anhtld commited on
Commit
8a0e8b2
·
verified ·
1 Parent(s): c09fae3

Upload per-dim trainmax test seed2

Browse files
Files changed (23) hide show
  1. workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/command.txt +1 -0
  2. workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/config.yaml +22 -0
  3. workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/data_hash.txt +1 -0
  4. workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/eval.log +535 -0
  5. workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/git_hash.txt +1 -0
  6. workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/measured_candidates.json +0 -0
  7. workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/measured_metrics/metrics.json +0 -0
  8. workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/measured_metrics/metrics_by_seed.json +1410 -0
  9. workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/measured_metrics/metrics_by_task.json +178 -0
  10. workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/measured_metrics/report.md +40 -0
  11. workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/measured_metrics/table.tex +41 -0
  12. workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/metrics.json +2814 -0
  13. workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/metrics_by_seed.json +1410 -0
  14. workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/metrics_by_task.json +178 -0
  15. workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/report.md +13 -0
  16. workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/run.log +535 -0
  17. workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/source_data_hash.txt +1 -0
  18. workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/source_index.json +0 -0
  19. workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/source_split_hash.txt +1 -0
  20. workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/split_hash.txt +1 -0
  21. workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/table.tex +41 -0
  22. workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/target_index.json +911 -0
  23. workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/train.log +1 -0
workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/command.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ python scripts/eval_ctt_generated_rollout.py --checkpoint runs/ctt_residual_base_context_obs_seed2/model.pt --source-index data/cil_charts_rgb_refs/train/index.json --target-index data/cil_charts_rgb_refs/test/index.json --out-dir runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2 --k 8 --pool-size 0 --neighbors 8 --max-target-charts 48 --group-batch-size 1 --device cpu --sim-backend physx_cpu --render-backend cpu --restore-tolerance 1e-5 --delta-scale 1.0 --execution-action-scale 1.0 --execution-action-scale-vector 0.381401261606,0.371117178171,0.354308917074,0.237850081835,0.248271283428,0.232648812617,0.215420039658 --execution-action-transform identity --bootstrap-samples 200 --disable-action-clipping
workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/config.yaml ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ bootstrap_samples: 200
2
+ checkpoint: runs/ctt_residual_base_context_obs_seed2/model.pt
3
+ delta_scale: 1.0
4
+ device: cpu
5
+ disable_action_clipping: True
6
+ exclude_self_source: False
7
+ execution_action_scale: 1.0
8
+ execution_action_scale_vector: 0.381401261606,0.371117178171,0.354308917074,0.237850081835,0.248271283428,0.232648812617,0.215420039658
9
+ execution_action_transform: identity
10
+ group_batch_size: 1
11
+ include_targets_without_positives: False
12
+ k: 8
13
+ max_target_charts: 48
14
+ neighbors: 8
15
+ out_dir: runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2
16
+ pool_size: 0
17
+ render_backend: cpu
18
+ restore_tolerance: 1e-05
19
+ sim_backend: physx_cpu
20
+ skip_metrics: False
21
+ source_index: data/cil_charts_rgb_refs/train/index.json
22
+ target_index: data/cil_charts_rgb_refs/test/index.json
workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/data_hash.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ 0c38d73dc9bc4ab756e12006671b5d12f1a23358b281e54b3f6a9ae3cbf6e9af
workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/eval.log ADDED
@@ -0,0 +1,535 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ start
2
+ importing gymnasium/mani_skill
3
+ imported gymnasium/mani_skill
4
+ loaded checkpoint=runs/ctt_residual_base_context_obs_seed2/model.pt
5
+ loaded charts source=301 target=48
6
+ generated proposals rows=48 total=384
7
+ env init sequential task=LiftPegUpright-v1 batch=1 sim=physx_cpu render=cpu
8
+ execute sequential task=LiftPegUpright-v1 chart=ms-5799584ed40e211c50a5 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
9
+ execute sequential task=LiftPegUpright-v1 chart=ms-5799584ed40e211c50a5 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
10
+ execute sequential task=LiftPegUpright-v1 chart=ms-5799584ed40e211c50a5 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
11
+ execute sequential task=LiftPegUpright-v1 chart=ms-5799584ed40e211c50a5 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
12
+ execute sequential task=LiftPegUpright-v1 chart=ms-5799584ed40e211c50a5 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
13
+ execute sequential task=LiftPegUpright-v1 chart=ms-5799584ed40e211c50a5 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
14
+ execute sequential task=LiftPegUpright-v1 chart=ms-5799584ed40e211c50a5 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
15
+ execute sequential task=LiftPegUpright-v1 chart=ms-5799584ed40e211c50a5 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
16
+ execute sequential task=LiftPegUpright-v1 chart=ms-5799584ed40e211c50a5 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
17
+ batch done sequential task=LiftPegUpright-v1
18
+ env init sequential task=LiftPegUpright-v1 batch=1 sim=physx_cpu render=cpu
19
+ execute sequential task=LiftPegUpright-v1 chart=ms-715db95b13ebc6c977c8 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
20
+ execute sequential task=LiftPegUpright-v1 chart=ms-715db95b13ebc6c977c8 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
21
+ execute sequential task=LiftPegUpright-v1 chart=ms-715db95b13ebc6c977c8 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
22
+ execute sequential task=LiftPegUpright-v1 chart=ms-715db95b13ebc6c977c8 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
23
+ execute sequential task=LiftPegUpright-v1 chart=ms-715db95b13ebc6c977c8 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
24
+ execute sequential task=LiftPegUpright-v1 chart=ms-715db95b13ebc6c977c8 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
25
+ execute sequential task=LiftPegUpright-v1 chart=ms-715db95b13ebc6c977c8 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
26
+ execute sequential task=LiftPegUpright-v1 chart=ms-715db95b13ebc6c977c8 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
27
+ execute sequential task=LiftPegUpright-v1 chart=ms-715db95b13ebc6c977c8 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
28
+ batch done sequential task=LiftPegUpright-v1
29
+ env init sequential task=LiftPegUpright-v1 batch=1 sim=physx_cpu render=cpu
30
+ execute sequential task=LiftPegUpright-v1 chart=ms-7b229b60d7f64fb624cf candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
31
+ execute sequential task=LiftPegUpright-v1 chart=ms-7b229b60d7f64fb624cf candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
32
+ execute sequential task=LiftPegUpright-v1 chart=ms-7b229b60d7f64fb624cf candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
33
+ execute sequential task=LiftPegUpright-v1 chart=ms-7b229b60d7f64fb624cf candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
34
+ execute sequential task=LiftPegUpright-v1 chart=ms-7b229b60d7f64fb624cf candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
35
+ execute sequential task=LiftPegUpright-v1 chart=ms-7b229b60d7f64fb624cf candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
36
+ execute sequential task=LiftPegUpright-v1 chart=ms-7b229b60d7f64fb624cf candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
37
+ execute sequential task=LiftPegUpright-v1 chart=ms-7b229b60d7f64fb624cf candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
38
+ execute sequential task=LiftPegUpright-v1 chart=ms-7b229b60d7f64fb624cf candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
39
+ batch done sequential task=LiftPegUpright-v1
40
+ env init sequential task=LiftPegUpright-v1 batch=1 sim=physx_cpu render=cpu
41
+ execute sequential task=LiftPegUpright-v1 chart=ms-99f263cb3f383b1f6243 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
42
+ execute sequential task=LiftPegUpright-v1 chart=ms-99f263cb3f383b1f6243 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
43
+ execute sequential task=LiftPegUpright-v1 chart=ms-99f263cb3f383b1f6243 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
44
+ execute sequential task=LiftPegUpright-v1 chart=ms-99f263cb3f383b1f6243 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
45
+ execute sequential task=LiftPegUpright-v1 chart=ms-99f263cb3f383b1f6243 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
46
+ execute sequential task=LiftPegUpright-v1 chart=ms-99f263cb3f383b1f6243 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
47
+ execute sequential task=LiftPegUpright-v1 chart=ms-99f263cb3f383b1f6243 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
48
+ execute sequential task=LiftPegUpright-v1 chart=ms-99f263cb3f383b1f6243 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
49
+ execute sequential task=LiftPegUpright-v1 chart=ms-99f263cb3f383b1f6243 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
50
+ batch done sequential task=LiftPegUpright-v1
51
+ env init sequential task=LiftPegUpright-v1 batch=1 sim=physx_cpu render=cpu
52
+ execute sequential task=LiftPegUpright-v1 chart=ms-ec2e4afb34b439731136 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
53
+ execute sequential task=LiftPegUpright-v1 chart=ms-ec2e4afb34b439731136 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
54
+ execute sequential task=LiftPegUpright-v1 chart=ms-ec2e4afb34b439731136 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
55
+ execute sequential task=LiftPegUpright-v1 chart=ms-ec2e4afb34b439731136 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
56
+ execute sequential task=LiftPegUpright-v1 chart=ms-ec2e4afb34b439731136 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
57
+ execute sequential task=LiftPegUpright-v1 chart=ms-ec2e4afb34b439731136 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
58
+ execute sequential task=LiftPegUpright-v1 chart=ms-ec2e4afb34b439731136 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
59
+ execute sequential task=LiftPegUpright-v1 chart=ms-ec2e4afb34b439731136 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
60
+ execute sequential task=LiftPegUpright-v1 chart=ms-ec2e4afb34b439731136 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
61
+ batch done sequential task=LiftPegUpright-v1
62
+ env init sequential task=PickCube-v1 batch=1 sim=physx_cpu render=cpu
63
+ execute sequential task=PickCube-v1 chart=ms-a9ac13384fe5afc835f1 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
64
+ execute sequential task=PickCube-v1 chart=ms-a9ac13384fe5afc835f1 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
65
+ execute sequential task=PickCube-v1 chart=ms-a9ac13384fe5afc835f1 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
66
+ execute sequential task=PickCube-v1 chart=ms-a9ac13384fe5afc835f1 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
67
+ execute sequential task=PickCube-v1 chart=ms-a9ac13384fe5afc835f1 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
68
+ execute sequential task=PickCube-v1 chart=ms-a9ac13384fe5afc835f1 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
69
+ execute sequential task=PickCube-v1 chart=ms-a9ac13384fe5afc835f1 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
70
+ execute sequential task=PickCube-v1 chart=ms-a9ac13384fe5afc835f1 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
71
+ execute sequential task=PickCube-v1 chart=ms-a9ac13384fe5afc835f1 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
72
+ batch done sequential task=PickCube-v1
73
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
74
+ execute sequential task=PullCube-v1 chart=ms-06babf453b22ba8f2407 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
75
+ execute sequential task=PullCube-v1 chart=ms-06babf453b22ba8f2407 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
76
+ execute sequential task=PullCube-v1 chart=ms-06babf453b22ba8f2407 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
77
+ execute sequential task=PullCube-v1 chart=ms-06babf453b22ba8f2407 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
78
+ execute sequential task=PullCube-v1 chart=ms-06babf453b22ba8f2407 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
79
+ execute sequential task=PullCube-v1 chart=ms-06babf453b22ba8f2407 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
80
+ execute sequential task=PullCube-v1 chart=ms-06babf453b22ba8f2407 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
81
+ execute sequential task=PullCube-v1 chart=ms-06babf453b22ba8f2407 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
82
+ execute sequential task=PullCube-v1 chart=ms-06babf453b22ba8f2407 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
83
+ batch done sequential task=PullCube-v1
84
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
85
+ execute sequential task=PullCube-v1 chart=ms-0e1ff8d16372796dac97 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
86
+ execute sequential task=PullCube-v1 chart=ms-0e1ff8d16372796dac97 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
87
+ execute sequential task=PullCube-v1 chart=ms-0e1ff8d16372796dac97 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
88
+ execute sequential task=PullCube-v1 chart=ms-0e1ff8d16372796dac97 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
89
+ execute sequential task=PullCube-v1 chart=ms-0e1ff8d16372796dac97 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
90
+ execute sequential task=PullCube-v1 chart=ms-0e1ff8d16372796dac97 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
91
+ execute sequential task=PullCube-v1 chart=ms-0e1ff8d16372796dac97 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
92
+ execute sequential task=PullCube-v1 chart=ms-0e1ff8d16372796dac97 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
93
+ execute sequential task=PullCube-v1 chart=ms-0e1ff8d16372796dac97 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
94
+ batch done sequential task=PullCube-v1
95
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
96
+ execute sequential task=PullCube-v1 chart=ms-223b0a8319beb051a1ee candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
97
+ execute sequential task=PullCube-v1 chart=ms-223b0a8319beb051a1ee candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
98
+ execute sequential task=PullCube-v1 chart=ms-223b0a8319beb051a1ee candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
99
+ execute sequential task=PullCube-v1 chart=ms-223b0a8319beb051a1ee candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
100
+ execute sequential task=PullCube-v1 chart=ms-223b0a8319beb051a1ee candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
101
+ execute sequential task=PullCube-v1 chart=ms-223b0a8319beb051a1ee candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
102
+ execute sequential task=PullCube-v1 chart=ms-223b0a8319beb051a1ee candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
103
+ execute sequential task=PullCube-v1 chart=ms-223b0a8319beb051a1ee candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
104
+ execute sequential task=PullCube-v1 chart=ms-223b0a8319beb051a1ee candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
105
+ batch done sequential task=PullCube-v1
106
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
107
+ execute sequential task=PullCube-v1 chart=ms-34726202b475d29cdf38 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
108
+ execute sequential task=PullCube-v1 chart=ms-34726202b475d29cdf38 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
109
+ execute sequential task=PullCube-v1 chart=ms-34726202b475d29cdf38 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
110
+ execute sequential task=PullCube-v1 chart=ms-34726202b475d29cdf38 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
111
+ execute sequential task=PullCube-v1 chart=ms-34726202b475d29cdf38 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
112
+ execute sequential task=PullCube-v1 chart=ms-34726202b475d29cdf38 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
113
+ execute sequential task=PullCube-v1 chart=ms-34726202b475d29cdf38 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
114
+ execute sequential task=PullCube-v1 chart=ms-34726202b475d29cdf38 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
115
+ execute sequential task=PullCube-v1 chart=ms-34726202b475d29cdf38 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
116
+ batch done sequential task=PullCube-v1
117
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
118
+ execute sequential task=PullCube-v1 chart=ms-4cc40587ec490257b1d2 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
119
+ execute sequential task=PullCube-v1 chart=ms-4cc40587ec490257b1d2 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
120
+ execute sequential task=PullCube-v1 chart=ms-4cc40587ec490257b1d2 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
121
+ execute sequential task=PullCube-v1 chart=ms-4cc40587ec490257b1d2 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
122
+ execute sequential task=PullCube-v1 chart=ms-4cc40587ec490257b1d2 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
123
+ execute sequential task=PullCube-v1 chart=ms-4cc40587ec490257b1d2 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
124
+ execute sequential task=PullCube-v1 chart=ms-4cc40587ec490257b1d2 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
125
+ execute sequential task=PullCube-v1 chart=ms-4cc40587ec490257b1d2 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
126
+ execute sequential task=PullCube-v1 chart=ms-4cc40587ec490257b1d2 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
127
+ batch done sequential task=PullCube-v1
128
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
129
+ execute sequential task=PullCube-v1 chart=ms-62f8c18a15c566ec4b00 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
130
+ execute sequential task=PullCube-v1 chart=ms-62f8c18a15c566ec4b00 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
131
+ execute sequential task=PullCube-v1 chart=ms-62f8c18a15c566ec4b00 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
132
+ execute sequential task=PullCube-v1 chart=ms-62f8c18a15c566ec4b00 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
133
+ execute sequential task=PullCube-v1 chart=ms-62f8c18a15c566ec4b00 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
134
+ execute sequential task=PullCube-v1 chart=ms-62f8c18a15c566ec4b00 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
135
+ execute sequential task=PullCube-v1 chart=ms-62f8c18a15c566ec4b00 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
136
+ execute sequential task=PullCube-v1 chart=ms-62f8c18a15c566ec4b00 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
137
+ execute sequential task=PullCube-v1 chart=ms-62f8c18a15c566ec4b00 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
138
+ batch done sequential task=PullCube-v1
139
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
140
+ execute sequential task=PullCube-v1 chart=ms-6661b157c41ad4d6ab85 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
141
+ execute sequential task=PullCube-v1 chart=ms-6661b157c41ad4d6ab85 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
142
+ execute sequential task=PullCube-v1 chart=ms-6661b157c41ad4d6ab85 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
143
+ execute sequential task=PullCube-v1 chart=ms-6661b157c41ad4d6ab85 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
144
+ execute sequential task=PullCube-v1 chart=ms-6661b157c41ad4d6ab85 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
145
+ execute sequential task=PullCube-v1 chart=ms-6661b157c41ad4d6ab85 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
146
+ execute sequential task=PullCube-v1 chart=ms-6661b157c41ad4d6ab85 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
147
+ execute sequential task=PullCube-v1 chart=ms-6661b157c41ad4d6ab85 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
148
+ execute sequential task=PullCube-v1 chart=ms-6661b157c41ad4d6ab85 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
149
+ batch done sequential task=PullCube-v1
150
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
151
+ execute sequential task=PullCube-v1 chart=ms-6b789356d65b3312fbae candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
152
+ execute sequential task=PullCube-v1 chart=ms-6b789356d65b3312fbae candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
153
+ execute sequential task=PullCube-v1 chart=ms-6b789356d65b3312fbae candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
154
+ execute sequential task=PullCube-v1 chart=ms-6b789356d65b3312fbae candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
155
+ execute sequential task=PullCube-v1 chart=ms-6b789356d65b3312fbae candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
156
+ execute sequential task=PullCube-v1 chart=ms-6b789356d65b3312fbae candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
157
+ execute sequential task=PullCube-v1 chart=ms-6b789356d65b3312fbae candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
158
+ execute sequential task=PullCube-v1 chart=ms-6b789356d65b3312fbae candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
159
+ execute sequential task=PullCube-v1 chart=ms-6b789356d65b3312fbae candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
160
+ batch done sequential task=PullCube-v1
161
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
162
+ execute sequential task=PullCube-v1 chart=ms-7923b8196b397863cf4f candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
163
+ execute sequential task=PullCube-v1 chart=ms-7923b8196b397863cf4f candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
164
+ execute sequential task=PullCube-v1 chart=ms-7923b8196b397863cf4f candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
165
+ execute sequential task=PullCube-v1 chart=ms-7923b8196b397863cf4f candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
166
+ execute sequential task=PullCube-v1 chart=ms-7923b8196b397863cf4f candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
167
+ execute sequential task=PullCube-v1 chart=ms-7923b8196b397863cf4f candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
168
+ execute sequential task=PullCube-v1 chart=ms-7923b8196b397863cf4f candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
169
+ execute sequential task=PullCube-v1 chart=ms-7923b8196b397863cf4f candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
170
+ execute sequential task=PullCube-v1 chart=ms-7923b8196b397863cf4f candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
171
+ batch done sequential task=PullCube-v1
172
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
173
+ execute sequential task=PullCube-v1 chart=ms-7df455efd413d77e6ce3 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
174
+ execute sequential task=PullCube-v1 chart=ms-7df455efd413d77e6ce3 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
175
+ execute sequential task=PullCube-v1 chart=ms-7df455efd413d77e6ce3 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
176
+ execute sequential task=PullCube-v1 chart=ms-7df455efd413d77e6ce3 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
177
+ execute sequential task=PullCube-v1 chart=ms-7df455efd413d77e6ce3 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
178
+ execute sequential task=PullCube-v1 chart=ms-7df455efd413d77e6ce3 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
179
+ execute sequential task=PullCube-v1 chart=ms-7df455efd413d77e6ce3 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
180
+ execute sequential task=PullCube-v1 chart=ms-7df455efd413d77e6ce3 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
181
+ execute sequential task=PullCube-v1 chart=ms-7df455efd413d77e6ce3 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
182
+ batch done sequential task=PullCube-v1
183
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
184
+ execute sequential task=PullCube-v1 chart=ms-7e295b5859c07c90487f candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
185
+ execute sequential task=PullCube-v1 chart=ms-7e295b5859c07c90487f candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
186
+ execute sequential task=PullCube-v1 chart=ms-7e295b5859c07c90487f candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
187
+ execute sequential task=PullCube-v1 chart=ms-7e295b5859c07c90487f candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
188
+ execute sequential task=PullCube-v1 chart=ms-7e295b5859c07c90487f candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
189
+ execute sequential task=PullCube-v1 chart=ms-7e295b5859c07c90487f candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
190
+ execute sequential task=PullCube-v1 chart=ms-7e295b5859c07c90487f candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
191
+ execute sequential task=PullCube-v1 chart=ms-7e295b5859c07c90487f candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
192
+ execute sequential task=PullCube-v1 chart=ms-7e295b5859c07c90487f candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
193
+ batch done sequential task=PullCube-v1
194
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
195
+ execute sequential task=PullCube-v1 chart=ms-8577816da2917a2bb63d candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
196
+ execute sequential task=PullCube-v1 chart=ms-8577816da2917a2bb63d candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
197
+ execute sequential task=PullCube-v1 chart=ms-8577816da2917a2bb63d candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
198
+ execute sequential task=PullCube-v1 chart=ms-8577816da2917a2bb63d candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
199
+ execute sequential task=PullCube-v1 chart=ms-8577816da2917a2bb63d candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
200
+ execute sequential task=PullCube-v1 chart=ms-8577816da2917a2bb63d candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
201
+ execute sequential task=PullCube-v1 chart=ms-8577816da2917a2bb63d candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
202
+ execute sequential task=PullCube-v1 chart=ms-8577816da2917a2bb63d candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
203
+ execute sequential task=PullCube-v1 chart=ms-8577816da2917a2bb63d candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
204
+ batch done sequential task=PullCube-v1
205
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
206
+ execute sequential task=PullCube-v1 chart=ms-8e56b0eaa2604db6749e candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
207
+ execute sequential task=PullCube-v1 chart=ms-8e56b0eaa2604db6749e candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
208
+ execute sequential task=PullCube-v1 chart=ms-8e56b0eaa2604db6749e candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
209
+ execute sequential task=PullCube-v1 chart=ms-8e56b0eaa2604db6749e candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
210
+ execute sequential task=PullCube-v1 chart=ms-8e56b0eaa2604db6749e candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
211
+ execute sequential task=PullCube-v1 chart=ms-8e56b0eaa2604db6749e candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
212
+ execute sequential task=PullCube-v1 chart=ms-8e56b0eaa2604db6749e candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
213
+ execute sequential task=PullCube-v1 chart=ms-8e56b0eaa2604db6749e candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
214
+ execute sequential task=PullCube-v1 chart=ms-8e56b0eaa2604db6749e candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
215
+ batch done sequential task=PullCube-v1
216
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
217
+ execute sequential task=PullCube-v1 chart=ms-972c5be72d0b24be7cd5 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
218
+ execute sequential task=PullCube-v1 chart=ms-972c5be72d0b24be7cd5 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
219
+ execute sequential task=PullCube-v1 chart=ms-972c5be72d0b24be7cd5 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
220
+ execute sequential task=PullCube-v1 chart=ms-972c5be72d0b24be7cd5 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
221
+ execute sequential task=PullCube-v1 chart=ms-972c5be72d0b24be7cd5 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
222
+ execute sequential task=PullCube-v1 chart=ms-972c5be72d0b24be7cd5 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
223
+ execute sequential task=PullCube-v1 chart=ms-972c5be72d0b24be7cd5 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
224
+ execute sequential task=PullCube-v1 chart=ms-972c5be72d0b24be7cd5 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
225
+ execute sequential task=PullCube-v1 chart=ms-972c5be72d0b24be7cd5 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
226
+ batch done sequential task=PullCube-v1
227
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
228
+ execute sequential task=PullCube-v1 chart=ms-9757fc494e79b5543bb6 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
229
+ execute sequential task=PullCube-v1 chart=ms-9757fc494e79b5543bb6 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
230
+ execute sequential task=PullCube-v1 chart=ms-9757fc494e79b5543bb6 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
231
+ execute sequential task=PullCube-v1 chart=ms-9757fc494e79b5543bb6 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
232
+ execute sequential task=PullCube-v1 chart=ms-9757fc494e79b5543bb6 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
233
+ execute sequential task=PullCube-v1 chart=ms-9757fc494e79b5543bb6 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
234
+ execute sequential task=PullCube-v1 chart=ms-9757fc494e79b5543bb6 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
235
+ execute sequential task=PullCube-v1 chart=ms-9757fc494e79b5543bb6 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
236
+ execute sequential task=PullCube-v1 chart=ms-9757fc494e79b5543bb6 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
237
+ batch done sequential task=PullCube-v1
238
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
239
+ execute sequential task=PullCube-v1 chart=ms-a8731059713697ce5e2a candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
240
+ execute sequential task=PullCube-v1 chart=ms-a8731059713697ce5e2a candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
241
+ execute sequential task=PullCube-v1 chart=ms-a8731059713697ce5e2a candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
242
+ execute sequential task=PullCube-v1 chart=ms-a8731059713697ce5e2a candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
243
+ execute sequential task=PullCube-v1 chart=ms-a8731059713697ce5e2a candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
244
+ execute sequential task=PullCube-v1 chart=ms-a8731059713697ce5e2a candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
245
+ execute sequential task=PullCube-v1 chart=ms-a8731059713697ce5e2a candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
246
+ execute sequential task=PullCube-v1 chart=ms-a8731059713697ce5e2a candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
247
+ execute sequential task=PullCube-v1 chart=ms-a8731059713697ce5e2a candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
248
+ batch done sequential task=PullCube-v1
249
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
250
+ execute sequential task=PullCube-v1 chart=ms-ac94b48febe523da15a5 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
251
+ execute sequential task=PullCube-v1 chart=ms-ac94b48febe523da15a5 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
252
+ execute sequential task=PullCube-v1 chart=ms-ac94b48febe523da15a5 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
253
+ execute sequential task=PullCube-v1 chart=ms-ac94b48febe523da15a5 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
254
+ execute sequential task=PullCube-v1 chart=ms-ac94b48febe523da15a5 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
