128/1526 (8.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 352/1600 (22.0%), 2: 256/1600 (16.0%), 3: 224/1600 (14.0%), 4: 480/1600 (30.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 117 chunks / 234.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 625 rows | mean depth 19 decisions | mean travel 155 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 49% 44% 25%>7% 18% 18 973 90%/8%/0%/2% 9% 128/0/0/0 0175A_5mm_v1 352 54% 43% 26%>3% 22% 11 819 88%/8%/1%/4% 5% 150/4/18/1 1447_5mm_v1 256 47% 50% 22%>3% 20% 14 1069 91%/6%/0%/3% 3% 83/2/0/45 paris4_afv_central 224 43% 52% 20%>5% 15% 10 903 83%/9%/1%/6% 9% 104/4/46/13 PHerc0125_3mm_v2 480 39% 43% 26%>18% 8% 75 2758 79%/9%/4%/8% 22% 160/1/0/45 gradients: max 42, clipped 0/50 updates Step 27,150 | last 50 updates / 1,600 crops | lr 6.64e-07 points (35/curve, tolerance 2 vox): 46,891 right / 3,098 wrong | 93.8% correct | 6,011 unknown loss 0.8310 | geometry 0.2685 | confidence 1.1250 | mean error 0.598 vox speed: 1369 ms/update | data wait 1 ms/update | 23.1 crops/s | 23.1 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1465 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 280, from the previous prediction 1185 confidence-labeled crops: terminal 36/1529 (2.4%) | recoverable 117/1529 (7.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 128/1600 (8.0%), 1: 512/1600 (32.0%), 2: 352/1600 (22.0%), 3: 288/1600 (18.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 117 chunks / 234.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 596 rows | mean depth 21 decisions | mean travel 170 vox | 7 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 128 48% 47% 28%>5% 23% 6 710 91%/6%/0%/3% 9% 60/0/0/0 0175A_5mm_v1 512 54% 43% 25%>3% 21% 20 1022 88%/8%/1%/3% 4% 165/3/68/55 1447_5mm_v1 352 49% 45% 22%>6% 16% 20 1054 87%/7%/1%/5% 6% 159/3/0/0 paris4_afv_central 288 47% 43% 29%>10% 19% 30 2743 82%/7%/5%/7% 14% 66/0/1/58 PHerc0125_3mm_v2 320 38% 46% 23%>17% 7% 46 2030 82%/7%/5%/7% 21% 146/8/0/0 gradients: max 37, clipped 0/50 updates Step 27,200 | last 50 updates / 1,600 crops | lr 6.41e-07 points (35/curve, tolerance 2 vox): 47,590 right / 3,885 wrong | 92.5% correct | 4,525 unknown loss 0.8993 | geometry 0.2970 | confidence 1.2047 | mean error 0.651 vox speed: 1336 ms/update | data wait 1 ms/update | 23.7 crops/s | 23.7 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1468 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 245, from the previous prediction 1223 confidence-labeled crops: terminal 36/1542 (2.3%) | recoverable 108/1542 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 224/1600 (14.0%), 2: 384/1600 (24.0%), 3: 416/1600 (26.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 117 chunks / 234.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 639 rows | mean depth 19 decisions | mean travel 164 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 49% 41% 29%>10% 19% 29 1365 88%/7%/0%/5% 13% 145/0/0/0 0175A_5mm_v1 224 50% 45% 27%>5% 22% 10 644 87%/8%/3%/3% 7% 100/0/0/0 1447_5mm_v1 384 49% 48% 22%>3% 19% 11 519 90%/4%/2%/4% 6% 166/2/39/18 paris4_afv_central 416 45% 47% 24%>8% 17% 28 1686 87%/6%/3%/4% 9% 138/0/0/57 PHerc0125_3mm_v2 224 42% 40% 29%>17% 12% 32 2269 80%/10%/5%/5% 22% 90/3/0/0 gradients: max 70, clipped 1/50 updates Step 27,250 | last 50 updates / 1,600 crops | lr 6.18e-07 points (35/curve, tolerance 2 vox): 48,509 right / 2,518 wrong | 95.1% correct | 4,973 unknown loss 0.8221 | geometry 0.2513 | confidence 1.1417 | mean error 0.606 vox speed: 1352 ms/update | data wait 1 ms/update | 23.4 crops/s | 23.4 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1471 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 248, from the previous prediction 1223 confidence-labeled crops: terminal 21/1536 (1.4%) | recoverable 110/1536 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 320/1600 (20.0%), 2: 416/1600 (26.0%), 3: 256/1600 (16.0%), 4: 160/1600 (10.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 117 chunks / 234.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 586 rows | mean depth 20 decisions | mean travel 160 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 50% 43% 29%>7% 22% 27 1241 88%/9%/0%/3% 12% 164/1/40/29 0175A_5mm_v1 320 50% 43% 27%>8% 19% 23 1216 88%/5%/2%/6% 6% 137/2/0/0 1447_5mm_v1 416 49% 44% 26%>6% 20% 28 2430 89%/6%/1%/3% 5% 110/0/0/75 paris4_afv_central 256 50% 45% 25%>5% 20% 11 974 89%/7%/2%/3% 6% 114/2/0/0 PHerc0125_3mm_v2 160 50% 38% 27%>12% 15% 18 1834 82%/6%/3%/9% 14% 61/1/0/0 gradients: max 40, clipped 0/50 updates Step 27,300 | last 50 updates / 1,600 crops | lr 5.96e-07 points (35/curve, tolerance 2 vox): 48,110 right / 2,698 wrong | 94.7% correct | 5,192 unknown loss 0.7858 | geometry 0.2319 | confidence 1.1078 | mean error 0.589 vox speed: 1337 ms/update | data wait 1 ms/update | 23.7 crops/s | 23.7 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1485 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 231, from the previous prediction 1254 confidence-labeled crops: terminal 27/1531 (1.8%) | recoverable 94/1531 (6.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 384/1600 (24.0%), 2: 320/1600 (20.0%), 3: 352/1600 (22.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 117 chunks / 234.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 670 rows | mean depth 23 decisions | mean travel 192 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 50% 45% 25%>5% 20% 14 850 91%/8%/0%/1% 7% 137/0/0/21 0175A_5mm_v1 384 51% 44% 26%>5% 21% 16 663 88%/7%/1%/4% 4% 169/3/1/0 1447_5mm_v1 320 48% 44% 24%>8% 17% 25 1307 88%/6%/2%/4% 6% 141/0/0/0 paris4_afv_central 352 44% 51% 22%>6% 16% 16 1188 87%/4%/1%/8% 7% 135/1/0/43 PHerc0125_3mm_v2 192 36% 46% 25%>17% 8% 33 2118 83%/5%/6%/6% 21% 88/0/33/1 gradients: max 44, clipped 0/50 updates Step 27,350 | last 50 updates / 1,600 crops | lr 5.74e-07 points (35/curve, tolerance 2 vox): 47,789 right / 2,908 wrong | 94.3% correct | 5,303 unknown loss 0.8320 | geometry 0.2828 | confidence 1.0984 | mean error 0.646 vox speed: 1357 ms/update | data wait 1 ms/update | 23.3 crops/s | 23.3 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1473 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 283, from the previous prediction 1190 confidence-labeled crops: terminal 26/1525 (1.7%) | recoverable 105/1525 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 224/1600 (14.0%), 2: 416/1600 (26.0%), 3: 480/1600 (30.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 117 chunks / 234.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 623 rows | mean depth 15 decisions | mean travel 125 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 47% 47% 23%>7% 17% 16 1058 90%/8%/0%/2% 11% 109/0/0/26 0175A_5mm_v1 224 50% 46% 27%>4% 23% 9 604 86%/8%/0%/5% 6% 103/2/0/0 1447_5mm_v1 416 50% 46% 26%>3% 22% 14 945 91%/5%/1%/4% 2% 156/2/0/36 paris4_afv_central 480 43% 47% 23%>10% 14% 49 2079 85%/7%/3%/6% 8% 177/6/0/49 PHerc0125_3mm_v2 192 39% 48% 20%>14% 7% 28 2068 81%/7%/4%/8% 22% 78/0/43/14 gradients: max 43, clipped 0/50 updates Step 27,400 | last 50 updates / 1,600 crops | lr 5.53e-07 points (35/curve, tolerance 2 vox): 46,758 right / 3,505 wrong | 93.0% correct | 5,737 unknown loss 0.9704 | geometry 0.3404 | confidence 1.2600 | mean error 0.698 vox speed: 1362 ms/update | data wait 1 ms/update | 23.3 crops/s | 23.3 decisions/s | 1.27 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1456 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 214, from the previous prediction 1242 confidence-labeled crops: terminal 24/1520 (1.6%) | recoverable 112/1520 (7.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 288/1600 (18.0%), 2: 224/1600 (14.0%), 3: 256/1600 (16.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 117 chunks / 234.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 651 rows | mean depth 18 decisions | mean travel 150 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 47% 45% 27%>8% 19% 27 1094 88%/8%/0%/4% 11% 201/3/0/2 0175A_5mm_v1 288 57% 43% 29%>0% 29% 0 154 90%/5%/0%/5% 4% 125/2/0/0 1447_5mm_v1 224 53% 47% 21%>0% 21% 0 143 91%/7%/0%/3% 3% 106/2/0/0 paris4_afv_central 256 44% 49% 25%>7% 18% 18 1432 86%/5%/2%/7% 7% 107/1/0/18 PHerc0125_3mm_v2 384 39% 42% 31%>20% 11% 71 3321 78%/9%/5%/9% 24% 112/1/31/49 gradients: max 59, clipped 0/50 updates Step 27,450 | last 50 updates / 1,600 crops | lr 5.32e-07 points (35/curve, tolerance 2 vox): 46,967 right / 3,369 wrong | 93.3% correct | 5,664 unknown loss 1.0736 | geometry 0.3237 | confidence 1.4998 | mean error 0.708 vox speed: 1354 ms/update | data wait 1 ms/update | 23.4 crops/s | 23.4 decisions/s | 1.24 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1459 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 300, from the previous prediction 1159 confidence-labeled crops: terminal 11/1528 (0.7%) | recoverable 120/1528 (7.9%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 480/1600 (30.0%), 1: 224/1600 (14.0%), 2: 224/1600 (14.0%), 3: 384/1600 (24.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 542 rows | mean depth 24 decisions | mean travel 199 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 480 46% 43% 22%>10% 12% 36 2292 88%/8%/0%/4% 16% 115/0/0/93 0175A_5mm_v1 224 54% 42% 25%>4% 22% 8 968 88%/6%/0%/6% 3% 95/1/0/0 1447_5mm_v1 224 56% 38% 32%>6% 25% 13 752 88%/9%/0%/3% 5% 85/2/0/0 paris4_afv_central 384 44% 49% 24%>7% 17% 26 1312 85%/7%/2%/6% 9% 170/2/0/17 PHerc0125_3mm_v2 288 49% 44% 23%>7% 16% 31 2234 88%/7%/1%/5% 17% 77/1/0/49 gradients: max 59, clipped 0/50 updates Step 27,500 | last 50 updates / 1,600 crops | lr 5.12e-07 points (35/curve, tolerance 2 vox): 47,213 right / 3,772 wrong | 92.6% correct | 5,015 unknown loss 1.0240 | geometry 0.3076 | confidence 1.4327 | mean error 0.697 vox speed: 1361 ms/update | data wait 1 ms/update | 23.3 crops/s | 23.3 decisions/s | 1.24 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1461 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 251, from the previous prediction 1210 confidence-labeled crops: terminal 31/1534 (2.0%) | recoverable 98/1534 (6.