255
+ execute sequential task=PullCube-v1 chart=ms-ac94b48febe523da15a5 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
256
+ execute sequential task=PullCube-v1 chart=ms-ac94b48febe523da15a5 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
257
+ execute sequential task=PullCube-v1 chart=ms-ac94b48febe523da15a5 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
258
+ execute sequential task=PullCube-v1 chart=ms-ac94b48febe523da15a5 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
259
+ batch done sequential task=PullCube-v1
260
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
261
+ execute sequential task=PullCube-v1 chart=ms-b49043f15f4846de5b13 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
262
+ execute sequential task=PullCube-v1 chart=ms-b49043f15f4846de5b13 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
263
+ execute sequential task=PullCube-v1 chart=ms-b49043f15f4846de5b13 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
264
+ execute sequential task=PullCube-v1 chart=ms-b49043f15f4846de5b13 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
265
+ execute sequential task=PullCube-v1 chart=ms-b49043f15f4846de5b13 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
266
+ execute sequential task=PullCube-v1 chart=ms-b49043f15f4846de5b13 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
267
+ execute sequential task=PullCube-v1 chart=ms-b49043f15f4846de5b13 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
268
+ execute sequential task=PullCube-v1 chart=ms-b49043f15f4846de5b13 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
269
+ execute sequential task=PullCube-v1 chart=ms-b49043f15f4846de5b13 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
270
+ batch done sequential task=PullCube-v1
271
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
272
+ execute sequential task=PullCube-v1 chart=ms-b7d7021b782f73816083 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
273
+ execute sequential task=PullCube-v1 chart=ms-b7d7021b782f73816083 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
274
+ execute sequential task=PullCube-v1 chart=ms-b7d7021b782f73816083 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
275
+ execute sequential task=PullCube-v1 chart=ms-b7d7021b782f73816083 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
276
+ execute sequential task=PullCube-v1 chart=ms-b7d7021b782f73816083 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
277
+ execute sequential task=PullCube-v1 chart=ms-b7d7021b782f73816083 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
278
+ execute sequential task=PullCube-v1 chart=ms-b7d7021b782f73816083 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
279
+ execute sequential task=PullCube-v1 chart=ms-b7d7021b782f73816083 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
280
+ execute sequential task=PullCube-v1 chart=ms-b7d7021b782f73816083 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
281
+ batch done sequential task=PullCube-v1
282
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
283
+ execute sequential task=PullCube-v1 chart=ms-b8c3ec3ea915c5d35b3e candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
284
+ execute sequential task=PullCube-v1 chart=ms-b8c3ec3ea915c5d35b3e candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
285
+ execute sequential task=PullCube-v1 chart=ms-b8c3ec3ea915c5d35b3e candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
286
+ execute sequential task=PullCube-v1 chart=ms-b8c3ec3ea915c5d35b3e candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
287
+ execute sequential task=PullCube-v1 chart=ms-b8c3ec3ea915c5d35b3e candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
288
+ execute sequential task=PullCube-v1 chart=ms-b8c3ec3ea915c5d35b3e candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
289
+ execute sequential task=PullCube-v1 chart=ms-b8c3ec3ea915c5d35b3e candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
290
+ execute sequential task=PullCube-v1 chart=ms-b8c3ec3ea915c5d35b3e candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
291
+ execute sequential task=PullCube-v1 chart=ms-b8c3ec3ea915c5d35b3e candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
292
+ batch done sequential task=PullCube-v1
293
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
294
+ execute sequential task=PullCube-v1 chart=ms-b8fc1568ce66c45c8d8b candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
295
+ execute sequential task=PullCube-v1 chart=ms-b8fc1568ce66c45c8d8b candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
296
+ execute sequential task=PullCube-v1 chart=ms-b8fc1568ce66c45c8d8b candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
297
+ execute sequential task=PullCube-v1 chart=ms-b8fc1568ce66c45c8d8b candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
298
+ execute sequential task=PullCube-v1 chart=ms-b8fc1568ce66c45c8d8b candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
299
+ execute sequential task=PullCube-v1 chart=ms-b8fc1568ce66c45c8d8b candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
300
+ execute sequential task=PullCube-v1 chart=ms-b8fc1568ce66c45c8d8b candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
301
+ execute sequential task=PullCube-v1 chart=ms-b8fc1568ce66c45c8d8b candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
302
+ execute sequential task=PullCube-v1 chart=ms-b8fc1568ce66c45c8d8b candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
303
+ batch done sequential task=PullCube-v1
304
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
305
+ execute sequential task=PullCube-v1 chart=ms-da555a03dd382d93e119 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
306
+ execute sequential task=PullCube-v1 chart=ms-da555a03dd382d93e119 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
307
+ execute sequential task=PullCube-v1 chart=ms-da555a03dd382d93e119 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
308
+ execute sequential task=PullCube-v1 chart=ms-da555a03dd382d93e119 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
309
+ execute sequential task=PullCube-v1 chart=ms-da555a03dd382d93e119 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
310
+ execute sequential task=PullCube-v1 chart=ms-da555a03dd382d93e119 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
311
+ execute sequential task=PullCube-v1 chart=ms-da555a03dd382d93e119 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
312
+ execute sequential task=PullCube-v1 chart=ms-da555a03dd382d93e119 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
313
+ execute sequential task=PullCube-v1 chart=ms-da555a03dd382d93e119 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
314
+ batch done sequential task=PullCube-v1
315
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
316
+ execute sequential task=PullCube-v1 chart=ms-f5dd1c43ebc4c0cfc3f1 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
317
+ execute sequential task=PullCube-v1 chart=ms-f5dd1c43ebc4c0cfc3f1 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
318
+ execute sequential task=PullCube-v1 chart=ms-f5dd1c43ebc4c0cfc3f1 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
319
+ execute sequential task=PullCube-v1 chart=ms-f5dd1c43ebc4c0cfc3f1 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
320
+ execute sequential task=PullCube-v1 chart=ms-f5dd1c43ebc4c0cfc3f1 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
321
+ execute sequential task=PullCube-v1 chart=ms-f5dd1c43ebc4c0cfc3f1 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
322
+ execute sequential task=PullCube-v1 chart=ms-f5dd1c43ebc4c0cfc3f1 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
323
+ execute sequential task=PullCube-v1 chart=ms-f5dd1c43ebc4c0cfc3f1 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
324
+ execute sequential task=PullCube-v1 chart=ms-f5dd1c43ebc4c0cfc3f1 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
325
+ batch done sequential task=PullCube-v1
326
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
327
+ execute sequential task=PullCube-v1 chart=ms-f9740a08a98cb251b761 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
328
+ execute sequential task=PullCube-v1 chart=ms-f9740a08a98cb251b761 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
329
+ execute sequential task=PullCube-v1 chart=ms-f9740a08a98cb251b761 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
330
+ execute sequential task=PullCube-v1 chart=ms-f9740a08a98cb251b761 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
331
+ execute sequential task=PullCube-v1 chart=ms-f9740a08a98cb251b761 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
332
+ execute sequential task=PullCube-v1 chart=ms-f9740a08a98cb251b761 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
333
+ execute sequential task=PullCube-v1 chart=ms-f9740a08a98cb251b761 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
334
+ execute sequential task=PullCube-v1 chart=ms-f9740a08a98cb251b761 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
335
+ execute sequential task=PullCube-v1 chart=ms-f9740a08a98cb251b761 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
336
+ batch done sequential task=PullCube-v1
337
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
338
+ execute sequential task=PullCube-v1 chart=ms-fa8b4ba41ecac8030271 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
339
+ execute sequential task=PullCube-v1 chart=ms-fa8b4ba41ecac8030271 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
340
+ execute sequential task=PullCube-v1 chart=ms-fa8b4ba41ecac8030271 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
341
+ execute sequential task=PullCube-v1 chart=ms-fa8b4ba41ecac8030271 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
342
+ execute sequential task=PullCube-v1 chart=ms-fa8b4ba41ecac8030271 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
343
+ execute sequential task=PullCube-v1 chart=ms-fa8b4ba41ecac8030271 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
344
+ execute sequential task=PullCube-v1 chart=ms-fa8b4ba41ecac8030271 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
345
+ execute sequential task=PullCube-v1 chart=ms-fa8b4ba41ecac8030271 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
346
+ execute sequential task=PullCube-v1 chart=ms-fa8b4ba41ecac8030271 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
347
+ batch done sequential task=PullCube-v1
348
+ env init sequential task=PushCube-v1 batch=1 sim=physx_cpu render=cpu
349
+ execute sequential task=PushCube-v1 chart=ms-05f34a90fa7355262b8e candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
350
+ execute sequential task=PushCube-v1 chart=ms-05f34a90fa7355262b8e candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
351
+ execute sequential task=PushCube-v1 chart=ms-05f34a90fa7355262b8e candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
352
+ execute sequential task=PushCube-v1 chart=ms-05f34a90fa7355262b8e candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
353
+ execute sequential task=PushCube-v1 chart=ms-05f34a90fa7355262b8e candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
354
+ execute sequential task=PushCube-v1 chart=ms-05f34a90fa7355262b8e candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
355
+ execute sequential task=PushCube-v1 chart=ms-05f34a90fa7355262b8e candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
356
+ execute sequential task=PushCube-v1 chart=ms-05f34a90fa7355262b8e candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
357
+ execute sequential task=PushCube-v1 chart=ms-05f34a90fa7355262b8e candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
358
+ batch done sequential task=PushCube-v1
359
+ env init sequential task=PushCube-v1 batch=1 sim=physx_cpu render=cpu
360
+ execute sequential task=PushCube-v1 chart=ms-2413d1155c87f0442485 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
361
+ execute sequential task=PushCube-v1 chart=ms-2413d1155c87f0442485 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
362
+ execute sequential task=PushCube-v1 chart=ms-2413d1155c87f0442485 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
363
+ execute sequential task=PushCube-v1 chart=ms-2413d1155c87f0442485 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
364
+ execute sequential task=PushCube-v1 chart=ms-2413d1155c87f0442485 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
365
+ execute sequential task=PushCube-v1 chart=ms-2413d1155c87f0442485 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
366
+ execute sequential task=PushCube-v1 chart=ms-2413d1155c87f0442485 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
367
+ execute sequential task=PushCube-v1 chart=ms-2413d1155c87f0442485 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
368
+ execute sequential task=PushCube-v1 chart=ms-2413d1155c87f0442485 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
369
+ batch done sequential task=PushCube-v1
370
+ env init sequential task=PushCube-v1 batch=1 sim=physx_cpu render=cpu
371
+ execute sequential task=PushCube-v1 chart=ms-54346a8f79bd9dca1dfe candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
372
+ execute sequential task=PushCube-v1 chart=ms-54346a8f79bd9dca1dfe candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
373
+ execute sequential task=PushCube-v1 chart=ms-54346a8f79bd9dca1dfe candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
374
+ execute sequential task=PushCube-v1 chart=ms-54346a8f79bd9dca1dfe candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
375
+ execute sequential task=PushCube-v1 chart=ms-54346a8f79bd9dca1dfe candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
376
+ execute sequential task=PushCube-v1 chart=ms-54346a8f79bd9dca1dfe candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
377
+ execute sequential task=PushCube-v1 chart=ms-54346a8f79bd9dca1dfe candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
378
+ execute sequential task=PushCube-v1 chart=ms-54346a8f79bd9dca1dfe candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
379
+ execute sequential task=PushCube-v1 chart=ms-54346a8f79bd9dca1dfe candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
380
+ batch done sequential task=PushCube-v1
381
+ env init sequential task=PushCube-v1 batch=1 sim=physx_cpu render=cpu
382
+ execute sequential task=PushCube-v1 chart=ms-64f98d163b6bff29332b candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
383
+ execute sequential task=PushCube-v1 chart=ms-64f98d163b6bff29332b candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
384
+ execute sequential task=PushCube-v1 chart=ms-64f98d163b6bff29332b candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
385
+ execute sequential task=PushCube-v1 chart=ms-64f98d163b6bff29332b candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
386
+ execute sequential task=PushCube-v1 chart=ms-64f98d163b6bff29332b candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
387
+ execute sequential task=PushCube-v1 chart=ms-64f98d163b6bff29332b candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
388
+ execute sequential task=PushCube-v1 chart=ms-64f98d163b6bff29332b candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
389
+ execute sequential task=PushCube-v1 chart=ms-64f98d163b6bff29332b candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
390
+ execute sequential task=PushCube-v1 chart=ms-64f98d163b6bff29332b candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
391
+ batch done sequential task=PushCube-v1
392
+ env init sequential task=PushCube-v1 batch=1 sim=physx_cpu render=cpu
393
+ execute sequential task=PushCube-v1 chart=ms-c2ae2b78ab7891e74c80 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
394
+ execute sequential task=PushCube-v1 chart=ms-c2ae2b78ab7891e74c80 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
395
+ execute sequential task=PushCube-v1 chart=ms-c2ae2b78ab7891e74c80 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
396
+ execute sequential task=PushCube-v1 chart=ms-c2ae2b78ab7891e74c80 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
397
+ execute sequential task=PushCube-v1 chart=ms-c2ae2b78ab7891e74c80 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
398
+ execute sequential task=PushCube-v1 chart=ms-c2ae2b78ab7891e74c80 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
399
+ execute sequential task=PushCube-v1 chart=ms-c2ae2b78ab7891e74c80 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
400
+ execute sequential task=PushCube-v1 chart=ms-c2ae2b78ab7891e74c80 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
401
+ execute sequential task=PushCube-v1 chart=ms-c2ae2b78ab7891e74c80 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
402
+ batch done sequential task=PushCube-v1
403
+ env init sequential task=PushCube-v1 batch=1 sim=physx_cpu render=cpu
404
+ execute sequential task=PushCube-v1 chart=ms-d4c37ed44fb338c1d551 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
405
+ execute sequential task=PushCube-v1 chart=ms-d4c37ed44fb338c1d551 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
406
+ execute sequential task=PushCube-v1 chart=ms-d4c37ed44fb338c1d551 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
407
+ execute sequential task=PushCube-v1 chart=ms-d4c37ed44fb338c1d551 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
408
+ execute sequential task=PushCube-v1 chart=ms-d4c37ed44fb338c1d551 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
409
+ execute sequential task=PushCube-v1 chart=ms-d4c37ed44fb338c1d551 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
410
+ execute sequential task=PushCube-v1 chart=ms-d4c37ed44fb338c1d551 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
411
+ execute sequential task=PushCube-v1 chart=ms-d4c37ed44fb338c1d551 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
412
+ execute sequential task=PushCube-v1 chart=ms-d4c37ed44fb338c1d551 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
413
+ batch done sequential task=PushCube-v1
414
+ env init sequential task=StackCube-v1 batch=1 sim=physx_cpu render=cpu
415
+ execute sequential task=StackCube-v1 chart=ms-146ff6c3f22b3a49af3a candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
416
+ execute sequential task=StackCube-v1 chart=ms-146ff6c3f22b3a49af3a candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
417
+ execute sequential task=StackCube-v1 chart=ms-146ff6c3f22b3a49af3a candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
418
+ execute sequential task=StackCube-v1 chart=ms-146ff6c3f22b3a49af3a candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
419
+ execute sequential task=StackCube-v1 chart=ms-146ff6c3f22b3a49af3a candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
420
+ execute sequential task=StackCube-v1 chart=ms-146ff6c3f22b3a49af3a candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
421
+ execute sequential task=StackCube-v1 chart=ms-146ff6c3f22b3a49af3a candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
422
+ execute sequential task=StackCube-v1 chart=ms-146ff6c3f22b3a49af3a candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
423
+ execute sequential task=StackCube-v1 chart=ms-146ff6c3f22b3a49af3a candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
424
+ batch done sequential task=StackCube-v1
425
+ env init sequential task=StackCube-v1 batch=1 sim=physx_cpu render=cpu
426
+ execute sequential task=StackCube-v1 chart=ms-15fc8cc6f74cf589056f candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
427
+ execute sequential task=StackCube-v1 chart=ms-15fc8cc6f74cf589056f candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
428
+ execute sequential task=StackCube-v1 chart=ms-15fc8cc6f74cf589056f candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
429
+ execute sequential task=StackCube-v1 chart=ms-15fc8cc6f74cf589056f candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
430
+ execute sequential task=StackCube-v1 chart=ms-15fc8cc6f74cf589056f candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
431
+ execute sequential task=StackCube-v1 chart=ms-15fc8cc6f74cf589056f candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
432
+ execute sequential task=StackCube-v1 chart=ms-15fc8cc6f74cf589056f candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
433
+ execute sequential task=StackCube-v1 chart=ms-15fc8cc6f74cf589056f candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
434
+ execute sequential task=StackCube-v1 chart=ms-15fc8cc6f74cf589056f candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
435
+ batch done sequential task=StackCube-v1
436
+ env init sequential task=StackCube-v1 batch=1 sim=physx_cpu render=cpu
437
+ execute sequential task=StackCube-v1 chart=ms-36bc1364d2439099e9ed candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
438
+ execute sequential task=StackCube-v1 chart=ms-36bc1364d2439099e9ed candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
439
+ execute sequential task=StackCube-v1 chart=ms-36bc1364d2439099e9ed candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
440
+ execute sequential task=StackCube-v1 chart=ms-36bc1364d2439099e9ed candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
441
+ execute sequential task=StackCube-v1 chart=ms-36bc1364d2439099e9ed candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
442
+ execute sequential task=StackCube-v1 chart=ms-36bc1364d2439099e9ed candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
443
+ execute sequential task=StackCube-v1 chart=ms-36bc1364d2439099e9ed candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
444
+ execute sequential task=StackCube-v1 chart=ms-36bc1364d2439099e9ed candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
445
+ execute sequential task=StackCube-v1 chart=ms-36bc1364d2439099e9ed candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
446
+ batch done sequential task=StackCube-v1
447
+ env init sequential task=StackCube-v1 batch=1 sim=physx_cpu render=cpu
448
+ execute sequential task=StackCube-v1 chart=ms-420280f15a522ce21c72 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
449
+ execute sequential task=StackCube-v1 chart=ms-420280f15a522ce21c72 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
450
+ execute sequential task=StackCube-v1 chart=ms-420280f15a522ce21c72 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
451
+ execute sequential task=StackCube-v1 chart=ms-420280f15a522ce21c72 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
452
+ execute sequential task=StackCube-v1 chart=ms-420280f15a522ce21c72 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
453
+ execute sequential task=StackCube-v1 chart=ms-420280f15a522ce21c72 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
454
+ execute sequential task=StackCube-v1 chart=ms-420280f15a522ce21c72 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
455
+ execute sequential task=StackCube-v1 chart=ms-420280f15a522ce21c72 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
456
+ execute sequential task=StackCube-v1 chart=ms-420280f15a522ce21c72 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
457
+ batch done sequential task=StackCube-v1
458
+ env init sequential task=StackCube-v1 batch=1 sim=physx_cpu render=cpu
459
+ execute sequential task=StackCube-v1 chart=ms-4aed1197d0906ae51c34 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
460
+ execute sequential task=StackCube-v1 chart=ms-4aed1197d0906ae51c34 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
461
+ execute sequential task=StackCube-v1 chart=ms-4aed1197d0906ae51c34 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
462
+ execute sequential task=StackCube-v1 chart=ms-4aed1197d0906ae51c34 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
463
+ execute sequential task=StackCube-v1 chart=ms-4aed1197d0906ae51c34 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
464
+ execute sequential task=StackCube-v1 chart=ms-4aed1197d0906ae51c34 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
465
+ execute sequential task=StackCube-v1 chart=ms-4aed1197d0906ae51c34 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
466
+ execute sequential task=StackCube-v1 chart=ms-4aed1197d0906ae51c34 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
467
+ execute sequential task=StackCube-v1 chart=ms-4aed1197d0906ae51c34 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
468
+ batch done sequential task=StackCube-v1
469
+ env init sequential task=StackCube-v1 batch=1 sim=physx_cpu render=cpu
470
+ execute sequential task=StackCube-v1 chart=ms-55e78c04643bd06fd5f8 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
471
+ execute sequential task=StackCube-v1 chart=ms-55e78c04643bd06fd5f8 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
472
+ execute sequential task=StackCube-v1 chart=ms-55e78c04643bd06fd5f8 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
473
+ execute sequential task=StackCube-v1 chart=ms-55e78c04643bd06fd5f8 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
474
+ execute sequential task=StackCube-v1 chart=ms-55e78c04643bd06fd5f8 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
475
+ execute sequential task=StackCube-v1 chart=ms-55e78c04643bd06fd5f8 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
476
+ execute sequential task=StackCube-v1 chart=ms-55e78c04643bd06fd5f8 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
477
+ execute sequential task=StackCube-v1 chart=ms-55e78c04643bd06fd5f8 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
478
+ execute sequential task=StackCube-v1 chart=ms-55e78c04643bd06fd5f8 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
479
+ batch done sequential task=StackCube-v1
480
+ env init sequential task=StackCube-v1 batch=1 sim=physx_cpu render=cpu
481
+ execute sequential task=StackCube-v1 chart=ms-77536801924a2a78688e candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
482
+ execute sequential task=StackCube-v1 chart=ms-77536801924a2a78688e candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
483
+ execute sequential task=StackCube-v1 chart=ms-77536801924a2a78688e candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
484
+ execute sequential task=StackCube-v1 chart=ms-77536801924a2a78688e candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
485
+ execute sequential task=StackCube-v1 chart=ms-77536801924a2a78688e candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
486
+ execute sequential task=StackCube-v1 chart=ms-77536801924a2a78688e candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
487
+ execute sequential task=StackCube-v1 chart=ms-77536801924a2a78688e candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
488
+ execute sequential task=StackCube-v1 chart=ms-77536801924a2a78688e candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
489
+ execute sequential task=StackCube-v1 chart=ms-77536801924a2a78688e candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
490
+ batch done sequential task=StackCube-v1
491
+ env init sequential task=StackCube-v1 batch=1 sim=physx_cpu render=cpu
492
+ execute sequential task=StackCube-v1 chart=ms-9dd1ff4c297a81b13071 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
493
+ execute sequential task=StackCube-v1 chart=ms-9dd1ff4c297a81b13071 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
494
+ execute sequential task=StackCube-v1 chart=ms-9dd1ff4c297a81b13071 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
495
+ execute sequential task=StackCube-v1 chart=ms-9dd1ff4c297a81b13071 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
496
+ execute sequential task=StackCube-v1 chart=ms-9dd1ff4c297a81b13071 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
497
+ execute sequential task=StackCube-v1 chart=ms-9dd1ff4c297a81b13071 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
498
+ execute sequential task=StackCube-v1 chart=ms-9dd1ff4c297a81b13071 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
499
+ execute sequential task=StackCube-v1 chart=ms-9dd1ff4c297a81b13071 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
500
+ execute sequential task=StackCube-v1 chart=ms-9dd1ff4c297a81b13071 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
501
+ batch done sequential task=StackCube-v1
502
+ env init sequential task=StackCube-v1 batch=1 sim=physx_cpu render=cpu
503
+ execute sequential task=StackCube-v1 chart=ms-d7254aa722729aa1a031 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
504
+ execute sequential task=StackCube-v1 chart=ms-d7254aa722729aa1a031 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
505
+ execute sequential task=StackCube-v1 chart=ms-d7254aa722729aa1a031 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
506
+ execute sequential task=StackCube-v1 chart=ms-d7254aa722729aa1a031 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
507
+ execute sequential task=StackCube-v1 chart=ms-d7254aa722729aa1a031 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
508
+ execute sequential task=StackCube-v1 chart=ms-d7254aa722729aa1a031 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
509
+ execute sequential task=StackCube-v1 chart=ms-d7254aa722729aa1a031 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
510
+ execute sequential task=StackCube-v1 chart=ms-d7254aa722729aa1a031 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
511
+ execute sequential task=StackCube-v1 chart=ms-d7254aa722729aa1a031 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
512
+ batch done sequential task=StackCube-v1
513
+ env init sequential task=StackCube-v1 batch=1 sim=physx_cpu render=cpu
514
+ execute sequential task=StackCube-v1 chart=ms-d776afa14e6322cadedd candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
515
+ execute sequential task=StackCube-v1 chart=ms-d776afa14e6322cadedd candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
516
+ execute sequential task=StackCube-v1 chart=ms-d776afa14e6322cadedd candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
517
+ execute sequential task=StackCube-v1 chart=ms-d776afa14e6322cadedd candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
518
+ execute sequential task=StackCube-v1 chart=ms-d776afa14e6322cadedd candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
519
+ execute sequential task=StackCube-v1 chart=ms-d776afa14e6322cadedd candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
520
+ execute sequential task=StackCube-v1 chart=ms-d776afa14e6322cadedd candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
521
+ execute sequential task=StackCube-v1 chart=ms-d776afa14e6322cadedd candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
522
+ execute sequential task=StackCube-v1 chart=ms-d776afa14e6322cadedd candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
523
+ batch done sequential task=StackCube-v1
524
+ env init sequential task=StackCube-v1 batch=1 sim=physx_cpu render=cpu
525
+ execute sequential task=StackCube-v1 chart=ms-e887dbfa0133af46829d candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
526
+ execute sequential task=StackCube-v1 chart=ms-e887dbfa0133af46829d candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
527
+ execute sequential task=StackCube-v1 chart=ms-e887dbfa0133af46829d candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
528
+ execute sequential task=StackCube-v1 chart=ms-e887dbfa0133af46829d candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
529
+ execute sequential task=StackCube-v1 chart=ms-e887dbfa0133af46829d candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
530
+ execute sequential task=StackCube-v1 chart=ms-e887dbfa0133af46829d candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
531
+ execute sequential task=StackCube-v1 chart=ms-e887dbfa0133af46829d candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
532
+ execute sequential task=StackCube-v1 chart=ms-e887dbfa0133af46829d candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
533
+ execute sequential task=StackCube-v1 chart=ms-e887dbfa0133af46829d candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
534
+ batch done sequential task=StackCube-v1
535
+ rollout complete rows=48
workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/git_hash.txt ADDED
@@ -0,0 +1 @@
 
 
1
+
workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/measured_candidates.json ADDED
The diff for this file is too large to render. See raw diff
 
workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/measured_metrics/metrics.json ADDED
The diff for this file is too large to render. See raw diff
 
workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/measured_metrics/metrics_by_seed.json ADDED
@@ -0,0 +1,1410 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "1073161778": {
3
+ "any_generated_unsafe_known_at_8": 0.0,
4
+ "base_safety_label_known": 1.0,
5
+ "base_success": 0.0,
6
+ "base_unsafe_known": 0.0,
7
+ "base_utility": 0.725517988204956,
8
+ "branch_car_at_8": 1.2767786383628845,
9
+ "generated_safety_label_coverage_at_8": 1.0,
10
+ "generated_unsafe_rate_known_at_8": 0.0,
11
+ "hidden_chart_oracle_success_at_8": 1.0,
12
+ "hidden_chart_oracle_utility_at_8": 2.0,
13
+ "ncar_to_hidden_chart_oracle_at_8": 1.0018019999721972,
14
+ "ncar_to_proposal_oracle_at_8": 1.0018019999721972,
15
+ "outcome_ptr_at_8": 1.0,
16
+ "pairwise_causal_calibration_ece": 0.14432341808925672,
17
+ "proposal_oracle_safety_label_known_at_8": 1.0,
18
+ "proposal_oracle_success_at_8": 1.0,
19
+ "proposal_oracle_success_gain_over_base_at_8": 1.0,
20
+ "proposal_oracle_unsafe_known_at_8": 0.0,
21
+ "proposal_oracle_utility_at_8": 2.0,
22
+ "proposal_oracle_utility_gain_over_base_at_8": 1.274482011795044,
23
+ "selected_safety_label_known_at_8": 1.0,
24
+ "selected_success_at_8": 0.0,
25
+ "selected_success_gain_over_base_at_8": 0.0,
26
+ "selected_unsafe_known_at_8": 0.0,
27
+ "selected_utility_at_8": 0.7232213616371155,
28
+ "selected_utility_gain_over_base_at_8": -0.002296626567840576,
29
+ "selector_gap_fraction_to_hidden_at_8": 1.0018020078326613,
30
+ "selector_regret_at_8": 1.2767786383628845,
31
+ "success_selector_gap_at_8": 1.0,
32
+ "success_support_gap_at_8": 0.0,
33
+ "success_total_car_to_hidden_at_8": 1.0,
34
+ "support_gap_at_8": 0.0,
35
+ "support_gap_fraction_to_hidden_at_8": 0.0,
36
+ "total_car_to_hidden_at_8": 1.2767786383628845
37
+ },
38
+ "1113371261": {
39
+ "any_generated_unsafe_known_at_8": 0.0,
40
+ "base_safety_label_known": 1.0,
41
+ "base_success": 0.0,
42
+ "base_unsafe_known": 0.0,
43
+ "base_utility": 0.11109928786754608,
44
+ "branch_car_at_8": 0.007540524005889893,
45
+ "generated_safety_label_coverage_at_8": 1.0,
46
+ "generated_unsafe_rate_known_at_8": 0.0,
47
+ "hidden_chart_oracle_success_at_8": 1.0,
48
+ "hidden_chart_oracle_utility_at_8": 2.0,
49
+ "ncar_to_hidden_chart_oracle_at_8": 1.0096749460109702,
50
+ "ncar_to_proposal_oracle_at_8": 0.7024563821726059,
51
+ "outcome_ptr_at_8": 0.0,
52
+ "pairwise_causal_calibration_ece": 0.18981168502744894,
53
+ "proposal_oracle_safety_label_known_at_8": 1.0,
54
+ "proposal_oracle_success_at_8": 0.0,
55
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
56
+ "proposal_oracle_unsafe_known_at_8": 0.0,
57
+ "proposal_oracle_utility_at_8": 0.1003647893667221,
58
+ "proposal_oracle_utility_gain_over_base_at_8": -0.010734498500823975,
59
+ "selected_safety_label_known_at_8": 1.0,
60
+ "selected_success_at_8": 0.0,
61
+ "selected_success_gain_over_base_at_8": 0.0,
62
+ "selected_unsafe_known_at_8": 0.0,
63
+ "selected_utility_at_8": 0.09282426536083221,
64
+ "selected_utility_gain_over_base_at_8": -0.018275022506713867,
65
+ "selector_gap_fraction_to_hidden_at_8": 0.003992017133275946,
66
+ "selector_regret_at_8": 0.007540524005889893,
67
+ "success_selector_gap_at_8": 0.0,
68
+ "success_support_gap_at_8": 1.0,
69
+ "success_total_car_to_hidden_at_8": 1.0,
70
+ "support_gap_at_8": 1.899635210633278,
71
+ "support_gap_fraction_to_hidden_at_8": 1.0056829342229987,
72
+ "total_car_to_hidden_at_8": 1.9071757346391678
73
+ },
74
+ "1151410346": {
75
+ "any_generated_unsafe_known_at_8": 0.0,
76
+ "base_safety_label_known": 1.0,
77
+ "base_success": 0.0,
78
+ "base_unsafe_known": 0.0,
79
+ "base_utility": 0.2284735143184662,
80
+ "branch_car_at_8": 0.3237016499042511,
81
+ "generated_safety_label_coverage_at_8": 1.0,
82
+ "generated_unsafe_rate_known_at_8": 0.0,
83
+ "hidden_chart_oracle_success_at_8": 1.0,
84
+ "hidden_chart_oracle_utility_at_8": 2.0,
85
+ "ncar_to_hidden_chart_oracle_at_8": 0.9938836701698183,
86
+ "ncar_to_proposal_oracle_at_8": 0.9676112241545515,
87
+ "outcome_ptr_at_8": 1.0,
88
+ "pairwise_causal_calibration_ece": 0.5181919511721376,
89
+ "proposal_oracle_safety_label_known_at_8": 1.0,
90
+ "proposal_oracle_success_at_8": 0.0,
91
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
92
+ "proposal_oracle_unsafe_known_at_8": 0.0,
93
+ "proposal_oracle_utility_at_8": 0.5630103945732117,
94
+ "proposal_oracle_utility_gain_over_base_at_8": 0.3345368802547455,
95
+ "selected_safety_label_known_at_8": 1.0,
96
+ "selected_success_at_8": 0.0,
97
+ "selected_success_gain_over_base_at_8": 0.0,
98
+ "selected_unsafe_known_at_8": 0.0,
99
+ "selected_utility_at_8": 0.23930874466896057,
100
+ "selected_utility_gain_over_base_at_8": 0.010835230350494385,
101
+ "selector_gap_fraction_to_hidden_at_8": 0.18272470240811445,
102
+ "selector_regret_at_8": 0.3237016499042511,
103
+ "success_selector_gap_at_8": 0.0,
104
+ "success_support_gap_at_8": 1.0,
105
+ "success_total_car_to_hidden_at_8": 1.0,
106
+ "support_gap_at_8": 1.4369896054267883,
107
+ "support_gap_fraction_to_hidden_at_8": 0.8111589733720272,
108
+ "total_car_to_hidden_at_8": 1.7606912553310394
109
+ },
110
+ "1261930456": {
111
+ "any_generated_unsafe_known_at_8": 0.0,
112
+ "base_safety_label_known": 1.0,
113
+ "base_success": 1.0,
114
+ "base_unsafe_known": 0.0,
115
+ "base_utility": 2.0,
116
+ "branch_car_at_8": 0.011454477906227112,
117
+ "generated_safety_label_coverage_at_8": 1.0,
118
+ "generated_unsafe_rate_known_at_8": 0.0,
119
+ "hidden_chart_oracle_success_at_8": 1.0,
120
+ "hidden_chart_oracle_utility_at_8": 2.0,
121
+ "ncar_to_proposal_oracle_at_8": 0.005982870275641443,
122
+ "outcome_ptr_at_8": 0.0,
123
+ "pairwise_causal_calibration_ece": 0.14694969164298693,
124
+ "proposal_oracle_safety_label_known_at_8": 1.0,
125
+ "proposal_oracle_success_at_8": 0.0,
126
+ "proposal_oracle_success_gain_over_base_at_8": -1.0,
127
+ "proposal_oracle_unsafe_known_at_8": 0.0,
128
+ "proposal_oracle_utility_at_8": 0.08545441925525665,
129
+ "proposal_oracle_utility_gain_over_base_at_8": -1.9145455807447433,
130
+ "selected_safety_label_known_at_8": 1.0,
131
+ "selected_success_at_8": 0.0,
132
+ "selected_success_gain_over_base_at_8": -1.0,
133
+ "selected_unsafe_known_at_8": 0.0,
134
+ "selected_utility_at_8": 0.07399994134902954,
135
+ "selected_utility_gain_over_base_at_8": -1.9260000586509705,
136
+ "selector_regret_at_8": 0.011454477906227112,
137
+ "success_selector_gap_at_8": 0.0,
138
+ "success_support_gap_at_8": 1.0,
139
+ "success_total_car_to_hidden_at_8": 1.0,
140
+ "support_gap_at_8": 1.9145455807447433,
141
+ "total_car_to_hidden_at_8": 1.9260000586509705
142
+ },
143
+ "1324133122": {
144
+ "any_generated_unsafe_known_at_8": 0.0,
145
+ "base_safety_label_known": 1.0,
146
+ "base_success": 0.0,
147
+ "base_unsafe_known": 0.0,
148
+ "base_utility": 0.084227055311203,
149
+ "branch_car_at_8": 0.010064825415611267,
150
+ "generated_safety_label_coverage_at_8": 1.0,
151
+ "generated_unsafe_rate_known_at_8": 0.0,
152
+ "hidden_chart_oracle_success_at_8": 1.0,
153
+ "hidden_chart_oracle_utility_at_8": 2.0,
154
+ "ncar_to_hidden_chart_oracle_at_8": 0.9979853617028601,
155
+ "ncar_to_proposal_oracle_at_8": 0.7228185479224526,
156
+ "outcome_ptr_at_8": 1.0,
157
+ "pairwise_causal_calibration_ece": 0.2599006738407752,
158
+ "proposal_oracle_safety_label_known_at_8": 1.0,
159
+ "proposal_oracle_success_at_8": 0.0,
160
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
161
+ "proposal_oracle_unsafe_known_at_8": 0.0,
162
+ "proposal_oracle_utility_at_8": 0.09815146028995514,
163
+ "proposal_oracle_utility_gain_over_base_at_8": 0.013924404978752136,
164
+ "selected_safety_label_known_at_8": 1.0,
165
+ "selected_success_at_8": 0.0,
166
+ "selected_success_gain_over_base_at_8": 0.0,
167
+ "selected_unsafe_known_at_8": 0.0,
168
+ "selected_utility_at_8": 0.08808663487434387,
169
+ "selected_utility_gain_over_base_at_8": 0.003859579563140869,
170
+ "selector_gap_fraction_to_hidden_at_8": 0.00525366298940307,
171
+ "selector_regret_at_8": 0.010064825415611267,
172
+ "success_selector_gap_at_8": 0.0,
173
+ "success_support_gap_at_8": 1.0,
174
+ "success_total_car_to_hidden_at_8": 1.0,
175
+ "support_gap_at_8": 1.9018485397100449,
176
+ "support_gap_fraction_to_hidden_at_8": 0.9927317039227662,
177
+ "total_car_to_hidden_at_8": 1.9119133651256561
178
+ },
179
+ "1324642411": {
180
+ "any_generated_unsafe_known_at_8": 0.0,
181
+ "base_safety_label_known": 1.0,
182
+ "base_success": 1.0,
183
+ "base_unsafe_known": 0.0,
184
+ "base_utility": 2.0,
185
+ "branch_car_at_8": 1.858040377497673,
186
+ "generated_safety_label_coverage_at_8": 1.0,
187
+ "generated_unsafe_rate_known_at_8": 0.0,
188
+ "hidden_chart_oracle_success_at_8": 1.0,
189
+ "hidden_chart_oracle_utility_at_8": 2.0,
190
+ "outcome_ptr_at_8": 0.0,
191
+ "pairwise_causal_calibration_ece": 0.37631581765020666,
192
+ "proposal_oracle_safety_label_known_at_8": 1.0,
193
+ "proposal_oracle_success_at_8": 1.0,
194
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
195
+ "proposal_oracle_unsafe_known_at_8": 0.0,
196
+ "proposal_oracle_utility_at_8": 2.0,
197
+ "proposal_oracle_utility_gain_over_base_at_8": 0.0,
198
+ "selected_safety_label_known_at_8": 1.0,
199
+ "selected_success_at_8": 0.0,
200
+ "selected_success_gain_over_base_at_8": -1.0,
201
+ "selected_unsafe_known_at_8": 0.0,
202
+ "selected_utility_at_8": 0.14195962250232697,
203
+ "selected_utility_gain_over_base_at_8": -1.858040377497673,
204
+ "selector_regret_at_8": 1.858040377497673,
205
+ "success_selector_gap_at_8": 1.0,
206
+ "success_support_gap_at_8": 0.0,
207
+ "success_total_car_to_hidden_at_8": 1.0,
208
+ "support_gap_at_8": 0.0,
209
+ "total_car_to_hidden_at_8": 1.858040377497673
210
+ },
211
+ "1375492786": {
212
+ "any_generated_unsafe_known_at_8": 1.0,
213
+ "base_safety_label_known": 1.0,
214
+ "base_success": 1.0,
215
+ "base_unsafe_known": 0.0,
216
+ "base_utility": 2.0,
217
+ "branch_car_at_8": 0.027821257710456848,
218
+ "generated_safety_label_coverage_at_8": 1.0,
219
+ "generated_unsafe_rate_known_at_8": 1.0,
220
+ "hidden_chart_oracle_success_at_8": 1.0,
221
+ "hidden_chart_oracle_utility_at_8": 2.0,
222
+ "ncar_to_proposal_oracle_at_8": 0.015409403984710279,
223
+ "outcome_ptr_at_8": 0.0,
224
+ "pairwise_causal_calibration_ece": 0.16288447414215274,
225
+ "proposal_oracle_safety_label_known_at_8": 1.0,
226
+ "proposal_oracle_success_at_8": 0.0,
227
+ "proposal_oracle_success_gain_over_base_at_8": -1.0,
228
+ "proposal_oracle_unsafe_known_at_8": 1.0,
229
+ "proposal_oracle_utility_at_8": 0.19452734291553497,
230
+ "proposal_oracle_utility_gain_over_base_at_8": -1.805472657084465,
231
+ "selected_safety_label_known_at_8": 1.0,
232
+ "selected_success_at_8": 0.0,
233
+ "selected_success_gain_over_base_at_8": -1.0,
234
+ "selected_unsafe_known_at_8": 1.0,
235
+ "selected_utility_at_8": 0.16670608520507812,
236
+ "selected_utility_gain_over_base_at_8": -1.8332939147949219,
237
+ "selector_regret_at_8": 0.027821257710456848,
238
+ "success_selector_gap_at_8": 0.0,
239
+ "success_support_gap_at_8": 1.0,
240
+ "success_total_car_to_hidden_at_8": 1.0,
241
+ "support_gap_at_8": 1.805472657084465,
242
+ "total_car_to_hidden_at_8": 1.8332939147949219
243
+ },
244
+ "141984615": {
245
+ "any_generated_unsafe_known_at_8": 1.0,
246
+ "base_safety_label_known": 1.0,
247
+ "base_success": 0.0,
248
+ "base_unsafe_known": 0.0,
249
+ "base_utility": 0.16783273220062256,
250
+ "branch_car_at_8": 0.002475142478942871,
251
+ "generated_safety_label_coverage_at_8": 1.0,
252
+ "generated_unsafe_rate_known_at_8": 1.0,
253
+ "hidden_chart_oracle_success_at_8": 0.0,
254
+ "hidden_chart_oracle_utility_at_8": 0.2857772707939148,
255
+ "ncar_to_hidden_chart_oracle_at_8": 2.0668632118652317,
256
+ "ncar_to_proposal_oracle_at_8": 0.020065103905333725,
257
+ "outcome_ptr_at_8": 0.0,
258
+ "pairwise_causal_calibration_ece": 0.15133945279036964,
259
+ "proposal_oracle_safety_label_known_at_8": 1.0,
260
+ "proposal_oracle_success_at_8": 0.0,
261
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
262
+ "proposal_oracle_unsafe_known_at_8": 1.0,
263
+ "proposal_oracle_utility_at_8": 0.04447716474533081,
264
+ "proposal_oracle_utility_gain_over_base_at_8": -0.12335556745529175,
265
+ "selected_safety_label_known_at_8": 1.0,
266
+ "selected_success_at_8": 0.0,
267
+ "selected_success_gain_over_base_at_8": 0.0,
268
+ "selected_unsafe_known_at_8": 1.0,
269
+ "selected_utility_at_8": 0.04200202226638794,
270
+ "selected_utility_gain_over_base_at_8": -0.12583070993423462,
271
+ "selector_gap_fraction_to_hidden_at_8": 0.02098564722422542,
272
+ "selector_regret_at_8": 0.002475142478942871,
273
+ "success_selector_gap_at_8": 0.0,
274
+ "success_support_gap_at_8": 0.0,
275
+ "success_total_car_to_hidden_at_8": 0.0,
276
+ "support_gap_at_8": 0.24130010604858398,
277
+ "support_gap_fraction_to_hidden_at_8": 2.0458777398812704,
278
+ "total_car_to_hidden_at_8": 0.24377524852752686
279
+ },
280
+ "1461069713": {
281
+ "any_generated_unsafe_known_at_8": 1.0,
282
+ "base_safety_label_known": 1.0,
283
+ "base_success": 0.0,
284
+ "base_unsafe_known": 0.0,
285
+ "base_utility": 0.22852759063243866,
286
+ "branch_car_at_8": 0.026291392743587494,
287
+ "generated_safety_label_coverage_at_8": 1.0,
288
+ "generated_unsafe_rate_known_at_8": 0.875,
289
+ "hidden_chart_oracle_success_at_8": 0.0,
290
+ "hidden_chart_oracle_utility_at_8": 0.1593424528837204,
291
+ "ncar_to_hidden_chart_oracle_at_8": 1.0147308339002874,
292
+ "ncar_to_proposal_oracle_at_8": 0.2324654665550531,
293
+ "outcome_ptr_at_8": 0.0,
294
+ "pairwise_causal_calibration_ece": 0.3576754247076027,
295
+ "proposal_oracle_safety_label_known_at_8": 1.0,
296
+ "proposal_oracle_success_at_8": 0.0,
297
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
298
+ "proposal_oracle_unsafe_known_at_8": 0.0,
299
+ "proposal_oracle_utility_at_8": 0.11542954295873642,
300
+ "proposal_oracle_utility_gain_over_base_at_8": -0.11309804767370224,
301
+ "selected_safety_label_known_at_8": 1.0,
302
+ "selected_success_at_8": 0.0,
303
+ "selected_success_gain_over_base_at_8": 0.0,
304
+ "selected_unsafe_known_at_8": 1.0,
305
+ "selected_utility_at_8": 0.08913815021514893,
306
+ "selected_utility_gain_over_base_at_8": -0.13938944041728973,
307
+ "selector_gap_fraction_to_hidden_at_8": 0.3800150378984332,
308
+ "selector_regret_at_8": 0.026291392743587494,
309
+ "success_selector_gap_at_8": 0.0,
310
+ "success_support_gap_at_8": 0.0,
311
+ "success_total_car_to_hidden_at_8": 0.0,
312
+ "support_gap_at_8": 0.04391290992498398,
313
+ "support_gap_fraction_to_hidden_at_8": 0.6347159426707584,
314
+ "total_car_to_hidden_at_8": 0.07020430266857147
315
+ },
316
+ "1480856810": {
317
+ "any_generated_unsafe_known_at_8": 0.0,
318
+ "base_safety_label_known": 1.0,
319
+ "base_success": 1.0,
320
+ "base_unsafe_known": 0.0,
321
+ "base_utility": 2.0,
322
+ "branch_car_at_8": 0.043351709842681885,
323
+ "generated_safety_label_coverage_at_8": 1.0,
324
+ "generated_unsafe_rate_known_at_8": 0.0,
325
+ "hidden_chart_oracle_success_at_8": 0.0,
326
+ "hidden_chart_oracle_utility_at_8": 0.20948036015033722,
327
+ "ncar_to_hidden_chart_oracle_at_8": 0.031613708638168724,
328
+ "ncar_to_proposal_oracle_at_8": 0.02403390675499265,
329
+ "outcome_ptr_at_8": 0.0,
330
+ "pairwise_causal_calibration_ece": 0.33394552923262694,
331
+ "proposal_oracle_safety_label_known_at_8": 1.0,
332
+ "proposal_oracle_success_at_8": 0.0,
333
+ "proposal_oracle_success_gain_over_base_at_8": -1.0,
334
+ "proposal_oracle_unsafe_known_at_8": 0.0,
335
+ "proposal_oracle_utility_at_8": 0.19622710347175598,
336
+ "proposal_oracle_utility_gain_over_base_at_8": -1.803772896528244,
337
+ "selected_safety_label_known_at_8": 1.0,
338
+ "selected_success_at_8": 0.0,
339
+ "selected_success_gain_over_base_at_8": -1.0,
340
+ "selected_unsafe_known_at_8": 0.0,
341
+ "selected_utility_at_8": 0.1528753936290741,
342
+ "selected_utility_gain_over_base_at_8": -1.847124606370926,
343
+ "selector_gap_fraction_to_hidden_at_8": 0.024211803589220513,
344
+ "selector_regret_at_8": 0.043351709842681885,
345
+ "success_selector_gap_at_8": 0.0,
346
+ "success_support_gap_at_8": 0.0,
347
+ "success_total_car_to_hidden_at_8": 0.0,
348
+ "support_gap_at_8": 0.013253256678581238,
349
+ "support_gap_fraction_to_hidden_at_8": 0.007401905225509852,
350
+ "total_car_to_hidden_at_8": 0.05660496652126312
351
+ },
352
+ "1509980664": {
353
+ "any_generated_unsafe_known_at_8": 0.0,
354
+ "base_safety_label_known": 1.0,
355
+ "base_success": 0.0,
356
+ "base_unsafe_known": 0.0,
357
+ "base_utility": 0.8320798277854919,
358
+ "branch_car_at_8": 0.014023125171661377,
359
+ "generated_safety_label_coverage_at_8": 1.0,
360
+ "generated_unsafe_rate_known_at_8": 0.0,
361
+ "hidden_chart_oracle_success_at_8": 1.0,
362
+ "hidden_chart_oracle_utility_at_8": 2.0,
363
+ "ncar_to_hidden_chart_oracle_at_8": 1.0061738317414306,
364
+ "ncar_to_proposal_oracle_at_8": 2.058415470840539,
365
+ "outcome_ptr_at_8": 1.0,
366
+ "pairwise_causal_calibration_ece": 0.14023834058751442,
367
+ "proposal_oracle_safety_label_known_at_8": 1.0,
368
+ "proposal_oracle_success_at_8": 0.0,
369
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
370
+ "proposal_oracle_unsafe_known_at_8": 0.0,
371
+ "proposal_oracle_utility_at_8": 0.83889240026474,
372
+ "proposal_oracle_utility_gain_over_base_at_8": 0.006812572479248047,
373
+ "selected_safety_label_known_at_8": 1.0,
374
+ "selected_success_at_8": 0.0,
375
+ "selected_success_gain_over_base_at_8": 0.0,
376
+ "selected_unsafe_known_at_8": 0.0,
377
+ "selected_utility_at_8": 0.8248692750930786,
378
+ "selected_utility_gain_over_base_at_8": -0.00721055269241333,
379
+ "selector_gap_fraction_to_hidden_at_8": 0.01200692093970083,
380
+ "selector_regret_at_8": 0.014023125171661377,
381
+ "success_selector_gap_at_8": 0.0,
382
+ "success_support_gap_at_8": 1.0,
383
+ "success_total_car_to_hidden_at_8": 1.0,
384
+ "support_gap_at_8": 1.16110759973526,
385
+ "support_gap_fraction_to_hidden_at_8": 0.9941669194168206,
386
+ "total_car_to_hidden_at_8": 1.1751307249069214
387
+ },
388
+ "1522272686": {
389
+ "any_generated_unsafe_known_at_8": 0.0,
390
+ "base_safety_label_known": 1.0,
391
+ "base_success": 1.0,
392
+ "base_unsafe_known": 0.0,
393
+ "base_utility": 2.0,
394
+ "branch_car_at_8": 1.4563714861869812,
395
+ "generated_safety_label_coverage_at_8": 1.0,
396
+ "generated_unsafe_rate_known_at_8": 0.0,
397
+ "hidden_chart_oracle_success_at_8": 1.0,
398
+ "hidden_chart_oracle_utility_at_8": 2.0,
399
+ "outcome_ptr_at_8": 0.0,
400
+ "pairwise_causal_calibration_ece": 0.12267671398942098,
401
+ "proposal_oracle_safety_label_known_at_8": 1.0,
402
+ "proposal_oracle_success_at_8": 1.0,
403
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
404
+ "proposal_oracle_unsafe_known_at_8": 0.0,
405
+ "proposal_oracle_utility_at_8": 2.0,
406
+ "proposal_oracle_utility_gain_over_base_at_8": 0.0,
407
+ "selected_safety_label_known_at_8": 1.0,
408
+ "selected_success_at_8": 0.0,
409
+ "selected_success_gain_over_base_at_8": -1.0,
410
+ "selected_unsafe_known_at_8": 0.0,
411
+ "selected_utility_at_8": 0.5436285138130188,
412
+ "selected_utility_gain_over_base_at_8": -1.4563714861869812,
413
+ "selector_regret_at_8": 1.4563714861869812,
414
+ "success_selector_gap_at_8": 1.0,
415
+ "success_support_gap_at_8": 0.0,
416
+ "success_total_car_to_hidden_at_8": 1.0,
417
+ "support_gap_at_8": 0.0,
418
+ "total_car_to_hidden_at_8": 1.4563714861869812
419
+ },
420
+ "1549324805": {
421
+ "any_generated_unsafe_known_at_8": 0.0,
422
+ "base_safety_label_known": 1.0,
423
+ "base_success": 0.0,
424
+ "base_unsafe_known": 0.0,
425
+ "base_utility": 0.17249532043933868,
426
+ "branch_car_at_8": 0.09635330736637115,
427
+ "generated_safety_label_coverage_at_8": 1.0,
428
+ "generated_unsafe_rate_known_at_8": 0.0,
429
+ "hidden_chart_oracle_success_at_8": 0.0,
430
+ "hidden_chart_oracle_utility_at_8": 0.2974560856819153,
431
+ "ncar_to_hidden_chart_oracle_at_8": 1.4386390413214543,
432
+ "ncar_to_proposal_oracle_at_8": 2.319495421176755,
433
+ "outcome_ptr_at_8": 1.0,
434
+ "pairwise_causal_calibration_ece": 0.3814015792041445,
435
+ "proposal_oracle_safety_label_known_at_8": 1.0,
436
+ "proposal_oracle_success_at_8": 0.0,
437
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
438
+ "proposal_oracle_unsafe_known_at_8": 0.0,
439
+ "proposal_oracle_utility_at_8": 0.21403594315052032,
440
+ "proposal_oracle_utility_gain_over_base_at_8": 0.04154062271118164,
441
+ "selected_safety_label_known_at_8": 1.0,
442
+ "selected_success_at_8": 0.0,
443
+ "selected_success_gain_over_base_at_8": 0.0,
444
+ "selected_unsafe_known_at_8": 0.0,
445
+ "selected_utility_at_8": 0.11768263578414917,
446
+ "selected_utility_gain_over_base_at_8": -0.054812684655189514,
447
+ "selector_gap_fraction_to_hidden_at_8": 0.7710684804092547,
448
+ "selector_regret_at_8": 0.09635330736637115,
449
+ "success_selector_gap_at_8": 0.0,
450
+ "success_support_gap_at_8": 0.0,
451
+ "success_total_car_to_hidden_at_8": 0.0,
452
+ "support_gap_at_8": 0.08342014253139496,
453
+ "support_gap_fraction_to_hidden_at_8": 0.6675706760394587,
454
+ "total_car_to_hidden_at_8": 0.1797734498977661
455
+ },
456
+ "1578990937": {
457
+ "any_generated_unsafe_known_at_8": 0.0,
458
+ "base_safety_label_known": 1.0,
459
+ "base_success": 0.0,
460
+ "base_unsafe_known": 0.0,
461
+ "base_utility": 0.21426597237586975,
462
+ "branch_car_at_8": 0.025389984250068665,
463
+ "generated_safety_label_coverage_at_8": 1.0,
464
+ "generated_unsafe_rate_known_at_8": 0.0,
465
+ "hidden_chart_oracle_success_at_8": 1.0,
466
+ "hidden_chart_oracle_utility_at_8": 2.0,
467
+ "ncar_to_hidden_chart_oracle_at_8": 1.0039557654834124,
468
+ "ncar_to_proposal_oracle_at_8": 1.3854594546391021,
469
+ "outcome_ptr_at_8": 1.0,
470
+ "pairwise_causal_calibration_ece": 0.25510759849883147,
471
+ "proposal_oracle_safety_label_known_at_8": 1.0,
472
+ "proposal_oracle_success_at_8": 0.0,
473
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
474
+ "proposal_oracle_unsafe_known_at_8": 0.0,
475
+ "proposal_oracle_utility_at_8": 0.23259200155735016,
476
+ "proposal_oracle_utility_gain_over_base_at_8": 0.018326029181480408,
477
+ "selected_safety_label_known_at_8": 1.0,
478
+ "selected_success_at_8": 0.0,
479
+ "selected_success_gain_over_base_at_8": 0.0,
480
+ "selected_unsafe_known_at_8": 0.0,
481
+ "selected_utility_at_8": 0.2072020173072815,
482
+ "selected_utility_gain_over_base_at_8": -0.007063955068588257,
483
+ "selector_gap_fraction_to_hidden_at_8": 0.014218233990786039,
484
+ "selector_regret_at_8": 0.025389984250068665,
485
+ "success_selector_gap_at_8": 0.0,
486
+ "success_support_gap_at_8": 1.0,
487
+ "success_total_car_to_hidden_at_8": 1.0,
488
+ "support_gap_at_8": 1.7674079984426498,
489
+ "support_gap_fraction_to_hidden_at_8": 0.9897375371147166,
490
+ "total_car_to_hidden_at_8": 1.7927979826927185
491
+ },
492
+ "1747178030": {
493
+ "any_generated_unsafe_known_at_8": 1.0,
494
+ "base_safety_label_known": 1.0,
495
+ "base_success": 1.0,
496
+ "base_unsafe_known": 0.0,
497
+ "base_utility": 2.0,
498
+ "branch_car_at_8": 0.0,
499
+ "generated_safety_label_coverage_at_8": 1.0,
500
+ "generated_unsafe_rate_known_at_8": 1.0,
501
+ "hidden_chart_oracle_success_at_8": 1.0,
502
+ "hidden_chart_oracle_utility_at_8": 2.0,
503
+ "outcome_ptr_at_8": 0.0,
504
+ "pairwise_causal_calibration_ece": 0.0,
505
+ "proposal_oracle_safety_label_known_at_8": 1.0,
506
+ "proposal_oracle_success_at_8": 1.0,
507
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
508
+ "proposal_oracle_unsafe_known_at_8": 1.0,
509
+ "proposal_oracle_utility_at_8": 2.0,
510
+ "proposal_oracle_utility_gain_over_base_at_8": 0.0,
511
+ "selected_safety_label_known_at_8": 1.0,
512
+ "selected_success_at_8": 1.0,
513
+ "selected_success_gain_over_base_at_8": 0.0,
514
+ "selected_unsafe_known_at_8": 1.0,
515
+ "selected_utility_at_8": 2.0,
516
+ "selected_utility_gain_over_base_at_8": 0.0,
517
+ "selector_regret_at_8": 0.0,
518
+ "success_selector_gap_at_8": 0.0,
519
+ "success_support_gap_at_8": 0.0,
520
+ "success_total_car_to_hidden_at_8": 0.0,
521
+ "support_gap_at_8": 0.0,
522
+ "total_car_to_hidden_at_8": 0.0
523
+ },
524
+ "1790188787": {
525
+ "any_generated_unsafe_known_at_8": 0.0,
526
+ "base_safety_label_known": 1.0,
527
+ "base_success": 1.0,
528
+ "base_unsafe_known": 0.0,
529
+ "base_utility": 2.0,
530
+ "branch_car_at_8": 0.0,
531
+ "generated_safety_label_coverage_at_8": 1.0,
532
+ "generated_unsafe_rate_known_at_8": 0.0,
533
+ "hidden_chart_oracle_success_at_8": 1.0,
534
+ "hidden_chart_oracle_utility_at_8": 2.0,
535
+ "outcome_ptr_at_8": 0.0,
536
+ "pairwise_causal_calibration_ece": 0.25625757964299634,
537
+ "proposal_oracle_safety_label_known_at_8": 1.0,
538
+ "proposal_oracle_success_at_8": 1.0,
539
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
540
+ "proposal_oracle_unsafe_known_at_8": 0.0,
541
+ "proposal_oracle_utility_at_8": 2.0,
542
+ "proposal_oracle_utility_gain_over_base_at_8": 0.0,
543
+ "selected_safety_label_known_at_8": 1.0,
544
+ "selected_success_at_8": 1.0,
545
+ "selected_success_gain_over_base_at_8": 0.0,
546
+ "selected_unsafe_known_at_8": 0.0,
547
+ "selected_utility_at_8": 2.0,
548
+ "selected_utility_gain_over_base_at_8": 0.0,
549
+ "selector_regret_at_8": 0.0,
550
+ "success_selector_gap_at_8": 0.0,
551
+ "success_support_gap_at_8": 0.0,
552
+ "success_total_car_to_hidden_at_8": 0.0,
553
+ "support_gap_at_8": 0.0,
554
+ "total_car_to_hidden_at_8": 0.0
555
+ },
556
+ "1822210471": {
557
+ "any_generated_unsafe_known_at_8": 0.0,
558
+ "base_safety_label_known": 1.0,
559
+ "base_success": 0.0,
560
+ "base_unsafe_known": 0.0,
561
+ "base_utility": 0.09442546218633652,
562
+ "branch_car_at_8": 0.0011331625282764435,
563
+ "generated_safety_label_coverage_at_8": 1.0,
564
+ "generated_unsafe_rate_known_at_8": 0.0,
565
+ "hidden_chart_oracle_success_at_8": 1.0,
566
+ "hidden_chart_oracle_utility_at_8": 2.0,
567
+ "ncar_to_hidden_chart_oracle_at_8": 1.019279876796129,
568
+ "ncar_to_proposal_oracle_at_8": 0.031824954963663915,
569
+ "outcome_ptr_at_8": 0.0,
570
+ "pairwise_causal_calibration_ece": 0.15125404295167477,
571
+ "proposal_oracle_safety_label_known_at_8": 1.0,
572
+ "proposal_oracle_success_at_8": 0.0,
573
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
574
+ "proposal_oracle_unsafe_known_at_8": 0.0,
575
+ "proposal_oracle_utility_at_8": 0.058819372206926346,
576
+ "proposal_oracle_utility_gain_over_base_at_8": -0.03560608997941017,
577
+ "selected_safety_label_known_at_8": 1.0,
578
+ "selected_success_at_8": 0.0,
579
+ "selected_success_gain_over_base_at_8": 0.0,
580
+ "selected_unsafe_known_at_8": 0.0,
581
+ "selected_utility_at_8": 0.0576862096786499,
582
+ "selected_utility_gain_over_base_at_8": -0.036739252507686615,
583
+ "selector_gap_fraction_to_hidden_at_8": 0.0005946566275893689,
584
+ "selector_regret_at_8": 0.0011331625282764435,
585
+ "success_selector_gap_at_8": 0.0,
586
+ "success_support_gap_at_8": 1.0,
587
+ "success_total_car_to_hidden_at_8": 1.0,
588
+ "support_gap_at_8": 1.9411806277930737,
589
+ "support_gap_fraction_to_hidden_at_8": 1.018685225517477,
590
+ "total_car_to_hidden_at_8": 1.94231379032135
591
+ },
592
+ "1845489013": {
593
+ "any_generated_unsafe_known_at_8": 1.0,
594
+ "base_safety_label_known": 1.0,
595
+ "base_success": 0.0,
596
+ "base_unsafe_known": 0.0,
597
+ "base_utility": 0.11255800724029541,
598
+ "branch_car_at_8": 0.009949073195457458,
599
+ "generated_safety_label_coverage_at_8": 1.0,
600
+ "generated_unsafe_rate_known_at_8": 0.75,
601
+ "hidden_chart_oracle_success_at_8": 1.0,
602
+ "hidden_chart_oracle_utility_at_8": 2.0,
603
+ "ncar_to_hidden_chart_oracle_at_8": 1.019238671785175,
604
+ "ncar_to_proposal_oracle_at_8": 0.37739026714244994,
605
+ "outcome_ptr_at_8": 0.0,
606
+ "pairwise_causal_calibration_ece": 0.1601652398891617,
607
+ "proposal_oracle_safety_label_known_at_8": 1.0,
608
+ "proposal_oracle_success_at_8": 0.0,
609
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
610
+ "proposal_oracle_unsafe_known_at_8": 1.0,
611
+ "proposal_oracle_utility_at_8": 0.0861951932311058,
612
+ "proposal_oracle_utility_gain_over_base_at_8": -0.026362814009189606,
613
+ "selected_safety_label_known_at_8": 1.0,
614
+ "selected_success_at_8": 0.0,
615
+ "selected_success_gain_over_base_at_8": 0.0,
616
+ "selected_unsafe_known_at_8": 1.0,
617
+ "selected_utility_at_8": 0.07624612003564835,
618
+ "selected_utility_gain_over_base_at_8": -0.036311887204647064,
619
+ "selector_gap_fraction_to_hidden_at_8": 0.0052711941525209575,
620
+ "selector_regret_at_8": 0.009949073195457458,
621
+ "success_selector_gap_at_8": 0.0,
622
+ "success_support_gap_at_8": 1.0,
623
+ "success_total_car_to_hidden_at_8": 1.0,
624
+ "support_gap_at_8": 1.9138048067688942,
625
+ "support_gap_fraction_to_hidden_at_8": 1.0139674830327599,
626
+ "total_car_to_hidden_at_8": 1.9237538799643517
627
+ },
628
+ "1999022574": {
629
+ "any_generated_unsafe_known_at_8": 0.0,
630
+ "base_safety_label_known": 1.0,
631
+ "base_success": 0.0,
632
+ "base_unsafe_known": 0.0,
633
+ "base_utility": 0.8279688954353333,
634
+ "branch_car_at_8": 0.6338560879230499,
635
+ "generated_safety_label_coverage_at_8": 1.0,
636
+ "generated_unsafe_rate_known_at_8": 0.0,
637
+ "hidden_chart_oracle_success_at_8": 1.0,
638
+ "hidden_chart_oracle_utility_at_8": 2.0,
639
+ "ncar_to_hidden_chart_oracle_at_8": 1.5672404802573083,
640
+ "ncar_to_proposal_oracle_at_8": 20.468479398320074,
641
+ "outcome_ptr_at_8": 0.0,
642
+ "pairwise_causal_calibration_ece": 0.15346652783738673,
643
+ "proposal_oracle_safety_label_known_at_8": 1.0,
644
+ "proposal_oracle_success_at_8": 0.0,
645
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
646
+ "proposal_oracle_unsafe_known_at_8": 0.0,
647
+ "proposal_oracle_utility_at_8": 0.7970014810562134,
648
+ "proposal_oracle_utility_gain_over_base_at_8": -0.030967414379119873,
649
+ "selected_safety_label_known_at_8": 1.0,
650
+ "selected_success_at_8": 0.0,
651
+ "selected_success_gain_over_base_at_8": 0.0,
652
+ "selected_unsafe_known_at_8": 0.0,
653
+ "selected_utility_at_8": 0.16314539313316345,
654
+ "selected_utility_gain_over_base_at_8": -0.6648235023021698,
655
+ "selector_gap_fraction_to_hidden_at_8": 0.5408184863476693,
656
+ "selector_regret_at_8": 0.6338560879230499,
657
+ "success_selector_gap_at_8": 0.0,
658
+ "success_support_gap_at_8": 1.0,
659
+ "success_total_car_to_hidden_at_8": 1.0,
660
+ "support_gap_at_8": 1.2029985189437866,
661
+ "support_gap_fraction_to_hidden_at_8": 1.0264220072816432,
662
+ "total_car_to_hidden_at_8": 1.8368546068668365
663
+ },
664
+ "2012211287": {
665
+ "any_generated_unsafe_known_at_8": 0.0,
666
+ "base_safety_label_known": 1.0,
667
+ "base_success": 0.0,
668
+ "base_unsafe_known": 0.0,
669
+ "base_utility": 0.7282466888427734,
670
+ "branch_car_at_8": 0.00025147199630737305,
671
+ "generated_safety_label_coverage_at_8": 1.0,
672
+ "generated_unsafe_rate_known_at_8": 0.0,
673
+ "hidden_chart_oracle_success_at_8": 1.0,
674
+ "hidden_chart_oracle_utility_at_8": 2.0,
675
+ "ncar_to_hidden_chart_oracle_at_8": 1.0011445108113346,
676
+ "ncar_to_proposal_oracle_at_8": 0.20884931245481586,
677
+ "outcome_ptr_at_8": 0.0,
678
+ "pairwise_causal_calibration_ece": 0.19320969574917243,
679
+ "proposal_oracle_safety_label_known_at_8": 1.0,
680
+ "proposal_oracle_success_at_8": 0.0,
681
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
682
+ "proposal_oracle_unsafe_known_at_8": 0.0,
683
+ "proposal_oracle_utility_at_8": 0.7270426154136658,
684
+ "proposal_oracle_utility_gain_over_base_at_8": -0.001204073429107666,
685
+ "selected_safety_label_known_at_8": 1.0,
686
+ "selected_success_at_8": 0.0,
687
+ "selected_success_gain_over_base_at_8": 0.0,
688
+ "selected_unsafe_known_at_8": 0.0,
689
+ "selected_utility_at_8": 0.7267911434173584,
690
+ "selected_utility_gain_over_base_at_8": -0.001455545425415039,
691
+ "selector_gap_fraction_to_hidden_at_8": 0.00019773645887231634,
692
+ "selector_regret_at_8": 0.00025147199630737305,
693
+ "success_selector_gap_at_8": 0.0,
694
+ "success_support_gap_at_8": 1.0,
695
+ "success_total_car_to_hidden_at_8": 1.0,
696
+ "support_gap_at_8": 1.2729573845863342,
697
+ "support_gap_fraction_to_hidden_at_8": 1.0009467822246219,
698
+ "total_car_to_hidden_at_8": 1.2732088565826416
699
+ },
700
+ "2038594191": {
701
+ "any_generated_unsafe_known_at_8": 1.0,
702
+ "base_safety_label_known": 1.0,
703
+ "base_success": 0.0,
704
+ "base_unsafe_known": 0.0,
705
+ "base_utility": 0.09504059702157974,
706
+ "branch_car_at_8": 0.1589493229985237,
707
+ "generated_safety_label_coverage_at_8": 1.0,
708
+ "generated_unsafe_rate_known_at_8": 0.375,
709
+ "hidden_chart_oracle_success_at_8": 1.0,
710
+ "hidden_chart_oracle_utility_at_8": 2.0,
711
+ "ncar_to_hidden_chart_oracle_at_8": 1.002720058258962,
712
+ "ncar_to_proposal_oracle_at_8": 1.0336975829706272,
713
+ "outcome_ptr_at_8": 1.0,
714
+ "pairwise_causal_calibration_ece": 0.5070876425629134,
715
+ "proposal_oracle_safety_label_known_at_8": 1.0,
716
+ "proposal_oracle_success_at_8": 0.0,
717
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
718
+ "proposal_oracle_unsafe_known_at_8": 1.0,
719
+ "proposal_oracle_utility_at_8": 0.24880830943584442,
720
+ "proposal_oracle_utility_gain_over_base_at_8": 0.15376771241426468,