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 320/1600 (20.0%), 2: 288/1600 (18.0%), 3: 192/1600 (12.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 658 rows | mean depth 17 decisions | mean travel 143 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 45% 48% 23%>7% 16% 27 1553 89%/8%/0%/3% 14% 132/0/0/53 0175A_5mm_v1 320 48% 48% 22%>4% 18% 13 873 90%/6%/2%/2% 5% 150/2/8/3 1447_5mm_v1 288 49% 48% 23%>3% 20% 7 670 89%/5%/1%/5% 2% 139/2/0/0 paris4_afv_central 192 49% 43% 24%>7% 17% 14 1356 87%/6%/3%/4% 7% 83/2/0/0 PHerc0125_3mm_v2 416 41% 44% 27%>16% 11% 63 2337 83%/6%/4%/7% 20% 154/4/0/27 gradients: max 49, clipped 0/50 updates Step 27,550 | last 50 updates / 1,600 crops | lr 4.91e-07 points (35/curve, tolerance 2 vox): 47,843 right / 2,776 wrong | 94.5% correct | 5,381 unknown loss 0.8126 | geometry 0.2435 | confidence 1.1382 | mean error 0.601 vox speed: 1329 ms/update | data wait 1 ms/update | 23.8 crops/s | 23.8 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1494 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 271, from the previous prediction 1223 confidence-labeled crops: terminal 24/1532 (1.6%) | recoverable 104/1532 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 352/1600 (22.0%), 2: 416/1600 (26.0%), 3: 256/1600 (16.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 605 rows | mean depth 19 decisions | mean travel 159 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 50% 46% 19%>4% 16% 8 608 90%/7%/0%/3% 9% 104/0/0/0 0175A_5mm_v1 352 49% 46% 26%>5% 22% 15 718 89%/7%/1%/4% 5% 132/4/0/31 1447_5mm_v1 416 54% 44% 25%>2% 23% 8 329 88%/7%/1%/5% 4% 184/6/0/0 paris4_afv_central 256 55% 39% 30%>6% 24% 14 1275 87%/7%/2%/4% 7% 99/3/0/0 PHerc0125_3mm_v2 352 39% 46% 26%>15% 11% 47 3141 84%/5%/4%/7% 20% 86/3/0/76 gradients: max 41, clipped 0/50 updates Step 27,559 | dagger_states: 6995 | dagger_source_step: 27000 | dagger_caches: 4 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/0175A_5mm_v1/decisions_027000.npz" | dagger_fibers: 384 | dagger_supply: {"states": 6995, "traces": 384, "fibers": 384, "rows": {"terminal": 191, "premature_stop": 126, "recoverable": 10, "pre_excursion": 185, "ordinary": 6477}, "events": {"terminal": 25, "premature_stop": 126, "recoverable": 6, "pre_excursion": 32, "ordinary": 379}, "supervision": {"following": 6787, "recoverable": 9, "terminal": 191, "unknown": 8}, "departed_traces": 33, "terminal_traces": 25, "stop_reasons": {"after_failure": 24, "confidence": 129, "unannotated": 231}, "traced_voxels": 106161.18659110522, "seconds": 770.2279464825988, "voxels_per_second": 137.83086822013104} | dagger_coverage: {"requested": 384, "seeds": 384, "skipped": [], "epoch": 0} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 0 | dagger_launch_interval: 5000 | dagger_publication_age: 559 | dataset: "0175A_5mm_v1" Step 27,600 | last 50 updates / 1,600 crops | lr 4.72e-07 points (35/curve, tolerance 2 vox): 48,469 right / 2,945 wrong | 94.3% correct | 4,586 unknown loss 0.8861 | geometry 0.2837 | confidence 1.2049 | mean error 0.671 vox speed: 932 ms/update | data wait 1 ms/update | 33.9 crops/s | 33.9 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1498 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 307, from the previous prediction 1191 confidence-labeled crops: terminal 15/1542 (1.0%) | recoverable 119/1542 (7.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 224/1600 (14.0%), 2: 384/1600 (24.0%), 3: 256/1600 (16.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 591 rows | mean depth 22 decisions | mean travel 185 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 44% 47% 24%>9% 15% 31 1179 90%/7%/0%/3% 12% 210/1/0/0 0175A_5mm_v1 224 58% 40% 25%>2% 22% 6 682 89%/8%/0%/3% 3% 54/1/0/35 1447_5mm_v1 384 57% 42% 25%>1% 24% 3 468 88%/9%/0%/3% 1% 81/1/106/82 paris4_afv_central 256 40% 55% 18%>5% 12% 12 886 88%/7%/2%/3% 8% 140/4/0/0 PHerc0125_3mm_v2 288 43% 46% 23%>10% 12% 38 2172 82%/7%/3%/8% 20% 106/3/0/27 gradients: max 50, clipped 0/50 updates Step 27,650 | last 50 updates / 1,600 crops | lr 4.52e-07 points (35/curve, tolerance 2 vox): 46,388 right / 4,031 wrong | 92.0% correct | 5,581 unknown loss 0.9573 | geometry 0.2926 | confidence 1.3294 | mean error 0.677 vox speed: 892 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1457 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 298, from the previous prediction 1159 confidence-labeled crops: terminal 41/1522 (2.7%) | recoverable 112/1522 (7.4%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 160/1600 (10.0%), 1: 224/1600 (14.0%), 2: 288/1600 (18.0%), 3: 448/1600 (28.0%), 4: 480/1600 (30.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 1557 chunk references / 1 windows live chains: 572 rows | mean depth 17 decisions | mean travel 138 vox | 6 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 160 44% 47% 23%>9% 14% 9 1175 86%/8%/1%/5% 13% 75/0/0/0 0175A_5mm_v1 224 47% 49% 26%>4% 21% 9 617 87%/8%/0%/4% 5% 109/2/0/0 1447_5mm_v1 288 57% 43% 27%>0% 27% 0 140 87%/9%/0%/4% 2% 78/1/0/45 paris4_afv_central 448 48% 44% 26%>8% 18% 35 2214 86%/6%/3%/4% 10% 137/3/74/59 PHerc0125_3mm_v2 480 42% 43% 26%>15% 12% 61 2405 83%/6%/5%/6% 20% 173/3/0/32 gradients: max 52, clipped 0/50 updates Step 27,700 | last 50 updates / 1,600 crops | lr 4.33e-07 points (35/curve, tolerance 2 vox): 47,862 right / 2,981 wrong | 94.1% correct | 5,157 unknown loss 0.7781 | geometry 0.2436 | confidence 1.0690 | mean error 0.598 vox speed: 891 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1459 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 276, from the previous prediction 1183 confidence-labeled crops: terminal 26/1532 (1.7%) | recoverable 112/1532 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 320/1600 (20.0%), 2: 416/1600 (26.0%), 3: 416/1600 (26.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 591 rows | mean depth 20 decisions | mean travel 161 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 39% 42% 29%>19% 9% 24 2230 82%/9%/3%/6% 16% 80/0/0/0 0175A_5mm_v1 320 48% 47% 25%>5% 20% 16 729 88%/6%/2%/4% 6% 149/3/0/0 1447_5mm_v1 416 47% 49% 23%>4% 19% 16 673 90%/6%/1%/3% 4% 197/5/0/6 paris4_afv_central 416 51% 42% 27%>6% 21% 27 2152 88%/6%/1%/4% 9% 110/1/0/65 PHerc0125_3mm_v2 256 50% 39% 28%>11% 17% 31 3009 81%/8%/2%/9% 16% 55/0/0/46 gradients: max 50, clipped 0/50 updates Step 27,750 | last 50 updates / 1,600 crops | lr 4.15e-07 points (35/curve, tolerance 2 vox): 46,224 right / 4,059 wrong | 91.9% correct | 5,717 unknown loss 0.9737 | geometry 0.3285 | confidence 1.2905 | mean error 0.708 vox speed: 893 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.29 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1443 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 226, from the previous prediction 1217 confidence-labeled crops: terminal 32/1523 (2.1%) | recoverable 108/1523 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 352/1600 (22.0%), 2: 288/1600 (18.0%), 3: 256/1600 (16.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 651 rows | mean depth 17 decisions | mean travel 145 vox | 6 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 42% 41% 29%>17% 12% 32 2097 87%/6%/2%/5% 19% 118/0/0/0 0175A_5mm_v1 352 51% 47% 28%>2% 26% 9 1003 88%/9%/0%/3% 4% 149/1/2/15 1447_5mm_v1 288 45% 49% 21%>6% 15% 16 880 87%/8%/2%/4% 6% 142/1/0/0 paris4_afv_central 256 52% 42% 25%>6% 19% 16 1543 87%/5%/3%/4% 8% 97/2/0/11 PHerc0125_3mm_v2 416 43% 45% 26%>12% 14% 51 2343 83%/6%/3%/8% 22% 145/6/46/41 gradients: max 48, clipped 0/50 updates Step 27,800 | last 50 updates / 1,600 crops | lr 3.97e-07 points (35/curve, tolerance 2 vox): 46,923 right / 3,078 wrong | 93.8% correct | 5,999 unknown loss 0.8898 | geometry 0.2828 | confidence 1.2140 | mean error 0.638 vox speed: 892 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.24 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1473 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 309, from the previous prediction 1164 confidence-labeled crops: terminal 23/1517 (1.5%) | recoverable 106/1517 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 352/1600 (22.0%), 2: 480/1600 (30.0%), 3: 32/1600 (2.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 570 rows | mean depth 18 decisions | mean travel 154 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 49% 45% 22%>6% 16% 19 1282 88%/6%/1%/5% 11% 126/0/128/31 0175A_5mm_v1 352 45% 43% 28%>11% 16% 33 1929 85%/6%/2%/7% 11% 134/5/0/19 1447_5mm_v1 480 53% 45% 25%>2% 24% 9 442 89%/7%/1%/4% 3% 131/3/0/85 paris4_afv_central 32 56% 44% 31%>0% 31% 0 187 91%/6%/0%/3% 1% 14/0/17/0 PHerc0125_3mm_v2 384 45% 43% 26%>12% 13% 41 1867 82%/7%/3%/8% 18% 165/3/0/0 gradients: max 34, clipped 0/50 updates Step 27,850 | last 50 updates / 1,600 crops | lr 3.79e-07 points (35/curve, tolerance 2 vox): 46,478 right / 3,613 wrong | 92.8% correct | 5,909 unknown loss 0.9612 | geometry 0.2897 | confidence 1.3429 | mean error 0.668 vox speed: 891 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.27 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1454 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 289, from the previous prediction 1165 confidence-labeled crops: terminal 32/1525 (2.1%) | recoverable 103/1525 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 352/1600 (22.0%), 2: 192/1600 (12.0%), 3: 352/1600 (22.