721
+ "selected_safety_label_known_at_8": 1.0,
722
+ "selected_success_at_8": 0.0,
723
+ "selected_success_gain_over_base_at_8": 0.0,
724
+ "selected_unsafe_known_at_8": 0.0,
725
+ "selected_utility_at_8": 0.08985898643732071,
726
+ "selected_utility_gain_over_base_at_8": -0.005181610584259033,
727
+ "selector_gap_fraction_to_hidden_at_8": 0.08343974299400034,
728
+ "selector_regret_at_8": 0.1589493229985237,
729
+ "success_selector_gap_at_8": 0.0,
730
+ "success_support_gap_at_8": 1.0,
731
+ "success_total_car_to_hidden_at_8": 1.0,
732
+ "support_gap_at_8": 1.7511916905641556,
733
+ "support_gap_fraction_to_hidden_at_8": 0.919280320528696,
734
+ "total_car_to_hidden_at_8": 1.9101410135626793
735
+ },
736
+ "2054899126": {
737
+ "any_generated_unsafe_known_at_8": 1.0,
738
+ "base_safety_label_known": 1.0,
739
+ "base_success": 1.0,
740
+ "base_unsafe_known": 0.0,
741
+ "base_utility": 2.0,
742
+ "branch_car_at_8": 0.005500897765159607,
743
+ "generated_safety_label_coverage_at_8": 1.0,
744
+ "generated_unsafe_rate_known_at_8": 0.9375,
745
+ "hidden_chart_oracle_success_at_8": 1.0,
746
+ "hidden_chart_oracle_utility_at_8": 2.0,
747
+ "ncar_to_proposal_oracle_at_8": 0.005904623170763854,
748
+ "outcome_ptr_at_8": 0.0,
749
+ "pairwise_causal_calibration_ece": 0.09662639932716256,
750
+ "proposal_oracle_safety_label_known_at_8": 1.0,
751
+ "proposal_oracle_success_at_8": 0.5,
752
+ "proposal_oracle_success_gain_over_base_at_8": -0.5,
753
+ "proposal_oracle_unsafe_known_at_8": 1.0,
754
+ "proposal_oracle_utility_at_8": 1.0683744624257088,
755
+ "proposal_oracle_utility_gain_over_base_at_8": -0.9316255375742912,
756
+ "selected_safety_label_known_at_8": 1.0,
757
+ "selected_success_at_8": 0.5,
758
+ "selected_success_gain_over_base_at_8": -0.5,
759
+ "selected_unsafe_known_at_8": 1.0,
760
+ "selected_utility_at_8": 1.0628735646605492,
761
+ "selected_utility_gain_over_base_at_8": -0.9371264353394508,
762
+ "selector_regret_at_8": 0.005500897765159607,
763
+ "success_selector_gap_at_8": 0.0,
764
+ "success_support_gap_at_8": 0.5,
765
+ "success_total_car_to_hidden_at_8": 0.5,
766
+ "support_gap_at_8": 0.9316255375742912,
767
+ "total_car_to_hidden_at_8": 0.9371264353394508
768
+ },
769
+ "2061635888": {
770
+ "any_generated_unsafe_known_at_8": 0.0,
771
+ "base_safety_label_known": 1.0,
772
+ "base_success": 0.0,
773
+ "base_unsafe_known": 0.0,
774
+ "base_utility": 0.08620692789554596,
775
+ "branch_car_at_8": 0.0,
776
+ "generated_safety_label_coverage_at_8": 1.0,
777
+ "generated_unsafe_rate_known_at_8": 0.0,
778
+ "hidden_chart_oracle_success_at_8": 1.0,
779
+ "hidden_chart_oracle_utility_at_8": 2.0,
780
+ "ncar_to_hidden_chart_oracle_at_8": 1.0037531653219327,
781
+ "ncar_to_proposal_oracle_at_8": 0.0,
782
+ "outcome_ptr_at_8": 0.0,
783
+ "pairwise_causal_calibration_ece": 0.1024944972766888,
784
+ "proposal_oracle_safety_label_known_at_8": 1.0,
785
+ "proposal_oracle_success_at_8": 0.0,
786
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
787
+ "proposal_oracle_unsafe_known_at_8": 0.0,
788
+ "proposal_oracle_utility_at_8": 0.07902413606643677,
789
+ "proposal_oracle_utility_gain_over_base_at_8": -0.007182791829109192,
790
+ "selected_safety_label_known_at_8": 1.0,
791
+ "selected_success_at_8": 0.0,
792
+ "selected_success_gain_over_base_at_8": 0.0,
793
+ "selected_unsafe_known_at_8": 0.0,
794
+ "selected_utility_at_8": 0.07902413606643677,
795
+ "selected_utility_gain_over_base_at_8": -0.007182791829109192,
796
+ "selector_gap_fraction_to_hidden_at_8": 0.0,
797
+ "selector_regret_at_8": 0.0,
798
+ "success_selector_gap_at_8": 0.0,
799
+ "success_support_gap_at_8": 1.0,
800
+ "success_total_car_to_hidden_at_8": 1.0,
801
+ "support_gap_at_8": 1.9209758639335632,
802
+ "support_gap_fraction_to_hidden_at_8": 1.0037531705667693,
803
+ "total_car_to_hidden_at_8": 1.9209758639335632
804
+ },
805
+ "211245842": {
806
+ "any_generated_unsafe_known_at_8": 0.0,
807
+ "base_safety_label_known": 1.0,
808
+ "base_success": 0.0,
809
+ "base_unsafe_known": 0.0,
810
+ "base_utility": 0.18430444598197937,
811
+ "branch_car_at_8": 0.041060686111450195,
812
+ "generated_safety_label_coverage_at_8": 1.0,
813
+ "generated_unsafe_rate_known_at_8": 0.0,
814
+ "hidden_chart_oracle_success_at_8": 1.0,
815
+ "hidden_chart_oracle_utility_at_8": 2.0,
816
+ "ncar_to_hidden_chart_oracle_at_8": 1.0061024029871146,
817
+ "ncar_to_proposal_oracle_at_8": 1.3695761057579783,
818
+ "outcome_ptr_at_8": 1.0,
819
+ "pairwise_causal_calibration_ece": 0.11952033712097218,
820
+ "proposal_oracle_safety_label_known_at_8": 1.0,
821
+ "proposal_oracle_success_at_8": 0.0,
822
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
823
+ "proposal_oracle_unsafe_known_at_8": 0.0,
824
+ "proposal_oracle_utility_at_8": 0.2142850160598755,
825
+ "proposal_oracle_utility_gain_over_base_at_8": 0.029980570077896118,
826
+ "selected_safety_label_known_at_8": 1.0,
827
+ "selected_success_at_8": 0.0,
828
+ "selected_success_gain_over_base_at_8": 0.0,
829
+ "selected_unsafe_known_at_8": 0.0,
830
+ "selected_utility_at_8": 0.1732243299484253,
831
+ "selected_utility_gain_over_base_at_8": -0.011080116033554077,
832
+ "selector_gap_fraction_to_hidden_at_8": 0.022614301180936124,
833
+ "selector_regret_at_8": 0.041060686111450195,
834
+ "success_selector_gap_at_8": 0.0,
835
+ "success_support_gap_at_8": 1.0,
836
+ "success_total_car_to_hidden_at_8": 1.0,
837
+ "support_gap_at_8": 1.7857149839401245,
838
+ "support_gap_fraction_to_hidden_at_8": 0.9834881073473187,
839
+ "total_car_to_hidden_at_8": 1.8267756700515747
840
+ },
841
+ "278008524": {
842
+ "any_generated_unsafe_known_at_8": 0.5,
843
+ "base_safety_label_known": 1.0,
844
+ "base_success": 0.5,
845
+ "base_unsafe_known": 0.0,
846
+ "base_utility": 1.1250072121620178,
847
+ "branch_car_at_8": 0.043415673077106476,
848
+ "generated_safety_label_coverage_at_8": 1.0,
849
+ "generated_unsafe_rate_known_at_8": 0.1875,
850
+ "hidden_chart_oracle_success_at_8": 0.5,
851
+ "hidden_chart_oracle_utility_at_8": 1.1139086410403252,
852
+ "ncar_to_hidden_chart_oracle_at_8": 0.026316678586339484,
853
+ "ncar_to_proposal_oracle_at_8": 1.3556794180383116,
854
+ "outcome_ptr_at_8": 0.5,
855
+ "pairwise_causal_calibration_ece": 0.05319894873106218,
856
+ "proposal_oracle_safety_label_known_at_8": 1.0,
857
+ "proposal_oracle_success_at_8": 0.5,
858
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
859
+ "proposal_oracle_unsafe_known_at_8": 0.5,
860
+ "proposal_oracle_utility_at_8": 1.157032236456871,
861
+ "proposal_oracle_utility_gain_over_base_at_8": 0.03202502429485321,
862
+ "selected_safety_label_known_at_8": 1.0,
863
+ "selected_success_at_8": 0.5,
864
+ "selected_success_gain_over_base_at_8": 0.0,
865
+ "selected_unsafe_known_at_8": 0.0,
866
+ "selected_utility_at_8": 1.1136165633797646,
867
+ "selected_utility_gain_over_base_at_8": -0.011390648782253265,
868
+ "selector_gap_fraction_to_hidden_at_8": 3.911825459427708,
869
+ "selector_regret_at_8": 0.043415673077106476,
870
+ "success_selector_gap_at_8": 0.0,
871
+ "success_support_gap_at_8": 0.0,
872
+ "success_total_car_to_hidden_at_8": 0.0,
873
+ "support_gap_at_8": 0.0,
874
+ "support_gap_fraction_to_hidden_at_8": 0.0,
875
+ "total_car_to_hidden_at_8": 0.0002920776605606079
876
+ },
877
+ "307068879": {
878
+ "any_generated_unsafe_known_at_8": 0.0,
879
+ "base_safety_label_known": 1.0,
880
+ "base_success": 0.0,
881
+ "base_unsafe_known": 0.0,
882
+ "base_utility": 0.20915235579013824,
883
+ "branch_car_at_8": 0.0018310099840164185,
884
+ "generated_safety_label_coverage_at_8": 1.0,
885
+ "generated_unsafe_rate_known_at_8": 0.0,
886
+ "hidden_chart_oracle_success_at_8": 1.0,
887
+ "hidden_chart_oracle_utility_at_8": 2.0,
888
+ "ncar_to_hidden_chart_oracle_at_8": 1.004161799291129,
889
+ "ncar_to_proposal_oracle_at_8": 0.32567740299816206,
890
+ "outcome_ptr_at_8": 0.0,
891
+ "pairwise_causal_calibration_ece": 0.07483884431566282,
892
+ "proposal_oracle_safety_label_known_at_8": 1.0,
893
+ "proposal_oracle_success_at_8": 0.0,
894
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
895
+ "proposal_oracle_unsafe_known_at_8": 0.0,
896
+ "proposal_oracle_utility_at_8": 0.2035302072763443,
897
+ "proposal_oracle_utility_gain_over_base_at_8": -0.005622148513793945,
898
+ "selected_safety_label_known_at_8": 1.0,
899
+ "selected_success_at_8": 0.0,
900
+ "selected_success_gain_over_base_at_8": 0.0,
901
+ "selected_unsafe_known_at_8": 0.0,
902
+ "selected_utility_at_8": 0.20169919729232788,
903
+ "selected_utility_gain_over_base_at_8": -0.007453158497810364,
904
+ "selector_gap_fraction_to_hidden_at_8": 0.0010224264414320273,
905
+ "selector_regret_at_8": 0.0018310099840164185,
906
+ "success_selector_gap_at_8": 0.0,
907
+ "success_support_gap_at_8": 1.0,
908
+ "success_total_car_to_hidden_at_8": 1.0,
909
+ "support_gap_at_8": 1.7964697927236557,
910
+ "support_gap_fraction_to_hidden_at_8": 1.003139378456884,
911
+ "total_car_to_hidden_at_8": 1.7983008027076721
912
+ },
913
+ "412102563": {
914
+ "any_generated_unsafe_known_at_8": 0.0,
915
+ "base_safety_label_known": 1.0,
916
+ "base_success": 1.0,
917
+ "base_unsafe_known": 0.0,
918
+ "base_utility": 2.0,
919
+ "branch_car_at_8": 0.6269922256469727,
920
+ "generated_safety_label_coverage_at_8": 1.0,
921
+ "generated_unsafe_rate_known_at_8": 0.0,
922
+ "hidden_chart_oracle_success_at_8": 1.0,
923
+ "hidden_chart_oracle_utility_at_8": 2.0,
924
+ "ncar_to_proposal_oracle_at_8": 0.5347509829497624,
925
+ "outcome_ptr_at_8": 0.0,
926
+ "pairwise_causal_calibration_ece": 0.2630363607222759,
927
+ "proposal_oracle_safety_label_known_at_8": 1.0,
928
+ "proposal_oracle_success_at_8": 0.0,
929
+ "proposal_oracle_success_gain_over_base_at_8": -1.0,
930
+ "proposal_oracle_unsafe_known_at_8": 0.0,
931
+ "proposal_oracle_utility_at_8": 0.8275061845779419,
932
+ "proposal_oracle_utility_gain_over_base_at_8": -1.172493815422058,
933
+ "selected_safety_label_known_at_8": 1.0,
934
+ "selected_success_at_8": 0.0,
935
+ "selected_success_gain_over_base_at_8": -1.0,
936
+ "selected_unsafe_known_at_8": 0.0,
937
+ "selected_utility_at_8": 0.20051395893096924,
938
+ "selected_utility_gain_over_base_at_8": -1.7994860410690308,
939
+ "selector_regret_at_8": 0.6269922256469727,
940
+ "success_selector_gap_at_8": 0.0,
941
+ "success_support_gap_at_8": 1.0,
942
+ "success_total_car_to_hidden_at_8": 1.0,
943
+ "support_gap_at_8": 1.172493815422058,
944
+ "total_car_to_hidden_at_8": 1.7994860410690308
945
+ },
946
+ "440577711": {
947
+ "any_generated_unsafe_known_at_8": 0.0,
948
+ "base_safety_label_known": 1.0,
949
+ "base_success": 1.0,
950
+ "base_unsafe_known": 0.0,
951
+ "base_utility": 2.0,
952
+ "branch_car_at_8": 0.0,
953
+ "generated_safety_label_coverage_at_8": 1.0,
954
+ "generated_unsafe_rate_known_at_8": 0.0,
955
+ "hidden_chart_oracle_success_at_8": 1.0,
956
+ "hidden_chart_oracle_utility_at_8": 2.0,
957
+ "outcome_ptr_at_8": 0.0,
958
+ "pairwise_causal_calibration_ece": 0.0,
959
+ "proposal_oracle_safety_label_known_at_8": 1.0,
960
+ "proposal_oracle_success_at_8": 1.0,
961
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
962
+ "proposal_oracle_unsafe_known_at_8": 0.0,
963
+ "proposal_oracle_utility_at_8": 2.0,
964
+ "proposal_oracle_utility_gain_over_base_at_8": 0.0,
965
+ "selected_safety_label_known_at_8": 1.0,
966
+ "selected_success_at_8": 1.0,
967
+ "selected_success_gain_over_base_at_8": 0.0,
968
+ "selected_unsafe_known_at_8": 0.0,
969
+ "selected_utility_at_8": 2.0,
970
+ "selected_utility_gain_over_base_at_8": 0.0,
971
+ "selector_regret_at_8": 0.0,
972
+ "success_selector_gap_at_8": 0.0,
973
+ "success_support_gap_at_8": 0.0,
974
+ "success_total_car_to_hidden_at_8": 0.0,
975
+ "support_gap_at_8": 0.0,
976
+ "total_car_to_hidden_at_8": 0.0
977
+ },
978
+ "441046872": {
979
+ "any_generated_unsafe_known_at_8": 0.0,
980
+ "base_safety_label_known": 1.0,
981
+ "base_success": 0.0,
982
+ "base_unsafe_known": 0.0,
983
+ "base_utility": 0.2097676396369934,
984
+ "branch_car_at_8": 0.000804021954536438,
985
+ "generated_safety_label_coverage_at_8": 1.0,
986
+ "generated_unsafe_rate_known_at_8": 0.0,
987
+ "hidden_chart_oracle_success_at_8": 1.0,
988
+ "hidden_chart_oracle_utility_at_8": 2.0,
989
+ "ncar_to_hidden_chart_oracle_at_8": 1.003523045662944,
990
+ "ncar_to_proposal_oracle_at_8": 0.14610429876487893,
991
+ "outcome_ptr_at_8": 0.0,
992
+ "pairwise_causal_calibration_ece": 0.1123926329678895,
993
+ "proposal_oracle_safety_label_known_at_8": 1.0,
994
+ "proposal_oracle_success_at_8": 0.0,
995
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
996
+ "proposal_oracle_unsafe_known_at_8": 0.0,
997
+ "proposal_oracle_utility_at_8": 0.2042645812034607,
998
+ "proposal_oracle_utility_gain_over_base_at_8": -0.005503058433532715,
999
+ "selected_safety_label_known_at_8": 1.0,
1000
+ "selected_success_at_8": 0.0,
1001
+ "selected_success_gain_over_base_at_8": 0.0,
1002
+ "selected_unsafe_known_at_8": 0.0,
1003
+ "selected_utility_at_8": 0.20346055924892426,
1004
+ "selected_utility_gain_over_base_at_8": -0.006307080388069153,
1005
+ "selector_gap_fraction_to_hidden_at_8": 0.00044911597641627144,
1006
+ "selector_regret_at_8": 0.000804021954536438,
1007
+ "success_selector_gap_at_8": 0.0,
1008
+ "success_support_gap_at_8": 1.0,
1009
+ "success_total_car_to_hidden_at_8": 1.0,
1010
+ "support_gap_at_8": 1.7957354187965393,
1011
+ "support_gap_fraction_to_hidden_at_8": 1.003073935292074,
1012
+ "total_car_to_hidden_at_8": 1.7965394407510757
1013
+ },
1014
+ "509430392": {
1015
+ "any_generated_unsafe_known_at_8": 0.0,
1016
+ "base_safety_label_known": 1.0,
1017
+ "base_success": 0.0,
1018
+ "base_unsafe_known": 0.0,
1019
+ "base_utility": 0.04073460400104523,
1020
+ "branch_car_at_8": 0.0,
1021
+ "generated_safety_label_coverage_at_8": 1.0,
1022
+ "generated_unsafe_rate_known_at_8": 0.0,
1023
+ "hidden_chart_oracle_success_at_8": 1.0,
1024
+ "hidden_chart_oracle_utility_at_8": 2.0,
1025
+ "ncar_to_hidden_chart_oracle_at_8": 0.0,
1026
+ "ncar_to_proposal_oracle_at_8": 0.0,
1027
+ "outcome_ptr_at_8": 1.0,
1028
+ "pairwise_causal_calibration_ece": 0.11637116801567593,
1029
+ "proposal_oracle_safety_label_known_at_8": 1.0,
1030
+ "proposal_oracle_success_at_8": 1.0,
1031
+ "proposal_oracle_success_gain_over_base_at_8": 1.0,
1032
+ "proposal_oracle_unsafe_known_at_8": 0.0,
1033
+ "proposal_oracle_utility_at_8": 2.0,
1034
+ "proposal_oracle_utility_gain_over_base_at_8": 1.9592653959989548,
1035
+ "selected_safety_label_known_at_8": 1.0,
1036
+ "selected_success_at_8": 1.0,
1037
+ "selected_success_gain_over_base_at_8": 1.0,
1038
+ "selected_unsafe_known_at_8": 0.0,
1039
+ "selected_utility_at_8": 2.0,
1040
+ "selected_utility_gain_over_base_at_8": 1.9592653959989548,
1041
+ "selector_gap_fraction_to_hidden_at_8": 0.0,
1042
+ "selector_regret_at_8": 0.0,
1043
+ "success_selector_gap_at_8": 0.0,
1044
+ "success_support_gap_at_8": 0.0,
1045
+ "success_total_car_to_hidden_at_8": 0.0,
1046
+ "support_gap_at_8": 0.0,
1047
+ "support_gap_fraction_to_hidden_at_8": 0.0,
1048
+ "total_car_to_hidden_at_8": 0.0
1049
+ },
1050
+ "585904130": {
1051
+ "any_generated_unsafe_known_at_8": 1.0,
1052
+ "base_safety_label_known": 1.0,
1053
+ "base_success": 0.0,
1054
+ "base_unsafe_known": 0.0,
1055
+ "base_utility": 0.07103653997182846,
1056
+ "branch_car_at_8": 0.16070519387722015,
1057
+ "generated_safety_label_coverage_at_8": 1.0,
1058
+ "generated_unsafe_rate_known_at_8": 0.875,
1059
+ "hidden_chart_oracle_success_at_8": 1.0,
1060
+ "hidden_chart_oracle_utility_at_8": 2.0,
1061
+ "ncar_to_hidden_chart_oracle_at_8": 0.9677841286713383,
1062
+ "ncar_to_proposal_oracle_at_8": 0.7211412341552217,
1063
+ "outcome_ptr_at_8": 1.0,
1064
+ "pairwise_causal_calibration_ece": 0.537235764832106,
1065
+ "proposal_oracle_safety_label_known_at_8": 1.0,
1066
+ "proposal_oracle_success_at_8": 0.0,
1067
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
1068
+ "proposal_oracle_unsafe_known_at_8": 1.0,
1069
+ "proposal_oracle_utility_at_8": 0.2938849627971649,
1070
+ "proposal_oracle_utility_gain_over_base_at_8": 0.22284842282533646,
1071
+ "selected_safety_label_known_at_8": 1.0,
1072
+ "selected_success_at_8": 0.0,
1073
+ "selected_success_gain_over_base_at_8": 0.0,
1074
+ "selected_unsafe_known_at_8": 0.0,
1075
+ "selected_utility_at_8": 0.13317976891994476,
1076
+ "selected_utility_gain_over_base_at_8": 0.0621432289481163,
1077
+ "selector_gap_fraction_to_hidden_at_8": 0.08331168381741826,
1078
+ "selector_regret_at_8": 0.16070519387722015,
1079
+ "success_selector_gap_at_8": 0.0,
1080
+ "success_support_gap_at_8": 1.0,
1081
+ "success_total_car_to_hidden_at_8": 1.0,
1082
+ "support_gap_at_8": 1.706115037202835,
1083
+ "support_gap_fraction_to_hidden_at_8": 0.8844724498710401,
1084
+ "total_car_to_hidden_at_8": 1.8668202310800552
1085
+ },
1086
+ "608016290": {
1087
+ "any_generated_unsafe_known_at_8": 0.0,
1088
+ "base_safety_label_known": 1.0,
1089
+ "base_success": 0.0,
1090
+ "base_unsafe_known": 0.0,
1091
+ "base_utility": 0.525360107421875,
1092
+ "branch_car_at_8": 0.1574377715587616,
1093
+ "generated_safety_label_coverage_at_8": 1.0,
1094
+ "generated_unsafe_rate_known_at_8": 0.0,
1095
+ "hidden_chart_oracle_success_at_8": 1.0,
1096
+ "hidden_chart_oracle_utility_at_8": 2.0,
1097
+ "ncar_to_hidden_chart_oracle_at_8": 1.076145372525531,
1098
+ "ncar_to_proposal_oracle_at_8": 3.4869346102323533,
1099
+ "outcome_ptr_at_8": 1.0,
1100
+ "pairwise_causal_calibration_ece": 0.098162172719824,
1101
+ "proposal_oracle_safety_label_known_at_8": 1.0,
1102
+ "proposal_oracle_success_at_8": 0.0,
1103
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
1104
+ "proposal_oracle_unsafe_known_at_8": 0.0,
1105
+ "proposal_oracle_utility_at_8": 0.5705108642578125,
1106
+ "proposal_oracle_utility_gain_over_base_at_8": 0.0451507568359375,
1107
+ "selected_safety_label_known_at_8": 1.0,
1108
+ "selected_success_at_8": 0.0,
1109
+ "selected_success_gain_over_base_at_8": 0.0,
1110
+ "selected_unsafe_known_at_8": 0.0,
1111
+ "selected_utility_at_8": 0.4130730926990509,
1112
+ "selected_utility_gain_over_base_at_8": -0.1122870147228241,
1113
+ "selector_gap_fraction_to_hidden_at_8": 0.10676353756001532,
1114
+ "selector_regret_at_8": 0.1574377715587616,
1115
+ "success_selector_gap_at_8": 0.0,
1116
+ "success_support_gap_at_8": 1.0,
1117
+ "success_total_car_to_hidden_at_8": 1.0,
1118
+ "support_gap_at_8": 1.4294891357421875,
1119
+ "support_gap_fraction_to_hidden_at_8": 0.9693818422631982,
1120
+ "total_car_to_hidden_at_8": 1.586926907300949
1121
+ },
1122
+ "626486813": {
1123
+ "any_generated_unsafe_known_at_8": 0.5,
1124
+ "base_safety_label_known": 1.0,
1125
+ "base_success": 0.0,
1126
+ "base_unsafe_known": 0.0,
1127
+ "base_utility": 0.04174089431762695,
1128
+ "branch_car_at_8": 0.0030257832258939743,
1129
+ "generated_safety_label_coverage_at_8": 1.0,
1130
+ "generated_unsafe_rate_known_at_8": 0.3125,
1131
+ "hidden_chart_oracle_success_at_8": 0.0,
1132
+ "hidden_chart_oracle_utility_at_8": 0.16558238118886948,
1133
+ "ncar_to_hidden_chart_oracle_at_8": 0.8734881395113818,
1134
+ "ncar_to_proposal_oracle_at_8": 1.0887571854184497,
1135
+ "outcome_ptr_at_8": 1.0,
1136
+ "pairwise_causal_calibration_ece": 0.22930626964762768,
1137
+ "proposal_oracle_safety_label_known_at_8": 1.0,
1138
+ "proposal_oracle_success_at_8": 0.0,
1139
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
1140
+ "proposal_oracle_unsafe_known_at_8": 0.0,
1141
+ "proposal_oracle_utility_at_8": 0.08006523177027702,
1142
+ "proposal_oracle_utility_gain_over_base_at_8": 0.03832433745265007,
1143
+ "selected_safety_label_known_at_8": 1.0,
1144
+ "selected_success_at_8": 0.0,
1145
+ "selected_success_gain_over_base_at_8": 0.0,
1146
+ "selected_unsafe_known_at_8": 0.0,
1147
+ "selected_utility_at_8": 0.07703944854438305,
1148
+ "selected_utility_gain_over_base_at_8": 0.035298554226756096,
1149
+ "selector_gap_fraction_to_hidden_at_8": 0.07621781757550655,
1150
+ "selector_regret_at_8": 0.0030257832258939743,
1151
+ "success_selector_gap_at_8": 0.0,
1152
+ "success_support_gap_at_8": 0.0,
1153
+ "success_total_car_to_hidden_at_8": 0.0,
1154
+ "support_gap_at_8": 0.08551714941859245,
1155
+ "support_gap_fraction_to_hidden_at_8": 0.797270536737948,
1156
+ "total_car_to_hidden_at_8": 0.08854293264448643
1157
+ },
1158
+ "636680137": {
1159
+ "any_generated_unsafe_known_at_8": 0.0,
1160
+ "base_safety_label_known": 1.0,
1161
+ "base_success": 0.0,
1162
+ "base_unsafe_known": 0.0,
1163
+ "base_utility": 0.11957359313964844,
1164
+ "branch_car_at_8": 0.04479973018169403,
1165
+ "generated_safety_label_coverage_at_8": 1.0,
1166
+ "generated_unsafe_rate_known_at_8": 0.0,
1167
+ "hidden_chart_oracle_success_at_8": 0.0,
1168
+ "hidden_chart_oracle_utility_at_8": 0.17056432366371155,
1169
+ "ncar_to_hidden_chart_oracle_at_8": 1.7778231310760573,
1170
+ "ncar_to_proposal_oracle_at_8": 8.71937375135354,
1171
+ "outcome_ptr_at_8": 1.0,
1172
+ "pairwise_causal_calibration_ece": 0.17961355132445345,
1173
+ "proposal_oracle_safety_label_known_at_8": 1.0,
1174
+ "proposal_oracle_success_at_8": 0.0,
1175
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
1176
+ "proposal_oracle_unsafe_known_at_8": 0.0,
1177
+ "proposal_oracle_utility_at_8": 0.1247115358710289,
1178
+ "proposal_oracle_utility_gain_over_base_at_8": 0.005137942731380463,
1179
+ "selected_safety_label_known_at_8": 1.0,
1180
+ "selected_success_at_8": 0.0,
1181
+ "selected_success_gain_over_base_at_8": 0.0,
1182
+ "selected_unsafe_known_at_8": 0.0,
1183
+ "selected_utility_at_8": 0.07991180568933487,
1184
+ "selected_utility_gain_over_base_at_8": -0.03966178745031357,
1185
+ "selector_gap_fraction_to_hidden_at_8": 0.8785857688497427,
1186
+ "selector_regret_at_8": 0.04479973018169403,
1187
+ "success_selector_gap_at_8": 0.0,
1188
+ "success_support_gap_at_8": 0.0,
1189
+ "success_total_car_to_hidden_at_8": 0.0,
1190
+ "support_gap_at_8": 0.04585278779268265,
1191
+ "support_gap_fraction_to_hidden_at_8": 0.8992377108824552,
1192
+ "total_car_to_hidden_at_8": 0.09065251797437668
1193
+ },
1194
+ "655884605": {
1195
+ "any_generated_unsafe_known_at_8": 0.0,
1196
+ "base_safety_label_known": 1.0,
1197
+ "base_success": 0.0,
1198
+ "base_unsafe_known": 0.0,
1199
+ "base_utility": 0.1440753936767578,
1200
+ "branch_car_at_8": 0.0026672184467315674,
1201
+ "generated_safety_label_coverage_at_8": 1.0,
1202
+ "generated_unsafe_rate_known_at_8": 0.0,
1203
+ "hidden_chart_oracle_success_at_8": 1.0,
1204
+ "hidden_chart_oracle_utility_at_8": 2.0,
1205
+ "ncar_to_hidden_chart_oracle_at_8": 0.9861618738608336,
1206
+ "ncar_to_proposal_oracle_at_8": 0.09408265091957171,
1207
+ "outcome_ptr_at_8": 1.0,
1208
+ "pairwise_causal_calibration_ece": 0.030740674705558885,
1209
+ "proposal_oracle_safety_label_known_at_8": 1.0,
1210
+ "proposal_oracle_success_at_8": 0.0,
1211
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
1212
+ "proposal_oracle_unsafe_known_at_8": 0.0,
1213
+ "proposal_oracle_utility_at_8": 0.17242512106895447,
1214
+ "proposal_oracle_utility_gain_over_base_at_8": 0.028349727392196655,
1215
+ "selected_safety_label_known_at_8": 1.0,
1216
+ "selected_success_at_8": 0.0,
1217
+ "selected_success_gain_over_base_at_8": 0.0,
1218
+ "selected_unsafe_known_at_8": 0.0,
1219
+ "selected_utility_at_8": 0.1697579026222229,
1220
+ "selected_utility_gain_over_base_at_8": 0.025682508945465088,
1221
+ "selector_gap_fraction_to_hidden_at_8": 0.0014371372832949142,
1222
+ "selector_regret_at_8": 0.0026672184467315674,
1223
+ "success_selector_gap_at_8": 0.0,
1224
+ "success_support_gap_at_8": 1.0,
1225
+ "success_total_car_to_hidden_at_8": 1.0,
1226
+ "support_gap_at_8": 1.8275748789310455,
1227
+ "support_gap_fraction_to_hidden_at_8": 0.9847247418911267,
1228
+ "total_car_to_hidden_at_8": 1.830242097377777
1229
+ },
1230
+ "672328262": {
1231
+ "any_generated_unsafe_known_at_8": 0.0,
1232
+ "base_safety_label_known": 1.0,
1233
+ "base_success": 0.0,
1234
+ "base_unsafe_known": 0.0,
1235
+ "base_utility": 0.10593423992395401,
1236
+ "branch_car_at_8": 0.003098629415035248,
1237
+ "generated_safety_label_coverage_at_8": 1.0,
1238
+ "generated_unsafe_rate_known_at_8": 0.0,
1239
+ "hidden_chart_oracle_success_at_8": 1.0,
1240
+ "hidden_chart_oracle_utility_at_8": 2.0,
1241
+ "ncar_to_hidden_chart_oracle_at_8": 1.019464283192897,
1242
+ "ncar_to_proposal_oracle_at_8": 0.09176223920635006,
1243
+ "outcome_ptr_at_8": 0.0,
1244
+ "pairwise_causal_calibration_ece": 0.11992738854206779,
1245
+ "proposal_oracle_safety_label_known_at_8": 1.0,
1246
+ "proposal_oracle_success_at_8": 0.0,
1247
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
1248
+ "proposal_oracle_unsafe_known_at_8": 0.0,
1249
+ "proposal_oracle_utility_at_8": 0.07216622680425644,
1250
+ "proposal_oracle_utility_gain_over_base_at_8": -0.03376801311969757,
1251
+ "selected_safety_label_known_at_8": 1.0,
1252
+ "selected_success_at_8": 0.0,
1253
+ "selected_success_gain_over_base_at_8": 0.0,
1254
+ "selected_unsafe_known_at_8": 0.0,
1255
+ "selected_utility_at_8": 0.06906759738922119,
1256
+ "selected_utility_gain_over_base_at_8": -0.03686664253473282,
1257
+ "selector_gap_fraction_to_hidden_at_8": 0.001635967177248819,
1258
+ "selector_regret_at_8": 0.003098629415035248,
1259
+ "success_selector_gap_at_8": 0.0,
1260
+ "success_support_gap_at_8": 1.0,
1261
+ "success_total_car_to_hidden_at_8": 1.0,
1262
+ "support_gap_at_8": 1.9278337731957436,
1263
+ "support_gap_fraction_to_hidden_at_8": 1.0178283213980606,
1264
+ "total_car_to_hidden_at_8": 1.9309324026107788
1265
+ },
1266
+ "701667486": {
1267
+ "any_generated_unsafe_known_at_8": 0.0,
1268
+ "base_safety_label_known": 1.0,
1269
+ "base_success": 0.0,
1270
+ "base_unsafe_known": 0.0,
1271
+ "base_utility": 0.7912740707397461,
1272
+ "branch_car_at_8": 1.2084707021713257,
1273
+ "generated_safety_label_coverage_at_8": 1.0,
1274
+ "generated_unsafe_rate_known_at_8": 0.0,
1275
+ "hidden_chart_oracle_success_at_8": 1.0,
1276
+ "hidden_chart_oracle_utility_at_8": 2.0,
1277
+ "ncar_to_hidden_chart_oracle_at_8": 0.9997888379154961,
1278
+ "ncar_to_proposal_oracle_at_8": 0.9997888379154961,
1279
+ "outcome_ptr_at_8": 1.0,
1280
+ "pairwise_causal_calibration_ece": 0.19081260775351527,
1281
+ "proposal_oracle_safety_label_known_at_8": 1.0,
1282
+ "proposal_oracle_success_at_8": 1.0,
1283
+ "proposal_oracle_success_gain_over_base_at_8": 1.0,
1284
+ "proposal_oracle_unsafe_known_at_8": 0.0,
1285
+ "proposal_oracle_utility_at_8": 2.0,
1286
+ "proposal_oracle_utility_gain_over_base_at_8": 1.208725929260254,
1287
+ "selected_safety_label_known_at_8": 1.0,
1288
+ "selected_success_at_8": 0.0,
1289
+ "selected_success_gain_over_base_at_8": 0.0,
1290
+ "selected_unsafe_known_at_8": 0.0,
1291
+ "selected_utility_at_8": 0.7915292978286743,
1292
+ "selected_utility_gain_over_base_at_8": 0.00025522708892822266,
1293
+ "selector_gap_fraction_to_hidden_at_8": 0.9997888461869231,
1294
+ "selector_regret_at_8": 1.2084707021713257,
1295
+ "success_selector_gap_at_8": 1.0,
1296
+ "success_support_gap_at_8": 0.0,
1297
+ "success_total_car_to_hidden_at_8": 1.0,
1298
+ "support_gap_at_8": 0.0,
1299
+ "support_gap_fraction_to_hidden_at_8": 0.0,
1300
+ "total_car_to_hidden_at_8": 1.2084707021713257
1301
+ },
1302
+ "747260437": {
1303
+ "any_generated_unsafe_known_at_8": 1.0,
1304
+ "base_safety_label_known": 1.0,
1305
+ "base_success": 1.0,
1306
+ "base_unsafe_known": 0.0,
1307
+ "base_utility": 2.0,
1308
+ "branch_car_at_8": 0.02974964678287506,
1309
+ "generated_safety_label_coverage_at_8": 1.0,
1310
+ "generated_unsafe_rate_known_at_8": 1.0,
1311
+ "hidden_chart_oracle_success_at_8": 0.0,
1312
+ "hidden_chart_oracle_utility_at_8": 0.2786632478237152,
1313
+ "ncar_to_hidden_chart_oracle_at_8": 0.07153763725445794,
1314
+ "ncar_to_proposal_oracle_at_8": 0.016393451416906714,
1315
+ "outcome_ptr_at_8": 0.0,
1316
+ "pairwise_causal_calibration_ece": 0.30546034643967757,
1317
+ "proposal_oracle_safety_label_known_at_8": 1.0,
1318
+ "proposal_oracle_success_at_8": 0.0,
1319
+ "proposal_oracle_success_gain_over_base_at_8": -1.0,
1320
+ "proposal_oracle_unsafe_known_at_8": 1.0,
1321
+ "proposal_oracle_utility_at_8": 0.18527252972126007,
1322
+ "proposal_oracle_utility_gain_over_base_at_8": -1.81472747027874,
1323
+ "selected_safety_label_known_at_8": 1.0,
1324
+ "selected_success_at_8": 0.0,
1325
+ "selected_success_gain_over_base_at_8": -1.0,
1326
+ "selected_unsafe_known_at_8": 1.0,
1327
+ "selected_utility_at_8": 0.155522882938385,
1328
+ "selected_utility_gain_over_base_at_8": -1.844477117061615,
1329
+ "selector_gap_fraction_to_hidden_at_8": 0.017282874338947683,
1330
+ "selector_regret_at_8": 0.02974964678287506,
1331
+ "success_selector_gap_at_8": 0.0,
1332
+ "success_support_gap_at_8": 0.0,
1333
+ "success_total_car_to_hidden_at_8": 0.0,
1334
+ "support_gap_at_8": 0.09339071810245514,
1335
+ "support_gap_fraction_to_hidden_at_8": 0.05425476333110376,
1336
+ "total_car_to_hidden_at_8": 0.1231403648853302
1337
+ },
1338
+ "750448296": {
1339
+ "any_generated_unsafe_known_at_8": 1.0,
1340
+ "base_safety_label_known": 1.0,
1341
+ "base_success": 0.3333333333333333,
1342
+ "base_unsafe_known": 0.0,
1343
+ "base_utility": 0.7605033268531164,
1344
+ "branch_car_at_8": 0.05147687097390493,
1345
+ "generated_safety_label_coverage_at_8": 1.0,
1346
+ "generated_unsafe_rate_known_at_8": 0.9583333333333334,
1347
+ "hidden_chart_oracle_success_at_8": 0.6666666666666666,
1348
+ "hidden_chart_oracle_utility_at_8": 1.4079964955647786,
1349
+ "ncar_to_hidden_chart_oracle_at_8": 3.6663220458717465,
1350
+ "ncar_to_proposal_oracle_at_8": 0.22983336179896974,
1351
+ "outcome_ptr_at_8": 0.3333333333333333,
1352
+ "pairwise_causal_calibration_ece": 0.22679962996282277,
1353
+ "proposal_oracle_safety_label_known_at_8": 1.0,
1354
+ "proposal_oracle_success_at_8": 0.0,
1355
+ "proposal_oracle_success_gain_over_base_at_8": -0.3333333333333333,
1356
+ "proposal_oracle_unsafe_known_at_8": 0.6666666666666666,
1357
+ "proposal_oracle_utility_at_8": 0.2481476068496704,
1358
+ "proposal_oracle_utility_gain_over_base_at_8": -0.512355720003446,
1359
+ "selected_safety_label_known_at_8": 1.0,
1360
+ "selected_success_at_8": 0.0,
1361
+ "selected_success_gain_over_base_at_8": -0.3333333333333333,
1362
+ "selected_unsafe_known_at_8": 1.0,
1363
+ "selected_utility_at_8": 0.19667073587576547,
1364
+ "selected_utility_gain_over_base_at_8": -0.5638325909773508,
1365
+ "selector_gap_fraction_to_hidden_at_8": 0.7611347320927621,
1366
+ "selector_regret_at_8": 0.05147687097390493,
1367
+ "success_selector_gap_at_8": 0.0,
1368
+ "success_support_gap_at_8": 0.6666666666666666,
1369
+ "success_total_car_to_hidden_at_8": 0.6666666666666666,
1370
+ "support_gap_at_8": 1.1598488887151082,
1371
+ "support_gap_fraction_to_hidden_at_8": 2.905190365647023,
1372
+ "total_car_to_hidden_at_8": 1.2113257596890132
1373
+ },
1374
+ "767747424": {
1375
+ "any_generated_unsafe_known_at_8": 0.75,
1376
+ "base_safety_label_known": 1.0,
1377
+ "base_success": 0.5,
1378
+ "base_unsafe_known": 0.0,
1379
+ "base_utility": 1.0747087560594082,
1380
+ "branch_car_at_8": 0.45911610033363104,
1381
+ "generated_safety_label_coverage_at_8": 1.0,
1382
+ "generated_unsafe_rate_known_at_8": 0.75,
1383
+ "hidden_chart_oracle_success_at_8": 0.25,
1384
+ "hidden_chart_oracle_utility_at_8": 0.6737305596470833,
1385
+ "ncar_to_hidden_chart_oracle_at_8": 1.15659739162032,
1386
+ "ncar_to_proposal_oracle_at_8": 0.5383285029875181,
1387
+ "outcome_ptr_at_8": 0.25,
1388
+ "pairwise_causal_calibration_ece": 0.14681410264927175,
1389
+ "proposal_oracle_safety_label_known_at_8": 1.0,
1390
+ "proposal_oracle_success_at_8": 0.25,
1391
+ "proposal_oracle_success_gain_over_base_at_8": -0.25,
1392
+ "proposal_oracle_unsafe_known_at_8": 0.75,
1393
+ "proposal_oracle_utility_at_8": 0.5809952197596431,
1394
+ "proposal_oracle_utility_gain_over_base_at_8": -0.4937135362997651,
1395
+ "selected_safety_label_known_at_8": 1.0,
1396
+ "selected_success_at_8": 0.0,
1397
+ "selected_success_gain_over_base_at_8": -0.5,
1398
+ "selected_unsafe_known_at_8": 0.75,
1399
+ "selected_utility_at_8": 0.12187911942601204,
1400
+ "selected_utility_gain_over_base_at_8": -0.9528296366333961,
1401
+ "selector_gap_fraction_to_hidden_at_8": 0.7611605820457173,
1402
+ "selector_regret_at_8": 0.45911610033363104,
1403
+ "success_selector_gap_at_8": 0.25,
1404
+ "success_support_gap_at_8": 0.0,
1405
+ "success_total_car_to_hidden_at_8": 0.25,
1406
+ "support_gap_at_8": 0.09709237236529589,
1407
+ "support_gap_fraction_to_hidden_at_8": 0.5478765791717298,
1408
+ "total_car_to_hidden_at_8": 0.5518514402210712
1409
+ }
1410
+ }
workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/measured_metrics/metrics_by_task.json ADDED
@@ -0,0 +1,178 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "LiftPegUpright-v1": {
3
+ "any_generated_unsafe_known_at_8": 0.0,
4
+ "base_safety_label_known": 1.0,
5
+ "base_success": 0.0,
6
+ "base_unsafe_known": 0.0,
7
+ "base_utility": 0.4796089559793472,
8
+ "branch_car_at_8": 0.2874205827713013,
9
+ "generated_safety_label_coverage_at_8": 1.0,
10
+ "generated_unsafe_rate_known_at_8": 0.0,
11
+ "hidden_chart_oracle_success_at_8": 1.0,
12
+ "hidden_chart_oracle_utility_at_8": 2.0,
13
+ "ncar_to_hidden_chart_oracle_at_8": 1.0173553456526272,
14
+ "ncar_to_proposal_oracle_at_8": 1.0338735248844815,
15
+ "outcome_ptr_at_8": 0.4,
16
+ "pairwise_causal_calibration_ece": 0.1245853527683611,
17
+ "proposal_oracle_safety_label_known_at_8": 1.0,
18
+ "proposal_oracle_success_at_8": 0.2,
19
+ "proposal_oracle_success_gain_over_base_at_8": 0.2,
20
+ "proposal_oracle_unsafe_known_at_8": 0.0,
21
+ "proposal_oracle_utility_at_8": 0.7410696536302567,
22
+ "proposal_oracle_utility_gain_over_base_at_8": 0.2614606976509094,
23
+ "selected_safety_label_known_at_8": 1.0,
24
+ "selected_success_at_8": 0.0,
25
+ "selected_success_gain_over_base_at_8": 0.0,
26
+ "selected_unsafe_known_at_8": 0.0,
27
+ "selected_utility_at_8": 0.4536490708589554,
28
+ "selected_utility_gain_over_base_at_8": -0.025959885120391844,
29
+ "selector_gap_fraction_to_hidden_at_8": 0.22204696485387942,
30
+ "selector_regret_at_8": 0.2874205827713013,
31
+ "success_selector_gap_at_8": 0.2,
32
+ "success_support_gap_at_8": 0.8,
33
+ "success_total_car_to_hidden_at_8": 1.0,
34
+ "support_gap_at_8": 1.2589303463697434,
35
+ "support_gap_fraction_to_hidden_at_8": 0.7953083876473557,
36
+ "total_car_to_hidden_at_8": 1.5463509291410447
37
+ },
38
+ "PickCube-v1": {
39
+ "any_generated_unsafe_known_at_8": 0.0,
40
+ "base_safety_label_known": 1.0,
41
+ "base_success": 1.0,
42
+ "base_unsafe_known": 0.0,
43
+ "base_utility": 2.0,
44
+ "branch_car_at_8": 1.4563714861869812,
45
+ "generated_safety_label_coverage_at_8": 1.0,
46
+ "generated_unsafe_rate_known_at_8": 0.0,
47
+ "hidden_chart_oracle_success_at_8": 1.0,
48
+ "hidden_chart_oracle_utility_at_8": 2.0,
49
+ "outcome_ptr_at_8": 0.0,
50
+ "pairwise_causal_calibration_ece": 0.12267671398942098,
51
+ "proposal_oracle_safety_label_known_at_8": 1.0,
52
+ "proposal_oracle_success_at_8": 1.0,
53
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
54
+ "proposal_oracle_unsafe_known_at_8": 0.0,
55
+ "proposal_oracle_utility_at_8": 2.0,
56
+ "proposal_oracle_utility_gain_over_base_at_8": 0.0,
57
+ "selected_safety_label_known_at_8": 1.0,
58
+ "selected_success_at_8": 0.0,
59
+ "selected_success_gain_over_base_at_8": -1.0,
60
+ "selected_unsafe_known_at_8": 0.0,
61
+ "selected_utility_at_8": 0.5436285138130188,
62
+ "selected_utility_gain_over_base_at_8": -1.4563714861869812,
63
+ "selector_regret_at_8": 1.4563714861869812,
64
+ "success_selector_gap_at_8": 1.0,
65
+ "success_support_gap_at_8": 0.0,
66
+ "success_total_car_to_hidden_at_8": 1.0,
67
+ "support_gap_at_8": 0.0,
68
+ "total_car_to_hidden_at_8": 1.4563714861869812
69
+ },
70
+ "PullCube-v1": {
71
+ "any_generated_unsafe_known_at_8": 0.72,
72
+ "base_safety_label_known": 1.0,
73
+ "base_success": 0.36,
74
+ "base_unsafe_known": 0.0,
75
+ "base_utility": 0.8032506120204925,
76
+ "branch_car_at_8": 0.10624422326683998,
77
+ "generated_safety_label_coverage_at_8": 1.0,
78
+ "generated_unsafe_rate_known_at_8": 0.625,
79
+ "hidden_chart_oracle_success_at_8": 0.52,
80
+ "hidden_chart_oracle_utility_at_8": 1.1467878860235214,
81
+ "ncar_to_hidden_chart_oracle_at_8": 1.3318783387964193,
82
+ "ncar_to_proposal_oracle_at_8": 0.9248858515826692,
83
+ "outcome_ptr_at_8": 0.36,
84
+ "pairwise_causal_calibration_ece": 0.20159853773134337,
85
+ "proposal_oracle_safety_label_known_at_8": 1.0,
86
+ "proposal_oracle_success_at_8": 0.16,
87
+ "proposal_oracle_success_gain_over_base_at_8": -0.2,
88
+ "proposal_oracle_unsafe_known_at_8": 0.6,
89
+ "proposal_oracle_utility_at_8": 0.45270782738924026,
90
+ "proposal_oracle_utility_gain_over_base_at_8": -0.35054278463125227,
91
+ "selected_safety_label_known_at_8": 1.0,
92
+ "selected_success_at_8": 0.12,
93
+ "selected_success_gain_over_base_at_8": -0.24,
94
+ "selected_unsafe_known_at_8": 0.56,
95
+ "selected_utility_at_8": 0.3464636041224003,
96
+ "selected_utility_gain_over_base_at_8": -0.4567870078980923,
97
+ "selector_gap_fraction_to_hidden_at_8": 0.5617890754661544,
98
+ "selector_regret_at_8": 0.10624422326683998,
99
+ "success_selector_gap_at_8": 0.04,
100
+ "success_support_gap_at_8": 0.36,
101
+ "success_total_car_to_hidden_at_8": 0.4,
102
+ "support_gap_at_8": 0.6982270714640617,
103
+ "support_gap_fraction_to_hidden_at_8": 1.011357898635456,
104
+ "total_car_to_hidden_at_8": 0.8003242819011211
105
+ },
106
+ "PushCube-v1": {
107
+ "any_generated_unsafe_known_at_8": 0.0,
108
+ "base_safety_label_known": 1.0,
109
+ "base_success": 0.8333333333333334,
110
+ "base_unsafe_known": 0.0,
111
+ "base_utility": 1.6734557673335075,
112
+ "branch_car_at_8": 0.31880776087443036,
113
+ "generated_safety_label_coverage_at_8": 1.0,
114
+ "generated_unsafe_rate_known_at_8": 0.0,
115
+ "hidden_chart_oracle_success_at_8": 0.8333333333333334,
116
+ "hidden_chart_oracle_utility_at_8": 1.7015800600250561,
117
+ "ncar_to_hidden_chart_oracle_at_8": 0.015806854319084362,
118
+ "ncar_to_proposal_oracle_at_8": 0.010005592343544699,
119
+ "outcome_ptr_at_8": 0.16666666666666666,
120
+ "pairwise_causal_calibration_ece": 0.20497329769741546,
121
+ "proposal_oracle_safety_label_known_at_8": 1.0,
122
+ "proposal_oracle_success_at_8": 0.6666666666666666,
123
+ "proposal_oracle_success_gain_over_base_at_8": -0.16666666666666666,
124
+ "proposal_oracle_unsafe_known_at_8": 0.0,
125
+ "proposal_oracle_utility_at_8": 1.3802802537878354,
126
+ "proposal_oracle_utility_gain_over_base_at_8": -0.2931755135456721,
127
+ "selected_safety_label_known_at_8": 1.0,
128
+ "selected_success_at_8": 0.5,
129
+ "selected_success_gain_over_base_at_8": -0.3333333333333333,
130
+ "selected_unsafe_known_at_8": 0.0,
131
+ "selected_utility_at_8": 1.0614724929134052,
132
+ "selected_utility_gain_over_base_at_8": -0.6119832744201025,
133
+ "selector_gap_fraction_to_hidden_at_8": 0.012105901794610257,
134
+ "selector_regret_at_8": 0.31880776087443036,
135
+ "success_selector_gap_at_8": 0.16666666666666666,
136
+ "success_support_gap_at_8": 0.16666666666666666,
137
+ "success_total_car_to_hidden_at_8": 0.3333333333333333,
138
+ "support_gap_at_8": 0.32129980623722076,
139
+ "support_gap_fraction_to_hidden_at_8": 0.003700952612754926,
140
+ "total_car_to_hidden_at_8": 0.6401075671116511
141
+ },
142
+ "StackCube-v1": {
143
+ "any_generated_unsafe_known_at_8": 0.0,
144
+ "base_safety_label_known": 1.0,
145
+ "base_success": 0.09090909090909091,
146
+ "base_unsafe_known": 0.0,
147
+ "base_utility": 0.5003613992170854,
148
+ "branch_car_at_8": 0.26306972991336475,
149
+ "generated_safety_label_coverage_at_8": 1.0,
150
+ "generated_unsafe_rate_known_at_8": 0.0,
151
+ "hidden_chart_oracle_success_at_8": 1.0,
152
+ "hidden_chart_oracle_utility_at_8": 2.0,
153
+ "ncar_to_hidden_chart_oracle_at_8": 1.0574720335451175,
154
+ "ncar_to_proposal_oracle_at_8": 2.6639490050538304,
155
+ "outcome_ptr_at_8": 0.6363636363636364,
156
+ "pairwise_causal_calibration_ece": 0.2021201140493732,
157
+ "proposal_oracle_safety_label_known_at_8": 1.0,
158
+ "proposal_oracle_success_at_8": 0.09090909090909091,
159
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
160
+ "proposal_oracle_unsafe_known_at_8": 0.0,
161
+ "proposal_oracle_utility_at_8": 0.5384775440801274,
162
+ "proposal_oracle_utility_gain_over_base_at_8": 0.03811614486304196,
163
+ "selected_safety_label_known_at_8": 1.0,
164
+ "selected_success_at_8": 0.0,
165
+ "selected_success_gain_over_base_at_8": -0.09090909090909091,
166
+ "selected_unsafe_known_at_8": 0.0,
167
+ "selected_utility_at_8": 0.27540781416676263,
168
+ "selected_utility_gain_over_base_at_8": -0.22495358505032279,
169
+ "selector_gap_fraction_to_hidden_at_8": 0.1782854308460104,
170
+ "selector_regret_at_8": 0.26306972991336475,
171
+ "success_selector_gap_at_8": 0.09090909090909091,
172
+ "success_support_gap_at_8": 0.9090909090909091,
173
+ "success_total_car_to_hidden_at_8": 1.0,
174
+ "support_gap_at_8": 1.4615224559198727,
175
+ "support_gap_fraction_to_hidden_at_8": 0.8791866095136187,
176
+ "total_car_to_hidden_at_8": 1.7245921858332374
177
+ }
178
+ }
workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/measured_metrics/report.md ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Metric Evaluation (measured)
2
+
3
+ K: `8`
4
+
5
+ | Metric | N | Micro mean | 95% CI | Task macro | Seed macro |
6
+ | --- | ---: | ---: | ---: | ---: | ---: |
7
+ | any_generated_unsafe_known_at_8 | 48 | 0.3750 | [0.2500, 0.5208] | 0.1440 | 0.2938 |
8
+ | base_safety_label_known | 48 | 1.0000 | [1.0000, 1.0000] | 1.0000 | 1.0000 |
9
+ | base_success | 48 | 0.3333 | [0.2083, 0.4583] | 0.4568 | 0.3083 |
10
+ | base_unsafe_known | 48 | 0.0000 | [0.0000, 0.0000] | 0.0000 | 0.0000 |
11
+ | base_utility | 48 | 0.8338 | [0.5920, 1.0708] | 1.0913 | 0.8028 |
12
+ | branch_car_at_8 | 48 | 0.2158 | [0.1011, 0.3661] | 0.4864 | 0.2206 |
13
+ | generated_safety_label_coverage_at_8 | 48 | 1.0000 | [1.0000, 1.0000] | 1.0000 | 1.0000 |
14
+ | generated_unsafe_rate_known_at_8 | 48 | 0.3255 | [0.2214, 0.4505] | 0.1250 | 0.2505 |
15
+ | hidden_chart_oracle_success_at_8 | 48 | 0.7292 | [0.6042, 0.8333] | 0.8707 | 0.7854 |
16
+ | hidden_chart_oracle_utility_at_8 | 48 | 1.5183 | [1.2906, 1.7056] | 1.7697 | 1.6191 |
17
+ | ncar_to_hidden_chart_oracle_at_8 | 35 | 1.1333 | [0.8365, 1.5441] | 0.8556 | 1.0585 |
18
+ | ncar_to_proposal_oracle_at_8 | 40 | 1.3481 | [0.5300, 2.5628] | 1.1582 | 1.4657 |
19
+ | outcome_ptr_at_8 | 48 | 0.3958 | [0.2708, 0.5417] | 0.3126 | 0.4021 |
20
+ | pairwise_causal_calibration_ece | 48 | 0.1925 | [0.1548, 0.2249] | 0.1712 | 0.1991 |
21
+ | proposal_oracle_safety_label_known_at_8 | 48 | 1.0000 | [1.0000, 1.0000] | 1.0000 | 1.0000 |
22
+ | proposal_oracle_success_at_8 | 48 | 0.2292 | [0.1042, 0.3333] | 0.4235 | 0.2313 |
23
+ | proposal_oracle_success_gain_over_base_at_8 | 48 | -0.1042 | [-0.2292, 0.0208] | -0.0333 | -0.0771 |
24
+ | proposal_oracle_unsafe_known_at_8 | 48 | 0.3125 | [0.1875, 0.4583] | 0.1200 | 0.2479 |
25
+ | proposal_oracle_utility_at_8 | 48 | 0.6506 | [0.4362, 0.8420] | 1.0225 | 0.6671 |
26
+ | proposal_oracle_utility_gain_over_base_at_8 | 48 | -0.1833 | [-0.4133, 0.0378] | -0.0688 | -0.1357 |
27
+ | selected_safety_label_known_at_8 | 48 | 1.0000 | [1.0000, 1.0000] | 1.0000 | 1.0000 |
28
+ | selected_success_at_8 | 48 | 0.1250 | [0.0417, 0.2292] | 0.1240 | 0.1250 |
29
+ | selected_success_gain_over_base_at_8 | 48 | -0.2083 | [-0.3125, -0.0833] | -0.3328 | -0.1833 |
30
+ | selected_unsafe_known_at_8 | 48 | 0.2917 | [0.1875, 0.4375] | 0.1120 | 0.2188 |
31
+ | selected_utility_at_8 | 48 | 0.4348 | [0.2639, 0.6220] | 0.5361 | 0.4465 |
32
+ | selected_utility_gain_over_base_at_8 | 48 | -0.3990 | [-0.5913, -0.1621] | -0.5552 | -0.3563 |
33
+ | selector_gap_fraction_to_hidden_at_8 | 35 | 0.3723 | [0.1653, 0.6711] | 0.2436 | 0.3442 |
34
+ | selector_regret_at_8 | 48 | 0.2158 | [0.1011, 0.3661] | 0.4864 | 0.2206 |
35
+ | success_selector_gap_at_8 | 48 | 0.1042 | [0.0208, 0.2083] | 0.2995 | 0.1062 |
36
+ | success_support_gap_at_8 | 48 | 0.5000 | [0.3333, 0.6667] | 0.4472 | 0.5542 |
37
+ | success_total_car_to_hidden_at_8 | 48 | 0.6042 | [0.4583, 0.7292] | 0.7467 | 0.6604 |
38
+ | support_gap_at_8 | 48 | 0.8699 | [0.5969, 1.1405] | 0.7480 | 0.9532 |
39
+ | support_gap_fraction_to_hidden_at_8 | 35 | 0.8852 | [0.6005, 1.2127] | 0.6724 | 0.8446 |
40
+ | total_car_to_hidden_at_8 | 48 | 1.0835 | [0.8389, 1.2861] | 1.2335 | 1.1726 |
workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/measured_metrics/table.tex ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ % Auto-generated by scripts/eval_metrics.py
2
+ \begin{tabular}{lrrrr}
3
+ \toprule
4
+ Metric & N & Micro mean & CI low & CI high \\
5
+ \midrule
6
+ any\_generated\_unsafe\_known\_at\_8 & 48 & 0.3750 & 0.2500 & 0.5208 \\
7
+ base\_safety\_label\_known & 48 & 1.0000 & 1.0000 & 1.0000 \\
8
+ base\_success & 48 & 0.3333 & 0.2083 & 0.4583 \\
9
+ base\_unsafe\_known & 48 & 0.0000 & 0.0000 & 0.0000 \\
10
+ base\_utility & 48 & 0.8338 & 0.5920 & 1.0708 \\
11
+ branch\_car\_at\_8 & 48 & 0.2158 & 0.1011 & 0.3661 \\
12
+ generated\_safety\_label\_coverage\_at\_8 & 48 & 1.0000 & 1.0000 & 1.0000 \\
13
+ generated\_unsafe\_rate\_known\_at\_8 & 48 & 0.3255 & 0.2214 & 0.4505 \\
14
+ hidden\_chart\_oracle\_success\_at\_8 & 48 & 0.7292 & 0.6042 & 0.8333 \\
15
+ hidden\_chart\_oracle\_utility\_at\_8 & 48 & 1.5183 & 1.2906 & 1.7056 \\
16
+ ncar\_to\_hidden\_chart\_oracle\_at\_8 & 35 & 1.1333 & 0.8365 & 1.5441 \\
17
+ ncar\_to\_proposal\_oracle\_at\_8 & 40 & 1.3481 & 0.5300 & 2.5628 \\
18
+ outcome\_ptr\_at\_8 & 48 & 0.3958 & 0.2708 & 0.5417 \\
19
+ pairwise\_causal\_calibration\_ece & 48 & 0.1925 & 0.1548 & 0.2249 \\
20
+ proposal\_oracle\_safety\_label\_known\_at\_8 & 48 & 1.0000 & 1.0000 & 1.0000 \\
21
+ proposal\_oracle\_success\_at\_8 & 48 & 0.2292 & 0.1042 & 0.3333 \\
22
+ proposal\_oracle\_success\_gain\_over\_base\_at\_8 & 48 & -0.1042 & -0.2292 & 0.0208 \\
23
+ proposal\_oracle\_unsafe\_known\_at\_8 & 48 & 0.3125 & 0.1875 & 0.4583 \\
24
+ proposal\_oracle\_utility\_at\_8 & 48 & 0.6506 & 0.4362 & 0.8420 \\
25
+ proposal\_oracle\_utility\_gain\_over\_base\_at\_8 & 48 & -0.1833 & -0.4133 & 0.0378 \\
26
+ selected\_safety\_label\_known\_at\_8 & 48 & 1.0000 & 1.0000 & 1.0000 \\
27
+ selected\_success\_at\_8 & 48 & 0.1250 & 0.0417 & 0.2292 \\
28
+ selected\_success\_gain\_over\_base\_at\_8 & 48 & -0.2083 & -0.3125 & -0.0833 \\
29
+ selected\_unsafe\_known\_at\_8 & 48 & 0.2917 & 0.1875 & 0.4375 \\
30
+ selected\_utility\_at\_8 & 48 & 0.4348 & 0.2639 & 0.6220 \\
31
+ selected\_utility\_gain\_over\_base\_at\_8 & 48 & -0.3990 & -0.5913 & -0.1621 \\
32
+ selector\_gap\_fraction\_to\_hidden\_at\_8 & 35 & 0.3723 & 0.1653 & 0.6711 \\
33
+ selector\_regret\_at\_8 & 48 & 0.2158 & 0.1011 & 0.3661 \\
34
+ success\_selector\_gap\_at\_8 & 48 & 0.1042 & 0.0208 & 0.2083 \\
35
+ success\_support\_gap\_at\_8 & 48 & 0.5000 & 0.3333 & 0.6667 \\
36
+ success\_total\_car\_to\_hidden\_at\_8 & 48 & 0.6042 & 0.4583 & 0.7292 \\
37
+ support\_gap\_at\_8 & 48 & 0.8699 & 0.5969 & 1.1405 \\
38
+ support\_gap\_fraction\_to\_hidden\_at\_8 & 35 & 0.8852 & 0.6005 & 1.2127 \\
39
+ total\_car\_to\_hidden\_at\_8 & 48 & 1.0835 & 0.8389 & 1.2861 \\
40
+ \bottomrule
41
+ \end{tabular}
workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/metrics.json ADDED
@@ -0,0 +1,2814 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "checkpoint": "runs/ctt_residual_base_context_obs_seed2/model.pt",
3
+ "k": 8,
4
+ "measured_metric_summary": {
5
+ "any_generated_unsafe_known_at_8": {
6
+ "macro_by_seed": {
7
+ "ci": {
8
+ "confidence": 0.95,
9
+ "high": 0.4,
10
+ "low": 0.1625,
11
+ "mean": 0.29375,
12
+ "n": 40,
13
+ "num_samples": 200
14
+ },
15
+ "mean": 0.29375,
16
+ "per_seed": {
17
+ "1073161778": 0.0,
18
+ "1113371261": 0.0,
19
+ "1151410346": 0.0,
20
+ "1261930456": 0.0,
21
+ "1324133122": 0.0,
22
+ "1324642411": 0.0,
23
+ "1375492786": 1.0,
24
+ "141984615": 1.0,
25
+ "1461069713": 1.0,
26
+ "1480856810": 0.0,
27
+ "1509980664": 0.0,
28
+ "1522272686": 0.0,
29
+ "1549324805": 0.0,
30
+ "1578990937": 0.0,
31
+ "1747178030": 1.0,
32
+ "1790188787": 0.0,
33
+ "1822210471": 0.0,
34
+ "1845489013": 1.0,
35
+ "1999022574": 0.0,
36
+ "2012211287": 0.0,
37
+ "2038594191": 1.0,
38
+ "2054899126": 1.0,
39
+ "2061635888": 0.0,
40
+ "211245842": 0.0,
41
+ "278008524": 0.5,
42
+ "307068879": 0.0,
43
+ "412102563": 0.0,
44
+ "440577711": 0.0,
45
+ "441046872": 0.0,
46
+ "509430392": 0.0,
47
+ "585904130": 1.0,
48
+ "608016290": 0.0,
49
+ "626486813": 0.5,
50
+ "636680137": 0.0,
51
+ "655884605": 0.0,
52
+ "672328262": 0.0,
53
+ "701667486": 0.0,
54
+ "747260437": 1.0,
55
+ "750448296": 1.0,
56
+ "767747424": 0.75
57
+ }
58
+ },
59
+ "macro_by_task": {
60
+ "ci": {
61
+ "confidence": 0.95,
62
+ "high": 0.43200000000000005,
63
+ "low": 0.0,
64
+ "mean": 0.144,
65
+ "n": 5,
66
+ "num_samples": 200
67
+ },
68
+ "mean": 0.144,
69
+ "per_task": {
70
+ "LiftPegUpright-v1": 0.0,
71
+ "PickCube-v1": 0.0,
72
+ "PullCube-v1": 0.72,
73
+ "PushCube-v1": 0.0,
74
+ "StackCube-v1": 0.0
75
+ }
76
+ },
77
+ "metric": "any_generated_unsafe_known_at_8",
78
+ "micro": {
79
+ "confidence": 0.95,
80
+ "high": 0.5208333333333334,
81
+ "low": 0.25,
82
+ "mean": 0.375,
83
+ "n": 48,
84
+ "num_samples": 200
85
+ },
86
+ "n": 48
87
+ },
88
+ "base_safety_label_known": {
89
+ "macro_by_seed": {
90
+ "ci": {
91
+ "confidence": 0.95,
92
+ "high": 1.0,
93
+ "low": 1.0,
94
+ "mean": 1.0,
95
+ "n": 40,
96
+ "num_samples": 200
97
+ },
98
+ "mean": 1.0,
99
+ "per_seed": {
100
+ "1073161778": 1.0,
101
+ "1113371261": 1.0,
102
+ "1151410346": 1.0,
103
+ "1261930456": 1.0,
104
+ "1324133122": 1.0,
105
+ "1324642411": 1.0,
106
+ "1375492786": 1.0,
107
+ "141984615": 1.0,
108
+ "1461069713": 1.0,
109
+ "1480856810": 1.0,
110
+ "1509980664": 1.0,
111
+ "1522272686": 1.0,
112
+ "1549324805": 1.0,
113
+ "1578990937": 1.0,
114
+ "1747178030": 1.0,
115
+ "1790188787": 1.0,
116
+ "1822210471": 1.0,
117
+ "1845489013": 1.0,
118
+ "1999022574": 1.0,
119
+ "2012211287": 1.0,
120
+ "2038594191": 1.0,
121
+ "2054899126": 1.0,
122
+ "2061635888": 1.0,
123
+ "211245842": 1.0,
124
+ "278008524": 1.0,
125
+ "307068879": 1.0,
126
+ "412102563": 1.0,
127
+ "440577711": 1.0,
128
+ "441046872": 1.0,
129
+ "509430392": 1.0,
130
+ "585904130": 1.0,
131
+ "608016290": 1.0,
132
+ "626486813": 1.0,
133
+ "636680137": 1.0,
134
+ "655884605": 1.0,
135
+ "672328262": 1.0,
136
+ "701667486": 1.0,
137
+ "747260437": 1.0,
138
+ "750448296": 1.0,
139
+ "767747424": 1.0
140
+ }
141
+ },
142
+ "macro_by_task": {
143
+ "ci": {
144
+ "confidence": 0.95,
145
+ "high": 1.0,
146
+ "low": 1.0,
147
+ "mean": 1.0,
148
+ "n": 5,
149
+ "num_samples": 200
150
+ },
151
+ "mean": 1.0,
152
+ "per_task": {
153
+ "LiftPegUpright-v1": 1.0,
154
+ "PickCube-v1": 1.0,
155
+ "PullCube-v1": 1.0,
156
+ "PushCube-v1": 1.0,
157
+ "StackCube-v1": 1.0
158
+ }
159
+ },
160
+ "metric": "base_safety_label_known",
161
+ "micro": {
162
+ "confidence": 0.95,
163
+ "high": 1.0,
164
+ "low": 1.0,
165
+ "mean": 1.0,
166
+ "n": 48,
167
+ "num_samples": 200
168
+ },
169
+ "n": 48
170
+ },
171
+ "base_success": {
172
+ "macro_by_seed": {
173
+ "ci": {
174
+ "confidence": 0.95,
175
+ "high": 0.4375,
176
+ "low": 0.1625,
177
+ "mean": 0.30833333333333335,
178
+ "n": 40,
179
+ "num_samples": 200
180
+ },
181
+ "mean": 0.30833333333333335,
182
+ "per_seed": {
183
+ "1073161778": 0.0,
184
+ "1113371261": 0.0,
185
+ "1151410346": 0.0,
186
+ "1261930456": 1.0,
187
+ "1324133122": 0.0,
188
+ "1324642411": 1.0,
189
+ "1375492786": 1.0,
190
+ "141984615": 0.0,
191
+ "1461069713": 0.0,
192
+ "1480856810": 1.0,
193
+ "1509980664": 0.0,
194
+ "1522272686": 1.0,
195
+ "1549324805": 0.0,
196
+ "1578990937": 0.0,
197
+ "1747178030": 1.0,
198
+ "1790188787": 1.0,
199
+ "1822210471": 0.0,
200
+ "1845489013": 0.0,
201
+ "1999022574": 0.0,
202
+ "2012211287": 0.0,
203
+ "2038594191": 0.0,
204
+ "2054899126": 1.0,
205
+ "2061635888": 0.0,
206
+ "211245842": 0.0,
207
+ "278008524": 0.5,
208
+ "307068879": 0.0,
209
+ "412102563": 1.0,
210
+ "440577711": 1.0,
211
+ "441046872": 0.0,
212
+ "509430392": 0.0,
213
+ "585904130": 0.0,
214
+ "608016290": 0.0,
215
+ "626486813": 0.0,
216
+ "636680137": 0.0,
217
+ "655884605": 0.0,
218
+ "672328262": 0.0,
219
+ "701667486": 0.0,
220
+ "747260437": 1.0,
221
+ "750448296": 0.3333333333333333,
222
+ "767747424": 0.5
223
+ }
224
+ },
225
+ "macro_by_task": {
226
+ "ci": {
227
+ "confidence": 0.95,
228
+ "high": 0.772,
229
+ "low": 0.09018181818181818,
230
+ "mean": 0.45684848484848484,
231
+ "n": 5,
232
+ "num_samples": 200
233
+ },
234
+ "mean": 0.45684848484848484,
235
+ "per_task": {
236
+ "LiftPegUpright-v1": 0.0,
237
+ "PickCube-v1": 1.0,
238
+ "PullCube-v1": 0.36,
239
+ "PushCube-v1": 0.8333333333333334,
240
+ "StackCube-v1": 0.09090909090909091
241
+ }
242
+ },
243
+ "metric": "base_success",
244
+ "micro": {
245
+ "confidence": 0.95,
246
+ "high": 0.4583333333333333,
247
+ "low": 0.20833333333333334,
248
+ "mean": 0.3333333333333333,
249
+ "n": 48,
250
+ "num_samples": 200
251
+ },
252
+ "n": 48
253
+ },
254
+ "base_unsafe_known": {
255
+ "macro_by_seed": {
256
+ "ci": {
257
+ "confidence": 0.95,
258
+ "high": 0.0,
259
+ "low": 0.0,
260
+ "mean": 0.0,
261
+ "n": 40,
262
+ "num_samples": 200
263
+ },
264
+ "mean": 0.0,
265
+ "per_seed": {
266
+ "1073161778": 0.0,
267
+ "1113371261": 0.0,
268
+ "1151410346": 0.0,
269
+ "1261930456": 0.0,
270
+ "1324133122": 0.0,
271
+ "1324642411": 0.0,
272
+ "1375492786": 0.0,
273
+ "141984615": 0.0,
274
+ "1461069713": 0.0,
275
+ "1480856810": 0.0,
276
+ "1509980664": 0.0,
277
+ "1522272686": 0.0,
278
+ "1549324805": 0.0,
279
+ "1578990937": 0.0,
280
+ "1747178030": 0.0,
281
+ "1790188787": 0.0,
282
+ "1822210471": 0.0,
283
+ "1845489013": 0.0,
284
+ "1999022574": 0.0,
285
+ "2012211287": 0.0,
286
+ "2038594191": 0.0,
287
+ "2054899126": 0.0,
288
+ "2061635888": 0.0,
289
+ "211245842": 0.0,
290
+ "278008524": 0.0,
291
+ "307068879": 0.0,
292
+ "412102563": 0.0,
293
+ "440577711": 0.0,
294
+ "441046872": 0.0,
295
+ "509430392": 0.0,
296
+ "585904130": 0.0,
297
+ "608016290": 0.0,
298
+ "626486813": 0.0,
299
+ "636680137": 0.0,
300
+ "655884605": 0.0,
301
+ "672328262": 0.0,
302
+ "701667486": 0.0,
303
+ "747260437": 0.0,
304
+ "750448296": 0.0,
305
+ "767747424": 0.0
306
+ }
307
+ },
308
+ "macro_by_task": {
309
+ "ci": {
310
+ "confidence": 0.95,
311
+ "high": 0.0,
312
+ "low": 0.0,
313
+ "mean": 0.0,
314
+ "n": 5,
315
+ "num_samples": 200
316
+ },
317
+ "mean": 0.0,
318
+ "per_task": {
319
+ "LiftPegUpright-v1": 0.0,
320
+ "PickCube-v1": 0.0,
321
+ "PullCube-v1": 0.0,
322
+ "PushCube-v1": 0.0,
323
+ "StackCube-v1": 0.0
324
+ }
325
+ },
326
+ "metric": "base_unsafe_known",
327
+ "micro": {
328
+ "confidence": 0.95,
329
+ "high": 0.0,
330
+ "low": 0.0,
331
+ "mean": 0.0,
332
+ "n": 48,
333
+ "num_samples": 200
334
+ },
335
+ "n": 48
336
+ },
337
+ "base_utility": {
338
+ "macro_by_seed": {
339
+ "ci": {
340
+ "confidence": 0.95,
341
+ "high": 1.0251519351576763,
342
+ "low": 0.537800694660594,
343
+ "mean": 0.8028034761858482,
344
+ "n": 40,
345
+ "num_samples": 200
346
+ },
347
+ "mean": 0.8028034761858482,
348
+ "per_seed": {
349
+ "1073161778": 0.725517988204956,
350
+ "1113371261": 0.11109928786754608,
351
+ "1151410346": 0.2284735143184662,
352
+ "1261930456": 2.0,
353
+ "1324133122": 0.084227055311203,
354
+ "1324642411": 2.0,
355
+ "1375492786": 2.0,
356
+ "141984615": 0.16783273220062256,
357
+ "1461069713": 0.22852759063243866,
358
+ "1480856810": 2.0,
359
+ "1509980664": 0.8320798277854919,
360
+ "1522272686": 2.0,
361
+ "1549324805": 0.17249532043933868,
362
+ "1578990937": 0.21426597237586975,
363
+ "1747178030": 2.0,
364
+ "1790188787": 2.0,
365
+ "1822210471": 0.09442546218633652,
366
+ "1845489013": 0.11255800724029541,
367
+ "1999022574": 0.8279688954353333,
368
+ "2012211287": 0.7282466888427734,
369
+ "2038594191": 0.09504059702157974,
370
+ "2054899126": 2.0,
371
+ "2061635888": 0.08620692789554596,
372
+ "211245842": 0.18430444598197937,
373
+ "278008524": 1.1250072121620178,
374
+ "307068879": 0.20915235579013824,
375
+ "412102563": 2.0,
376
+ "440577711": 2.0,
377
+ "441046872": 0.2097676396369934,
378
+ "509430392": 0.04073460400104523,
379
+ "585904130": 0.07103653997182846,
380
+ "608016290": 0.525360107421875,
381
+ "626486813": 0.04174089431762695,
382
+ "636680137": 0.11957359313964844,
383
+ "655884605": 0.1440753936767578,
384
+ "672328262": 0.10593423992395401,
385
+ "701667486": 0.7912740707397461,
386
+ "747260437": 2.0,
387
+ "750448296": 0.7605033268531164,
388
+ "767747424": 1.0747087560594082
389
+ }
390
+ },
391
+ "macro_by_task": {
392
+ "ci": {
393
+ "confidence": 0.95,
394
+ "high": 1.630612944662571,
395
+ "low": 0.5484877758351239,
396
+ "mean": 1.0913353469100866,
397
+ "n": 5,
398
+ "num_samples": 200
399
+ },
400
+ "mean": 1.0913353469100866,
401
+ "per_task": {
402
+ "LiftPegUpright-v1": 0.4796089559793472,
403
+ "PickCube-v1": 2.0,
404
+ "PullCube-v1": 0.8032506120204925,
405
+ "PushCube-v1": 1.6734557673335075,
406
+ "StackCube-v1": 0.5003613992170854
407
+ }
408
+ },
409
+ "metric": "base_utility",
410
+ "micro": {
411
+ "confidence": 0.95,
412
+ "high": 1.0708436407148838,
413
+ "low": 0.5920010643700758,
414
+ "mean": 0.8338337515791258,
415
+ "n": 48,
416
+ "num_samples": 200
417
+ },
418
+ "n": 48
419
+ },
420
+ "branch_car_at_8": {
421
+ "macro_by_seed": {
422
+ "ci": {
423
+ "confidence": 0.95,
424
+ "high": 0.3860186778940261,
425
+ "low": 0.10536204695235937,
426
+ "mean": 0.22059872947478049,
427
+ "n": 40,
428
+ "num_samples": 200
429
+ },
430
+ "mean": 0.22059872947478049,
431
+ "per_seed": {
432
+ "1073161778": 1.2767786383628845,
433
+ "1113371261": 0.007540524005889893,
434
+ "1151410346": 0.3237016499042511,
435
+ "1261930456": 0.011454477906227112,
436
+ "1324133122": 0.010064825415611267,
437
+ "1324642411": 1.858040377497673,
438
+ "1375492786": 0.027821257710456848,
439
+ "141984615": 0.002475142478942871,
440
+ "1461069713": 0.026291392743587494,
441
+ "1480856810": 0.043351709842681885,
442
+ "1509980664": 0.014023125171661377,
443
+ "1522272686": 1.4563714861869812,
444
+ "1549324805": 0.09635330736637115,
445
+ "1578990937": 0.025389984250068665,
446
+ "1747178030": 0.0,
447
+ "1790188787": 0.0,
448
+ "1822210471": 0.0011331625282764435,
449
+ "1845489013": 0.009949073195457458,
450
+ "1999022574": 0.6338560879230499,
451
+ "2012211287": 0.00025147199630737305,
452
+ "2038594191": 0.1589493229985237,
453
+ "2054899126": 0.005500897765159607,
454
+ "2061635888": 0.0,
455
+ "211245842": 0.041060686111450195,
456
+ "278008524": 0.043415673077106476,
457
+ "307068879": 0.0018310099840164185,
458
+ "412102563": 0.6269922256469727,
459
+ "440577711": 0.0,
460
+ "441046872": 0.000804021954536438,
461
+ "509430392": 0.0,
462
+ "585904130": 0.16070519387722015,
463
+ "608016290": 0.1574377715587616,
464
+ "626486813": 0.0030257832258939743,
465
+ "636680137": 0.04479973018169403,
466
+ "655884605": 0.0026672184467315674,
467
+ "672328262": 0.003098629415035248,
468
+ "701667486": 1.2084707021713257,
469
+ "747260437": 0.02974964678287506,
470
+ "750448296": 0.05147687097390493,
471
+ "767747424": 0.45911610033363104
472
+ }
473
+ },
474
+ "macro_by_task": {
475
+ "ci": {
476
+ "confidence": 0.95,
477
+ "high": 0.9839209542491218,
478
+ "low": 0.18012203211766303,
479
+ "mean": 0.48638275660258345,
480
+ "n": 5,
481
+ "num_samples": 200
482
+ },
483
+ "mean": 0.48638275660258345,
484
+ "per_task": {
485
+ "LiftPegUpright-v1": 0.2874205827713013,
486
+ "PickCube-v1": 1.4563714861869812,
487
+ "PullCube-v1": 0.10624422326683998,
488
+ "PushCube-v1": 0.31880776087443036,
489
+ "StackCube-v1": 0.26306972991336475
490
+ }
491
+ },
492
+ "metric": "branch_car_at_8",
493
+ "micro": {
494
+ "confidence": 0.95,
495
+ "high": 0.3661268416326493,
496
+ "low": 0.10114351854038735,
497
+ "mean": 0.2157540328335017,
498
+ "n": 48,
499
+ "num_samples": 200
500
+ },
501
+ "n": 48
502
+ },
503
+ "generated_safety_label_coverage_at_8": {
504
+ "macro_by_seed": {
505
+ "ci": {
506
+ "confidence": 0.95,
507
+ "high": 1.0,
508
+ "low": 1.0,
509
+ "mean": 1.0,
510
+ "n": 40,
511
+ "num_samples": 200
512
+ },
513
+ "mean": 1.0,
514
+ "per_seed": {
515
+ "1073161778": 1.0,
516
+ "1113371261": 1.0,
517
+ "1151410346": 1.0,
518
+ "1261930456": 1.0,
519
+ "1324133122": 1.0,
520
+ "1324642411": 1.0,
521
+ "1375492786": 1.0,
522
+ "141984615": 1.0,
523
+ "1461069713": 1.0,
524
+ "1480856810": 1.0,
525
+ "1509980664": 1.0,
526
+ "1522272686": 1.0,
527
+ "1549324805": 1.0,
528
+ "1578990937": 1.0,
529
+ "1747178030": 1.0,
530
+ "1790188787": 1.0,
531
+ "1822210471": 1.0,
532
+ "1845489013": 1.0,
533
+ "1999022574": 1.0,
534
+ "2012211287": 1.0,
535
+ "2038594191": 1.0,
536
+ "2054899126": 1.0,
537
+ "2061635888": 1.0,
538
+ "211245842": 1.0,
539
+ "278008524": 1.0,
540
+ "307068879": 1.0,
541
+ "412102563": 1.0,
542
+ "440577711": 1.0,
543
+ "441046872": 1.0,
544
+ "509430392": 1.0,
545
+ "585904130": 1.0,
546
+ "608016290": 1.0,
547
+ "626486813": 1.0,
548
+ "636680137": 1.0,
549
+ "655884605": 1.0,
550
+ "672328262": 1.0,
551
+ "701667486": 1.0,
552
+ "747260437": 1.0,
553
+ "750448296": 1.0,
554
+ "767747424": 1.0
555
+ }
556
+ },
557
+ "macro_by_task": {
558
+ "ci": {
559
+ "confidence": 0.95,
560
+ "high": 1.0,
561
+ "low": 1.0,
562
+ "mean": 1.0,
563
+ "n": 5,
564
+ "num_samples": 200
565
+ },
566
+ "mean": 1.0,
567
+ "per_task": {
568
+ "LiftPegUpright-v1": 1.0,
569
+ "PickCube-v1": 1.0,
570
+ "PullCube-v1": 1.0,
571
+ "PushCube-v1": 1.0,
572
+ "StackCube-v1": 1.0
573
+ }
574
+ },
575
+ "metric": "generated_safety_label_coverage_at_8",
576
+ "micro": {
577
+ "confidence": 0.95,
578
+ "high": 1.0,
579
+ "low": 1.0,
580
+ "mean": 1.0,
581
+ "n": 48,
582
+ "num_samples": 200
583
+ },
584
+ "n": 48
585
+ },
586
+ "generated_unsafe_rate_known_at_8": {
587
+ "macro_by_seed": {
588
+ "ci": {
589
+ "confidence": 0.95,
590
+ "high": 0.34427083333333336,
591
+ "low": 0.1234375,
592
+ "mean": 0.25052083333333336,
593
+ "n": 40,
594
+ "num_samples": 200
595
+ },
596
+ "mean": 0.25052083333333336,
597
+ "per_seed": {
598
+ "1073161778": 0.0,
599
+ "1113371261": 0.0,
600
+ "1151410346": 0.0,
601
+ "1261930456": 0.0,
602
+ "1324133122": 0.0,
603
+ "1324642411": 0.0,
604
+ "1375492786": 1.0,
605
+ "141984615": 1.0,
606
+ "1461069713": 0.875,
607
+ "1480856810": 0.0,
608
+ "1509980664": 0.0,
609
+ "1522272686": 0.0,
610
+ "1549324805": 0.0,
611
+ "1578990937": 0.0,
612
+ "1747178030": 1.0,
613
+ "1790188787": 0.0,
614
+ "1822210471": 0.0,
615
+ "1845489013": 0.75,
616
+ "1999022574": 0.0,
617
+ "2012211287": 0.0,
618
+ "2038594191": 0.375,
619
+ "2054899126": 0.9375,
620
+ "2061635888": 0.0,
621
+ "211245842": 0.0,
622
+ "278008524": 0.1875,
623
+ "307068879": 0.0,
624
+ "412102563": 0.0,
625
+ "440577711": 0.0,
626
+ "441046872": 0.0,
627
+ "509430392": 0.0,
628
+ "585904130": 0.875,
629
+ "608016290": 0.0,
630
+ "626486813": 0.3125,
631
+ "636680137": 0.0,
632
+ "655884605": 0.0,
633
+ "672328262": 0.0,
634
+ "701667486": 0.0,
635
+ "747260437": 1.0,
636
+ "750448296": 0.9583333333333334,
637
+ "767747424": 0.75
638
+ }
639
+ },
640
+ "macro_by_task": {
641
+ "ci": {
642
+ "confidence": 0.95,
643
+ "high": 0.375,
644
+ "low": 0.0,
645
+ "mean": 0.125,
646
+ "n": 5,
647
+ "num_samples": 200
648
+ },
649
+ "mean": 0.125,
650
+ "per_task": {
651
+ "LiftPegUpright-v1": 0.0,
652
+ "PickCube-v1": 0.0,
653
+ "PullCube-v1": 0.625,
654
+ "PushCube-v1": 0.0,
655
+ "StackCube-v1": 0.0
656
+ }
657
+ },
658
+ "metric": "generated_unsafe_rate_known_at_8",
659
+ "micro": {
660
+ "confidence": 0.95,
661
+ "high": 0.4505208333333333,
662
+ "low": 0.22135416666666666,
663
+ "mean": 0.3255208333333333,
664
+ "n": 48,
665
+ "num_samples": 200
666
+ },
667
+ "n": 48
668
+ },
669
+ "hidden_chart_oracle_success_at_8": {
670
+ "macro_by_seed": {
671
+ "ci": {
672
+ "confidence": 0.95,
673
+ "high": 0.9,
674
+ "low": 0.6729166666666667,
675
+ "mean": 0.7854166666666667,
676
+ "n": 40,
677
+ "num_samples": 200
678
+ },
679
+ "mean": 0.7854166666666667,
680
+ "per_seed": {
681
+ "1073161778": 1.0,
682
+ "1113371261": 1.0,
683
+ "1151410346": 1.0,
684
+ "1261930456": 1.0,
685
+ "1324133122": 1.0,
686
+ "1324642411": 1.0,
687
+ "1375492786": 1.0,
688
+ "141984615": 0.0,
689
+ "1461069713": 0.0,
690
+ "1480856810": 0.0,
691
+ "1509980664": 1.0,
692
+ "1522272686": 1.0,
693
+ "1549324805": 0.0,
694
+ "1578990937": 1.0,
695
+ "1747178030": 1.0,
696
+ "1790188787": 1.0,
697
+ "1822210471": 1.0,
698
+ "1845489013": 1.0,
699
+ "1999022574": 1.0,
700
+ "2012211287": 1.0,
701
+ "2038594191": 1.0,
702
+ "2054899126": 1.0,
703
+ "2061635888": 1.0,
704
+ "211245842": 1.0,
705
+ "278008524": 0.5,
706
+ "307068879": 1.0,
707
+ "412102563": 1.0,
708
+ "440577711": 1.0,
709
+ "441046872": 1.0,
710
+ "509430392": 1.0,
711
+ "585904130": 1.0,
712
+ "608016290": 1.0,
713
+ "626486813": 0.0,
714
+ "636680137": 0.0,
715
+ "655884605": 1.0,
716
+ "672328262": 1.0,
717
+ "701667486": 1.0,
718
+ "747260437": 0.0,
719
+ "750448296": 0.6666666666666666,
720
+ "767747424": 0.25
721
+ }
722
+ },
723
+ "macro_by_task": {
724
+ "ci": {
725
+ "confidence": 0.95,
726
+ "high": 1.0,
727
+ "low": 0.6786666666666668,
728
+ "mean": 0.8706666666666667,
729
+ "n": 5,
730
+ "num_samples": 200
731
+ },
732
+ "mean": 0.8706666666666667,
733
+ "per_task": {
734
+ "LiftPegUpright-v1": 1.0,
735
+ "PickCube-v1": 1.0,
736
+ "PullCube-v1": 0.52,
737
+ "PushCube-v1": 0.8333333333333334,
738
+ "StackCube-v1": 1.0
739
+ }
740
+ },
741
+ "metric": "hidden_chart_oracle_success_at_8",
742
+ "micro": {
743
+ "confidence": 0.95,
744
+ "high": 0.8333333333333334,
745
+ "low": 0.6041666666666666,
746
+ "mean": 0.7291666666666666,
747
+ "n": 48,
748
+ "num_samples": 200
749
+ },
750
+ "n": 48
751
+ },
752
+ "hidden_chart_oracle_utility_at_8": {
753
+ "macro_by_seed": {
754
+ "ci": {
755
+ "confidence": 0.95,
756
+ "high": 1.8218825578689575,
757
+ "low": 1.4154964266344905,
758
+ "mean": 1.619062545460959,