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 368 chunk references / 1 windows live chains: 617 rows | mean depth 19 decisions | mean travel 159 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 49% 41% 26%>9% 17% 22 2107 89%/9%/0%/2% 12% 67/0/0/52 0175A_5mm_v1 352 47% 50% 20%>2% 18% 8 399 90%/6%/1%/4% 3% 168/3/0/9 1447_5mm_v1 192 43% 49% 22%>8% 15% 14 1105 88%/5%/2%/6% 5% 95/1/0/0 paris4_afv_central 352 42% 51% 19%>7% 12% 20 1185 88%/6%/2%/5% 9% 172/3/85/9 PHerc0125_3mm_v2 416 40% 46% 25%>14% 11% 65 3100 81%/6%/5%/8% 23% 115/0/0/76 gradients: max 53, clipped 0/50 updates Step 27,900 | last 50 updates / 1,600 crops | lr 3.62e-07 points (35/curve, tolerance 2 vox): 47,064 right / 3,300 wrong | 93.4% correct | 5,636 unknown loss 0.9839 | geometry 0.3028 | confidence 1.3622 | mean error 0.684 vox speed: 894 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.27 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1485 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 288, from the previous prediction 1197 confidence-labeled crops: terminal 20/1519 (1.3%) | recoverable 118/1519 (7.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 192/1600 (12.0%), 2: 288/1600 (18.0%), 3: 256/1600 (16.0%), 4: 576/1600 (36.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 1149 chunk references / 1 windows live chains: 596 rows | mean depth 18 decisions | mean travel 153 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 38% 46% 25%>16% 9% 29 1854 85%/9%/1%/4% 17% 133/0/0/0 0175A_5mm_v1 192 56% 44% 22%>0% 22% 0 153 94%/4%/0%/3% 2% 85/5/0/0 1447_5mm_v1 288 43% 45% 24%>12% 12% 30 1331 87%/4%/3%/6% 7% 129/2/0/0 paris4_afv_central 256 53% 47% 21%>0% 21% 0 147 89%/6%/0%/5% 6% 104/1/0/16 PHerc0125_3mm_v2 576 51% 44% 28%>5% 23% 35 1673 83%/10%/1%/6% 14% 145/1/0/109 gradients: max 43, clipped 0/50 updates Step 27,950 | last 50 updates / 1,600 crops | lr 3.45e-07 points (35/curve, tolerance 2 vox): 47,262 right / 2,904 wrong | 94.2% correct | 5,834 unknown loss 0.8133 | geometry 0.2710 | confidence 1.0846 | mean error 0.642 vox speed: 891 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.25 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1490 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 263, from the previous prediction 1227 confidence-labeled crops: terminal 23/1525 (1.5%) | recoverable 116/1525 (7.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 288/1600 (18.0%), 2: 480/1600 (30.0%), 3: 224/1600 (14.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 616 rows | mean depth 14 decisions | mean travel 121 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 46% 45% 27%>8% 18% 16 1231 91%/7%/0%/3% 12% 87/0/0/0 0175A_5mm_v1 288 52% 44% 26%>3% 22% 8 566 89%/6%/1%/4% 5% 128/1/0/0 1447_5mm_v1 480 52% 47% 25%>1% 24% 6 463 86%/9%/0%/4% 1% 214/7/6/11 paris4_afv_central 224 50% 46% 29%>3% 26% 11 2137 89%/6%/0%/4% 6% 44/1/0/60 PHerc0125_3mm_v2 416 39% 47% 25%>14% 11% 58 2540 82%/6%/5%/7% 21% 143/3/0/53 gradients: max 37, clipped 0/50 updates Step 28,000 | last 50 updates / 1,600 crops | lr 3.28e-07 points (35/curve, tolerance 2 vox): 47,075 right / 3,005 wrong | 94.0% correct | 5,920 unknown loss 0.8374 | geometry 0.2667 | confidence 1.1413 | mean error 0.615 vox speed: 893 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.25 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1475 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 207, from the previous prediction 1268 confidence-labeled crops: terminal 24/1513 (1.6%) | recoverable 110/1513 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 384/1600 (24.0%), 2: 256/1600 (16.0%), 3: 320/1600 (20.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 679 rows | mean depth 20 decisions | mean travel 164 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 43% 46% 25%>11% 14% 18 1451 88%/8%/0%/4% 12% 103/0/2/0 0175A_5mm_v1 384 54% 45% 26%>1% 25% 5 311 86%/8%/0%/6% 3% 166/2/29/6 1447_5mm_v1 256 54% 42% 26%>4% 22% 10 557 88%/7%/0%/4% 3% 108/1/0/0 paris4_afv_central 320 48% 47% 22%>5% 18% 14 942 89%/5%/1%/5% 7% 151/3/0/0 PHerc0125_3mm_v2 416 42% 41% 24%>17% 7% 58 2647 80%/6%/4%/9% 21% 151/4/0/20 gradients: max 49, clipped 0/50 updates Step 28,000 | dagger_launched: true Step 28,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 250, "stopped": 0, "mean_example_error": 0.4122824336091677, "geometry_examples": 30, "examples": 32, "inference_and_measurement_seconds": 1.9839247018098831, "total_diagnostic_seconds": 3.777825240045786} Step 28,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 6/8 (75.0%) | false stops 1 1.5-3: recovered 6/8 (75.0%) | false stops 1 3-4.5: recovered 6/8 (75.0%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 0 entire proposed 8-point prefix correct (distance): 22/28 (78.6%) gate @ 0.50: false stops 0/23 (0.0%) | accepted wrong 5/27 (18.5%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.611 1.569 following 16 464 1.158 1.174 recoverable 12 348 2.215 2.096 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 1.158 1.174 332 32 812 1.611 1.569 improved/worsened/compared states: 3/2/28 Step 28,000 | batch_diagnostics_seconds: 3.7803662605583668 | recovery_seconds: 14.499374981969595 Step 28,050 | last 50 updates / 1,600 crops | lr 3.12e-07 points (35/curve, tolerance 2 vox): 48,772 right / 2,348 wrong | 95.4% correct | 4,880 unknown loss 0.7647 | geometry 0.2534 | confidence 1.0227 | mean error 0.594 vox speed: 1325 ms/update | data wait 1 ms/update | 18.5 crops/s | 18.5 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1534 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 289, from the previous prediction 1245 confidence-labeled crops: terminal 17/1536 (1.1%) | recoverable 120/1536 (7.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 480/1600 (30.0%), 2: 384/1600 (24.0%), 3: 256/1600 (16.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 594 rows | mean depth 20 decisions | mean travel 163 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 43% 50% 22%>7% 15% 14 1088 91%/7%/0%/2% 12% 119/0/16/8 0175A_5mm_v1 480 54% 44% 27%>1% 25% 6 488 86%/7%/0%/7% 3% 110/4/0/102 1447_5mm_v1 384 52% 45% 24%>4% 21% 13 616 89%/8%/2%/2% 5% 170/3/0/1 paris4_afv_central 256 52% 48% 23%>0% 22% 1 244 86%/8%/1%/5% 5% 102/3/0/20 PHerc0125_3mm_v2 224 47% 42% 29%>11% 17% 24 1821 86%/7%/3%/5% 19% 93/3/0/0 gradients: max 52, clipped 0/50 updates Step 28,100 | last 50 updates / 1,600 crops | lr 2.96e-07 points (35/curve, tolerance 2 vox): 48,711 right / 2,532 wrong | 95.1% correct | 4,757 unknown loss 0.8208 | geometry 0.2842 | confidence 1.0731 | mean error 0.626 vox speed: 1363 ms/update | data wait 1 ms/update | 23.3 crops/s | 23.3 decisions/s | 1.25 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1529 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 251, from the previous prediction 1278 confidence-labeled crops: terminal 15/1538 (1.0%) | recoverable 106/1538 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 224/1600 (14.0%), 2: 288/1600 (18.0%), 3: 384/1600 (24.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 627 rows | mean depth 16 decisions | mean travel 138 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 45% 47% 27%>8% 19% 24 1778 88%/8%/1%/3% 13% 116/0/63/64 0175A_5mm_v1 224 51% 45% 26%>4% 22% 9 632 88%/5%/0%/7% 4% 100/1/0/0 1447_5mm_v1 288 60% 40% 27%>0% 27% 0 143 90%/7%/0%/2% 1% 114/1/0/0 paris4_afv_central 384 55% 44% 23%>1% 22% 4 358 88%/7%/1%/4% 5% 156/4/0/13 PHerc0125_3mm_v2 320 49% 44% 26%>7% 19% 19 1228 87%/5%/2%/6% 13% 141/4/0/0 gradients: max 41, clipped 0/50 updates Step 28,150 | last 50 updates / 1,600 crops | lr 2.81e-07 points (35/curve, tolerance 2 vox): 48,638 right / 2,591 wrong | 94.9% correct | 4,771 unknown loss 0.7450 | geometry 0.2207 | confidence 1.0487 | mean error 0.566 vox speed: 1371 ms/update | data wait 1 ms/update | 23.0 crops/s | 23.0 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1508 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 201, from the previous prediction 1307 confidence-labeled crops: terminal 24/1538 (1.6%) | recoverable 100/1538 (6.5%) | target connections crossing a neighbor 1 datasets (IDs from dataset_configuration): 0: 160/1600 (10.0%), 1: 480/1600 (30.0%), 2: 384/1600 (24.0%), 3: 256/1600 (16.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 662 rows | mean depth 17 decisions | mean travel 145 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 160 52% 46% 22%>2% 21% 2 419 92%/6%/0%/1% 4% 73/0/0/0 0175A_5mm_v1 480 49% 42% 25%>9% 16% 30 1218 87%/6%/2%/5% 9% 201/5/0/0 1447_5mm_v1 384 56% 42% 25%>2% 23% 9 410 90%/7%/1%/3% 3% 160/3/0/0 paris4_afv_central 256 49% 48% 23%>2% 20% 6 690 89%/6%/1%/4% 7% 93/1/0/31 PHerc0125_3mm_v2 320 46% 45% 24%>9% 15% 26 1600 84%/6%/2%/8% 15% 135/2/67/8 gradients: max 47, clipped 0/50 updates Step 28,200 | last 50 updates / 1,600 crops | lr 2.66e-07 points (35/curve, tolerance 2 vox): 48,077 right / 2,907 wrong | 94.3% correct | 5,016 unknown loss 0.8701 | geometry 0.2649 | confidence 1.2105 | mean error 0.630 vox speed: 1367 ms/update | data wait 1 ms/update | 23.2 crops/s | 23.2 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1481 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 226, from the previous prediction 1255 confidence-labeled crops: terminal 20/1537 (1.3%) | recoverable 112/1537 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 288/1600 (18.0%), 2: 320/1600 (20.0%), 3: 384/1600 (24.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 673 rows | mean depth 19 decisions | mean travel 158 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 43% 48% 25%>9% 16% 30 1288 90%/8%/0%/2% 12% 169/2/0/0 0175A_5mm_v1 288 49% 47% 24%>5% 20% 11 684 90%/7%/0%/3% 4% 134/2/0/0 1447_5mm_v1 320 52% 43% 29%>5% 24% 14 636 87%/9%/2%/2% 5% 139/1/0/0 paris4_afv_central 384 49% 44% 24%>6% 18% 26 1555 88%/5%/2%/5% 7% 151/4/0/19 PHerc0125_3mm_v2 256 48% 43% 26%>9% 16% 30 2329 84%/6%/3%/7% 17% 80/2/61/29 gradients: max 52, clipped 0/50 updates Step 28,250 | last 50 updates / 1,600 crops | lr 2.51e-07 points (35/curve, tolerance 2 vox): 50,229 right / 1,738 wrong | 96.7% correct | 4,033 unknown loss 0.6727 | geometry 0.2094 | confidence 0.9266 | mean error 0.566 vox speed: 1369 ms/update | data wait 1 ms/update | 23.1 crops/s | 23.1 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1563 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 283, from the previous prediction 1280 confidence-labeled crops: terminal 9/1555 (0.6%) | recoverable 109/1555 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 384/1600 (24.0%), 2: 352/1600 (22.0%), 3: 480/1600 (30.0%), 4: 128/1600 (8.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 631 rows | mean depth 15 decisions | mean travel 127 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 54% 43% 23%>2% 21% 6 505 91%/7%/0%/1% 7% 111/0/0/0 0175A_5mm_v1 384 48% 47% 25%>5% 20% 16 774 89%/7%/1%/3% 7% 163/3/0/19 1447_5mm_v1 352 57% 42% 24%>2% 23% 5 361 90%/6%/1%/3% 3% 133/2/12/14 paris4_afv_central 480 55% 44% 24%>1% 22% 6 462 89%/7%/0%/4% 5% 161/1/0/49 PHerc0125_3mm_v2 128 51% 49% 19%>0% 19% 0 149 86%/9%/0%/5% 8% 63/2/0/0 gradients: max 54, clipped 0/50 updates Step 28,300 | last 50 updates / 1,600 crops | lr 2.37e-07 points (35/curve, tolerance 2 vox): 47,962 right / 2,819 wrong | 94.4% correct | 5,219 unknown loss 0.8829 | geometry 0.2615 | confidence 1.2428 | mean error 0.626 vox speed: 1371 ms/update | data wait 1 ms/update | 23.1 crops/s | 23.1 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1485 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 268, from the previous prediction 1217 confidence-labeled crops: terminal 21/1532 (1.4%) | recoverable 108/1532 (7.0%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 448/1600 (28.0%), 2: 320/1600 (20.0%), 3: 288/1600 (18.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 625 rows | mean depth 18 decisions | mean travel 150 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 40% 50% 23%>10% 12% 17 1639 90%/6%/0%/4% 12% 79/1/13/17 0175A_5mm_v1 448 46% 44% 24%>9% 14% 34 1621 86%/7%/1%/5% 8% 161/3/0/38 1447_5mm_v1 320 46% 45% 23%>9% 14% 24 1005 87%/7%/2%/4% 7% 145/3/0/0 paris4_afv_central 288 53% 44% 23%>3% 20% 8 985 89%/6%/1%/4% 4% 78/2/0/50 PHerc0125_3mm_v2 352 52% 46% 24%>2% 22% 6 447 88%/7%/1%/4% 13% 162/4/0/0 gradients: max 42, clipped 0/50 updates Step 28,350 | last 50 updates / 1,600 crops | lr 2.24e-07 points (35/curve, tolerance 2 vox): 46,955 right / 3,336 wrong | 93.4% correct | 5,709 unknown loss 0.9019 | geometry 0.2891 | confidence 1.2256 | mean error 0.663 vox speed: 1372 ms/update | data wait 1 ms/update | 23.0 crops/s | 23.0 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1492 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 281, from the previous prediction 1211 confidence-labeled crops: terminal 23/1521 (1.5%) | recoverable 101/1521 (6.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 352/1600 (22.0%), 2: 224/1600 (14.0%), 3: 256/1600 (16.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 615 rows | mean depth 19 decisions | mean travel 155 vox | 8 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 48% 44% 23%>8% 16% 18 1192 91%/7%/0%/3% 11% 141/2/0/0 0175A_5mm_v1 352 49% 47% 22%>4% 19% 12 775 89%/7%/0%/4% 5% 130/3/0/35 1447_5mm_v1 224 54% 46% 24%>0% 24% 0 152 92%/6%/0%/2% 2% 102/0/0/0 paris4_afv_central 256 48% 44% 27%>9% 19% 19 1667 86%/6%/2%/6% 7% 112/0/0/0 PHerc0125_3mm_v2 448 48% 44% 23%>9% 15% 42 2318 81%/6%/4%/8% 20% 130/1/0/66 gradients: max 41, clipped 0/50 updates Step 28,400 | last 50 updates / 1,600 crops | lr 2.10e-07 points (35/curve, tolerance 2 vox): 47,624 right / 3,221 wrong | 93.7% correct | 5,155 unknown loss 0.9284 | geometry 0.3047 | confidence 1.2475 | mean error 0.672 vox speed: 1378 ms/update | data wait 1 ms/update | 23.0 crops/s | 23.0 decisions/s | 1.25 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1500 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 283, from the previous prediction 1217 confidence-labeled crops: terminal 24/1543 (1.6%) | recoverable 112/1543 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 160/1600 (10.0%), 2: 448/1600 (28.0%), 3: 192/1600 (12.0%), 4: 512/1600 (32.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 577 rows | mean depth 15 decisions | mean travel 125 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 56% 41% 27%>3% 23% 10 746 88%/10%/0%/3% 7% 110/1/0/8 0175A_5mm_v1 160 51% 43% 31%>6% 25% 9 878 88%/8%/1%/4% 6% 69/2/0/0 1447_5mm_v1 448 50% 44% 23%>6% 17% 22 757 87%/8%/2%/3% 6% 198/2/0/0 paris4_afv_central 192 49% 47% 18%>4% 14% 6 523 92%/6%/1%/2% 6% 90/2/0/0 PHerc0125_3mm_v2 512 50% 44% 22%>6% 16% 39 2638 87%/5%/2%/6% 18% 110/2/0/116 gradients: max 61, clipped 1/50 updates Step 28,450 | last 50 updates / 1,600 crops | lr 1.97e-07 points (35/curve, tolerance 2 vox): 48,356 right / 2,474 wrong | 95.1% correct | 5,170 unknown loss 0.8558 | geometry 0.2610 | confidence 1.1896 | mean error 0.598 vox speed: 1347 ms/update | data wait 1 ms/update | 23.5 crops/s | 23.5 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1483 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 285, from the previous prediction 1198 confidence-labeled crops: terminal 21/1538 (1.4%) | recoverable 117/1538 (7.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 192/1600 (12.0%), 2: 448/1600 (28.0%), 3: 384/1600 (24.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 600 rows | mean depth 21 decisions | mean travel 176 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 48% 44% 27%>8% 19% 19 1296 88%/7%/0%/5% 11% 122/0/0/5 0175A_5mm_v1 192 41% 43% 32%>17% 16% 29 1951 87%/5%/4%/4% 12% 81/2/52/1 1447_5mm_v1 448 48% 45% 23%>7% 16% 34 1499 86%/9%/1%/4% 6% 138/4/0/62 paris4_afv_central 384 54% 44% 25%>2% 22% 9 674 89%/8%/1%/3% 4% 156/2/63/13 PHerc0125_3mm_v2 288 45% 52% 27%>2% 24% 8 805 88%/7%/1%/4% 14% 103/0/0/47 gradients: max 58, clipped 0/50 updates Step 28,494 | dagger_states: 6484 | dagger_source_step: 28000 | dagger_caches: 4 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/1447_5mm_v1/decisions_028000.npz" | dagger_fibers: 384 | dagger_supply: {"states": 6484, "traces": 384, "fibers": 384, "rows": {"terminal": 199, "premature_stop": 144, "recoverable": 16, "pre_excursion": 174, "ordinary": 5946}, "events": {"terminal": 27, "premature_stop": 144, "recoverable": 5, "pre_excursion": 30, "ordinary": 376}, "supervision": {"following": 6264, "recoverable": 14, "terminal": 199, "unknown": 7}, "departed_traces": 29, "terminal_traces": 27, "stop_reasons": {"after_failure": 23, "confidence": 149, "unannotated": 212}, "traced_voxels": 97122.37284053776, "seconds": 690.9172544814646, "voxels_per_second": 140.57019449228892} | dagger_coverage: {"requested": 384, "seeds": 384, "skipped": [], "epoch": 0} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 0 | dagger_launch_interval: 5000 | dagger_publication_age: 494 | dataset: "1447_5mm_v1" Step 28,500 | last 50 updates / 1,600 crops | lr 1.85e-07 points (35/curve, tolerance 2 vox): 47,827 right / 3,179 wrong | 93.8% correct | 4,994 unknown loss 0.9486 | geometry 0.2821 | confidence 1.3329 | mean error 0.651 vox speed: 1269 ms/update | data wait 1 ms/update | 25.0 crops/s | 25.0 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1497 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 180, from the previous prediction 1317 confidence-labeled crops: terminal 20/1543 (1.3%) | recoverable 113/1543 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 352/1600 (22.0%), 2: 224/1600 (14.0%), 3: 256/1600 (16.0%), 4: 416/1600 (26.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 698 rows | mean depth 17 decisions | mean travel 141 vox | 7 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 53% 41% 26%>6% 20% 19 1232 92%/7%/0%/1% 11% 137/0/0/6 0175A_5mm_v1 352 45% 43% 25%>11% 14% 32 1477 83%/8%/3%/6% 9% 152/1/0/0 1447_5mm_v1 224 53% 44% 24%>3% 21% 5 444 92%/6%/0%/3% 2% 99/3/0/0 paris4_afv_central 256 50% 48% 24%>3% 21% 9 884 89%/7%/0%/4% 7% 117/3/27/5 PHerc0125_3mm_v2 416 46% 48% 25%>6% 19% 22 1029 85%/7%/2%/6% 16% 193/2/0/7 gradients: max 1e+02, clipped 1/50 updates Step 28,550 | last 50 updates / 1,600 crops | lr 1.73e-07 points (35/curve, tolerance 2 vox): 48,796 right / 2,254 wrong | 95.6% correct | 4,950 unknown loss 0.8429 | geometry 0.2377 | confidence 1.2103 | mean error 0.607 vox speed: 893 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1519 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 339, from the previous prediction 1180 confidence-labeled crops: terminal 17/1532 (1.1%) | recoverable 126/1532 (8.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 384/1600 (24.0%), 2: 256/1600 (16.0%), 3: 416/1600 (26.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 536 rows | mean depth 15 decisions | mean travel 125 vox | 7 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 53% 43% 29%>4% 25% 10 971 89%/8%/0%/2% 9% 103/0/0/35 0175A_5mm_v1 384 53% 43% 30%>4% 26% 17 1208 85%/9%/1%/5% 5% 95/0/66/70 1447_5mm_v1 256 50% 45% 28%>5% 23% 12 699 90%/7%/1%/3% 4% 115/3/0/0 paris4_afv_central 416 52% 44% 28%>5% 23% 24 1766 84%/8%/2%/6% 6% 124/3/0/57 PHerc0125_3mm_v2 224 51% 44% 19%>4% 14% 10 908 88%/5%/2%/5% 14% 99/3/0/0 gradients: max 55, clipped 0/50 updates Step 28,600 | last 50 updates / 1,600 crops | lr 1.61e-07 points (35/curve, tolerance 2 vox): 47,762 right / 3,064 wrong | 94.0% correct | 5,174 unknown loss 0.9725 | geometry 0.2942 | confidence 1.3565 | mean error 0.661 vox speed: 894 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1516 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 246, from the previous prediction 1270 confidence-labeled crops: terminal 17/1536 (1.1%) | recoverable 125/1536 (8.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 320/1600 (20.0%), 2: 224/1600 (14.0%), 3: 384/1600 (24.