759
+ "n": 40,
760
+ "num_samples": 200
761
+ },
762
+ "mean": 1.619062545460959,
763
+ "per_seed": {
764
+ "1073161778": 2.0,
765
+ "1113371261": 2.0,
766
+ "1151410346": 2.0,
767
+ "1261930456": 2.0,
768
+ "1324133122": 2.0,
769
+ "1324642411": 2.0,
770
+ "1375492786": 2.0,
771
+ "141984615": 0.2857772707939148,
772
+ "1461069713": 0.1593424528837204,
773
+ "1480856810": 0.20948036015033722,
774
+ "1509980664": 2.0,
775
+ "1522272686": 2.0,
776
+ "1549324805": 0.2974560856819153,
777
+ "1578990937": 2.0,
778
+ "1747178030": 2.0,
779
+ "1790188787": 2.0,
780
+ "1822210471": 2.0,
781
+ "1845489013": 2.0,
782
+ "1999022574": 2.0,
783
+ "2012211287": 2.0,
784
+ "2038594191": 2.0,
785
+ "2054899126": 2.0,
786
+ "2061635888": 2.0,
787
+ "211245842": 2.0,
788
+ "278008524": 1.1139086410403252,
789
+ "307068879": 2.0,
790
+ "412102563": 2.0,
791
+ "440577711": 2.0,
792
+ "441046872": 2.0,
793
+ "509430392": 2.0,
794
+ "585904130": 2.0,
795
+ "608016290": 2.0,
796
+ "626486813": 0.16558238118886948,
797
+ "636680137": 0.17056432366371155,
798
+ "655884605": 2.0,
799
+ "672328262": 2.0,
800
+ "701667486": 2.0,
801
+ "747260437": 0.2786632478237152,
802
+ "750448296": 1.4079964955647786,
803
+ "767747424": 0.6737305596470833
804
+ }
805
+ },
806
+ "macro_by_task": {
807
+ "ci": {
808
+ "confidence": 0.95,
809
+ "high": 2.0,
810
+ "low": 1.428388743619124,
811
+ "mean": 1.7696735892097153,
812
+ "n": 5,
813
+ "num_samples": 200
814
+ },
815
+ "mean": 1.7696735892097153,
816
+ "per_task": {
817
+ "LiftPegUpright-v1": 2.0,
818
+ "PickCube-v1": 2.0,
819
+ "PullCube-v1": 1.1467878860235214,
820
+ "PushCube-v1": 1.7015800600250561,
821
+ "StackCube-v1": 2.0
822
+ }
823
+ },
824
+ "metric": "hidden_chart_oracle_utility_at_8",
825
+ "micro": {
826
+ "confidence": 0.95,
827
+ "high": 1.7055829176679254,
828
+ "low": 1.2906390205025673,
829
+ "mean": 1.5183161981403828,
830
+ "n": 48,
831
+ "num_samples": 200
832
+ },
833
+ "n": 48
834
+ },
835
+ "ncar_to_hidden_chart_oracle_at_8": {
836
+ "macro_by_seed": {
837
+ "ci": {
838
+ "confidence": 0.95,
839
+ "high": 1.286984027248401,
840
+ "low": 0.8600994916401823,
841
+ "mean": 1.0585134161956213,
842
+ "n": 31,
843
+ "num_samples": 200
844
+ },
845
+ "mean": 1.0585134161956213,
846
+ "per_seed": {
847
+ "1073161778": 1.0018019999721972,
848
+ "1113371261": 1.0096749460109702,
849
+ "1151410346": 0.9938836701698183,
850
+ "1324133122": 0.9979853617028601,
851
+ "141984615": 2.0668632118652317,
852
+ "1461069713": 1.0147308339002874,
853
+ "1480856810": 0.031613708638168724,
854
+ "1509980664": 1.0061738317414306,
855
+ "1549324805": 1.4386390413214543,
856
+ "1578990937": 1.0039557654834124,
857
+ "1822210471": 1.019279876796129,
858
+ "1845489013": 1.019238671785175,
859
+ "1999022574": 1.5672404802573083,
860
+ "2012211287": 1.0011445108113346,
861
+ "2038594191": 1.002720058258962,
862
+ "2061635888": 1.0037531653219327,
863
+ "211245842": 1.0061024029871146,
864
+ "278008524": 0.026316678586339484,
865
+ "307068879": 1.004161799291129,
866
+ "441046872": 1.003523045662944,
867
+ "509430392": 0.0,
868
+ "585904130": 0.9677841286713383,
869
+ "608016290": 1.076145372525531,
870
+ "626486813": 0.8734881395113818,
871
+ "636680137": 1.7778231310760573,
872
+ "655884605": 0.9861618738608336,
873
+ "672328262": 1.019464283192897,
874
+ "701667486": 0.9997888379154961,
875
+ "747260437": 0.07153763725445794,
876
+ "750448296": 3.6663220458717465,
877
+ "767747424": 1.15659739162032
878
+ }
879
+ },
880
+ "macro_by_task": {
881
+ "ci": {
882
+ "confidence": 0.95,
883
+ "high": 1.1946751861707683,
884
+ "low": 0.3448247254384181,
885
+ "mean": 0.8556281430783121,
886
+ "n": 4,
887
+ "num_samples": 200
888
+ },
889
+ "mean": 0.8556281430783121,
890
+ "per_task": {
891
+ "LiftPegUpright-v1": 1.0173553456526272,
892
+ "PullCube-v1": 1.3318783387964193,
893
+ "PushCube-v1": 0.015806854319084362,
894
+ "StackCube-v1": 1.0574720335451175
895
+ }
896
+ },
897
+ "metric": "ncar_to_hidden_chart_oracle_at_8",
898
+ "micro": {
899
+ "confidence": 0.95,
900
+ "high": 1.5441353359191043,
901
+ "low": 0.8365329804496855,
902
+ "mean": 1.1333405963053722,
903
+ "n": 35,
904
+ "num_samples": 200
905
+ },
906
+ "n": 35
907
+ },
908
+ "ncar_to_proposal_oracle_at_8": {
909
+ "macro_by_seed": {
910
+ "ci": {
911
+ "confidence": 0.95,
912
+ "high": 2.8433985339611567,
913
+ "low": 0.48202118863384635,
914
+ "mean": 1.4657241550082802,
915
+ "n": 35,
916
+ "num_samples": 200
917
+ },
918
+ "mean": 1.4657241550082802,
919
+ "per_seed": {
920
+ "1073161778": 1.0018019999721972,
921
+ "1113371261": 0.7024563821726059,
922
+ "1151410346": 0.9676112241545515,
923
+ "1261930456": 0.005982870275641443,
924
+ "1324133122": 0.7228185479224526,
925
+ "1375492786": 0.015409403984710279,
926
+ "141984615": 0.020065103905333725,
927
+ "1461069713": 0.2324654665550531,
928
+ "1480856810": 0.02403390675499265,
929
+ "1509980664": 2.058415470840539,
930
+ "1549324805": 2.319495421176755,
931
+ "1578990937": 1.3854594546391021,
932
+ "1822210471": 0.031824954963663915,
933
+ "1845489013": 0.37739026714244994,
934
+ "1999022574": 20.468479398320074,
935
+ "2012211287": 0.20884931245481586,
936
+ "2038594191": 1.0336975829706272,
937
+ "2054899126": 0.005904623170763854,
938
+ "2061635888": 0.0,
939
+ "211245842": 1.3695761057579783,
940
+ "278008524": 1.3556794180383116,
941
+ "307068879": 0.32567740299816206,
942
+ "412102563": 0.5347509829497624,
943
+ "441046872": 0.14610429876487893,
944
+ "509430392": 0.0,
945
+ "585904130": 0.7211412341552217,
946
+ "608016290": 3.4869346102323533,
947
+ "626486813": 1.0887571854184497,
948
+ "636680137": 8.71937375135354,
949
+ "655884605": 0.09408265091957171,
950
+ "672328262": 0.09176223920635006,
951
+ "701667486": 0.9997888379154961,
952
+ "747260437": 0.016393451416906714,
953
+ "750448296": 0.22983336179896974,
954
+ "767747424": 0.5383285029875181
955
+ }
956
+ },
957
+ "macro_by_task": {
958
+ "ci": {
959
+ "confidence": 0.95,
960
+ "high": 2.22918321668604,
961
+ "low": 0.2659725754787789,
962
+ "mean": 1.1581784934661314,
963
+ "n": 4,
964
+ "num_samples": 200
965
+ },
966
+ "mean": 1.1581784934661314,
967
+ "per_task": {
968
+ "LiftPegUpright-v1": 1.0338735248844815,
969
+ "PullCube-v1": 0.9248858515826692,
970
+ "PushCube-v1": 0.010005592343544699,
971
+ "StackCube-v1": 2.6639490050538304
972
+ }
973
+ },
974
+ "metric": "ncar_to_proposal_oracle_at_8",
975
+ "micro": {
976
+ "confidence": 0.95,
977
+ "high": 2.5627814309791064,
978
+ "low": 0.5299536435770751,
979
+ "mean": 1.3481356585070305,
980
+ "n": 40,
981
+ "num_samples": 200
982
+ },
983
+ "n": 40
984
+ },
985
+ "outcome_ptr_at_8": {
986
+ "macro_by_seed": {
987
+ "ci": {
988
+ "confidence": 0.95,
989
+ "high": 0.5333333333333333,
990
+ "low": 0.27083333333333337,
991
+ "mean": 0.4020833333333334,
992
+ "n": 40,
993
+ "num_samples": 200
994
+ },
995
+ "mean": 0.4020833333333334,
996
+ "per_seed": {
997
+ "1073161778": 1.0,
998
+ "1113371261": 0.0,
999
+ "1151410346": 1.0,
1000
+ "1261930456": 0.0,
1001
+ "1324133122": 1.0,
1002
+ "1324642411": 0.0,
1003
+ "1375492786": 0.0,
1004
+ "141984615": 0.0,
1005
+ "1461069713": 0.0,
1006
+ "1480856810": 0.0,
1007
+ "1509980664": 1.0,
1008
+ "1522272686": 0.0,
1009
+ "1549324805": 1.0,
1010
+ "1578990937": 1.0,
1011
+ "1747178030": 0.0,
1012
+ "1790188787": 0.0,
1013
+ "1822210471": 0.0,
1014
+ "1845489013": 0.0,
1015
+ "1999022574": 0.0,
1016
+ "2012211287": 0.0,
1017
+ "2038594191": 1.0,
1018
+ "2054899126": 0.0,
1019
+ "2061635888": 0.0,
1020
+ "211245842": 1.0,
1021
+ "278008524": 0.5,
1022
+ "307068879": 0.0,
1023
+ "412102563": 0.0,
1024
+ "440577711": 0.0,
1025
+ "441046872": 0.0,
1026
+ "509430392": 1.0,
1027
+ "585904130": 1.0,
1028
+ "608016290": 1.0,
1029
+ "626486813": 1.0,
1030
+ "636680137": 1.0,
1031
+ "655884605": 1.0,
1032
+ "672328262": 0.0,
1033
+ "701667486": 1.0,
1034
+ "747260437": 0.0,
1035
+ "750448296": 0.3333333333333333,
1036
+ "767747424": 0.25
1037
+ }
1038
+ },
1039
+ "macro_by_task": {
1040
+ "ci": {
1041
+ "confidence": 0.95,
1042
+ "high": 0.4945454545454545,
1043
+ "low": 0.13866666666666666,
1044
+ "mean": 0.3126060606060606,
1045
+ "n": 5,
1046
+ "num_samples": 200
1047
+ },
1048
+ "mean": 0.3126060606060606,
1049
+ "per_task": {
1050
+ "LiftPegUpright-v1": 0.4,
1051
+ "PickCube-v1": 0.0,
1052
+ "PullCube-v1": 0.36,
1053
+ "PushCube-v1": 0.16666666666666666,
1054
+ "StackCube-v1": 0.6363636363636364
1055
+ }
1056
+ },
1057
+ "metric": "outcome_ptr_at_8",
1058
+ "micro": {
1059
+ "confidence": 0.95,
1060
+ "high": 0.5416666666666666,
1061
+ "low": 0.2708333333333333,
1062
+ "mean": 0.3958333333333333,
1063
+ "n": 48,
1064
+ "num_samples": 200
1065
+ },
1066
+ "n": 48
1067
+ },
1068
+ "pairwise_causal_calibration_ece": {
1069
+ "macro_by_seed": {
1070
+ "ci": {
1071
+ "confidence": 0.95,
1072
+ "high": 0.2388753711513772,
1073
+ "low": 0.16096659065885394,
1074
+ "mean": 0.19913886940657743,
1075
+ "n": 40,
1076
+ "num_samples": 200
1077
+ },
1078
+ "mean": 0.19913886940657743,
1079
+ "per_seed": {
1080
+ "1073161778": 0.14432341808925672,
1081
+ "1113371261": 0.18981168502744894,
1082
+ "1151410346": 0.5181919511721376,
1083
+ "1261930456": 0.14694969164298693,
1084
+ "1324133122": 0.2599006738407752,
1085
+ "1324642411": 0.37631581765020666,
1086
+ "1375492786": 0.16288447414215274,
1087
+ "141984615": 0.15133945279036964,
1088
+ "1461069713": 0.3576754247076027,
1089
+ "1480856810": 0.33394552923262694,
1090
+ "1509980664": 0.14023834058751442,
1091
+ "1522272686": 0.12267671398942098,
1092
+ "1549324805": 0.3814015792041445,
1093
+ "1578990937": 0.25510759849883147,
1094
+ "1747178030": 0.0,
1095
+ "1790188787": 0.25625757964299634,
1096
+ "1822210471": 0.15125404295167477,
1097
+ "1845489013": 0.1601652398891617,
1098
+ "1999022574": 0.15346652783738673,
1099
+ "2012211287": 0.19320969574917243,
1100
+ "2038594191": 0.5070876425629134,
1101
+ "2054899126": 0.09662639932716256,
1102
+ "2061635888": 0.1024944972766888,
1103
+ "211245842": 0.11952033712097218,
1104
+ "278008524": 0.05319894873106218,
1105
+ "307068879": 0.07483884431566282,
1106
+ "412102563": 0.2630363607222759,
1107
+ "440577711": 0.0,
1108
+ "441046872": 0.1123926329678895,
1109
+ "509430392": 0.11637116801567593,
1110
+ "585904130": 0.537235764832106,
1111
+ "608016290": 0.098162172719824,
1112
+ "626486813": 0.22930626964762768,
1113
+ "636680137": 0.17961355132445345,
1114
+ "655884605": 0.030740674705558885,
1115
+ "672328262": 0.11992738854206779,
1116
+ "701667486": 0.19081260775351527,
1117
+ "747260437": 0.30546034643967757,
1118
+ "750448296": 0.22679962996282277,
1119
+ "767747424": 0.14681410264927175
1120
+ }
1121
+ },
1122
+ "macro_by_task": {
1123
+ "ci": {
1124
+ "confidence": 0.95,
1125
+ "high": 0.20305275698137817,
1126
+ "low": 0.1395177584868079,
1127
+ "mean": 0.1711908032471828,
1128
+ "n": 5,
1129
+ "num_samples": 200
1130
+ },
1131
+ "mean": 0.1711908032471828,
1132
+ "per_task": {
1133
+ "LiftPegUpright-v1": 0.1245853527683611,
1134
+ "PickCube-v1": 0.12267671398942098,
1135
+ "PullCube-v1": 0.20159853773134337,
1136
+ "PushCube-v1": 0.20497329769741546,
1137
+ "StackCube-v1": 0.2021201140493732
1138
+ }
1139
+ },
1140
+ "metric": "pairwise_causal_calibration_ece",
1141
+ "micro": {
1142
+ "confidence": 0.95,
1143
+ "high": 0.22492210423240874,
1144
+ "low": 0.15483929696829266,
1145
+ "mean": 0.19247349920505022,
1146
+ "n": 48,
1147
+ "num_samples": 200
1148
+ },
1149
+ "n": 48
1150
+ },
1151
+ "proposal_oracle_safety_label_known_at_8": {
1152
+ "macro_by_seed": {
1153
+ "ci": {
1154
+ "confidence": 0.95,
1155
+ "high": 1.0,
1156
+ "low": 1.0,
1157
+ "mean": 1.0,
1158
+ "n": 40,
1159
+ "num_samples": 200
1160
+ },
1161
+ "mean": 1.0,
1162
+ "per_seed": {
1163
+ "1073161778": 1.0,
1164
+ "1113371261": 1.0,
1165
+ "1151410346": 1.0,
1166
+ "1261930456": 1.0,
1167
+ "1324133122": 1.0,
1168
+ "1324642411": 1.0,
1169
+ "1375492786": 1.0,
1170
+ "141984615": 1.0,
1171
+ "1461069713": 1.0,
1172
+ "1480856810": 1.0,
1173
+ "1509980664": 1.0,
1174
+ "1522272686": 1.0,
1175
+ "1549324805": 1.0,
1176
+ "1578990937": 1.0,
1177
+ "1747178030": 1.0,
1178
+ "1790188787": 1.0,
1179
+ "1822210471": 1.0,
1180
+ "1845489013": 1.0,
1181
+ "1999022574": 1.0,
1182
+ "2012211287": 1.0,
1183
+ "2038594191": 1.0,
1184
+ "2054899126": 1.0,
1185
+ "2061635888": 1.0,
1186
+ "211245842": 1.0,
1187
+ "278008524": 1.0,
1188
+ "307068879": 1.0,
1189
+ "412102563": 1.0,
1190
+ "440577711": 1.0,
1191
+ "441046872": 1.0,
1192
+ "509430392": 1.0,
1193
+ "585904130": 1.0,
1194
+ "608016290": 1.0,
1195
+ "626486813": 1.0,
1196
+ "636680137": 1.0,
1197
+ "655884605": 1.0,
1198
+ "672328262": 1.0,
1199
+ "701667486": 1.0,
1200
+ "747260437": 1.0,
1201
+ "750448296": 1.0,
1202
+ "767747424": 1.0
1203
+ }
1204
+ },
1205
+ "macro_by_task": {
1206
+ "ci": {
1207
+ "confidence": 0.95,
1208
+ "high": 1.0,
1209
+ "low": 1.0,
1210
+ "mean": 1.0,
1211
+ "n": 5,
1212
+ "num_samples": 200
1213
+ },
1214
+ "mean": 1.0,
1215
+ "per_task": {
1216
+ "LiftPegUpright-v1": 1.0,
1217
+ "PickCube-v1": 1.0,
1218
+ "PullCube-v1": 1.0,
1219
+ "PushCube-v1": 1.0,
1220
+ "StackCube-v1": 1.0
1221
+ }
1222
+ },
1223
+ "metric": "proposal_oracle_safety_label_known_at_8",
1224
+ "micro": {
1225
+ "confidence": 0.95,
1226
+ "high": 1.0,
1227
+ "low": 1.0,
1228
+ "mean": 1.0,
1229
+ "n": 48,
1230
+ "num_samples": 200
1231
+ },
1232
+ "n": 48
1233
+ },
1234
+ "proposal_oracle_success_at_8": {
1235
+ "macro_by_seed": {
1236
+ "ci": {
1237
+ "confidence": 0.95,
1238
+ "high": 0.35,
1239
+ "low": 0.09375,
1240
+ "mean": 0.23125,
1241
+ "n": 40,
1242
+ "num_samples": 200
1243
+ },
1244
+ "mean": 0.23125,
1245
+ "per_seed": {
1246
+ "1073161778": 1.0,
1247
+ "1113371261": 0.0,
1248
+ "1151410346": 0.0,
1249
+ "1261930456": 0.0,
1250
+ "1324133122": 0.0,
1251
+ "1324642411": 1.0,
1252
+ "1375492786": 0.0,
1253
+ "141984615": 0.0,
1254
+ "1461069713": 0.0,
1255
+ "1480856810": 0.0,
1256
+ "1509980664": 0.0,
1257
+ "1522272686": 1.0,
1258
+ "1549324805": 0.0,
1259
+ "1578990937": 0.0,
1260
+ "1747178030": 1.0,
1261
+ "1790188787": 1.0,
1262
+ "1822210471": 0.0,
1263
+ "1845489013": 0.0,
1264
+ "1999022574": 0.0,
1265
+ "2012211287": 0.0,
1266
+ "2038594191": 0.0,
1267
+ "2054899126": 0.5,
1268
+ "2061635888": 0.0,
1269
+ "211245842": 0.0,
1270
+ "278008524": 0.5,
1271
+ "307068879": 0.0,
1272
+ "412102563": 0.0,
1273
+ "440577711": 1.0,
1274
+ "441046872": 0.0,
1275
+ "509430392": 1.0,
1276
+ "585904130": 0.0,
1277
+ "608016290": 0.0,
1278
+ "626486813": 0.0,
1279
+ "636680137": 0.0,
1280
+ "655884605": 0.0,
1281
+ "672328262": 0.0,
1282
+ "701667486": 1.0,
1283
+ "747260437": 0.0,
1284
+ "750448296": 0.0,
1285
+ "767747424": 0.25
1286
+ }
1287
+ },
1288
+ "macro_by_task": {
1289
+ "ci": {
1290
+ "confidence": 0.95,
1291
+ "high": 0.7333333333333333,
1292
+ "low": 0.1461818181818182,
1293
+ "mean": 0.42351515151515146,
1294
+ "n": 5,
1295
+ "num_samples": 200
1296
+ },
1297
+ "mean": 0.42351515151515146,
1298
+ "per_task": {
1299
+ "LiftPegUpright-v1": 0.2,
1300
+ "PickCube-v1": 1.0,
1301
+ "PullCube-v1": 0.16,
1302
+ "PushCube-v1": 0.6666666666666666,
1303
+ "StackCube-v1": 0.09090909090909091
1304
+ }
1305
+ },
1306
+ "metric": "proposal_oracle_success_at_8",
1307
+ "micro": {
1308
+ "confidence": 0.95,
1309
+ "high": 0.3333333333333333,
1310
+ "low": 0.10416666666666667,
1311
+ "mean": 0.22916666666666666,
1312
+ "n": 48,
1313
+ "num_samples": 200
1314
+ },
1315
+ "n": 48
1316
+ },
1317
+ "proposal_oracle_success_gain_over_base_at_8": {
1318
+ "macro_by_seed": {
1319
+ "ci": {
1320
+ "confidence": 0.95,
1321
+ "high": 0.0625,
1322
+ "low": -0.19583333333333336,
1323
+ "mean": -0.07708333333333334,
1324
+ "n": 40,
1325
+ "num_samples": 200
1326
+ },
1327
+ "mean": -0.07708333333333334,
1328
+ "per_seed": {
1329
+ "1073161778": 1.0,
1330
+ "1113371261": 0.0,
1331
+ "1151410346": 0.0,
1332
+ "1261930456": -1.0,
1333
+ "1324133122": 0.0,
1334
+ "1324642411": 0.0,
1335
+ "1375492786": -1.0,
1336
+ "141984615": 0.0,
1337
+ "1461069713": 0.0,
1338
+ "1480856810": -1.0,
1339
+ "1509980664": 0.0,
1340
+ "1522272686": 0.0,
1341
+ "1549324805": 0.0,
1342
+ "1578990937": 0.0,
1343
+ "1747178030": 0.0,
1344
+ "1790188787": 0.0,
1345
+ "1822210471": 0.0,
1346
+ "1845489013": 0.0,
1347
+ "1999022574": 0.0,
1348
+ "2012211287": 0.0,
1349
+ "2038594191": 0.0,
1350
+ "2054899126": -0.5,
1351
+ "2061635888": 0.0,
1352
+ "211245842": 0.0,
1353
+ "278008524": 0.0,
1354
+ "307068879": 0.0,
1355
+ "412102563": -1.0,
1356
+ "440577711": 0.0,
1357
+ "441046872": 0.0,
1358
+ "509430392": 1.0,
1359
+ "585904130": 0.0,
1360
+ "608016290": 0.0,
1361
+ "626486813": 0.0,
1362
+ "636680137": 0.0,
1363
+ "655884605": 0.0,
1364
+ "672328262": 0.0,
1365
+ "701667486": 1.0,
1366
+ "747260437": -1.0,
1367
+ "750448296": -0.3333333333333333,
1368
+ "767747424": -0.25
1369
+ }
1370
+ },
1371
+ "macro_by_task": {
1372
+ "ci": {
1373
+ "confidence": 0.95,
1374
+ "high": 0.08,
1375
+ "low": -0.14666666666666667,
1376
+ "mean": -0.03333333333333333,
1377
+ "n": 5,
1378
+ "num_samples": 200
1379
+ },
1380
+ "mean": -0.03333333333333333,
1381
+ "per_task": {
1382
+ "LiftPegUpright-v1": 0.2,
1383
+ "PickCube-v1": 0.0,
1384
+ "PullCube-v1": -0.2,
1385
+ "PushCube-v1": -0.16666666666666666,
1386
+ "StackCube-v1": 0.0
1387
+ }
1388
+ },
1389
+ "metric": "proposal_oracle_success_gain_over_base_at_8",
1390
+ "micro": {
1391
+ "confidence": 0.95,
1392
+ "high": 0.020833333333333332,
1393
+ "low": -0.22916666666666666,
1394
+ "mean": -0.10416666666666667,
1395
+ "n": 48,
1396
+ "num_samples": 200
1397
+ },
1398
+ "n": 48
1399
+ },
1400
+ "proposal_oracle_unsafe_known_at_8": {
1401
+ "macro_by_seed": {
1402
+ "ci": {
1403
+ "confidence": 0.95,
1404
+ "high": 0.35416666666666663,
1405
+ "low": 0.11666666666666667,
1406
+ "mean": 0.24791666666666665,
1407
+ "n": 40,
1408
+ "num_samples": 200
1409
+ },
1410
+ "mean": 0.24791666666666665,
1411
+ "per_seed": {
1412
+ "1073161778": 0.0,
1413
+ "1113371261": 0.0,
1414
+ "1151410346": 0.0,
1415
+ "1261930456": 0.0,
1416
+ "1324133122": 0.0,
1417
+ "1324642411": 0.0,
1418
+ "1375492786": 1.0,
1419
+ "141984615": 1.0,
1420
+ "1461069713": 0.0,
1421
+ "1480856810": 0.0,
1422
+ "1509980664": 0.0,
1423
+ "1522272686": 0.0,
1424
+ "1549324805": 0.0,
1425
+ "1578990937": 0.0,
1426
+ "1747178030": 1.0,
1427
+ "1790188787": 0.0,
1428
+ "1822210471": 0.0,
1429
+ "1845489013": 1.0,
1430
+ "1999022574": 0.0,
1431
+ "2012211287": 0.0,
1432
+ "2038594191": 1.0,
1433
+ "2054899126": 1.0,
1434
+ "2061635888": 0.0,
1435
+ "211245842": 0.0,
1436
+ "278008524": 0.5,
1437
+ "307068879": 0.0,
1438
+ "412102563": 0.0,
1439
+ "440577711": 0.0,
1440
+ "441046872": 0.0,
1441
+ "509430392": 0.0,
1442
+ "585904130": 1.0,
1443
+ "608016290": 0.0,
1444
+ "626486813": 0.0,
1445
+ "636680137": 0.0,
1446
+ "655884605": 0.0,
1447
+ "672328262": 0.0,
1448
+ "701667486": 0.0,
1449
+ "747260437": 1.0,
1450
+ "750448296": 0.6666666666666666,
1451
+ "767747424": 0.75
1452
+ }
1453
+ },
1454
+ "macro_by_task": {
1455
+ "ci": {
1456
+ "confidence": 0.95,
1457
+ "high": 0.36,
1458
+ "low": 0.0,
1459
+ "mean": 0.12,
1460
+ "n": 5,
1461
+ "num_samples": 200
1462
+ },
1463
+ "mean": 0.12,
1464
+ "per_task": {
1465
+ "LiftPegUpright-v1": 0.0,
1466
+ "PickCube-v1": 0.0,
1467
+ "PullCube-v1": 0.6,
1468
+ "PushCube-v1": 0.0,
1469
+ "StackCube-v1": 0.0
1470
+ }
1471
+ },
1472
+ "metric": "proposal_oracle_unsafe_known_at_8",
1473
+ "micro": {
1474
+ "confidence": 0.95,
1475
+ "high": 0.4583333333333333,
1476
+ "low": 0.1875,
1477
+ "mean": 0.3125,
1478
+ "n": 48,
1479
+ "num_samples": 200
1480
+ },
1481
+ "n": 48
1482
+ },
1483
+ "proposal_oracle_utility_at_8": {
1484
+ "macro_by_seed": {
1485
+ "ci": {
1486
+ "confidence": 0.95,
1487
+ "high": 0.9234885451151058,
1488
+ "low": 0.41627855848055334,
1489
+ "mean": 0.6670806414214894,
1490
+ "n": 40,
1491
+ "num_samples": 200
1492
+ },
1493
+ "mean": 0.6670806414214894,
1494
+ "per_seed": {
1495
+ "1073161778": 2.0,
1496
+ "1113371261": 0.1003647893667221,
1497
+ "1151410346": 0.5630103945732117,
1498
+ "1261930456": 0.08545441925525665,
1499
+ "1324133122": 0.09815146028995514,
1500
+ "1324642411": 2.0,
1501
+ "1375492786": 0.19452734291553497,
1502
+ "141984615": 0.04447716474533081,
1503
+ "1461069713": 0.11542954295873642,
1504
+ "1480856810": 0.19622710347175598,
1505
+ "1509980664": 0.83889240026474,
1506
+ "1522272686": 2.0,
1507
+ "1549324805": 0.21403594315052032,
1508
+ "1578990937": 0.23259200155735016,
1509
+ "1747178030": 2.0,
1510
+ "1790188787": 2.0,
1511
+ "1822210471": 0.058819372206926346,
1512
+ "1845489013": 0.0861951932311058,
1513
+ "1999022574": 0.7970014810562134,
1514
+ "2012211287": 0.7270426154136658,
1515
+ "2038594191": 0.24880830943584442,
1516
+ "2054899126": 1.0683744624257088,
1517
+ "2061635888": 0.07902413606643677,
1518
+ "211245842": 0.2142850160598755,
1519
+ "278008524": 1.157032236456871,
1520
+ "307068879": 0.2035302072763443,
1521
+ "412102563": 0.8275061845779419,
1522
+ "440577711": 2.0,
1523
+ "441046872": 0.2042645812034607,
1524
+ "509430392": 2.0,
1525
+ "585904130": 0.2938849627971649,
1526
+ "608016290": 0.5705108642578125,
1527
+ "626486813": 0.08006523177027702,
1528
+ "636680137": 0.1247115358710289,
1529
+ "655884605": 0.17242512106895447,
1530
+ "672328262": 0.07216622680425644,
1531
+ "701667486": 2.0,
1532
+ "747260437": 0.18527252972126007,
1533
+ "750448296": 0.2481476068496704,
1534
+ "767747424": 0.5809952197596431
1535
+ }
1536
+ },
1537
+ "macro_by_task": {
1538
+ "ci": {
1539
+ "confidence": 0.95,
1540
+ "high": 1.5042242030302684,
1541
+ "low": 0.5618420226519758,
1542
+ "mean": 1.0225070557774918,
1543
+ "n": 5,
1544
+ "num_samples": 200
1545
+ },
1546
+ "mean": 1.0225070557774918,
1547
+ "per_task": {
1548
+ "LiftPegUpright-v1": 0.7410696536302567,
1549
+ "PickCube-v1": 2.0,
1550
+ "PullCube-v1": 0.45270782738924026,
1551
+ "PushCube-v1": 1.3802802537878354,
1552
+ "StackCube-v1": 0.5384775440801274
1553
+ }
1554
+ },
1555
+ "metric": "proposal_oracle_utility_at_8",
1556
+ "micro": {
1557
+ "confidence": 0.95,
1558
+ "high": 0.8419985768850893,
1559
+ "low": 0.43616288745154935,
1560
+ "mean": 0.6505828845935563,
1561
+ "n": 48,
1562
+ "num_samples": 200
1563
+ },
1564
+ "n": 48
1565
+ },
1566
+ "proposal_oracle_utility_gain_over_base_at_8": {
1567
+ "macro_by_seed": {
1568
+ "ci": {
1569
+ "confidence": 0.95,
1570
+ "high": 0.11130315090995282,
1571
+ "low": -0.35443208372065177,
1572
+ "mean": -0.1357228347643589,
1573
+ "n": 40,
1574
+ "num_samples": 200
1575
+ },
1576
+ "mean": -0.1357228347643589,
1577
+ "per_seed": {
1578
+ "1073161778": 1.274482011795044,
1579
+ "1113371261": -0.010734498500823975,
1580
+ "1151410346": 0.3345368802547455,
1581
+ "1261930456": -1.9145455807447433,
1582
+ "1324133122": 0.013924404978752136,
1583
+ "1324642411": 0.0,
1584
+ "1375492786": -1.805472657084465,
1585
+ "141984615": -0.12335556745529175,
1586
+ "1461069713": -0.11309804767370224,
1587
+ "1480856810": -1.803772896528244,
1588
+ "1509980664": 0.006812572479248047,
1589
+ "1522272686": 0.0,
1590
+ "1549324805": 0.04154062271118164,
1591
+ "1578990937": 0.018326029181480408,
1592
+ "1747178030": 0.0,
1593
+ "1790188787": 0.0,
1594
+ "1822210471": -0.03560608997941017,
1595
+ "1845489013": -0.026362814009189606,
1596
+ "1999022574": -0.030967414379119873,
1597
+ "2012211287": -0.001204073429107666,
1598
+ "2038594191": 0.15376771241426468,
1599
+ "2054899126": -0.9316255375742912,
1600
+ "2061635888": -0.007182791829109192,
1601
+ "211245842": 0.029980570077896118,
1602
+ "278008524": 0.03202502429485321,
1603
+ "307068879": -0.005622148513793945,
1604
+ "412102563": -1.172493815422058,
1605
+ "440577711": 0.0,
1606
+ "441046872": -0.005503058433532715,
1607
+ "509430392": 1.9592653959989548,
1608
+ "585904130": 0.22284842282533646,
1609
+ "608016290": 0.0451507568359375,
1610
+ "626486813": 0.03832433745265007,
1611
+ "636680137": 0.005137942731380463,
1612
+ "655884605": 0.028349727392196655,
1613
+ "672328262": -0.03376801311969757,
1614
+ "701667486": 1.208725929260254,
1615
+ "747260437": -1.81472747027874,
1616
+ "750448296": -0.512355720003446,
1617
+ "767747424": -0.4937135362997651
1618
+ }
1619
+ },
1620
+ "macro_by_task": {
1621
+ "ci": {
1622
+ "confidence": 0.95,
1623
+ "high": 0.09439109063690358,
1624
+ "low": -0.25748731927076973,
1625
+ "mean": -0.06882829113259462,
1626
+ "n": 5,
1627
+ "num_samples": 200
1628
+ },
1629
+ "mean": -0.06882829113259462,
1630
+ "per_task": {
1631
+ "LiftPegUpright-v1": 0.2614606976509094,
1632
+ "PickCube-v1": 0.0,
1633
+ "PullCube-v1": -0.35054278463125227,
1634
+ "PushCube-v1": -0.2931755135456721,
1635
+ "StackCube-v1": 0.03811614486304196
1636
+ }
1637
+ },
1638
+ "metric": "proposal_oracle_utility_gain_over_base_at_8",
1639
+ "micro": {
1640
+ "confidence": 0.95,
1641
+ "high": 0.03781239866899947,
1642
+ "low": -0.4132526982575655,
1643
+ "mean": -0.1832508669855694,
1644
+ "n": 48,
1645
+ "num_samples": 200
1646
+ },
1647
+ "n": 48
1648
+ },
1649
+ "selected_safety_label_known_at_8": {
1650
+ "macro_by_seed": {
1651
+ "ci": {
1652
+ "confidence": 0.95,
1653
+ "high": 1.0,
1654
+ "low": 1.0,
1655
+ "mean": 1.0,
1656
+ "n": 40,
1657
+ "num_samples": 200
1658
+ },
1659
+ "mean": 1.0,
1660
+ "per_seed": {
1661
+ "1073161778": 1.0,
1662
+ "1113371261": 1.0,
1663
+ "1151410346": 1.0,
1664
+ "1261930456": 1.0,
1665
+ "1324133122": 1.0,
1666
+ "1324642411": 1.0,
1667
+ "1375492786": 1.0,
1668
+ "141984615": 1.0,
1669
+ "1461069713": 1.0,
1670
+ "1480856810": 1.0,
1671
+ "1509980664": 1.0,
1672
+ "1522272686": 1.0,
1673
+ "1549324805": 1.0,
1674
+ "1578990937": 1.0,
1675
+ "1747178030": 1.0,
1676
+ "1790188787": 1.0,
1677
+ "1822210471": 1.0,
1678
+ "1845489013": 1.0,
1679
+ "1999022574": 1.0,
1680
+ "2012211287": 1.0,
1681
+ "2038594191": 1.0,
1682
+ "2054899126": 1.0,
1683
+ "2061635888": 1.0,
1684
+ "211245842": 1.0,
1685
+ "278008524": 1.0,
1686
+ "307068879": 1.0,
1687
+ "412102563": 1.0,
1688
+ "440577711": 1.0,
1689
+ "441046872": 1.0,
1690
+ "509430392": 1.0,
1691
+ "585904130": 1.0,
1692
+ "608016290": 1.0,
1693
+ "626486813": 1.0,
1694
+ "636680137": 1.0,
1695
+ "655884605": 1.0,
1696
+ "672328262": 1.0,
1697
+ "701667486": 1.0,
1698
+ "747260437": 1.0,
1699
+ "750448296": 1.0,
1700
+ "767747424": 1.0
1701
+ }
1702
+ },
1703
+ "macro_by_task": {
1704
+ "ci": {
1705
+ "confidence": 0.95,
1706
+ "high": 1.0,
1707
+ "low": 1.0,
1708
+ "mean": 1.0,
1709
+ "n": 5,
1710
+ "num_samples": 200
1711
+ },
1712
+ "mean": 1.0,
1713
+ "per_task": {
1714
+ "LiftPegUpright-v1": 1.0,
1715
+ "PickCube-v1": 1.0,
1716
+ "PullCube-v1": 1.0,
1717
+ "PushCube-v1": 1.0,
1718
+ "StackCube-v1": 1.0
1719
+ }
1720
+ },
1721
+ "metric": "selected_safety_label_known_at_8",
1722
+ "micro": {
1723
+ "confidence": 0.95,
1724
+ "high": 1.0,
1725
+ "low": 1.0,
1726
+ "mean": 1.0,
1727
+ "n": 48,
1728
+ "num_samples": 200
1729
+ },
1730
+ "n": 48
1731
+ },
1732
+ "selected_success_at_8": {
1733
+ "macro_by_seed": {
1734
+ "ci": {
1735
+ "confidence": 0.95,
1736
+ "high": 0.2375,
1737
+ "low": 0.05,
1738
+ "mean": 0.125,
1739
+ "n": 40,
1740
+ "num_samples": 200
1741
+ },
1742
+ "mean": 0.125,
1743
+ "per_seed": {
1744
+ "1073161778": 0.0,
1745
+ "1113371261": 0.0,
1746
+ "1151410346": 0.0,
1747
+ "1261930456": 0.0,
1748
+ "1324133122": 0.0,
1749
+ "1324642411": 0.0,
1750
+ "1375492786": 0.0,
1751
+ "141984615": 0.0,
1752
+ "1461069713": 0.0,
1753
+ "1480856810": 0.0,
1754
+ "1509980664": 0.0,
1755
+ "1522272686": 0.0,
1756
+ "1549324805": 0.0,
1757
+ "1578990937": 0.0,
1758
+ "1747178030": 1.0,
1759
+ "1790188787": 1.0,
1760
+ "1822210471": 0.0,
1761
+ "1845489013": 0.0,
1762
+ "1999022574": 0.0,
1763
+ "2012211287": 0.0,
1764
+ "2038594191": 0.0,
1765
+ "2054899126": 0.5,
1766
+ "2061635888": 0.0,
1767
+ "211245842": 0.0,
1768
+ "278008524": 0.5,
1769
+ "307068879": 0.0,
1770
+ "412102563": 0.0,
1771
+ "440577711": 1.0,
1772
+ "441046872": 0.0,
1773
+ "509430392": 1.0,
1774
+ "585904130": 0.0,
1775
+ "608016290": 0.0,
1776
+ "626486813": 0.0,
1777
+ "636680137": 0.0,
1778
+ "655884605": 0.0,
1779
+ "672328262": 0.0,
1780
+ "701667486": 0.0,
1781
+ "747260437": 0.0,
1782
+ "750448296": 0.0,
1783
+ "767747424": 0.0
1784
+ }
1785
+ },
1786
+ "macro_by_task": {
1787
+ "ci": {
1788
+ "confidence": 0.95,
1789
+ "high": 0.324,
1790
+ "low": 0.0,
1791
+ "mean": 0.124,
1792
+ "n": 5,
1793
+ "num_samples": 200
1794
+ },
1795
+ "mean": 0.124,
1796
+ "per_task": {
1797
+ "LiftPegUpright-v1": 0.0,
1798
+ "PickCube-v1": 0.0,
1799
+ "PullCube-v1": 0.12,
1800
+ "PushCube-v1": 0.5,
1801
+ "StackCube-v1": 0.0
1802
+ }
1803
+ },
1804
+ "metric": "selected_success_at_8",
1805
+ "micro": {
1806
+ "confidence": 0.95,
1807
+ "high": 0.22916666666666666,
1808
+ "low": 0.041666666666666664,
1809
+ "mean": 0.125,
1810
+ "n": 48,
1811
+ "num_samples": 200
1812
+ },
1813
+ "n": 48
1814
+ },
1815
+ "selected_success_gain_over_base_at_8": {
1816
+ "macro_by_seed": {
1817
+ "ci": {
1818
+ "confidence": 0.95,
1819
+ "high": -0.025,
1820
+ "low": -0.2875,
1821
+ "mean": -0.18333333333333332,
1822
+ "n": 40,
1823
+ "num_samples": 200
1824
+ },
1825
+ "mean": -0.18333333333333332,
1826
+ "per_seed": {
1827
+ "1073161778": 0.0,
1828
+ "1113371261": 0.0,
1829
+ "1151410346": 0.0,
1830
+ "1261930456": -1.0,
1831
+ "1324133122": 0.0,
1832
+ "1324642411": -1.0,
1833
+ "1375492786": -1.0,
1834
+ "141984615": 0.0,
1835
+ "1461069713": 0.0,
1836
+ "1480856810": -1.0,
1837
+ "1509980664": 0.0,
1838
+ "1522272686": -1.0,
1839
+ "1549324805": 0.0,
1840
+ "1578990937": 0.0,
1841
+ "1747178030": 0.0,
1842
+ "1790188787": 0.0,
1843
+ "1822210471": 0.0,
1844
+ "1845489013": 0.0,
1845
+ "1999022574": 0.0,
1846
+ "2012211287": 0.0,
1847
+ "2038594191": 0.0,
1848
+ "2054899126": -0.5,
1849
+ "2061635888": 0.0,
1850
+ "211245842": 0.0,
1851
+ "278008524": 0.0,
1852
+ "307068879": 0.0,
1853
+ "412102563": -1.0,
1854
+ "440577711": 0.0,
1855
+ "441046872": 0.0,
1856
+ "509430392": 1.0,
1857
+ "585904130": 0.0,
1858
+ "608016290": 0.0,
1859
+ "626486813": 0.0,
1860
+ "636680137": 0.0,
1861
+ "655884605": 0.0,
1862
+ "672328262": 0.0,
1863
+ "701667486": 0.0,
1864
+ "747260437": -1.0,
1865
+ "750448296": -0.3333333333333333,
1866
+ "767747424": -0.5
1867
+ }
1868
+ },
1869
+ "macro_by_task": {
1870
+ "ci": {
1871
+ "confidence": 0.95,
1872
+ "high": -0.07272727272727272,
1873
+ "low": -0.6661818181818182,
1874
+ "mean": -0.33284848484848484,
1875
+ "n": 5,
1876
+ "num_samples": 200
1877
+ },
1878
+ "mean": -0.33284848484848484,
1879
+ "per_task": {
1880
+ "LiftPegUpright-v1": 0.0,
1881
+ "PickCube-v1": -1.0,
1882
+ "PullCube-v1": -0.24,
1883
+ "PushCube-v1": -0.3333333333333333,
1884
+ "StackCube-v1": -0.09090909090909091
1885
+ }
1886
+ },
1887
+ "metric": "selected_success_gain_over_base_at_8",
1888
+ "micro": {
1889
+ "confidence": 0.95,
1890
+ "high": -0.08333333333333333,
1891
+ "low": -0.3125,
1892
+ "mean": -0.20833333333333334,
1893
+ "n": 48,
1894
+ "num_samples": 200
1895
+ },
1896
+ "n": 48
1897
+ },
1898
+ "selected_unsafe_known_at_8": {
1899
+ "macro_by_seed": {
1900
+ "ci": {
1901
+ "confidence": 0.95,
1902
+ "high": 0.31875,
1903
+ "low": 0.075,
1904
+ "mean": 0.21875,
1905
+ "n": 40,
1906
+ "num_samples": 200
1907
+ },
1908
+ "mean": 0.21875,
1909
+ "per_seed": {
1910
+ "1073161778": 0.0,
1911
+ "1113371261": 0.0,
1912
+ "1151410346": 0.0,
1913
+ "1261930456": 0.0,
1914
+ "1324133122": 0.0,
1915
+ "1324642411": 0.0,
1916
+ "1375492786": 1.0,
1917
+ "141984615": 1.0,
1918
+ "1461069713": 1.0,
1919
+ "1480856810": 0.0,
1920
+ "1509980664": 0.0,
1921
+ "1522272686": 0.0,
1922
+ "1549324805": 0.0,
1923
+ "1578990937": 0.0,
1924
+ "1747178030": 1.0,
1925
+ "1790188787": 0.0,
1926
+ "1822210471": 0.0,
1927
+ "1845489013": 1.0,
1928
+ "1999022574": 0.0,
1929
+ "2012211287": 0.0,
1930
+ "2038594191": 0.0,
1931
+ "2054899126": 1.0,
1932
+ "2061635888": 0.0,
1933
+ "211245842": 0.0,
1934
+ "278008524": 0.0,
1935
+ "307068879": 0.0,
1936
+ "412102563": 0.0,
1937
+ "440577711": 0.0,
1938
+ "441046872": 0.0,
1939
+ "509430392": 0.0,
1940
+ "585904130": 0.0,
1941
+ "608016290": 0.0,
1942
+ "626486813": 0.0,
1943
+ "636680137": 0.0,
1944
+ "655884605": 0.0,
1945
+ "672328262": 0.0,
1946
+ "701667486": 0.0,
1947
+ "747260437": 1.0,
1948
+ "750448296": 1.0,
1949
+ "767747424": 0.75
1950
+ }
1951
+ },
1952
+ "macro_by_task": {
1953
+ "ci": {
1954
+ "confidence": 0.95,
1955
+ "high": 0.336,
1956
+ "low": 0.0,
1957
+ "mean": 0.11200000000000002,
1958
+ "n": 5,
1959
+ "num_samples": 200
1960
+ },
1961
+ "mean": 0.11200000000000002,
1962
+ "per_task": {
1963
+ "LiftPegUpright-v1": 0.0,
1964
+ "PickCube-v1": 0.0,
1965
+ "PullCube-v1": 0.56,
1966
+ "PushCube-v1": 0.0,
1967
+ "StackCube-v1": 0.0
1968
+ }
1969
+ },
1970
+ "metric": "selected_unsafe_known_at_8",
1971
+ "micro": {
1972
+ "confidence": 0.95,
1973
+ "high": 0.4375,
1974
+ "low": 0.1875,
1975
+ "mean": 0.2916666666666667,
1976
+ "n": 48,
1977
+ "num_samples": 200
1978
+ },
1979
+ "n": 48
1980
+ },
1981
+ "selected_utility_at_8": {
1982
+ "macro_by_seed": {