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 644 rows | mean depth 19 decisions | mean travel 155 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 54% 45% 23%>1% 22% 2 250 90%/8%/0%/2% 6% 129/0/0/15 0175A_5mm_v1 320 45% 48% 22%>8% 15% 27 1969 84%/8%/2%/6% 9% 80/3/0/73 1447_5mm_v1 224 47% 52% 25%>1% 23% 3 315 87%/10%/0%/3% 1% 116/0/0/0 paris4_afv_central 384 54% 43% 26%>2% 24% 8 571 87%/8%/1%/4% 6% 167/4/0/0 PHerc0125_3mm_v2 352 46% 43% 26%>11% 15% 35 1872 85%/6%/3%/6% 21% 152/2/0/0 gradients: max 46, clipped 0/50 updates Step 28,650 | last 50 updates / 1,600 crops | lr 1.50e-07 points (35/curve, tolerance 2 vox): 47,688 right / 3,278 wrong | 93.6% correct | 5,034 unknown loss 0.9296 | geometry 0.3056 | confidence 1.2480 | mean error 0.664 vox speed: 890 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1484 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 215, from the previous prediction 1269 confidence-labeled crops: terminal 22/1539 (1.4%) | recoverable 110/1539 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 320/1600 (20.0%), 2: 288/1600 (18.0%), 3: 352/1600 (22.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 663 rows | mean depth 18 decisions | mean travel 150 vox | 6 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 44% 51% 23%>5% 18% 12 795 92%/7%/0%/1% 9% 146/2/0/0 0175A_5mm_v1 320 46% 41% 26%>12% 13% 26 1659 86%/6%/2%/7% 11% 132/0/0/0 1447_5mm_v1 288 56% 44% 27%>0% 27% 0 150 91%/8%/0%/1% 2% 127/0/0/0 paris4_afv_central 352 52% 44% 24%>3% 21% 16 1640 89%/6%/1%/5% 7% 106/4/0/50 PHerc0125_3mm_v2 352 45% 43% 25%>12% 13% 39 1987 82%/8%/4%/6% 18% 152/3/0/0 gradients: max 39, clipped 0/50 updates Step 28,700 | last 50 updates / 1,600 crops | lr 1.39e-07 points (35/curve, tolerance 2 vox): 47,914 right / 2,984 wrong | 94.1% correct | 5,102 unknown loss 0.8991 | geometry 0.2929 | confidence 1.2124 | mean error 0.662 vox speed: 896 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.23 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1500 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 230, from the previous prediction 1270 confidence-labeled crops: terminal 20/1534 (1.3%) | recoverable 102/1534 (6.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 224/1600 (14.0%), 2: 288/1600 (18.0%), 3: 384/1600 (24.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 641 rows | mean depth 17 decisions | mean travel 138 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 51% 43% 27%>7% 20% 19 1142 90%/7%/0%/2% 11% 150/2/5/0 0175A_5mm_v1 224 50% 49% 24%>2% 22% 4 374 90%/4%/1%/6% 5% 109/0/0/0 1447_5mm_v1 288 51% 41% 26%>7% 19% 20 1193 88%/7%/1%/4% 5% 119/2/0/0 paris4_afv_central 384 47% 46% 27%>7% 20% 26 1432 88%/7%/2%/4% 7% 124/2/0/51 PHerc0125_3mm_v2 352 52% 41% 29%>7% 22% 23 1400 85%/7%/2%/6% 14% 139/3/38/6 gradients: max 43, clipped 0/50 updates Step 28,750 | last 50 updates / 1,600 crops | lr 1.29e-07 points (35/curve, tolerance 2 vox): 48,969 right / 2,595 wrong | 95.0% correct | 4,436 unknown loss 0.9190 | geometry 0.2757 | confidence 1.2866 | mean error 0.654 vox speed: 893 ms/update | data wait 1 ms/update | 34.8 crops/s | 34.8 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1527 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 298, from the previous prediction 1229 confidence-labeled crops: terminal 12/1540 (0.8%) | recoverable 107/1540 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 320/1600 (20.0%), 2: 256/1600 (16.0%), 3: 384/1600 (24.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 604 rows | mean depth 18 decisions | mean travel 154 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 49% 48% 23%>3% 21% 11 900 91%/6%/0%/3% 9% 116/2/5/69 0175A_5mm_v1 320 50% 45% 26%>5% 21% 14 771 90%/6%/1%/3% 6% 145/3/7/0 1447_5mm_v1 256 47% 45% 25%>8% 17% 18 1069 88%/7%/1%/4% 7% 101/0/20/15 paris4_afv_central 384 56% 41% 27%>3% 24% 10 565 86%/8%/1%/5% 6% 159/5/0/0 PHerc0125_3mm_v2 256 53% 41% 22%>6% 16% 14 1455 88%/6%/2%/4% 15% 83/1/0/23 gradients: max 58, clipped 0/50 updates Step 28,800 | last 50 updates / 1,600 crops | lr 1.18e-07 points (35/curve, tolerance 2 vox): 48,325 right / 2,881 wrong | 94.4% correct | 4,794 unknown loss 0.8847 | geometry 0.2773 | confidence 1.2148 | mean error 0.652 vox speed: 898 ms/update | data wait 1 ms/update | 35.1 crops/s | 35.1 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1533 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 240, from the previous prediction 1293 confidence-labeled crops: terminal 22/1543 (1.4%) | recoverable 113/1543 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 320/1600 (20.0%), 2: 256/1600 (16.0%), 3: 352/1600 (22.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 376 chunk references / 1 windows live chains: 655 rows | mean depth 16 decisions | mean travel 136 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 48% 48% 24%>4% 20% 12 997 88%/9%/0%/3% 9% 110/0/0/59 0175A_5mm_v1 320 53% 44% 27%>3% 24% 8 544 88%/8%/1%/3% 6% 141/1/0/0 1447_5mm_v1 256 54% 43% 30%>4% 27% 10 633 86%/10%/0%/4% 4% 109/3/0/0 paris4_afv_central 352 51% 45% 23%>4% 19% 13 913 90%/5%/3%/3% 10% 146/2/0/11 PHerc0125_3mm_v2 320 48% 47% 25%>6% 19% 17 1127 87%/5%/2%/6% 15% 149/1/0/0 gradients: max 40, clipped 0/50 updates Step 28,850 | last 50 updates / 1,600 crops | lr 1.09e-07 points (35/curve, tolerance 2 vox): 49,045 right / 2,581 wrong | 95.0% correct | 4,374 unknown loss 0.7692 | geometry 0.2546 | confidence 1.0291 | mean error 0.613 vox speed: 898 ms/update | data wait 1 ms/update | 35.1 crops/s | 35.1 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1497 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 223, from the previous prediction 1274 confidence-labeled crops: terminal 23/1541 (1.5%) | recoverable 115/1541 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 288/1600 (18.0%), 2: 384/1600 (24.0%), 3: 512/1600 (32.0%), 4: 160/1600 (10.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 627 rows | mean depth 16 decisions | mean travel 134 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 48% 44% 27%>9% 19% 18 1504 88%/7%/1%/4% 11% 102/0/0/10 0175A_5mm_v1 288 51% 43% 24%>6% 18% 18 1154 87%/8%/1%/4% 6% 101/1/61/23 1447_5mm_v1 384 48% 44% 26%>8% 18% 29 1042 87%/7%/2%/4% 4% 160/3/42/9 paris4_afv_central 512 54% 45% 25%>1% 24% 5 393 89%/7%/0%/4% 5% 194/2/0/36 PHerc0125_3mm_v2 160 39% 44% 26%>18% 9% 25 2634 85%/7%/3%/5% 19% 70/3/0/0 gradients: max 64, clipped 1/50 updates Step 28,900 | last 50 updates / 1,600 crops | lr 9.96e-08 points (35/curve, tolerance 2 vox): 48,110 right / 2,511 wrong | 95.0% correct | 5,379 unknown loss 0.7536 | geometry 0.2279 | confidence 1.0515 | mean error 0.586 vox speed: 891 ms/update | data wait 1 ms/update | 35.0 crops/s | 35.0 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1500 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 312, from the previous prediction 1188 confidence-labeled crops: terminal 24/1527 (1.6%) | recoverable 114/1527 (7.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 512/1600 (32.0%), 2: 416/1600 (26.0%), 3: 224/1600 (14.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 118 chunks / 236.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 557 rows | mean depth 18 decisions | mean travel 151 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 48% 45% 25%>7% 18% 15 1164 92%/5%/0%/2% 10% 101/1/0/0 0175A_5mm_v1 512 54% 40% 26%>7% 20% 27 1641 85%/7%/2%/6% 7% 139/3/0/65 1447_5mm_v1 416 50% 43% 25%>7% 18% 25 1717 86%/7%/2%/4% 7% 104/0/0/74 paris4_afv_central 224 53% 45% 27%>2% 25% 5 598 85%/10%/0%/5% 3% 101/0/0/0 PHerc0125_3mm_v2 224 44% 50% 23%>6% 17% 13 1148 87%/6%/2%/5% 13% 112/2/16/0 gradients: max 46, clipped 0/50 updates Step 28,950 | last 50 updates / 1,600 crops | lr 9.08e-08 points (35/curve, tolerance 2 vox): 48,336 right / 2,522 wrong | 95.0% correct | 5,142 unknown loss 0.7897 | geometry 0.2394 | confidence 1.1006 | mean error 0.580 vox speed: 889 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1477 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 222, from the previous prediction 1255 confidence-labeled crops: terminal 24/1533 (1.6%) | recoverable 129/1533 (8.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 448/1600 (28.0%), 2: 416/1600 (26.0%), 3: 256/1600 (16.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 120 chunks / 240.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 636 chunk references / 1 windows live chains: 637 rows | mean depth 20 decisions | mean travel 163 vox | 5 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 58% 42% 26%>0% 25% 1 255 89%/9%/0%/2% 6% 93/0/0/0 0175A_5mm_v1 448 48% 48% 25%>4% 21% 22 982 87%/8%/1%/4% 5% 175/1/0/40 1447_5mm_v1 416 43% 48% 26%>10% 16% 37 1647 87%/8%/1%/5% 9% 147/1/0/51 paris4_afv_central 256 52% 43% 24%>5% 19% 12 630 85%/8%/2%/5% 8% 111/3/0/0 PHerc0125_3mm_v2 256 44% 43% 27%>13% 14% 31 1991 81%/7%/4%/8% 18% 111/2/0/0 gradients: max 57, clipped 0/50 updates Step 29,000 | last 50 updates / 1,600 crops | lr 8.23e-08 points (35/curve, tolerance 2 vox): 46,538 right / 3,370 wrong | 93.2% correct | 6,092 unknown loss 0.9134 | geometry 0.2868 | confidence 1.2533 | mean error 0.667 vox speed: 888 ms/update | data wait 1 ms/update | 35.5 crops/s | 35.5 decisions/s | 1.22 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1487 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 356, from the previous prediction 1131 confidence-labeled crops: terminal 26/1508 (1.7%) | recoverable 110/1508 (7.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 448/1600 (28.0%), 2: 256/1600 (16.0%), 3: 256/1600 (16.0%), 4: 448/1600 (28.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 120 chunks / 240.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 546 rows | mean depth 17 decisions | mean travel 149 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 50% 42% 29%>8% 21% 14 1484 90%/7%/0%/3% 11% 80/2/1/0 0175A_5mm_v1 448 47% 44% 26%>9% 17% 33 1893 81%/9%/2%/8% 10% 127/3/0/68 1447_5mm_v1 256 43% 49% 21%>8% 13% 19 1314 87%/6%/3%/4% 6% 126/0/0/0 paris4_afv_central 256 55% 42% 23%>2% 21% 6 429 88%/5%/1%/6% 6% 108/4/0/0 PHerc0125_3mm_v2 448 49% 46% 21%>5% 16% 25 1739 86%/6%/1%/7% 16% 105/4/19/102 gradients: max 55, clipped 0/50 updates Step 29,000 | dagger_launched: true Step 29,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 256, "stopped": 0, "mean_example_error": 0.5047818232199242, "geometry_examples": 29, "examples": 32, "inference_and_measurement_seconds": 1.833443146198988, "total_diagnostic_seconds": 3.5697269067168236} Step 29,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 6/8 (75.0%) | false stops 1 1.5-3: recovered 6/8 (75.0%) | false stops 1 3-4.5: recovered 6/8 (75.0%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 0 entire proposed 8-point prefix correct (distance): 22/28 (78.6%) gate @ 0.50: false stops 0/23 (0.0%) | accepted wrong 5/27 (18.5%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.607 1.567 following 16 464 1.155 1.172 recoverable 12 348 2.210 2.093 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 1.155 1.172 332 32 812 1.607 1.567 improved/worsened/compared states: 3/2/28 Step 29,000 | batch_diagnostics_seconds: 3.580671288073063 | recovery_seconds: 14.535640899091959 Step 29,050 | last 50 updates / 1,600 crops | lr 7.43e-08 points (35/curve, tolerance 2 vox): 48,409 right / 2,637 wrong | 94.8% correct | 4,954 unknown loss 0.8187 | geometry 0.2463 | confidence 1.1448 | mean error 0.595 vox speed: 1320 ms/update | data wait 1 ms/update | 18.6 crops/s | 18.6 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1491 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 226, from the previous prediction 1265 confidence-labeled crops: terminal 21/1542 (1.4%) | recoverable 100/1542 (6.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 288/1600 (18.0%), 2: 352/1600 (22.0%), 3: 320/1600 (20.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 120 chunks / 240.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 658 rows | mean depth 19 decisions | mean travel 159 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 51% 44% 26%>6% 20% 19 1135 89%/7%/0%/3% 9% 183/0/124/12 0175A_5mm_v1 288 48% 43% 24%>9% 15% 24 1468 86%/8%/1%/5% 8% 123/2/0/2 1447_5mm_v1 352 50% 40% 25%>10% 15% 26 1343 89%/6%/2%/4% 9% 140/2/0/0 paris4_afv_central 320 45% 48% 21%>7% 14% 20 863 89%/5%/2%/4% 8% 155/3/0/0 PHerc0125_3mm_v2 192 52% 48% 23%>0% 23% 0 145 90%/5%/1%/5% 9% 57/0/0/35 gradients: max 42, clipped 0/50 updates Step 29,100 | last 50 updates / 1,600 crops | lr 6.67e-08 points (35/curve, tolerance 2 vox): 48,700 right / 2,777 wrong | 94.6% correct | 4,523 unknown loss 0.8517 | geometry 0.2680 | confidence 1.1673 | mean error 0.645 vox speed: 1391 ms/update | data wait 1 ms/update | 22.8 crops/s | 22.8 decisions/s | 1.24 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1540 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 320, from the previous prediction 1220 confidence-labeled crops: terminal 17/1540 (1.1%) | recoverable 104/1540 (6.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 256/1600 (16.0%), 2: 288/1600 (18.0%), 3: 288/1600 (18.0%), 4: 352/1600 (22.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 120 chunks / 240.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 584 rows | mean depth 18 decisions | mean travel 152 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 50% 46% 24%>4% 20% 14 1352 90%/7%/0%/3% 11% 99/0/14/92 0175A_5mm_v1 256 47% 45% 23%>8% 15% 17 1175 87%/4%/2%/6% 8% 116/5/0/0 1447_5mm_v1 288 54% 40% 28%>6% 23% 16 952 88%/7%/2%/3% 3% 114/1/55/2 paris4_afv_central 288 50% 50% 23%>0% 23% 0 140 90%/6%/1%/3% 5% 100/2/44/43 PHerc0125_3mm_v2 352 54% 44% 26%>2% 24% 6 483 86%/8%/1%/5% 13% 155/1/0/0 gradients: max 78, clipped 1/50 updates Step 29,150 | last 50 updates / 1,600 crops | lr 5.95e-08 points (35/curve, tolerance 2 vox): 48,961 right / 2,599 wrong | 95.0% correct | 4,440 unknown loss 0.8716 | geometry 0.2942 | confidence 1.1547 | mean error 0.666 vox speed: 1397 ms/update | data wait 1 ms/update | 22.6 crops/s | 22.6 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1539 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 316, from the previous prediction 1223 confidence-labeled crops: terminal 12/1542 (0.8%) | recoverable 117/1542 (7.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 544/1600 (34.0%), 1: 256/1600 (16.0%), 2: 352/1600 (22.0%), 3: 256/1600 (16.0%), 4: 192/1600 (12.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 120 chunks / 240.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 568 rows | mean depth 18 decisions | mean travel 148 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 544 52% 45% 25%>3% 22% 20 1606 90%/8%/0%/2% 11% 124/0/0/119 0175A_5mm_v1 256 48% 46% 25%>6% 19% 15 992 88%/7%/2%/3% 7% 118/4/0/0 1447_5mm_v1 352 55% 42% 26%>2% 23% 8 448 86%/7%/1%/6% 3% 149/0/0/0 paris4_afv_central 256 55% 45% 26%>0% 26% 0 136 89%/7%/1%/4% 5% 88/1/0/27 PHerc0125_3mm_v2 192 49% 46% 24%>5% 20% 8 523 86%/7%/2%/5% 13% 89/1/0/0 gradients: max 55, clipped 0/50 updates Step 29,200 | last 50 updates / 1,600 crops | lr 5.27e-08 points (35/curve, tolerance 2 vox): 47,631 right / 2,729 wrong | 94.6% correct | 5,640 unknown loss 0.8443 | geometry 0.2610 | confidence 1.1666 | mean error 0.611 vox speed: 1404 ms/update | data wait 1 ms/update | 22.6 crops/s | 22.6 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1492 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 208, from the previous prediction 1284 confidence-labeled crops: terminal 24/1527 (1.6%) | recoverable 116/1527 (7.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 352/1600 (22.0%), 2: 320/1600 (20.0%), 3: 288/1600 (18.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 120 chunks / 240.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 659 rows | mean depth 20 decisions | mean travel 167 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 57% 41% 23%>2% 21% 9 640 88%/8%/0%/3% 8% 149/0/0/21 0175A_5mm_v1 352 51% 42% 26%>7% 19% 24 1178 88%/5%/3%/5% 8% 148/0/0/0 1447_5mm_v1 320 44% 45% 28%>11% 17% 32 1767 86%/8%/2%/4% 8% 134/1/23/10 paris4_afv_central 288 45% 45% 26%>9% 17% 24 1261 85%/6%/2%/7% 7% 131/3/0/0 PHerc0125_3mm_v2 224 53% 43% 29%>4% 25% 7 843 84%/9%/2%/5% 13% 97/3/23/0 gradients: max 44, clipped 0/50 updates Step 29,250 | last 50 updates / 1,600 crops | lr 4.64e-08 points (35/curve, tolerance 2 vox): 48,930 right / 2,857 wrong | 94.5% correct | 4,213 unknown loss 0.9483 | geometry 0.3144 | confidence 1.2678 | mean error 0.657 vox speed: 1396 ms/update | data wait 1 ms/update | 22.7 crops/s | 22.7 decisions/s | 1.16 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1509 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 208, from the previous prediction 1301 confidence-labeled crops: terminal 18/1551 (1.2%) | recoverable 121/1551 (7.8%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 448/1600 (28.0%), 2: 160/1600 (10.0%), 3: 384/1600 (24.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 123 chunks / 246.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 654 rows | mean depth 17 decisions | mean travel 140 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 53% 40% 26%>8% 18% 18 1460 89%/9%/1%/1% 11% 114/0/0/0 0175A_5mm_v1 448 46% 45% 25%>9% 16% 36 1565 89%/6%/2%/4% 8% 187/1/69/15 1447_5mm_v1 160 46% 42% 27%>11% 16% 18 1417 82%/11%/3%/4% 10% 68/0/0/0 paris4_afv_central 384 57% 43% 29%>0% 29% 0 144 87%/9%/1%/4% 4% 165/3/0/0 PHerc0125_3mm_v2 320 52% 47% 22%>1% 21% 4 439 91%/6%/1%/3% 13% 120/2/38/29 gradients: max 50, clipped 0/50 updates Step 29,300 | last 50 updates / 1,600 crops | lr 4.04e-08 points (35/curve, tolerance 2 vox): 48,544 right / 2,420 wrong | 95.3% correct | 5,036 unknown loss 0.8622 | geometry 0.2727 | confidence 1.1790 | mean error 0.631 vox speed: 1402 ms/update | data wait 1 ms/update | 22.6 crops/s | 22.6 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1521 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 321, from the previous prediction 1200 confidence-labeled crops: terminal 11/1525 (0.7%) | recoverable 100/1525 (6.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 384/1600 (24.0%), 1: 288/1600 (18.0%), 2: 192/1600 (12.0%), 3: 512/1600 (32.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 123 chunks / 246.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 584 rows | mean depth 20 decisions | mean travel 163 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 384 46% 46% 26%>8% 18% 25 1423 90%/8%/0%/2% 14% 176/1/34/2 0175A_5mm_v1 288 51% 41% 26%>7% 19% 20 1255 87%/6%/3%/5% 7% 113/4/54/6 1447_5mm_v1 192 48% 47% 24%>5% 19% 9 712 89%/6%/1%/4% 1% 90/1/0/0 paris4_afv_central 512 51% 47% 23%>3% 21% 15 681 86%/5%/0%/8% 6% 137/1/0/103 PHerc0125_3mm_v2 224 55% 45% 21%>0% 21% 0 137 89%/6%/0%/5% 10% 68/3/0/33 gradients: max 38, clipped 0/50 updates Step 29,350 | last 50 updates / 1,600 crops | lr 3.48e-08 points (35/curve, tolerance 2 vox): 47,270 right / 2,839 wrong | 94.3% correct | 5,891 unknown loss 0.9340 | geometry 0.2742 | confidence 1.3195 | mean error 0.645 vox speed: 1410 ms/update | data wait 1 ms/update | 22.5 crops/s | 22.5 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1480 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 309, from the previous prediction 1171 confidence-labeled crops: terminal 18/1525 (1.2%) | recoverable 108/1525 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 160/1600 (10.0%), 1: 288/1600 (18.0%), 2: 288/1600 (18.0%), 3: 352/1600 (22.0%), 4: 512/1600 (32.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 123 chunks / 246.