1983
+ "ci": {
1984
+ "confidence": 0.95,
1985
+ "high": 0.6775430737063288,
1986
+ "low": 0.2742155750592549,
1987
+ "mean": 0.446481911946709,
1988
+ "n": 40,
1989
+ "num_samples": 200
1990
+ },
1991
+ "mean": 0.446481911946709,
1992
+ "per_seed": {
1993
+ "1073161778": 0.7232213616371155,
1994
+ "1113371261": 0.09282426536083221,
1995
+ "1151410346": 0.23930874466896057,
1996
+ "1261930456": 0.07399994134902954,
1997
+ "1324133122": 0.08808663487434387,
1998
+ "1324642411": 0.14195962250232697,
1999
+ "1375492786": 0.16670608520507812,
2000
+ "141984615": 0.04200202226638794,
2001
+ "1461069713": 0.08913815021514893,
2002
+ "1480856810": 0.1528753936290741,
2003
+ "1509980664": 0.8248692750930786,
2004
+ "1522272686": 0.5436285138130188,
2005
+ "1549324805": 0.11768263578414917,
2006
+ "1578990937": 0.2072020173072815,
2007
+ "1747178030": 2.0,
2008
+ "1790188787": 2.0,
2009
+ "1822210471": 0.0576862096786499,
2010
+ "1845489013": 0.07624612003564835,
2011
+ "1999022574": 0.16314539313316345,
2012
+ "2012211287": 0.7267911434173584,
2013
+ "2038594191": 0.08985898643732071,
2014
+ "2054899126": 1.0628735646605492,
2015
+ "2061635888": 0.07902413606643677,
2016
+ "211245842": 0.1732243299484253,
2017
+ "278008524": 1.1136165633797646,
2018
+ "307068879": 0.20169919729232788,
2019
+ "412102563": 0.20051395893096924,
2020
+ "440577711": 2.0,
2021
+ "441046872": 0.20346055924892426,
2022
+ "509430392": 2.0,
2023
+ "585904130": 0.13317976891994476,
2024
+ "608016290": 0.4130730926990509,
2025
+ "626486813": 0.07703944854438305,
2026
+ "636680137": 0.07991180568933487,
2027
+ "655884605": 0.1697579026222229,
2028
+ "672328262": 0.06906759738922119,
2029
+ "701667486": 0.7915292978286743,
2030
+ "747260437": 0.155522882938385,
2031
+ "750448296": 0.19667073587576547,
2032
+ "767747424": 0.12187911942601204
2033
+ }
2034
+ },
2035
+ "macro_by_task": {
2036
+ "ci": {
2037
+ "confidence": 0.95,
2038
+ "high": 0.7969060307443143,
2039
+ "low": 0.33225244613127275,
2040
+ "mean": 0.5361242991749086,
2041
+ "n": 5,
2042
+ "num_samples": 200
2043
+ },
2044
+ "mean": 0.5361242991749086,
2045
+ "per_task": {
2046
+ "LiftPegUpright-v1": 0.4536490708589554,
2047
+ "PickCube-v1": 0.5436285138130188,
2048
+ "PullCube-v1": 0.3464636041224003,
2049
+ "PushCube-v1": 1.0614724929134052,
2050
+ "StackCube-v1": 0.27540781416676263
2051
+ }
2052
+ },
2053
+ "metric": "selected_utility_at_8",
2054
+ "micro": {
2055
+ "confidence": 0.95,
2056
+ "high": 0.6220090012066066,
2057
+ "low": 0.26391754842673737,
2058
+ "mean": 0.43482885176005465,
2059
+ "n": 48,
2060
+ "num_samples": 200
2061
+ },
2062
+ "n": 48
2063
+ },
2064
+ "selected_utility_gain_over_base_at_8": {
2065
+ "macro_by_seed": {
2066
+ "ci": {
2067
+ "confidence": 0.95,
2068
+ "high": -0.07615911271423101,
2069
+ "low": -0.5524518134072423,
2070
+ "mean": -0.35632156423913935,
2071
+ "n": 40,
2072
+ "num_samples": 200
2073
+ },
2074
+ "mean": -0.35632156423913935,
2075
+ "per_seed": {
2076
+ "1073161778": -0.002296626567840576,
2077
+ "1113371261": -0.018275022506713867,
2078
+ "1151410346": 0.010835230350494385,
2079
+ "1261930456": -1.9260000586509705,
2080
+ "1324133122": 0.003859579563140869,
2081
+ "1324642411": -1.858040377497673,
2082
+ "1375492786": -1.8332939147949219,
2083
+ "141984615": -0.12583070993423462,
2084
+ "1461069713": -0.13938944041728973,
2085
+ "1480856810": -1.847124606370926,
2086
+ "1509980664": -0.00721055269241333,
2087
+ "1522272686": -1.4563714861869812,
2088
+ "1549324805": -0.054812684655189514,
2089
+ "1578990937": -0.007063955068588257,
2090
+ "1747178030": 0.0,
2091
+ "1790188787": 0.0,
2092
+ "1822210471": -0.036739252507686615,
2093
+ "1845489013": -0.036311887204647064,
2094
+ "1999022574": -0.6648235023021698,
2095
+ "2012211287": -0.001455545425415039,
2096
+ "2038594191": -0.005181610584259033,
2097
+ "2054899126": -0.9371264353394508,
2098
+ "2061635888": -0.007182791829109192,
2099
+ "211245842": -0.011080116033554077,
2100
+ "278008524": -0.011390648782253265,
2101
+ "307068879": -0.007453158497810364,
2102
+ "412102563": -1.7994860410690308,
2103
+ "440577711": 0.0,
2104
+ "441046872": -0.006307080388069153,
2105
+ "509430392": 1.9592653959989548,
2106
+ "585904130": 0.0621432289481163,
2107
+ "608016290": -0.1122870147228241,
2108
+ "626486813": 0.035298554226756096,
2109
+ "636680137": -0.03966178745031357,
2110
+ "655884605": 0.025682508945465088,
2111
+ "672328262": -0.03686664253473282,
2112
+ "701667486": 0.00025522708892822266,
2113
+ "747260437": -1.844477117061615,
2114
+ "750448296": -0.5638325909773508,
2115
+ "767747424": -0.9528296366333961
2116
+ }
2117
+ },
2118
+ "macro_by_task": {
2119
+ "ci": {
2120
+ "confidence": 0.95,
2121
+ "high": -0.1851548450643366,
2122
+ "low": -1.0014115236202876,
2123
+ "mean": -0.5552110477351782,
2124
+ "n": 5,
2125
+ "num_samples": 200
2126
+ },
2127
+ "mean": -0.5552110477351782,
2128
+ "per_task": {
2129
+ "LiftPegUpright-v1": -0.025959885120391844,
2130
+ "PickCube-v1": -1.4563714861869812,
2131
+ "PullCube-v1": -0.4567870078980923,
2132
+ "PushCube-v1": -0.6119832744201025,
2133
+ "StackCube-v1": -0.22495358505032279
2134
+ }
2135
+ },
2136
+ "metric": "selected_utility_gain_over_base_at_8",
2137
+ "micro": {
2138
+ "confidence": 0.95,
2139
+ "high": -0.16205039977406463,
2140
+ "low": -0.5912865311062584,
2141
+ "mean": -0.3990048998190711,
2142
+ "n": 48,
2143
+ "num_samples": 200
2144
+ },
2145
+ "n": 48
2146
+ },
2147
+ "selector_gap_fraction_to_hidden_at_8": {
2148
+ "macro_by_seed": {
2149
+ "ci": {
2150
+ "confidence": 0.95,
2151
+ "high": 0.6215697813245239,
2152
+ "low": 0.1641170912956223,
2153
+ "mean": 0.34418808325644507,
2154
+ "n": 31,
2155
+ "num_samples": 200
2156
+ },
2157
+ "mean": 0.34418808325644507,
2158
+ "per_seed": {
2159
+ "1073161778": 1.0018020078326613,
2160
+ "1113371261": 0.003992017133275946,
2161
+ "1151410346": 0.18272470240811445,
2162
+ "1324133122": 0.00525366298940307,
2163
+ "141984615": 0.02098564722422542,
2164
+ "1461069713": 0.3800150378984332,
2165
+ "1480856810": 0.024211803589220513,
2166
+ "1509980664": 0.01200692093970083,
2167
+ "1549324805": 0.7710684804092547,
2168
+ "1578990937": 0.014218233990786039,
2169
+ "1822210471": 0.0005946566275893689,
2170
+ "1845489013": 0.0052711941525209575,
2171
+ "1999022574": 0.5408184863476693,
2172
+ "2012211287": 0.00019773645887231634,
2173
+ "2038594191": 0.08343974299400034,
2174
+ "2061635888": 0.0,
2175
+ "211245842": 0.022614301180936124,
2176
+ "278008524": 3.911825459427708,
2177
+ "307068879": 0.0010224264414320273,
2178
+ "441046872": 0.00044911597641627144,
2179
+ "509430392": 0.0,
2180
+ "585904130": 0.08331168381741826,
2181
+ "608016290": 0.10676353756001532,
2182
+ "626486813": 0.07621781757550655,
2183
+ "636680137": 0.8785857688497427,
2184
+ "655884605": 0.0014371372832949142,
2185
+ "672328262": 0.001635967177248819,
2186
+ "701667486": 0.9997888461869231,
2187
+ "747260437": 0.017282874338947683,
2188
+ "750448296": 0.7611347320927621,
2189
+ "767747424": 0.7611605820457173
2190
+ }
2191
+ },
2192
+ "macro_by_task": {
2193
+ "ci": {
2194
+ "confidence": 0.95,
2195
+ "high": 0.42436828204826843,
2196
+ "low": 0.09519566632031032,
2197
+ "mean": 0.24355684324016363,
2198
+ "n": 4,
2199
+ "num_samples": 200
2200
+ },
2201
+ "mean": 0.24355684324016363,
2202
+ "per_task": {
2203
+ "LiftPegUpright-v1": 0.22204696485387942,
2204
+ "PullCube-v1": 0.5617890754661544,
2205
+ "PushCube-v1": 0.012105901794610257,
2206
+ "StackCube-v1": 0.1782854308460104
2207
+ }
2208
+ },
2209
+ "metric": "selector_gap_fraction_to_hidden_at_8",
2210
+ "micro": {
2211
+ "confidence": 0.95,
2212
+ "high": 0.6711178484299387,
2213
+ "low": 0.16527934398793365,
2214
+ "mean": 0.3722715512774144,
2215
+ "n": 35,
2216
+ "num_samples": 200
2217
+ },
2218
+ "n": 35
2219
+ },
2220
+ "selector_regret_at_8": {
2221
+ "macro_by_seed": {
2222
+ "ci": {
2223
+ "confidence": 0.95,
2224
+ "high": 0.3860186778940261,
2225
+ "low": 0.10536204695235937,
2226
+ "mean": 0.22059872947478049,
2227
+ "n": 40,
2228
+ "num_samples": 200
2229
+ },
2230
+ "mean": 0.22059872947478049,
2231
+ "per_seed": {
2232
+ "1073161778": 1.2767786383628845,
2233
+ "1113371261": 0.007540524005889893,
2234
+ "1151410346": 0.3237016499042511,
2235
+ "1261930456": 0.011454477906227112,
2236
+ "1324133122": 0.010064825415611267,
2237
+ "1324642411": 1.858040377497673,
2238
+ "1375492786": 0.027821257710456848,
2239
+ "141984615": 0.002475142478942871,
2240
+ "1461069713": 0.026291392743587494,
2241
+ "1480856810": 0.043351709842681885,
2242
+ "1509980664": 0.014023125171661377,
2243
+ "1522272686": 1.4563714861869812,
2244
+ "1549324805": 0.09635330736637115,
2245
+ "1578990937": 0.025389984250068665,
2246
+ "1747178030": 0.0,
2247
+ "1790188787": 0.0,
2248
+ "1822210471": 0.0011331625282764435,
2249
+ "1845489013": 0.009949073195457458,
2250
+ "1999022574": 0.6338560879230499,
2251
+ "2012211287": 0.00025147199630737305,
2252
+ "2038594191": 0.1589493229985237,
2253
+ "2054899126": 0.005500897765159607,
2254
+ "2061635888": 0.0,
2255
+ "211245842": 0.041060686111450195,
2256
+ "278008524": 0.043415673077106476,
2257
+ "307068879": 0.0018310099840164185,
2258
+ "412102563": 0.6269922256469727,
2259
+ "440577711": 0.0,
2260
+ "441046872": 0.000804021954536438,
2261
+ "509430392": 0.0,
2262
+ "585904130": 0.16070519387722015,
2263
+ "608016290": 0.1574377715587616,
2264
+ "626486813": 0.0030257832258939743,
2265
+ "636680137": 0.04479973018169403,
2266
+ "655884605": 0.0026672184467315674,
2267
+ "672328262": 0.003098629415035248,
2268
+ "701667486": 1.2084707021713257,
2269
+ "747260437": 0.02974964678287506,
2270
+ "750448296": 0.05147687097390493,
2271
+ "767747424": 0.45911610033363104
2272
+ }
2273
+ },
2274
+ "macro_by_task": {
2275
+ "ci": {
2276
+ "confidence": 0.95,
2277
+ "high": 0.9839209542491218,
2278
+ "low": 0.18012203211766303,
2279
+ "mean": 0.48638275660258345,
2280
+ "n": 5,
2281
+ "num_samples": 200
2282
+ },
2283
+ "mean": 0.48638275660258345,
2284
+ "per_task": {
2285
+ "LiftPegUpright-v1": 0.2874205827713013,
2286
+ "PickCube-v1": 1.4563714861869812,
2287
+ "PullCube-v1": 0.10624422326683998,
2288
+ "PushCube-v1": 0.31880776087443036,
2289
+ "StackCube-v1": 0.26306972991336475
2290
+ }
2291
+ },
2292
+ "metric": "selector_regret_at_8",
2293
+ "micro": {
2294
+ "confidence": 0.95,
2295
+ "high": 0.3661268416326493,
2296
+ "low": 0.10114351854038735,
2297
+ "mean": 0.2157540328335017,
2298
+ "n": 48,
2299
+ "num_samples": 200
2300
+ },
2301
+ "n": 48
2302
+ },
2303
+ "success_selector_gap_at_8": {
2304
+ "macro_by_seed": {
2305
+ "ci": {
2306
+ "confidence": 0.95,
2307
+ "high": 0.20625,
2308
+ "low": 0.03125,
2309
+ "mean": 0.10625,
2310
+ "n": 40,
2311
+ "num_samples": 200
2312
+ },
2313
+ "mean": 0.10625,
2314
+ "per_seed": {
2315
+ "1073161778": 1.0,
2316
+ "1113371261": 0.0,
2317
+ "1151410346": 0.0,
2318
+ "1261930456": 0.0,
2319
+ "1324133122": 0.0,
2320
+ "1324642411": 1.0,
2321
+ "1375492786": 0.0,
2322
+ "141984615": 0.0,
2323
+ "1461069713": 0.0,
2324
+ "1480856810": 0.0,
2325
+ "1509980664": 0.0,
2326
+ "1522272686": 1.0,
2327
+ "1549324805": 0.0,
2328
+ "1578990937": 0.0,
2329
+ "1747178030": 0.0,
2330
+ "1790188787": 0.0,
2331
+ "1822210471": 0.0,
2332
+ "1845489013": 0.0,
2333
+ "1999022574": 0.0,
2334
+ "2012211287": 0.0,
2335
+ "2038594191": 0.0,
2336
+ "2054899126": 0.0,
2337
+ "2061635888": 0.0,
2338
+ "211245842": 0.0,
2339
+ "278008524": 0.0,
2340
+ "307068879": 0.0,
2341
+ "412102563": 0.0,
2342
+ "440577711": 0.0,
2343
+ "441046872": 0.0,
2344
+ "509430392": 0.0,
2345
+ "585904130": 0.0,
2346
+ "608016290": 0.0,
2347
+ "626486813": 0.0,
2348
+ "636680137": 0.0,
2349
+ "655884605": 0.0,
2350
+ "672328262": 0.0,
2351
+ "701667486": 1.0,
2352
+ "747260437": 0.0,
2353
+ "750448296": 0.0,
2354
+ "767747424": 0.25
2355
+ }
2356
+ },
2357
+ "macro_by_task": {
2358
+ "ci": {
2359
+ "confidence": 0.95,
2360
+ "high": 0.6515151515151516,
2361
+ "low": 0.08218181818181818,
2362
+ "mean": 0.2995151515151515,
2363
+ "n": 5,
2364
+ "num_samples": 200
2365
+ },
2366
+ "mean": 0.2995151515151515,
2367
+ "per_task": {
2368
+ "LiftPegUpright-v1": 0.2,
2369
+ "PickCube-v1": 1.0,
2370
+ "PullCube-v1": 0.04,
2371
+ "PushCube-v1": 0.16666666666666666,
2372
+ "StackCube-v1": 0.09090909090909091
2373
+ }
2374
+ },
2375
+ "metric": "success_selector_gap_at_8",
2376
+ "micro": {
2377
+ "confidence": 0.95,
2378
+ "high": 0.20833333333333334,
2379
+ "low": 0.020833333333333332,
2380
+ "mean": 0.10416666666666667,
2381
+ "n": 48,
2382
+ "num_samples": 200
2383
+ },
2384
+ "n": 48
2385
+ },
2386
+ "success_support_gap_at_8": {
2387
+ "macro_by_seed": {
2388
+ "ci": {
2389
+ "confidence": 0.95,
2390
+ "high": 0.7208333333333333,
2391
+ "low": 0.41666666666666663,
2392
+ "mean": 0.5541666666666667,
2393
+ "n": 40,
2394
+ "num_samples": 200
2395
+ },
2396
+ "mean": 0.5541666666666667,
2397
+ "per_seed": {
2398
+ "1073161778": 0.0,
2399
+ "1113371261": 1.0,
2400
+ "1151410346": 1.0,
2401
+ "1261930456": 1.0,
2402
+ "1324133122": 1.0,
2403
+ "1324642411": 0.0,
2404
+ "1375492786": 1.0,
2405
+ "141984615": 0.0,
2406
+ "1461069713": 0.0,
2407
+ "1480856810": 0.0,
2408
+ "1509980664": 1.0,
2409
+ "1522272686": 0.0,
2410
+ "1549324805": 0.0,
2411
+ "1578990937": 1.0,
2412
+ "1747178030": 0.0,
2413
+ "1790188787": 0.0,
2414
+ "1822210471": 1.0,
2415
+ "1845489013": 1.0,
2416
+ "1999022574": 1.0,
2417
+ "2012211287": 1.0,
2418
+ "2038594191": 1.0,
2419
+ "2054899126": 0.5,
2420
+ "2061635888": 1.0,
2421
+ "211245842": 1.0,
2422
+ "278008524": 0.0,
2423
+ "307068879": 1.0,
2424
+ "412102563": 1.0,
2425
+ "440577711": 0.0,
2426
+ "441046872": 1.0,
2427
+ "509430392": 0.0,
2428
+ "585904130": 1.0,
2429
+ "608016290": 1.0,
2430
+ "626486813": 0.0,
2431
+ "636680137": 0.0,
2432
+ "655884605": 1.0,
2433
+ "672328262": 1.0,
2434
+ "701667486": 0.0,
2435
+ "747260437": 0.0,
2436
+ "750448296": 0.6666666666666666,
2437
+ "767747424": 0.0
2438
+ }
2439
+ },
2440
+ "macro_by_task": {
2441
+ "ci": {
2442
+ "confidence": 0.95,
2443
+ "high": 0.7774545454545454,
2444
+ "low": 0.172,
2445
+ "mean": 0.4471515151515152,
2446
+ "n": 5,
2447
+ "num_samples": 200
2448
+ },
2449
+ "mean": 0.4471515151515152,
2450
+ "per_task": {
2451
+ "LiftPegUpright-v1": 0.8,
2452
+ "PickCube-v1": 0.0,
2453
+ "PullCube-v1": 0.36,
2454
+ "PushCube-v1": 0.16666666666666666,
2455
+ "StackCube-v1": 0.9090909090909091
2456
+ }
2457
+ },
2458
+ "metric": "success_support_gap_at_8",
2459
+ "micro": {
2460
+ "confidence": 0.95,
2461
+ "high": 0.6666666666666666,
2462
+ "low": 0.3333333333333333,
2463
+ "mean": 0.5,
2464
+ "n": 48,
2465
+ "num_samples": 200
2466
+ },
2467
+ "n": 48
2468
+ },
2469
+ "success_total_car_to_hidden_at_8": {
2470
+ "macro_by_seed": {
2471
+ "ci": {
2472
+ "confidence": 0.95,
2473
+ "high": 0.8041666666666666,
2474
+ "low": 0.5166666666666667,
2475
+ "mean": 0.6604166666666667,
2476
+ "n": 40,
2477
+ "num_samples": 200
2478
+ },
2479
+ "mean": 0.6604166666666667,
2480
+ "per_seed": {
2481
+ "1073161778": 1.0,
2482
+ "1113371261": 1.0,
2483
+ "1151410346": 1.0,
2484
+ "1261930456": 1.0,
2485
+ "1324133122": 1.0,
2486
+ "1324642411": 1.0,
2487
+ "1375492786": 1.0,
2488
+ "141984615": 0.0,
2489
+ "1461069713": 0.0,
2490
+ "1480856810": 0.0,
2491
+ "1509980664": 1.0,
2492
+ "1522272686": 1.0,
2493
+ "1549324805": 0.0,
2494
+ "1578990937": 1.0,
2495
+ "1747178030": 0.0,
2496
+ "1790188787": 0.0,
2497
+ "1822210471": 1.0,
2498
+ "1845489013": 1.0,
2499
+ "1999022574": 1.0,
2500
+ "2012211287": 1.0,
2501
+ "2038594191": 1.0,
2502
+ "2054899126": 0.5,
2503
+ "2061635888": 1.0,
2504
+ "211245842": 1.0,
2505
+ "278008524": 0.0,
2506
+ "307068879": 1.0,
2507
+ "412102563": 1.0,
2508
+ "440577711": 0.0,
2509
+ "441046872": 1.0,
2510
+ "509430392": 0.0,
2511
+ "585904130": 1.0,
2512
+ "608016290": 1.0,
2513
+ "626486813": 0.0,
2514
+ "636680137": 0.0,
2515
+ "655884605": 1.0,
2516
+ "672328262": 1.0,
2517
+ "701667486": 1.0,
2518
+ "747260437": 0.0,
2519
+ "750448296": 0.6666666666666666,
2520
+ "767747424": 0.25
2521
+ }
2522
+ },
2523
+ "macro_by_task": {
2524
+ "ci": {
2525
+ "confidence": 0.95,
2526
+ "high": 1.0,
2527
+ "low": 0.48,
2528
+ "mean": 0.7466666666666667,
2529
+ "n": 5,
2530
+ "num_samples": 200
2531
+ },
2532
+ "mean": 0.7466666666666667,
2533
+ "per_task": {
2534
+ "LiftPegUpright-v1": 1.0,
2535
+ "PickCube-v1": 1.0,
2536
+ "PullCube-v1": 0.4,
2537
+ "PushCube-v1": 0.3333333333333333,
2538
+ "StackCube-v1": 1.0
2539
+ }
2540
+ },
2541
+ "metric": "success_total_car_to_hidden_at_8",
2542
+ "micro": {
2543
+ "confidence": 0.95,
2544
+ "high": 0.7291666666666666,
2545
+ "low": 0.4583333333333333,
2546
+ "mean": 0.6041666666666666,
2547
+ "n": 48,
2548
+ "num_samples": 200
2549
+ },
2550
+ "n": 48
2551
+ },
2552
+ "support_gap_at_8": {
2553
+ "macro_by_seed": {
2554
+ "ci": {
2555
+ "confidence": 0.95,
2556
+ "high": 1.229019458956706,
2557
+ "low": 0.6996882142111038,
2558
+ "mean": 0.9531689197368299,
2559
+ "n": 40,
2560
+ "num_samples": 200
2561
+ },
2562
+ "mean": 0.9531689197368299,
2563
+ "per_seed": {
2564
+ "1073161778": 0.0,
2565
+ "1113371261": 1.899635210633278,
2566
+ "1151410346": 1.4369896054267883,
2567
+ "1261930456": 1.9145455807447433,
2568
+ "1324133122": 1.9018485397100449,
2569
+ "1324642411": 0.0,
2570
+ "1375492786": 1.805472657084465,
2571
+ "141984615": 0.24130010604858398,
2572
+ "1461069713": 0.04391290992498398,
2573
+ "1480856810": 0.013253256678581238,
2574
+ "1509980664": 1.16110759973526,
2575
+ "1522272686": 0.0,
2576
+ "1549324805": 0.08342014253139496,
2577
+ "1578990937": 1.7674079984426498,
2578
+ "1747178030": 0.0,
2579
+ "1790188787": 0.0,
2580
+ "1822210471": 1.9411806277930737,
2581
+ "1845489013": 1.9138048067688942,
2582
+ "1999022574": 1.2029985189437866,
2583
+ "2012211287": 1.2729573845863342,
2584
+ "2038594191": 1.7511916905641556,
2585
+ "2054899126": 0.9316255375742912,
2586
+ "2061635888": 1.9209758639335632,
2587
+ "211245842": 1.7857149839401245,
2588
+ "278008524": 0.0,
2589
+ "307068879": 1.7964697927236557,
2590
+ "412102563": 1.172493815422058,
2591
+ "440577711": 0.0,
2592
+ "441046872": 1.7957354187965393,
2593
+ "509430392": 0.0,
2594
+ "585904130": 1.706115037202835,
2595
+ "608016290": 1.4294891357421875,
2596
+ "626486813": 0.08551714941859245,
2597
+ "636680137": 0.04585278779268265,
2598
+ "655884605": 1.8275748789310455,
2599
+ "672328262": 1.9278337731957436,
2600
+ "701667486": 0.0,
2601
+ "747260437": 0.09339071810245514,
2602
+ "750448296": 1.1598488887151082,
2603
+ "767747424": 0.09709237236529589
2604
+ }
2605
+ },
2606
+ "macro_by_task": {
2607
+ "ci": {
2608
+ "confidence": 0.95,
2609
+ "high": 1.2683449571186847,
2610
+ "low": 0.31604603052139285,
2611
+ "mean": 0.7479959359981798,
2612
+ "n": 5,
2613
+ "num_samples": 200
2614
+ },
2615
+ "mean": 0.7479959359981798,
2616
+ "per_task": {
2617
+ "LiftPegUpright-v1": 1.2589303463697434,
2618
+ "PickCube-v1": 0.0,
2619
+ "PullCube-v1": 0.6982270714640617,
2620
+ "PushCube-v1": 0.32129980623722076,
2621
+ "StackCube-v1": 1.4615224559198727
2622
+ }
2623
+ },
2624
+ "metric": "support_gap_at_8",
2625
+ "micro": {
2626
+ "confidence": 0.95,
2627
+ "high": 1.1405460875636588,
2628
+ "low": 0.5969324799564978,
2629
+ "mean": 0.8698932160623372,
2630
+ "n": 48,
2631
+ "num_samples": 200
2632
+ },
2633
+ "n": 48
2634
+ },
2635
+ "support_gap_fraction_to_hidden_at_8": {
2636
+ "macro_by_seed": {
2637
+ "ci": {
2638
+ "confidence": 0.95,
2639
+ "high": 1.0556634999476413,
2640
+ "low": 0.6567003928727587,
2641
+ "mean": 0.8445818726873633,
2642
+ "n": 31,
2643
+ "num_samples": 200
2644
+ },
2645
+ "mean": 0.8445818726873633,
2646
+ "per_seed": {
2647
+ "1073161778": 0.0,
2648
+ "1113371261": 1.0056829342229987,
2649
+ "1151410346": 0.8111589733720272,
2650
+ "1324133122": 0.9927317039227662,
2651
+ "141984615": 2.0458777398812704,
2652
+ "1461069713": 0.6347159426707584,
2653
+ "1480856810": 0.007401905225509852,
2654
+ "1509980664": 0.9941669194168206,
2655
+ "1549324805": 0.6675706760394587,
2656
+ "1578990937": 0.9897375371147166,
2657
+ "1822210471": 1.018685225517477,
2658
+ "1845489013": 1.0139674830327599,
2659
+ "1999022574": 1.0264220072816432,
2660
+ "2012211287": 1.0009467822246219,
2661
+ "2038594191": 0.919280320528696,
2662
+ "2061635888": 1.0037531705667693,
2663
+ "211245842": 0.9834881073473187,
2664
+ "278008524": 0.0,
2665
+ "307068879": 1.003139378456884,
2666
+ "441046872": 1.003073935292074,
2667
+ "509430392": 0.0,
2668
+ "585904130": 0.8844724498710401,
2669
+ "608016290": 0.9693818422631982,
2670
+ "626486813": 0.797270536737948,
2671
+ "636680137": 0.8992377108824552,
2672
+ "655884605": 0.9847247418911267,
2673
+ "672328262": 1.0178283213980606,
2674
+ "701667486": 0.0,
2675
+ "747260437": 0.05425476333110376,
2676
+ "750448296": 2.905190365647023,
2677
+ "767747424": 0.5478765791717298
2678
+ }
2679
+ },
2680
+ "macro_by_task": {
2681
+ "ci": {
2682
+ "confidence": 0.95,
2683
+ "high": 0.9452722540745375,
2684
+ "low": 0.25561518911843023,
2685
+ "mean": 0.6723884621022964,
2686
+ "n": 4,
2687
+ "num_samples": 200
2688
+ },
2689
+ "mean": 0.6723884621022964,
2690
+ "per_task": {
2691
+ "LiftPegUpright-v1": 0.7953083876473557,
2692
+ "PullCube-v1": 1.011357898635456,
2693
+ "PushCube-v1": 0.003700952612754926,
2694
+ "StackCube-v1": 0.8791866095136187
2695
+ }
2696
+ },
2697
+ "metric": "support_gap_fraction_to_hidden_at_8",
2698
+ "micro": {
2699
+ "confidence": 0.95,
2700
+ "high": 1.2127122154906198,
2701
+ "low": 0.6005082137101172,
2702
+ "mean": 0.8851500604010483,
2703
+ "n": 35,
2704
+ "num_samples": 200
2705
+ },
2706
+ "n": 35
2707
+ },
2708
+ "total_car_to_hidden_at_8": {
2709
+ "macro_by_seed": {
2710
+ "ci": {
2711
+ "confidence": 0.95,
2712
+ "high": 1.4246697119592378,
2713
+ "low": 0.9299319049033026,
2714
+ "mean": 1.1725806335142503,
2715
+ "n": 40,
2716
+ "num_samples": 200
2717
+ },
2718
+ "mean": 1.1725806335142503,
2719
+ "per_seed": {
2720
+ "1073161778": 1.2767786383628845,
2721
+ "1113371261": 1.9071757346391678,
2722
+ "1151410346": 1.7606912553310394,
2723
+ "1261930456": 1.9260000586509705,
2724
+ "1324133122": 1.9119133651256561,
2725
+ "1324642411": 1.858040377497673,
2726
+ "1375492786": 1.8332939147949219,
2727
+ "141984615": 0.24377524852752686,
2728
+ "1461069713": 0.07020430266857147,
2729
+ "1480856810": 0.05660496652126312,
2730
+ "1509980664": 1.1751307249069214,
2731
+ "1522272686": 1.4563714861869812,
2732
+ "1549324805": 0.1797734498977661,
2733
+ "1578990937": 1.7927979826927185,
2734
+ "1747178030": 0.0,
2735
+ "1790188787": 0.0,
2736
+ "1822210471": 1.94231379032135,
2737
+ "1845489013": 1.9237538799643517,
2738
+ "1999022574": 1.8368546068668365,
2739
+ "2012211287": 1.2732088565826416,
2740
+ "2038594191": 1.9101410135626793,
2741
+ "2054899126": 0.9371264353394508,
2742
+ "2061635888": 1.9209758639335632,
2743
+ "211245842": 1.8267756700515747,
2744
+ "278008524": 0.0002920776605606079,
2745
+ "307068879": 1.7983008027076721,
2746
+ "412102563": 1.7994860410690308,
2747
+ "440577711": 0.0,
2748
+ "441046872": 1.7965394407510757,
2749
+ "509430392": 0.0,
2750
+ "585904130": 1.8668202310800552,
2751
+ "608016290": 1.586926907300949,
2752
+ "626486813": 0.08854293264448643,
2753
+ "636680137": 0.09065251797437668,
2754
+ "655884605": 1.830242097377777,
2755
+ "672328262": 1.9309324026107788,
2756
+ "701667486": 1.2084707021713257,
2757
+ "747260437": 0.1231403648853302,
2758
+ "750448296": 1.2113257596890132,
2759
+ "767747424": 0.5518514402210712
2760
+ }
2761
+ },
2762
+ "macro_by_task": {
2763
+ "ci": {
2764
+ "confidence": 0.95,
2765
+ "high": 1.599651543227109,
2766
+ "low": 0.867447036842505,
2767
+ "mean": 1.233549290034807,
2768
+ "n": 5,
2769
+ "num_samples": 200
2770
+ },
2771
+ "mean": 1.233549290034807,
2772
+ "per_task": {
2773
+ "LiftPegUpright-v1": 1.5463509291410447,
2774
+ "PickCube-v1": 1.4563714861869812,
2775
+ "PullCube-v1": 0.8003242819011211,
2776
+ "PushCube-v1": 0.6401075671116511,
2777
+ "StackCube-v1": 1.7245921858332374
2778
+ }
2779
+ },
2780
+ "metric": "total_car_to_hidden_at_8",
2781
+ "micro": {
2782
+ "confidence": 0.95,
2783
+ "high": 1.2861365747327607,
2784
+ "low": 0.8389217427466065,
2785
+ "mean": 1.083487346380328,
2786
+ "n": 48,
2787
+ "num_samples": 200
2788
+ },
2789
+ "n": 48
2790
+ }
2791
+ },
2792
+ "num_rows": 48,
2793
+ "report_type": "ctt_generated_measured_rollout",
2794
+ "schema_version": 1,
2795
+ "source_content_hash": "67f7b4a692f7c7e71da24378dc71e3399c5d35afc14c494869eba6087b765c42",
2796
+ "source_split_hash": "867a94a1c67686c082765671a7bf8d74a9d633085e1b2619c83197c60bf58372",
2797
+ "success_summary": {
2798
+ "base_success_rate": 0.3333333333333333,
2799
+ "base_utility_mean": 0.8338337515791258,
2800
+ "hidden_chart_oracle_success_rate": 0.7291666666666666,
2801
+ "hidden_chart_oracle_utility_mean": 1.5183161981403828,
2802
+ "max_restore_error": 1.4901161193847656e-08,
2803
+ "proposal_oracle_success_gain_over_base": -0.10416666666666667,
2804
+ "proposal_oracle_success_rate": 0.22916666666666666,
2805
+ "proposal_oracle_utility_mean": 0.6505828845935563,
2806
+ "selected_success_gain_over_base": -0.20833333333333334,
2807
+ "selected_success_rate": 0.125,
2808
+ "selected_utility_mean": 0.43482885176005465,
2809
+ "success_selector_gap": 0.10416666666666667,
2810
+ "success_support_gap": 0.5
2811
+ },
2812
+ "target_content_hash": "0c38d73dc9bc4ab756e12006671b5d12f1a23358b281e54b3f6a9ae3cbf6e9af",
2813
+ "target_split_hash": "84151df9e88c219b76751677558cc5bbbc19082bf358cea7330158e9e403e25f"
2814
+ }
workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/metrics_by_seed.json ADDED
@@ -0,0 +1,1410 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "1073161778": {
3
+ "any_generated_unsafe_known_at_8": 0.0,
4
+ "base_safety_label_known": 1.0,
5
+ "base_success": 0.0,
6
+ "base_unsafe_known": 0.0,
7
+ "base_utility": 0.725517988204956,
8
+ "branch_car_at_8": 1.2767786383628845,
9
+ "generated_safety_label_coverage_at_8": 1.0,
10
+ "generated_unsafe_rate_known_at_8": 0.0,
11
+ "hidden_chart_oracle_success_at_8": 1.0,
12
+ "hidden_chart_oracle_utility_at_8": 2.0,
13
+ "ncar_to_hidden_chart_oracle_at_8": 1.0018019999721972,
14
+ "ncar_to_proposal_oracle_at_8": 1.0018019999721972,
15
+ "outcome_ptr_at_8": 1.0,
16
+ "pairwise_causal_calibration_ece": 0.14432341808925672,
17
+ "proposal_oracle_safety_label_known_at_8": 1.0,
18
+ "proposal_oracle_success_at_8": 1.0,
19
+ "proposal_oracle_success_gain_over_base_at_8": 1.0,
20
+ "proposal_oracle_unsafe_known_at_8": 0.0,
21
+ "proposal_oracle_utility_at_8": 2.0,
22
+ "proposal_oracle_utility_gain_over_base_at_8": 1.274482011795044,
23
+ "selected_safety_label_known_at_8": 1.0,
24
+ "selected_success_at_8": 0.0,
25
+ "selected_success_gain_over_base_at_8": 0.0,
26
+ "selected_unsafe_known_at_8": 0.0,
27
+ "selected_utility_at_8": 0.7232213616371155,
28
+ "selected_utility_gain_over_base_at_8": -0.002296626567840576,
29
+ "selector_gap_fraction_to_hidden_at_8": 1.0018020078326613,
30
+ "selector_regret_at_8": 1.2767786383628845,
31
+ "success_selector_gap_at_8": 1.0,
32
+ "success_support_gap_at_8": 0.0,
33
+ "success_total_car_to_hidden_at_8": 1.0,
34
+ "support_gap_at_8": 0.0,
35
+ "support_gap_fraction_to_hidden_at_8": 0.0,
36
+ "total_car_to_hidden_at_8": 1.2767786383628845
37
+ },
38
+ "1113371261": {
39
+ "any_generated_unsafe_known_at_8": 0.0,
40
+ "base_safety_label_known": 1.0,
41
+ "base_success": 0.0,
42
+ "base_unsafe_known": 0.0,
43
+ "base_utility": 0.11109928786754608,
44
+ "branch_car_at_8": 0.007540524005889893,
45
+ "generated_safety_label_coverage_at_8": 1.0,
46
+ "generated_unsafe_rate_known_at_8": 0.0,
47
+ "hidden_chart_oracle_success_at_8": 1.0,
48
+ "hidden_chart_oracle_utility_at_8": 2.0,
49
+ "ncar_to_hidden_chart_oracle_at_8": 1.0096749460109702,
50
+ "ncar_to_proposal_oracle_at_8": 0.7024563821726059,
51
+ "outcome_ptr_at_8": 0.0,
52
+ "pairwise_causal_calibration_ece": 0.18981168502744894,
53
+ "proposal_oracle_safety_label_known_at_8": 1.0,
54
+ "proposal_oracle_success_at_8": 0.0,
55
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
56
+ "proposal_oracle_unsafe_known_at_8": 0.0,
57
+ "proposal_oracle_utility_at_8": 0.1003647893667221,
58
+ "proposal_oracle_utility_gain_over_base_at_8": -0.010734498500823975,
59
+ "selected_safety_label_known_at_8": 1.0,
60
+ "selected_success_at_8": 0.0,
61
+ "selected_success_gain_over_base_at_8": 0.0,
62
+ "selected_unsafe_known_at_8": 0.0,
63
+ "selected_utility_at_8": 0.09282426536083221,
64
+ "selected_utility_gain_over_base_at_8": -0.018275022506713867,
65
+ "selector_gap_fraction_to_hidden_at_8": 0.003992017133275946,
66
+ "selector_regret_at_8": 0.007540524005889893,
67
+ "success_selector_gap_at_8": 0.0,
68
+ "success_support_gap_at_8": 1.0,
69
+ "success_total_car_to_hidden_at_8": 1.0,
70
+ "support_gap_at_8": 1.899635210633278,
71
+ "support_gap_fraction_to_hidden_at_8": 1.0056829342229987,
72
+ "total_car_to_hidden_at_8": 1.9071757346391678
73
+ },
74
+ "1151410346": {
75
+ "any_generated_unsafe_known_at_8": 0.0,
76
+ "base_safety_label_known": 1.0,
77
+ "base_success": 0.0,
78
+ "base_unsafe_known": 0.0,
79
+ "base_utility": 0.2284735143184662,
80
+ "branch_car_at_8": 0.3237016499042511,
81
+ "generated_safety_label_coverage_at_8": 1.0,
82
+ "generated_unsafe_rate_known_at_8": 0.0,
83
+ "hidden_chart_oracle_success_at_8": 1.0,
84
+ "hidden_chart_oracle_utility_at_8": 2.0,
85
+ "ncar_to_hidden_chart_oracle_at_8": 0.9938836701698183,
86
+ "ncar_to_proposal_oracle_at_8": 0.9676112241545515,
87
+ "outcome_ptr_at_8": 1.0,
88
+ "pairwise_causal_calibration_ece": 0.5181919511721376,
89
+ "proposal_oracle_safety_label_known_at_8": 1.0,
90
+ "proposal_oracle_success_at_8": 0.0,
91
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
92
+ "proposal_oracle_unsafe_known_at_8": 0.0,
93
+ "proposal_oracle_utility_at_8": 0.5630103945732117,
94
+ "proposal_oracle_utility_gain_over_base_at_8": 0.3345368802547455,
95
+ "selected_safety_label_known_at_8": 1.0,
96
+ "selected_success_at_8": 0.0,
97
+ "selected_success_gain_over_base_at_8": 0.0,
98
+ "selected_unsafe_known_at_8": 0.0,
99
+ "selected_utility_at_8": 0.23930874466896057,
100
+ "selected_utility_gain_over_base_at_8": 0.010835230350494385,
101
+ "selector_gap_fraction_to_hidden_at_8": 0.18272470240811445,
102
+ "selector_regret_at_8": 0.3237016499042511,
103
+ "success_selector_gap_at_8": 0.0,
104
+ "success_support_gap_at_8": 1.0,
105
+ "success_total_car_to_hidden_at_8": 1.0,
106
+ "support_gap_at_8": 1.4369896054267883,
107
+ "support_gap_fraction_to_hidden_at_8": 0.8111589733720272,
108
+ "total_car_to_hidden_at_8": 1.7606912553310394
109
+ },
110
+ "1261930456": {
111
+ "any_generated_unsafe_known_at_8": 0.0,
112
+ "base_safety_label_known": 1.0,
113
+ "base_success": 1.0,
114
+ "base_unsafe_known": 0.0,
115
+ "base_utility": 2.0,
116
+ "branch_car_at_8": 0.011454477906227112,
117
+ "generated_safety_label_coverage_at_8": 1.0,
118
+ "generated_unsafe_rate_known_at_8": 0.0,
119
+ "hidden_chart_oracle_success_at_8": 1.0,
120
+ "hidden_chart_oracle_utility_at_8": 2.0,
121
+ "ncar_to_proposal_oracle_at_8": 0.005982870275641443,
122
+ "outcome_ptr_at_8": 0.0,
123
+ "pairwise_causal_calibration_ece": 0.14694969164298693,
124
+ "proposal_oracle_safety_label_known_at_8": 1.0,
125
+ "proposal_oracle_success_at_8": 0.0,
126
+ "proposal_oracle_success_gain_over_base_at_8": -1.0,
127
+ "proposal_oracle_unsafe_known_at_8": 0.0,
128
+ "proposal_oracle_utility_at_8": 0.08545441925525665,
129
+ "proposal_oracle_utility_gain_over_base_at_8": -1.9145455807447433,
130
+ "selected_safety_label_known_at_8": 1.0,
131
+ "selected_success_at_8": 0.0,
132
+ "selected_success_gain_over_base_at_8": -1.0,
133
+ "selected_unsafe_known_at_8": 0.0,
134
+ "selected_utility_at_8": 0.07399994134902954,
135
+ "selected_utility_gain_over_base_at_8": -1.9260000586509705,
136
+ "selector_regret_at_8": 0.011454477906227112,
137
+ "success_selector_gap_at_8": 0.0,
138
+ "success_support_gap_at_8": 1.0,
139
+ "success_total_car_to_hidden_at_8": 1.0,
140
+ "support_gap_at_8": 1.9145455807447433,
141
+ "total_car_to_hidden_at_8": 1.9260000586509705
142
+ },
143
+ "1324133122": {
144
+ "any_generated_unsafe_known_at_8": 0.0,
145
+ "base_safety_label_known": 1.0,
146
+ "base_success": 0.0,
147
+ "base_unsafe_known": 0.0,
148
+ "base_utility": 0.084227055311203,
149
+ "branch_car_at_8": 0.010064825415611267,
150
+ "generated_safety_label_coverage_at_8": 1.0,
151
+ "generated_unsafe_rate_known_at_8": 0.0,
152
+ "hidden_chart_oracle_success_at_8": 1.0,
153
+ "hidden_chart_oracle_utility_at_8": 2.0,
154
+ "ncar_to_hidden_chart_oracle_at_8": 0.9979853617028601,
155
+ "ncar_to_proposal_oracle_at_8": 0.7228185479224526,
156
+ "outcome_ptr_at_8": 1.0,
157
+ "pairwise_causal_calibration_ece": 0.2599006738407752,
158
+ "proposal_oracle_safety_label_known_at_8": 1.0,
159
+ "proposal_oracle_success_at_8": 0.0,
160
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
161
+ "proposal_oracle_unsafe_known_at_8": 0.0,