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 581 rows | mean depth 19 decisions | mean travel 159 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 160 54% 39% 28%>6% 21% 9 1290 90%/10%/0%/0% 7% 61/0/32/2 0175A_5mm_v1 288 40% 41% 28%>19% 8% 41 2701 82%/6%/4%/8% 13% 101/1/0/16 1447_5mm_v1 288 42% 49% 22%>9% 13% 23 1123 88%/7%/1%/5% 6% 142/3/0/0 paris4_afv_central 352 50% 48% 23%>2% 21% 10 962 91%/5%/0%/3% 5% 124/5/0/44 PHerc0125_3mm_v2 512 53% 45% 23%>2% 21% 17 855 86%/7%/1%/6% 13% 153/1/0/79 gradients: max 56, clipped 0/50 updates Step 29,380 | dagger_states: 5388 | dagger_source_step: 29000 | dagger_caches: 4 | dagger_cache: "/ephemeral/sean/fiber_sequence/runs/mixed_ct_afv_unified_scoring_v2_resenc/dagger/paris4_afv_central/decisions_029000.npz" | dagger_fibers: 384 | dagger_supply: {"states": 5388, "traces": 384, "fibers": 384, "rows": {"terminal": 212, "premature_stop": 183, "recoverable": 15, "pre_excursion": 197, "ordinary": 4776}, "events": {"terminal": 31, "premature_stop": 183, "recoverable": 6, "pre_excursion": 35, "ordinary": 371}, "supervision": {"following": 5154, "recoverable": 16, "terminal": 212, "unknown": 6}, "departed_traces": 34, "terminal_traces": 31, "stop_reasons": {"after_failure": 24, "confidence": 192, "unannotated": 168}, "traced_voxels": 78748.82335313254, "seconds": 544.8382945284247, "voxels_per_second": 144.5361387846135} | dagger_coverage: {"requested": 384, "seeds": 384, "skipped": [], "epoch": 0} | dagger_operating_policy: {"confidence": 0.7, "n_commit": 8, "max_recovery_distance": 6.0, "refinement_steps": 3, "gate": "full"} | dagger_busy_skips: 0 | dagger_launch_interval: 5000 | dagger_publication_age: 380 | dataset: "paris4_afv_central" Step 29,400 | last 50 updates / 1,600 crops | lr 2.97e-08 points (35/curve, tolerance 2 vox): 48,194 right / 2,402 wrong | 95.3% correct | 5,404 unknown loss 0.8402 | geometry 0.2442 | confidence 1.1918 | mean error 0.609 vox speed: 1172 ms/update | data wait 1 ms/update | 27.0 crops/s | 27.0 decisions/s | 1.19 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1494 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 200, from the previous prediction 1294 confidence-labeled crops: terminal 16/1523 (1.1%) | recoverable 117/1523 (7.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 224/1600 (14.0%), 2: 256/1600 (16.0%), 3: 512/1600 (32.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 123 chunks / 246.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 661 rows | mean depth 15 decisions | mean travel 125 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 53% 42% 28%>5% 23% 15 1021 87%/8%/0%/5% 7% 134/1/0/0 0175A_5mm_v1 224 49% 41% 27%>10% 17% 19 1593 85%/6%/3%/6% 7% 92/4/0/0 1447_5mm_v1 256 39% 46% 28%>16% 12% 35 1750 85%/7%/2%/6% 8% 117/2/0/0 paris4_afv_central 512 50% 48% 22%>3% 20% 18 610 87%/9%/1%/4% 7% 191/6/0/53 PHerc0125_3mm_v2 288 53% 44% 22%>3% 19% 8 388 89%/5%/1%/6% 10% 127/0/0/0 gradients: max 54, clipped 0/50 updates Step 29,450 | last 50 updates / 1,600 crops | lr 2.50e-08 points (35/curve, tolerance 2 vox): 48,940 right / 2,375 wrong | 95.4% correct | 4,685 unknown loss 0.7987 | geometry 0.2459 | confidence 1.1056 | mean error 0.614 vox speed: 893 ms/update | data wait 1 ms/update | 35.3 crops/s | 35.3 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1513 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 244, from the previous prediction 1269 confidence-labeled crops: terminal 20/1547 (1.3%) | recoverable 107/1547 (6.9%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 352/1600 (22.0%), 2: 224/1600 (14.0%), 3: 416/1600 (26.0%), 4: 256/1600 (16.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 123 chunks / 246.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 648 rows | mean depth 20 decisions | mean travel 166 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 50% 46% 26%>4% 22% 12 844 91%/6%/0%/3% 7% 154/0/43/8 0175A_5mm_v1 352 46% 47% 23%>7% 16% 19 1036 87%/8%/1%/3% 8% 167/4/0/0 1447_5mm_v1 224 47% 45% 20%>8% 12% 14 1109 91%/4%/2%/2% 5% 100/3/0/0 paris4_afv_central 416 56% 42% 22%>2% 20% 7 378 89%/7%/1%/3% 5% 146/5/0/30 PHerc0125_3mm_v2 256 50% 43% 25%>7% 18% 25 1677 85%/6%/3%/6% 13% 81/4/0/30 gradients: max 54, clipped 0/50 updates Step 29,500 | last 50 updates / 1,600 crops | lr 2.06e-08 points (35/curve, tolerance 2 vox): 48,982 right / 2,143 wrong | 95.8% correct | 4,875 unknown loss 0.6839 | geometry 0.2258 | confidence 0.9163 | mean error 0.574 vox speed: 895 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.21 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1524 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 282, from the previous prediction 1242 confidence-labeled crops: terminal 19/1532 (1.2%) | recoverable 131/1532 (8.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 192/1600 (12.0%), 1: 320/1600 (20.0%), 2: 512/1600 (32.0%), 3: 288/1600 (18.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 123 chunks / 246.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 607 rows | mean depth 20 decisions | mean travel 162 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 192 51% 40% 26%>9% 17% 15 1720 88%/8%/0%/4% 9% 73/1/32/4 0175A_5mm_v1 320 52% 46% 21%>2% 19% 4 461 90%/7%/0%/3% 3% 131/2/31/15 1447_5mm_v1 512 46% 48% 25%>7% 18% 35 1506 85%/9%/2%/5% 7% 156/2/20/88 paris4_afv_central 288 57% 43% 27%>0% 27% 0 146 86%/9%/0%/4% 4% 125/0/0/0 PHerc0125_3mm_v2 288 54% 42% 23%>3% 20% 10 474 84%/8%/2%/6% 13% 122/2/0/0 gradients: max 47, clipped 0/50 updates Step 29,550 | last 50 updates / 1,600 crops | lr 1.67e-08 points (35/curve, tolerance 2 vox): 47,603 right / 2,843 wrong | 94.4% correct | 5,554 unknown loss 0.8552 | geometry 0.2411 | confidence 1.2283 | mean error 0.612 vox speed: 892 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1482 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 239, from the previous prediction 1243 confidence-labeled crops: terminal 30/1529 (2.0%) | recoverable 103/1529 (6.7%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 288/1600 (18.0%), 1: 384/1600 (24.0%), 2: 256/1600 (16.0%), 3: 448/1600 (28.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 123 chunks / 246.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 634 rows | mean depth 17 decisions | mean travel 140 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 288 49% 49% 26%>3% 23% 6 626 92%/6%/0%/1% 8% 132/1/9/8 0175A_5mm_v1 384 47% 45% 23%>8% 15% 29 2019 87%/7%/2%/4% 9% 114/2/0/57 1447_5mm_v1 256 41% 38% 29%>21% 8% 35 2380 81%/5%/5%/9% 10% 98/1/0/0 paris4_afv_central 448 51% 45% 25%>4% 21% 18 582 89%/6%/1%/3% 7% 202/3/0/0 PHerc0125_3mm_v2 224 61% 39% 33%>0% 33% 0 146 84%/7%/0%/8% 12% 88/1/0/0 gradients: max 47, clipped 0/50 updates Step 29,600 | last 50 updates / 1,600 crops | lr 1.32e-08 points (35/curve, tolerance 2 vox): 48,970 right / 2,448 wrong | 95.2% correct | 4,582 unknown loss 0.8427 | geometry 0.2494 | confidence 1.1867 | mean error 0.628 vox speed: 892 ms/update | data wait 1 ms/update | 35.4 crops/s | 35.4 decisions/s | 1.15 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1537 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 204, from the previous prediction 1333 confidence-labeled crops: terminal 17/1542 (1.1%) | recoverable 99/1542 (6.4%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 416/1600 (26.0%), 1: 288/1600 (18.0%), 2: 352/1600 (22.0%), 3: 224/1600 (14.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 123 chunks / 246.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 662 rows | mean depth 20 decisions | mean travel 161 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 416 54% 44% 26%>3% 23% 12 701 90%/8%/0%/3% 7% 167/0/0/15 0175A_5mm_v1 288 52% 44% 24%>4% 20% 11 789 89%/6%/2%/3% 6% 126/1/0/0 1447_5mm_v1 352 50% 43% 29%>6% 23% 21 1018 88%/6%/2%/4% 4% 139/1/0/14 paris4_afv_central 224 52% 44% 31%>4% 26% 10 652 89%/7%/1%/3% 8% 98/1/0/0 PHerc0125_3mm_v2 320 56% 42% 20%>2% 18% 7 368 89%/4%/1%/6% 11% 132/0/62/3 gradients: max 83, clipped 2/50 updates Step 29,650 | last 50 updates / 1,600 crops | lr 1.01e-08 points (35/curve, tolerance 2 vox): 48,770 right / 2,191 wrong | 95.7% correct | 5,039 unknown loss 0.7613 | geometry 0.2304 | confidence 1.0618 | mean error 0.581 vox speed: 895 ms/update | data wait 1 ms/update | 35.1 crops/s | 35.1 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1497 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 225, from the previous prediction 1272 confidence-labeled crops: terminal 17/1534 (1.1%) | recoverable 109/1534 (7.1%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 320/1600 (20.0%), 1: 288/1600 (18.0%), 2: 448/1600 (28.0%), 3: 320/1600 (20.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 123 chunks / 246.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 627 rows | mean depth 20 decisions | mean travel 167 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 320 52% 43% 30%>5% 25% 17 1610 91%/8%/0%/1% 9% 86/0/0/51 0175A_5mm_v1 288 53% 46% 24%>1% 23% 3 311 91%/6%/0%/3% 2% 132/2/0/0 1447_5mm_v1 448 48% 44% 27%>9% 18% 37 1220 87%/7%/2%/4% 6% 195/3/0/0 paris4_afv_central 320 50% 40% 30%>10% 21% 26 1215 85%/7%/2%/6% 7% 129/4/0/0 PHerc0125_3mm_v2 224 55% 40% 27%>5% 21% 12 1312 86%/5%/1%/7% 11% 85/0/0/4 gradients: max 57, clipped 0/50 updates Step 29,700 | last 50 updates / 1,600 crops | lr 7.45e-09 points (35/curve, tolerance 2 vox): 48,712 right / 2,717 wrong | 94.7% correct | 4,571 unknown loss 0.8671 | geometry 0.2798 | confidence 1.1747 | mean error 0.652 vox speed: 895 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1514 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 190, from the previous prediction 1324 confidence-labeled crops: terminal 16/1542 (1.0%) | recoverable 131/1542 (8.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 352/1600 (22.0%), 1: 416/1600 (26.0%), 2: 224/1600 (14.0%), 3: 320/1600 (20.0%), 4: 288/1600 (18.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 123 chunks / 246.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 701 rows | mean depth 16 decisions | mean travel 138 vox | 4 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 352 50% 44% 28%>6% 22% 17 1164 89%/9%/0%/2% 10% 146/3/0/9 0175A_5mm_v1 416 53% 41% 28%>6% 21% 21 1139 87%/8%/1%/4% 7% 170/5/0/0 1447_5mm_v1 224 54% 46% 26%>0% 26% 0 137 89%/9%/0%/1% 1% 102/0/10/0 paris4_afv_central 320 48% 48% 27%>5% 22% 11 645 84%/9%/1%/6% 7% 153/1/0/0 PHerc0125_3mm_v2 288 46% 45% 27%>9% 19% 24 878 86%/6%/3%/6% 15% 130/0/0/0 gradients: max 52, clipped 0/50 updates Step 29,750 | last 50 updates / 1,600 crops | lr 5.18e-09 points (35/curve, tolerance 2 vox): 48,266 right / 2,295 wrong | 95.5% correct | 5,439 unknown loss 0.8398 | geometry 0.2397 | confidence 1.2001 | mean error 0.592 vox speed: 897 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1527 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 297, from the previous prediction 1230 confidence-labeled crops: terminal 22/1528 (1.4%) | recoverable 99/1528 (6.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 416/1600 (26.0%), 2: 192/1600 (12.0%), 3: 352/1600 (22.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 123 chunks / 246.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 793 chunk references / 1 windows live chains: 586 rows | mean depth 18 decisions | mean travel 145 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 60% 40% 28%>0% 28% 0 141 89%/9%/0%/2% 7% 103/1/0/0 0175A_5mm_v1 416 48% 47% 23%>5% 19% 24 1287 88%/6%/1%/5% 6% 140/0/0/56 1447_5mm_v1 192 48% 40% 28%>12% 15% 21 1738 81%/8%/4%/6% 9% 76/2/0/0 paris4_afv_central 352 48% 45% 25%>7% 18% 20 970 87%/6%/2%/6% 6% 140/4/123/18 PHerc0125_3mm_v2 384 58% 42% 25%>0% 25% 0 135 91%/4%/1%/4% 11% 127/4/0/33 gradients: max 42, clipped 0/50 updates Step 29,800 | last 50 updates / 1,600 crops | lr 3.32e-09 points (35/curve, tolerance 2 vox): 49,085 right / 2,333 wrong | 95.5% correct | 4,582 unknown loss 0.8835 | geometry 0.2619 | confidence 1.2430 | mean error 0.614 vox speed: 896 ms/update | data wait 1 ms/update | 35.2 crops/s | 35.2 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1522 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 298, from the previous prediction 1224 confidence-labeled crops: terminal 14/1544 (0.9%) | recoverable 118/1544 (7.6%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 288/1600 (18.0%), 2: 352/1600 (22.0%), 3: 384/1600 (24.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 123 chunks / 246.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 602 rows | mean depth 23 decisions | mean travel 189 vox | 2 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 51% 43% 29%>6% 23% 12 1252 87%/11%/0%/2% 12% 110/2/0/0 0175A_5mm_v1 288 49% 47% 25%>4% 22% 10 842 89%/9%/0%/2% 4% 118/6/0/18 1447_5mm_v1 352 41% 51% 20%>8% 11% 26 1075 89%/7%/2%/2% 7% 178/3/0/0 paris4_afv_central 384 57% 40% 26%>3% 23% 12 632 86%/7%/1%/6% 6% 138/2/5/14 PHerc0125_3mm_v2 320 48% 49% 22%>2% 19% 8 692 90%/5%/0%/5% 12% 58/0/0/99 gradients: max 56, clipped 0/50 updates Step 29,850 | last 50 updates / 1,600 crops | lr 1.88e-09 points (35/curve, tolerance 2 vox): 48,581 right / 2,630 wrong | 94.9% correct | 4,789 unknown loss 0.8247 | geometry 0.2531 | confidence 1.1433 | mean error 0.628 vox speed: 897 ms/update | data wait 1 ms/update | 35.0 crops/s | 35.0 decisions/s | 1.18 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1513 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 309, from the previous prediction 1204 confidence-labeled crops: terminal 23/1538 (1.5%) | recoverable 100/1538 (6.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 256/1600 (16.0%), 1: 224/1600 (14.0%), 2: 256/1600 (16.0%), 3: 640/1600 (40.0%), 4: 224/1600 (14.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 123 chunks / 246.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 555 rows | mean depth 17 decisions | mean travel 144 vox | 3 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 256 52% 45% 26%>2% 24% 5 590 91%/7%/0%/2% 9% 116/0/0/0 0175A_5mm_v1 224 52% 40% 24%>8% 16% 15 1371 88%/6%/4%/2% 9% 90/3/0/0 1447_5mm_v1 256 50% 43% 28%>7% 20% 19 1138 88%/9%/1%/3% 5% 109/0/0/0 paris4_afv_central 640 47% 47% 24%>6% 18% 39 1466 88%/5%/1%/5% 6% 147/4/0/154 PHerc0125_3mm_v2 224 58% 42% 25%>0% 25% 0 147 88%/7%/0%/4% 11% 93/0/0/0 gradients: max 69, clipped 1/50 updates Step 29,900 | last 50 updates / 1,600 crops | lr 8.39e-10 points (35/curve, tolerance 2 vox): 48,028 right / 2,910 wrong | 94.3% correct | 5,062 unknown loss 0.8949 | geometry 0.2416 | confidence 1.3065 | mean error 0.627 vox speed: 898 ms/update | data wait 1 ms/update | 34.9 crops/s | 34.9 decisions/s | 1.20 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1534 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 266, from the previous prediction 1268 confidence-labeled crops: terminal 28/1534 (1.8%) | recoverable 99/1534 (6.5%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 288/1600 (18.0%), 2: 288/1600 (18.0%), 3: 256/1600 (16.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 123 chunks / 246.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 636 rows | mean depth 20 decisions | mean travel 166 vox | 9 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 53% 46% 27%>1% 26% 3 331 88%/9%/0%/2% 7% 137/0/63/71 0175A_5mm_v1 288 43% 48% 24%>9% 15% 22 1393 85%/7%/3%/5% 9% 137/0/0/0 1447_5mm_v1 288 51% 44% 20%>5% 16% 12 901 90%/5%/1%/5% 4% 102/4/20/26 paris4_afv_central 256 52% 44% 23%>4% 20% 9 574 89%/4%/2%/4% 8% 113/2/0/0 PHerc0125_3mm_v2 320 49% 46% 26%>5% 21% 15 590 87%/4%/3%/6% 12% 147/1/1/0 gradients: max 45, clipped 0/50 updates Step 29,950 | last 50 updates / 1,600 crops | lr 2.14e-10 points (35/curve, tolerance 2 vox): 48,498 right / 2,845 wrong | 94.5% correct | 4,657 unknown loss 0.8796 | geometry 0.2708 | confidence 1.2176 | mean error 0.639 vox speed: 896 ms/update | data wait 1 ms/update | 35.0 crops/s | 35.0 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1496 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 259, from the previous prediction 1237 confidence-labeled crops: terminal 20/1542 (1.3%) | recoverable 128/1542 (8.3%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 448/1600 (28.0%), 1: 224/1600 (14.0%), 2: 352/1600 (22.0%), 3: 256/1600 (16.0%), 4: 320/1600 (20.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 126 chunks / 252.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 597 rows | mean depth 20 decisions | mean travel 166 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 448 52% 45% 26%>3% 23% 11 1272 89%/9%/0%/2% 7% 101/0/0/100 0175A_5mm_v1 224 55% 38% 33%>7% 26% 16 1405 88%/7%/2%/3% 5% 85/2/0/0 1447_5mm_v1 352 45% 49% 22%>6% 16% 17 948 86%/9%/2%/3% 7% 157/3/0/16 paris4_afv_central 256 41% 46% 25%>13% 12% 28 1408 84%/5%/2%/9% 10% 118/3/0/0 PHerc0125_3mm_v2 320 52% 42% 29%>6% 23% 17 731 86%/9%/1%/4% 15% 136/2/4/0 gradients: max 40, clipped 0/50 updates Step 30,000 | last 50 updates / 1,600 crops | lr 8.22e-14 points (35/curve, tolerance 2 vox): 46,690 right / 3,399 wrong | 93.2% correct | 5,911 unknown loss 0.8906 | geometry 0.2689 | confidence 1.2433 | mean error 0.629 vox speed: 896 ms/update | data wait 1 ms/update | 35.1 crops/s | 35.1 decisions/s | 1.17 attempts/decision | peak allocated VRAM 41.43 GiB (session) supervision: 1600 decisions / 1600 observations CT frames: current 0/1434 fallbacks (0 transported, 0 deterministic); mean gap 0.000; heading from the frame model 249, from the previous prediction 1185 confidence-labeled crops: terminal 35/1521 (2.3%) | recoverable 110/1521 (7.2%) | target connections crossing a neighbor 0 datasets (IDs from dataset_configuration): 0: 224/1600 (14.0%), 1: 416/1600 (26.0%), 2: 288/1600 (18.0%), 3: 288/1600 (18.0%), 4: 384/1600 (24.0%) remote CT prefetch (cumulative): alive=True | active 0 | completed 126 chunks / 252.0 MiB | promoted 13 | preempted 0 | deferred 0 | errors 0 | lookahead 0 chunk references / 0 windows live chains: 641 rows | mean depth 17 decisions | mean travel 145 vox | 1 terminal sampling (dagger = replayed rollout states; fallback = nothing to replay -> fresh; age = steps since the model that made the data) source rows fresh live dagger req>got fallback replay ep age labels follow/recov/term/unk hazard tgts live adv/cens/stale/empty paris4 224 47% 46% 29%>7% 22% 12 1280 87%/9%/0%/4% 9% 101/0/0/3 0175A_5mm_v1 416 43% 47% 26%>10% 16% 35 1528 87%/5%/2%/6% 8% 197/4/0/0 1447_5mm_v1 288 45% 42% 28%>12% 15% 32 2129 83%/9%/3%/5% 9% 102/2/0/19 paris4_afv_central 288 46% 47% 23%>7% 16% 20 958 89%/6%/0%/5% 6% 134/2/0/0 PHerc0125_3mm_v2 384 43% 45% 27%>11% 16% 35 1591 84%/7%/4%/6% 17% 107/2/0/66 gradients: max 50, clipped 0/50 updates Step 30,000 | split: "current_training_microbatch" | diagnostic_images: {"committed_points": 248, "stopped": 1, "mean_example_error": 0.8515957913228444, "geometry_examples": 28, "examples": 32, "inference_and_measurement_seconds": 1.904625866562128, "total_diagnostic_seconds": 3.920374494045973} Step 30,000 | monitor rollout @ 0.50 coverage 71.4% | precision 84.2% | diverged 21.9% Step 30,000 | monitor rollout @ 0.50 coverage 87.7% | precision 100.0% | diverged 0.0% Step 30,000 | monitor rollout @ 0.50 coverage 94.2% | precision 100.0% | diverged 0.0% Step 30,000 | monitor rollout @ 0.50 coverage 87.8% | precision 97.3% | diverged 3.1% Step 30,000 | monitor rollout @ 0.50 coverage 65.8% | precision 77.2% | diverged 9.4% Step 30,000 | monitor recovery | 32 states recovery @ 0.50: 0-1.5: recovered 6/8 (75.0%) | false stops 1 1.5-3: recovered 6/8 (75.0%) | false stops 1 3-4.5: recovered 6/8 (75.0%) | false stops 0 4.5-6: recovered 4/8 (50.0%) | false stops 0 entire proposed 8-point prefix correct (distance): 22/28 (78.6%) gate @ 0.50: false stops 0/23 (0.0%) | accepted wrong 5/27 (18.5%) | accepted unknown 4 | terminal continues n/a (0 known) refinement: mean lateral error in voxels state states GT pts initial final all 32 812 1.610 1.567 following 16 464 1.155 1.170 recoverable 12 348 2.215 2.096 unknown 4 0 -- -- displaced states GT pts initial final d<=3 16 464 1.155 1.170 332 32 812 1.610 1.567 improved/worsened/compared states: 3/2/28 Step 30,000 | batch_diagnostics_seconds: 3.923679731786251 | diagnostics_seconds: 592.5769517719746 | recovery_seconds: 15.0382163785398 Training | event: "remote_prefetch_shutdown" | remote_prefetch: {"submitted": 36, "deferred": 0, "cached": 0, "completed": 126, "completed_bytes": 264241152, "deduplicated": 14, "promoted": 13, "preempted": 0, "errors": 0, "active": 0, "lookahead_windows": 0, "lookahead_chunks": 0, "alive": false}