162
+ "proposal_oracle_utility_at_8": 0.09815146028995514,
163
+ "proposal_oracle_utility_gain_over_base_at_8": 0.013924404978752136,
164
+ "selected_safety_label_known_at_8": 1.0,
165
+ "selected_success_at_8": 0.0,
166
+ "selected_success_gain_over_base_at_8": 0.0,
167
+ "selected_unsafe_known_at_8": 0.0,
168
+ "selected_utility_at_8": 0.08808663487434387,
169
+ "selected_utility_gain_over_base_at_8": 0.003859579563140869,
170
+ "selector_gap_fraction_to_hidden_at_8": 0.00525366298940307,
171
+ "selector_regret_at_8": 0.010064825415611267,
172
+ "success_selector_gap_at_8": 0.0,
173
+ "success_support_gap_at_8": 1.0,
174
+ "success_total_car_to_hidden_at_8": 1.0,
175
+ "support_gap_at_8": 1.9018485397100449,
176
+ "support_gap_fraction_to_hidden_at_8": 0.9927317039227662,
177
+ "total_car_to_hidden_at_8": 1.9119133651256561
178
+ },
179
+ "1324642411": {
180
+ "any_generated_unsafe_known_at_8": 0.0,
181
+ "base_safety_label_known": 1.0,
182
+ "base_success": 1.0,
183
+ "base_unsafe_known": 0.0,
184
+ "base_utility": 2.0,
185
+ "branch_car_at_8": 1.858040377497673,
186
+ "generated_safety_label_coverage_at_8": 1.0,
187
+ "generated_unsafe_rate_known_at_8": 0.0,
188
+ "hidden_chart_oracle_success_at_8": 1.0,
189
+ "hidden_chart_oracle_utility_at_8": 2.0,
190
+ "outcome_ptr_at_8": 0.0,
191
+ "pairwise_causal_calibration_ece": 0.37631581765020666,
192
+ "proposal_oracle_safety_label_known_at_8": 1.0,
193
+ "proposal_oracle_success_at_8": 1.0,
194
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
195
+ "proposal_oracle_unsafe_known_at_8": 0.0,
196
+ "proposal_oracle_utility_at_8": 2.0,
197
+ "proposal_oracle_utility_gain_over_base_at_8": 0.0,
198
+ "selected_safety_label_known_at_8": 1.0,
199
+ "selected_success_at_8": 0.0,
200
+ "selected_success_gain_over_base_at_8": -1.0,
201
+ "selected_unsafe_known_at_8": 0.0,
202
+ "selected_utility_at_8": 0.14195962250232697,
203
+ "selected_utility_gain_over_base_at_8": -1.858040377497673,
204
+ "selector_regret_at_8": 1.858040377497673,
205
+ "success_selector_gap_at_8": 1.0,
206
+ "success_support_gap_at_8": 0.0,
207
+ "success_total_car_to_hidden_at_8": 1.0,
208
+ "support_gap_at_8": 0.0,
209
+ "total_car_to_hidden_at_8": 1.858040377497673
210
+ },
211
+ "1375492786": {
212
+ "any_generated_unsafe_known_at_8": 1.0,
213
+ "base_safety_label_known": 1.0,
214
+ "base_success": 1.0,
215
+ "base_unsafe_known": 0.0,
216
+ "base_utility": 2.0,
217
+ "branch_car_at_8": 0.027821257710456848,
218
+ "generated_safety_label_coverage_at_8": 1.0,
219
+ "generated_unsafe_rate_known_at_8": 1.0,
220
+ "hidden_chart_oracle_success_at_8": 1.0,
221
+ "hidden_chart_oracle_utility_at_8": 2.0,
222
+ "ncar_to_proposal_oracle_at_8": 0.015409403984710279,
223
+ "outcome_ptr_at_8": 0.0,
224
+ "pairwise_causal_calibration_ece": 0.16288447414215274,
225
+ "proposal_oracle_safety_label_known_at_8": 1.0,
226
+ "proposal_oracle_success_at_8": 0.0,
227
+ "proposal_oracle_success_gain_over_base_at_8": -1.0,
228
+ "proposal_oracle_unsafe_known_at_8": 1.0,
229
+ "proposal_oracle_utility_at_8": 0.19452734291553497,
230
+ "proposal_oracle_utility_gain_over_base_at_8": -1.805472657084465,
231
+ "selected_safety_label_known_at_8": 1.0,
232
+ "selected_success_at_8": 0.0,
233
+ "selected_success_gain_over_base_at_8": -1.0,
234
+ "selected_unsafe_known_at_8": 1.0,
235
+ "selected_utility_at_8": 0.16670608520507812,
236
+ "selected_utility_gain_over_base_at_8": -1.8332939147949219,
237
+ "selector_regret_at_8": 0.027821257710456848,
238
+ "success_selector_gap_at_8": 0.0,
239
+ "success_support_gap_at_8": 1.0,
240
+ "success_total_car_to_hidden_at_8": 1.0,
241
+ "support_gap_at_8": 1.805472657084465,
242
+ "total_car_to_hidden_at_8": 1.8332939147949219
243
+ },
244
+ "141984615": {
245
+ "any_generated_unsafe_known_at_8": 1.0,
246
+ "base_safety_label_known": 1.0,
247
+ "base_success": 0.0,
248
+ "base_unsafe_known": 0.0,
249
+ "base_utility": 0.16783273220062256,
250
+ "branch_car_at_8": 0.002475142478942871,
251
+ "generated_safety_label_coverage_at_8": 1.0,
252
+ "generated_unsafe_rate_known_at_8": 1.0,
253
+ "hidden_chart_oracle_success_at_8": 0.0,
254
+ "hidden_chart_oracle_utility_at_8": 0.2857772707939148,
255
+ "ncar_to_hidden_chart_oracle_at_8": 2.0668632118652317,
256
+ "ncar_to_proposal_oracle_at_8": 0.020065103905333725,
257
+ "outcome_ptr_at_8": 0.0,
258
+ "pairwise_causal_calibration_ece": 0.15133945279036964,
259
+ "proposal_oracle_safety_label_known_at_8": 1.0,
260
+ "proposal_oracle_success_at_8": 0.0,
261
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
262
+ "proposal_oracle_unsafe_known_at_8": 1.0,
263
+ "proposal_oracle_utility_at_8": 0.04447716474533081,
264
+ "proposal_oracle_utility_gain_over_base_at_8": -0.12335556745529175,
265
+ "selected_safety_label_known_at_8": 1.0,
266
+ "selected_success_at_8": 0.0,
267
+ "selected_success_gain_over_base_at_8": 0.0,
268
+ "selected_unsafe_known_at_8": 1.0,
269
+ "selected_utility_at_8": 0.04200202226638794,
270
+ "selected_utility_gain_over_base_at_8": -0.12583070993423462,
271
+ "selector_gap_fraction_to_hidden_at_8": 0.02098564722422542,
272
+ "selector_regret_at_8": 0.002475142478942871,
273
+ "success_selector_gap_at_8": 0.0,
274
+ "success_support_gap_at_8": 0.0,
275
+ "success_total_car_to_hidden_at_8": 0.0,
276
+ "support_gap_at_8": 0.24130010604858398,
277
+ "support_gap_fraction_to_hidden_at_8": 2.0458777398812704,
278
+ "total_car_to_hidden_at_8": 0.24377524852752686
279
+ },
280
+ "1461069713": {
281
+ "any_generated_unsafe_known_at_8": 1.0,
282
+ "base_safety_label_known": 1.0,
283
+ "base_success": 0.0,
284
+ "base_unsafe_known": 0.0,
285
+ "base_utility": 0.22852759063243866,
286
+ "branch_car_at_8": 0.026291392743587494,
287
+ "generated_safety_label_coverage_at_8": 1.0,
288
+ "generated_unsafe_rate_known_at_8": 0.875,
289
+ "hidden_chart_oracle_success_at_8": 0.0,
290
+ "hidden_chart_oracle_utility_at_8": 0.1593424528837204,
291
+ "ncar_to_hidden_chart_oracle_at_8": 1.0147308339002874,
292
+ "ncar_to_proposal_oracle_at_8": 0.2324654665550531,
293
+ "outcome_ptr_at_8": 0.0,
294
+ "pairwise_causal_calibration_ece": 0.3576754247076027,
295
+ "proposal_oracle_safety_label_known_at_8": 1.0,
296
+ "proposal_oracle_success_at_8": 0.0,
297
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
298
+ "proposal_oracle_unsafe_known_at_8": 0.0,
299
+ "proposal_oracle_utility_at_8": 0.11542954295873642,
300
+ "proposal_oracle_utility_gain_over_base_at_8": -0.11309804767370224,
301
+ "selected_safety_label_known_at_8": 1.0,
302
+ "selected_success_at_8": 0.0,
303
+ "selected_success_gain_over_base_at_8": 0.0,
304
+ "selected_unsafe_known_at_8": 1.0,
305
+ "selected_utility_at_8": 0.08913815021514893,
306
+ "selected_utility_gain_over_base_at_8": -0.13938944041728973,
307
+ "selector_gap_fraction_to_hidden_at_8": 0.3800150378984332,
308
+ "selector_regret_at_8": 0.026291392743587494,
309
+ "success_selector_gap_at_8": 0.0,
310
+ "success_support_gap_at_8": 0.0,
311
+ "success_total_car_to_hidden_at_8": 0.0,
312
+ "support_gap_at_8": 0.04391290992498398,
313
+ "support_gap_fraction_to_hidden_at_8": 0.6347159426707584,
314
+ "total_car_to_hidden_at_8": 0.07020430266857147
315
+ },
316
+ "1480856810": {
317
+ "any_generated_unsafe_known_at_8": 0.0,
318
+ "base_safety_label_known": 1.0,
319
+ "base_success": 1.0,
320
+ "base_unsafe_known": 0.0,
321
+ "base_utility": 2.0,
322
+ "branch_car_at_8": 0.043351709842681885,
323
+ "generated_safety_label_coverage_at_8": 1.0,
324
+ "generated_unsafe_rate_known_at_8": 0.0,
325
+ "hidden_chart_oracle_success_at_8": 0.0,
326
+ "hidden_chart_oracle_utility_at_8": 0.20948036015033722,
327
+ "ncar_to_hidden_chart_oracle_at_8": 0.031613708638168724,
328
+ "ncar_to_proposal_oracle_at_8": 0.02403390675499265,
329
+ "outcome_ptr_at_8": 0.0,
330
+ "pairwise_causal_calibration_ece": 0.33394552923262694,
331
+ "proposal_oracle_safety_label_known_at_8": 1.0,
332
+ "proposal_oracle_success_at_8": 0.0,
333
+ "proposal_oracle_success_gain_over_base_at_8": -1.0,
334
+ "proposal_oracle_unsafe_known_at_8": 0.0,
335
+ "proposal_oracle_utility_at_8": 0.19622710347175598,
336
+ "proposal_oracle_utility_gain_over_base_at_8": -1.803772896528244,
337
+ "selected_safety_label_known_at_8": 1.0,
338
+ "selected_success_at_8": 0.0,
339
+ "selected_success_gain_over_base_at_8": -1.0,
340
+ "selected_unsafe_known_at_8": 0.0,
341
+ "selected_utility_at_8": 0.1528753936290741,
342
+ "selected_utility_gain_over_base_at_8": -1.847124606370926,
343
+ "selector_gap_fraction_to_hidden_at_8": 0.024211803589220513,
344
+ "selector_regret_at_8": 0.043351709842681885,
345
+ "success_selector_gap_at_8": 0.0,
346
+ "success_support_gap_at_8": 0.0,
347
+ "success_total_car_to_hidden_at_8": 0.0,
348
+ "support_gap_at_8": 0.013253256678581238,
349
+ "support_gap_fraction_to_hidden_at_8": 0.007401905225509852,
350
+ "total_car_to_hidden_at_8": 0.05660496652126312
351
+ },
352
+ "1509980664": {
353
+ "any_generated_unsafe_known_at_8": 0.0,
354
+ "base_safety_label_known": 1.0,
355
+ "base_success": 0.0,
356
+ "base_unsafe_known": 0.0,
357
+ "base_utility": 0.8320798277854919,
358
+ "branch_car_at_8": 0.014023125171661377,
359
+ "generated_safety_label_coverage_at_8": 1.0,
360
+ "generated_unsafe_rate_known_at_8": 0.0,
361
+ "hidden_chart_oracle_success_at_8": 1.0,
362
+ "hidden_chart_oracle_utility_at_8": 2.0,
363
+ "ncar_to_hidden_chart_oracle_at_8": 1.0061738317414306,
364
+ "ncar_to_proposal_oracle_at_8": 2.058415470840539,
365
+ "outcome_ptr_at_8": 1.0,
366
+ "pairwise_causal_calibration_ece": 0.14023834058751442,
367
+ "proposal_oracle_safety_label_known_at_8": 1.0,
368
+ "proposal_oracle_success_at_8": 0.0,
369
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
370
+ "proposal_oracle_unsafe_known_at_8": 0.0,
371
+ "proposal_oracle_utility_at_8": 0.83889240026474,
372
+ "proposal_oracle_utility_gain_over_base_at_8": 0.006812572479248047,
373
+ "selected_safety_label_known_at_8": 1.0,
374
+ "selected_success_at_8": 0.0,
375
+ "selected_success_gain_over_base_at_8": 0.0,
376
+ "selected_unsafe_known_at_8": 0.0,
377
+ "selected_utility_at_8": 0.8248692750930786,
378
+ "selected_utility_gain_over_base_at_8": -0.00721055269241333,
379
+ "selector_gap_fraction_to_hidden_at_8": 0.01200692093970083,
380
+ "selector_regret_at_8": 0.014023125171661377,
381
+ "success_selector_gap_at_8": 0.0,
382
+ "success_support_gap_at_8": 1.0,
383
+ "success_total_car_to_hidden_at_8": 1.0,
384
+ "support_gap_at_8": 1.16110759973526,
385
+ "support_gap_fraction_to_hidden_at_8": 0.9941669194168206,
386
+ "total_car_to_hidden_at_8": 1.1751307249069214
387
+ },
388
+ "1522272686": {
389
+ "any_generated_unsafe_known_at_8": 0.0,
390
+ "base_safety_label_known": 1.0,
391
+ "base_success": 1.0,
392
+ "base_unsafe_known": 0.0,
393
+ "base_utility": 2.0,
394
+ "branch_car_at_8": 1.4563714861869812,
395
+ "generated_safety_label_coverage_at_8": 1.0,
396
+ "generated_unsafe_rate_known_at_8": 0.0,
397
+ "hidden_chart_oracle_success_at_8": 1.0,
398
+ "hidden_chart_oracle_utility_at_8": 2.0,
399
+ "outcome_ptr_at_8": 0.0,
400
+ "pairwise_causal_calibration_ece": 0.12267671398942098,
401
+ "proposal_oracle_safety_label_known_at_8": 1.0,
402
+ "proposal_oracle_success_at_8": 1.0,
403
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
404
+ "proposal_oracle_unsafe_known_at_8": 0.0,
405
+ "proposal_oracle_utility_at_8": 2.0,
406
+ "proposal_oracle_utility_gain_over_base_at_8": 0.0,
407
+ "selected_safety_label_known_at_8": 1.0,
408
+ "selected_success_at_8": 0.0,
409
+ "selected_success_gain_over_base_at_8": -1.0,
410
+ "selected_unsafe_known_at_8": 0.0,
411
+ "selected_utility_at_8": 0.5436285138130188,
412
+ "selected_utility_gain_over_base_at_8": -1.4563714861869812,
413
+ "selector_regret_at_8": 1.4563714861869812,
414
+ "success_selector_gap_at_8": 1.0,
415
+ "success_support_gap_at_8": 0.0,
416
+ "success_total_car_to_hidden_at_8": 1.0,
417
+ "support_gap_at_8": 0.0,
418
+ "total_car_to_hidden_at_8": 1.4563714861869812
419
+ },
420
+ "1549324805": {
421
+ "any_generated_unsafe_known_at_8": 0.0,
422
+ "base_safety_label_known": 1.0,
423
+ "base_success": 0.0,
424
+ "base_unsafe_known": 0.0,
425
+ "base_utility": 0.17249532043933868,
426
+ "branch_car_at_8": 0.09635330736637115,
427
+ "generated_safety_label_coverage_at_8": 1.0,
428
+ "generated_unsafe_rate_known_at_8": 0.0,
429
+ "hidden_chart_oracle_success_at_8": 0.0,
430
+ "hidden_chart_oracle_utility_at_8": 0.2974560856819153,
431
+ "ncar_to_hidden_chart_oracle_at_8": 1.4386390413214543,
432
+ "ncar_to_proposal_oracle_at_8": 2.319495421176755,
433
+ "outcome_ptr_at_8": 1.0,
434
+ "pairwise_causal_calibration_ece": 0.3814015792041445,
435
+ "proposal_oracle_safety_label_known_at_8": 1.0,
436
+ "proposal_oracle_success_at_8": 0.0,
437
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
438
+ "proposal_oracle_unsafe_known_at_8": 0.0,
439
+ "proposal_oracle_utility_at_8": 0.21403594315052032,
440
+ "proposal_oracle_utility_gain_over_base_at_8": 0.04154062271118164,
441
+ "selected_safety_label_known_at_8": 1.0,
442
+ "selected_success_at_8": 0.0,
443
+ "selected_success_gain_over_base_at_8": 0.0,
444
+ "selected_unsafe_known_at_8": 0.0,
445
+ "selected_utility_at_8": 0.11768263578414917,
446
+ "selected_utility_gain_over_base_at_8": -0.054812684655189514,
447
+ "selector_gap_fraction_to_hidden_at_8": 0.7710684804092547,
448
+ "selector_regret_at_8": 0.09635330736637115,
449
+ "success_selector_gap_at_8": 0.0,
450
+ "success_support_gap_at_8": 0.0,
451
+ "success_total_car_to_hidden_at_8": 0.0,
452
+ "support_gap_at_8": 0.08342014253139496,
453
+ "support_gap_fraction_to_hidden_at_8": 0.6675706760394587,
454
+ "total_car_to_hidden_at_8": 0.1797734498977661
455
+ },
456
+ "1578990937": {
457
+ "any_generated_unsafe_known_at_8": 0.0,
458
+ "base_safety_label_known": 1.0,
459
+ "base_success": 0.0,
460
+ "base_unsafe_known": 0.0,
461
+ "base_utility": 0.21426597237586975,
462
+ "branch_car_at_8": 0.025389984250068665,
463
+ "generated_safety_label_coverage_at_8": 1.0,
464
+ "generated_unsafe_rate_known_at_8": 0.0,
465
+ "hidden_chart_oracle_success_at_8": 1.0,
466
+ "hidden_chart_oracle_utility_at_8": 2.0,
467
+ "ncar_to_hidden_chart_oracle_at_8": 1.0039557654834124,
468
+ "ncar_to_proposal_oracle_at_8": 1.3854594546391021,
469
+ "outcome_ptr_at_8": 1.0,
470
+ "pairwise_causal_calibration_ece": 0.25510759849883147,
471
+ "proposal_oracle_safety_label_known_at_8": 1.0,
472
+ "proposal_oracle_success_at_8": 0.0,
473
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
474
+ "proposal_oracle_unsafe_known_at_8": 0.0,
475
+ "proposal_oracle_utility_at_8": 0.23259200155735016,
476
+ "proposal_oracle_utility_gain_over_base_at_8": 0.018326029181480408,
477
+ "selected_safety_label_known_at_8": 1.0,
478
+ "selected_success_at_8": 0.0,
479
+ "selected_success_gain_over_base_at_8": 0.0,
480
+ "selected_unsafe_known_at_8": 0.0,
481
+ "selected_utility_at_8": 0.2072020173072815,
482
+ "selected_utility_gain_over_base_at_8": -0.007063955068588257,
483
+ "selector_gap_fraction_to_hidden_at_8": 0.014218233990786039,
484
+ "selector_regret_at_8": 0.025389984250068665,
485
+ "success_selector_gap_at_8": 0.0,
486
+ "success_support_gap_at_8": 1.0,
487
+ "success_total_car_to_hidden_at_8": 1.0,
488
+ "support_gap_at_8": 1.7674079984426498,
489
+ "support_gap_fraction_to_hidden_at_8": 0.9897375371147166,
490
+ "total_car_to_hidden_at_8": 1.7927979826927185
491
+ },
492
+ "1747178030": {
493
+ "any_generated_unsafe_known_at_8": 1.0,
494
+ "base_safety_label_known": 1.0,
495
+ "base_success": 1.0,
496
+ "base_unsafe_known": 0.0,
497
+ "base_utility": 2.0,
498
+ "branch_car_at_8": 0.0,
499
+ "generated_safety_label_coverage_at_8": 1.0,
500
+ "generated_unsafe_rate_known_at_8": 1.0,
501
+ "hidden_chart_oracle_success_at_8": 1.0,
502
+ "hidden_chart_oracle_utility_at_8": 2.0,
503
+ "outcome_ptr_at_8": 0.0,
504
+ "pairwise_causal_calibration_ece": 0.0,
505
+ "proposal_oracle_safety_label_known_at_8": 1.0,
506
+ "proposal_oracle_success_at_8": 1.0,
507
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
508
+ "proposal_oracle_unsafe_known_at_8": 1.0,
509
+ "proposal_oracle_utility_at_8": 2.0,
510
+ "proposal_oracle_utility_gain_over_base_at_8": 0.0,
511
+ "selected_safety_label_known_at_8": 1.0,
512
+ "selected_success_at_8": 1.0,
513
+ "selected_success_gain_over_base_at_8": 0.0,
514
+ "selected_unsafe_known_at_8": 1.0,
515
+ "selected_utility_at_8": 2.0,
516
+ "selected_utility_gain_over_base_at_8": 0.0,
517
+ "selector_regret_at_8": 0.0,
518
+ "success_selector_gap_at_8": 0.0,
519
+ "success_support_gap_at_8": 0.0,
520
+ "success_total_car_to_hidden_at_8": 0.0,
521
+ "support_gap_at_8": 0.0,
522
+ "total_car_to_hidden_at_8": 0.0
523
+ },
524
+ "1790188787": {
525
+ "any_generated_unsafe_known_at_8": 0.0,
526
+ "base_safety_label_known": 1.0,
527
+ "base_success": 1.0,
528
+ "base_unsafe_known": 0.0,
529
+ "base_utility": 2.0,
530
+ "branch_car_at_8": 0.0,
531
+ "generated_safety_label_coverage_at_8": 1.0,
532
+ "generated_unsafe_rate_known_at_8": 0.0,
533
+ "hidden_chart_oracle_success_at_8": 1.0,
534
+ "hidden_chart_oracle_utility_at_8": 2.0,
535
+ "outcome_ptr_at_8": 0.0,
536
+ "pairwise_causal_calibration_ece": 0.25625757964299634,
537
+ "proposal_oracle_safety_label_known_at_8": 1.0,
538
+ "proposal_oracle_success_at_8": 1.0,
539
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
540
+ "proposal_oracle_unsafe_known_at_8": 0.0,
541
+ "proposal_oracle_utility_at_8": 2.0,
542
+ "proposal_oracle_utility_gain_over_base_at_8": 0.0,
543
+ "selected_safety_label_known_at_8": 1.0,
544
+ "selected_success_at_8": 1.0,
545
+ "selected_success_gain_over_base_at_8": 0.0,
546
+ "selected_unsafe_known_at_8": 0.0,
547
+ "selected_utility_at_8": 2.0,
548
+ "selected_utility_gain_over_base_at_8": 0.0,
549
+ "selector_regret_at_8": 0.0,
550
+ "success_selector_gap_at_8": 0.0,
551
+ "success_support_gap_at_8": 0.0,
552
+ "success_total_car_to_hidden_at_8": 0.0,
553
+ "support_gap_at_8": 0.0,
554
+ "total_car_to_hidden_at_8": 0.0
555
+ },
556
+ "1822210471": {
557
+ "any_generated_unsafe_known_at_8": 0.0,
558
+ "base_safety_label_known": 1.0,
559
+ "base_success": 0.0,
560
+ "base_unsafe_known": 0.0,
561
+ "base_utility": 0.09442546218633652,
562
+ "branch_car_at_8": 0.0011331625282764435,
563
+ "generated_safety_label_coverage_at_8": 1.0,
564
+ "generated_unsafe_rate_known_at_8": 0.0,
565
+ "hidden_chart_oracle_success_at_8": 1.0,
566
+ "hidden_chart_oracle_utility_at_8": 2.0,
567
+ "ncar_to_hidden_chart_oracle_at_8": 1.019279876796129,
568
+ "ncar_to_proposal_oracle_at_8": 0.031824954963663915,
569
+ "outcome_ptr_at_8": 0.0,
570
+ "pairwise_causal_calibration_ece": 0.15125404295167477,
571
+ "proposal_oracle_safety_label_known_at_8": 1.0,
572
+ "proposal_oracle_success_at_8": 0.0,
573
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
574
+ "proposal_oracle_unsafe_known_at_8": 0.0,
575
+ "proposal_oracle_utility_at_8": 0.058819372206926346,
576
+ "proposal_oracle_utility_gain_over_base_at_8": -0.03560608997941017,
577
+ "selected_safety_label_known_at_8": 1.0,
578
+ "selected_success_at_8": 0.0,
579
+ "selected_success_gain_over_base_at_8": 0.0,
580
+ "selected_unsafe_known_at_8": 0.0,
581
+ "selected_utility_at_8": 0.0576862096786499,
582
+ "selected_utility_gain_over_base_at_8": -0.036739252507686615,
583
+ "selector_gap_fraction_to_hidden_at_8": 0.0005946566275893689,
584
+ "selector_regret_at_8": 0.0011331625282764435,
585
+ "success_selector_gap_at_8": 0.0,
586
+ "success_support_gap_at_8": 1.0,
587
+ "success_total_car_to_hidden_at_8": 1.0,
588
+ "support_gap_at_8": 1.9411806277930737,
589
+ "support_gap_fraction_to_hidden_at_8": 1.018685225517477,
590
+ "total_car_to_hidden_at_8": 1.94231379032135
591
+ },
592
+ "1845489013": {
593
+ "any_generated_unsafe_known_at_8": 1.0,
594
+ "base_safety_label_known": 1.0,
595
+ "base_success": 0.0,
596
+ "base_unsafe_known": 0.0,
597
+ "base_utility": 0.11255800724029541,
598
+ "branch_car_at_8": 0.009949073195457458,
599
+ "generated_safety_label_coverage_at_8": 1.0,
600
+ "generated_unsafe_rate_known_at_8": 0.75,
601
+ "hidden_chart_oracle_success_at_8": 1.0,
602
+ "hidden_chart_oracle_utility_at_8": 2.0,
603
+ "ncar_to_hidden_chart_oracle_at_8": 1.019238671785175,
604
+ "ncar_to_proposal_oracle_at_8": 0.37739026714244994,
605
+ "outcome_ptr_at_8": 0.0,
606
+ "pairwise_causal_calibration_ece": 0.1601652398891617,
607
+ "proposal_oracle_safety_label_known_at_8": 1.0,
608
+ "proposal_oracle_success_at_8": 0.0,
609
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
610
+ "proposal_oracle_unsafe_known_at_8": 1.0,
611
+ "proposal_oracle_utility_at_8": 0.0861951932311058,
612
+ "proposal_oracle_utility_gain_over_base_at_8": -0.026362814009189606,
613
+ "selected_safety_label_known_at_8": 1.0,
614
+ "selected_success_at_8": 0.0,
615
+ "selected_success_gain_over_base_at_8": 0.0,
616
+ "selected_unsafe_known_at_8": 1.0,
617
+ "selected_utility_at_8": 0.07624612003564835,
618
+ "selected_utility_gain_over_base_at_8": -0.036311887204647064,
619
+ "selector_gap_fraction_to_hidden_at_8": 0.0052711941525209575,
620
+ "selector_regret_at_8": 0.009949073195457458,
621
+ "success_selector_gap_at_8": 0.0,
622
+ "success_support_gap_at_8": 1.0,
623
+ "success_total_car_to_hidden_at_8": 1.0,
624
+ "support_gap_at_8": 1.9138048067688942,
625
+ "support_gap_fraction_to_hidden_at_8": 1.0139674830327599,
626
+ "total_car_to_hidden_at_8": 1.9237538799643517
627
+ },
628
+ "1999022574": {
629
+ "any_generated_unsafe_known_at_8": 0.0,
630
+ "base_safety_label_known": 1.0,
631
+ "base_success": 0.0,
632
+ "base_unsafe_known": 0.0,
633
+ "base_utility": 0.8279688954353333,
634
+ "branch_car_at_8": 0.6338560879230499,
635
+ "generated_safety_label_coverage_at_8": 1.0,
636
+ "generated_unsafe_rate_known_at_8": 0.0,
637
+ "hidden_chart_oracle_success_at_8": 1.0,
638
+ "hidden_chart_oracle_utility_at_8": 2.0,
639
+ "ncar_to_hidden_chart_oracle_at_8": 1.5672404802573083,
640
+ "ncar_to_proposal_oracle_at_8": 20.468479398320074,
641
+ "outcome_ptr_at_8": 0.0,
642
+ "pairwise_causal_calibration_ece": 0.15346652783738673,
643
+ "proposal_oracle_safety_label_known_at_8": 1.0,
644
+ "proposal_oracle_success_at_8": 0.0,
645
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
646
+ "proposal_oracle_unsafe_known_at_8": 0.0,
647
+ "proposal_oracle_utility_at_8": 0.7970014810562134,
648
+ "proposal_oracle_utility_gain_over_base_at_8": -0.030967414379119873,
649
+ "selected_safety_label_known_at_8": 1.0,
650
+ "selected_success_at_8": 0.0,
651
+ "selected_success_gain_over_base_at_8": 0.0,
652
+ "selected_unsafe_known_at_8": 0.0,
653
+ "selected_utility_at_8": 0.16314539313316345,
654
+ "selected_utility_gain_over_base_at_8": -0.6648235023021698,
655
+ "selector_gap_fraction_to_hidden_at_8": 0.5408184863476693,
656
+ "selector_regret_at_8": 0.6338560879230499,
657
+ "success_selector_gap_at_8": 0.0,
658
+ "success_support_gap_at_8": 1.0,
659
+ "success_total_car_to_hidden_at_8": 1.0,
660
+ "support_gap_at_8": 1.2029985189437866,
661
+ "support_gap_fraction_to_hidden_at_8": 1.0264220072816432,
662
+ "total_car_to_hidden_at_8": 1.8368546068668365
663
+ },
664
+ "2012211287": {
665
+ "any_generated_unsafe_known_at_8": 0.0,
666
+ "base_safety_label_known": 1.0,
667
+ "base_success": 0.0,
668
+ "base_unsafe_known": 0.0,
669
+ "base_utility": 0.7282466888427734,
670
+ "branch_car_at_8": 0.00025147199630737305,
671
+ "generated_safety_label_coverage_at_8": 1.0,
672
+ "generated_unsafe_rate_known_at_8": 0.0,
673
+ "hidden_chart_oracle_success_at_8": 1.0,
674
+ "hidden_chart_oracle_utility_at_8": 2.0,
675
+ "ncar_to_hidden_chart_oracle_at_8": 1.0011445108113346,
676
+ "ncar_to_proposal_oracle_at_8": 0.20884931245481586,
677
+ "outcome_ptr_at_8": 0.0,
678
+ "pairwise_causal_calibration_ece": 0.19320969574917243,
679
+ "proposal_oracle_safety_label_known_at_8": 1.0,
680
+ "proposal_oracle_success_at_8": 0.0,
681
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
682
+ "proposal_oracle_unsafe_known_at_8": 0.0,
683
+ "proposal_oracle_utility_at_8": 0.7270426154136658,
684
+ "proposal_oracle_utility_gain_over_base_at_8": -0.001204073429107666,
685
+ "selected_safety_label_known_at_8": 1.0,
686
+ "selected_success_at_8": 0.0,
687
+ "selected_success_gain_over_base_at_8": 0.0,
688
+ "selected_unsafe_known_at_8": 0.0,
689
+ "selected_utility_at_8": 0.7267911434173584,
690
+ "selected_utility_gain_over_base_at_8": -0.001455545425415039,
691
+ "selector_gap_fraction_to_hidden_at_8": 0.00019773645887231634,
692
+ "selector_regret_at_8": 0.00025147199630737305,
693
+ "success_selector_gap_at_8": 0.0,
694
+ "success_support_gap_at_8": 1.0,
695
+ "success_total_car_to_hidden_at_8": 1.0,
696
+ "support_gap_at_8": 1.2729573845863342,
697
+ "support_gap_fraction_to_hidden_at_8": 1.0009467822246219,
698
+ "total_car_to_hidden_at_8": 1.2732088565826416
699
+ },
700
+ "2038594191": {
701
+ "any_generated_unsafe_known_at_8": 1.0,
702
+ "base_safety_label_known": 1.0,
703
+ "base_success": 0.0,
704
+ "base_unsafe_known": 0.0,
705
+ "base_utility": 0.09504059702157974,
706
+ "branch_car_at_8": 0.1589493229985237,
707
+ "generated_safety_label_coverage_at_8": 1.0,
708
+ "generated_unsafe_rate_known_at_8": 0.375,
709
+ "hidden_chart_oracle_success_at_8": 1.0,
710
+ "hidden_chart_oracle_utility_at_8": 2.0,
711
+ "ncar_to_hidden_chart_oracle_at_8": 1.002720058258962,
712
+ "ncar_to_proposal_oracle_at_8": 1.0336975829706272,
713
+ "outcome_ptr_at_8": 1.0,
714
+ "pairwise_causal_calibration_ece": 0.5070876425629134,
715
+ "proposal_oracle_safety_label_known_at_8": 1.0,
716
+ "proposal_oracle_success_at_8": 0.0,
717
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
718
+ "proposal_oracle_unsafe_known_at_8": 1.0,
719
+ "proposal_oracle_utility_at_8": 0.24880830943584442,
720
+ "proposal_oracle_utility_gain_over_base_at_8": 0.15376771241426468,
721
+ "selected_safety_label_known_at_8": 1.0,
722
+ "selected_success_at_8": 0.0,
723
+ "selected_success_gain_over_base_at_8": 0.0,
724
+ "selected_unsafe_known_at_8": 0.0,
725
+ "selected_utility_at_8": 0.08985898643732071,
726
+ "selected_utility_gain_over_base_at_8": -0.005181610584259033,
727
+ "selector_gap_fraction_to_hidden_at_8": 0.08343974299400034,
728
+ "selector_regret_at_8": 0.1589493229985237,
729
+ "success_selector_gap_at_8": 0.0,
730
+ "success_support_gap_at_8": 1.0,
731
+ "success_total_car_to_hidden_at_8": 1.0,
732
+ "support_gap_at_8": 1.7511916905641556,
733
+ "support_gap_fraction_to_hidden_at_8": 0.919280320528696,
734
+ "total_car_to_hidden_at_8": 1.9101410135626793
735
+ },
736
+ "2054899126": {
737
+ "any_generated_unsafe_known_at_8": 1.0,
738
+ "base_safety_label_known": 1.0,
739
+ "base_success": 1.0,
740
+ "base_unsafe_known": 0.0,
741
+ "base_utility": 2.0,
742
+ "branch_car_at_8": 0.005500897765159607,
743
+ "generated_safety_label_coverage_at_8": 1.0,
744
+ "generated_unsafe_rate_known_at_8": 0.9375,
745
+ "hidden_chart_oracle_success_at_8": 1.0,
746
+ "hidden_chart_oracle_utility_at_8": 2.0,
747
+ "ncar_to_proposal_oracle_at_8": 0.005904623170763854,
748
+ "outcome_ptr_at_8": 0.0,
749
+ "pairwise_causal_calibration_ece": 0.09662639932716256,
750
+ "proposal_oracle_safety_label_known_at_8": 1.0,
751
+ "proposal_oracle_success_at_8": 0.5,
752
+ "proposal_oracle_success_gain_over_base_at_8": -0.5,
753
+ "proposal_oracle_unsafe_known_at_8": 1.0,
754
+ "proposal_oracle_utility_at_8": 1.0683744624257088,
755
+ "proposal_oracle_utility_gain_over_base_at_8": -0.9316255375742912,
756
+ "selected_safety_label_known_at_8": 1.0,
757
+ "selected_success_at_8": 0.5,
758
+ "selected_success_gain_over_base_at_8": -0.5,
759
+ "selected_unsafe_known_at_8": 1.0,
760
+ "selected_utility_at_8": 1.0628735646605492,
761
+ "selected_utility_gain_over_base_at_8": -0.9371264353394508,
762
+ "selector_regret_at_8": 0.005500897765159607,
763
+ "success_selector_gap_at_8": 0.0,
764
+ "success_support_gap_at_8": 0.5,
765
+ "success_total_car_to_hidden_at_8": 0.5,
766
+ "support_gap_at_8": 0.9316255375742912,
767
+ "total_car_to_hidden_at_8": 0.9371264353394508
768
+ },
769
+ "2061635888": {
770
+ "any_generated_unsafe_known_at_8": 0.0,
771
+ "base_safety_label_known": 1.0,
772
+ "base_success": 0.0,
773
+ "base_unsafe_known": 0.0,
774
+ "base_utility": 0.08620692789554596,
775
+ "branch_car_at_8": 0.0,
776
+ "generated_safety_label_coverage_at_8": 1.0,
777
+ "generated_unsafe_rate_known_at_8": 0.0,
778
+ "hidden_chart_oracle_success_at_8": 1.0,
779
+ "hidden_chart_oracle_utility_at_8": 2.0,
780
+ "ncar_to_hidden_chart_oracle_at_8": 1.0037531653219327,
781
+ "ncar_to_proposal_oracle_at_8": 0.0,
782
+ "outcome_ptr_at_8": 0.0,
783
+ "pairwise_causal_calibration_ece": 0.1024944972766888,
784
+ "proposal_oracle_safety_label_known_at_8": 1.0,
785
+ "proposal_oracle_success_at_8": 0.0,
786
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
787
+ "proposal_oracle_unsafe_known_at_8": 0.0,
788
+ "proposal_oracle_utility_at_8": 0.07902413606643677,
789
+ "proposal_oracle_utility_gain_over_base_at_8": -0.007182791829109192,
790
+ "selected_safety_label_known_at_8": 1.0,
791
+ "selected_success_at_8": 0.0,
792
+ "selected_success_gain_over_base_at_8": 0.0,
793
+ "selected_unsafe_known_at_8": 0.0,
794
+ "selected_utility_at_8": 0.07902413606643677,
795
+ "selected_utility_gain_over_base_at_8": -0.007182791829109192,
796
+ "selector_gap_fraction_to_hidden_at_8": 0.0,
797
+ "selector_regret_at_8": 0.0,
798
+ "success_selector_gap_at_8": 0.0,
799
+ "success_support_gap_at_8": 1.0,
800
+ "success_total_car_to_hidden_at_8": 1.0,
801
+ "support_gap_at_8": 1.9209758639335632,
802
+ "support_gap_fraction_to_hidden_at_8": 1.0037531705667693,
803
+ "total_car_to_hidden_at_8": 1.9209758639335632
804
+ },
805
+ "211245842": {
806
+ "any_generated_unsafe_known_at_8": 0.0,
807
+ "base_safety_label_known": 1.0,
808
+ "base_success": 0.0,
809
+ "base_unsafe_known": 0.0,
810
+ "base_utility": 0.18430444598197937,
811
+ "branch_car_at_8": 0.041060686111450195,
812
+ "generated_safety_label_coverage_at_8": 1.0,
813
+ "generated_unsafe_rate_known_at_8": 0.0,
814
+ "hidden_chart_oracle_success_at_8": 1.0,
815
+ "hidden_chart_oracle_utility_at_8": 2.0,
816
+ "ncar_to_hidden_chart_oracle_at_8": 1.0061024029871146,
817
+ "ncar_to_proposal_oracle_at_8": 1.3695761057579783,
818
+ "outcome_ptr_at_8": 1.0,
819
+ "pairwise_causal_calibration_ece": 0.11952033712097218,
820
+ "proposal_oracle_safety_label_known_at_8": 1.0,
821
+ "proposal_oracle_success_at_8": 0.0,
822
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
823
+ "proposal_oracle_unsafe_known_at_8": 0.0,
824
+ "proposal_oracle_utility_at_8": 0.2142850160598755,
825
+ "proposal_oracle_utility_gain_over_base_at_8": 0.029980570077896118,
826
+ "selected_safety_label_known_at_8": 1.0,
827
+ "selected_success_at_8": 0.0,
828
+ "selected_success_gain_over_base_at_8": 0.0,
829
+ "selected_unsafe_known_at_8": 0.0,
830
+ "selected_utility_at_8": 0.1732243299484253,
831
+ "selected_utility_gain_over_base_at_8": -0.011080116033554077,
832
+ "selector_gap_fraction_to_hidden_at_8": 0.022614301180936124,
833
+ "selector_regret_at_8": 0.041060686111450195,
834
+ "success_selector_gap_at_8": 0.0,
835
+ "success_support_gap_at_8": 1.0,
836
+ "success_total_car_to_hidden_at_8": 1.0,
837
+ "support_gap_at_8": 1.7857149839401245,
838
+ "support_gap_fraction_to_hidden_at_8": 0.9834881073473187,
839
+ "total_car_to_hidden_at_8": 1.8267756700515747
840
+ },
841
+ "278008524": {
842
+ "any_generated_unsafe_known_at_8": 0.5,
843
+ "base_safety_label_known": 1.0,
844
+ "base_success": 0.5,
845
+ "base_unsafe_known": 0.0,
846
+ "base_utility": 1.1250072121620178,
847
+ "branch_car_at_8": 0.043415673077106476,
848
+ "generated_safety_label_coverage_at_8": 1.0,
849
+ "generated_unsafe_rate_known_at_8": 0.1875,
850
+ "hidden_chart_oracle_success_at_8": 0.5,
851
+ "hidden_chart_oracle_utility_at_8": 1.1139086410403252,
852
+ "ncar_to_hidden_chart_oracle_at_8": 0.026316678586339484,
853
+ "ncar_to_proposal_oracle_at_8": 1.3556794180383116,
854
+ "outcome_ptr_at_8": 0.5,
855
+ "pairwise_causal_calibration_ece": 0.05319894873106218,
856
+ "proposal_oracle_safety_label_known_at_8": 1.0,
857
+ "proposal_oracle_success_at_8": 0.5,
858
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
859
+ "proposal_oracle_unsafe_known_at_8": 0.5,
860
+ "proposal_oracle_utility_at_8": 1.157032236456871,
861
+ "proposal_oracle_utility_gain_over_base_at_8": 0.03202502429485321,
862
+ "selected_safety_label_known_at_8": 1.0,
863
+ "selected_success_at_8": 0.5,
864
+ "selected_success_gain_over_base_at_8": 0.0,
865
+ "selected_unsafe_known_at_8": 0.0,
866
+ "selected_utility_at_8": 1.1136165633797646,
867
+ "selected_utility_gain_over_base_at_8": -0.011390648782253265,
868
+ "selector_gap_fraction_to_hidden_at_8": 3.911825459427708,
869
+ "selector_regret_at_8": 0.043415673077106476,
870
+ "success_selector_gap_at_8": 0.0,
871
+ "success_support_gap_at_8": 0.0,
872
+ "success_total_car_to_hidden_at_8": 0.0,
873
+ "support_gap_at_8": 0.0,
874
+ "support_gap_fraction_to_hidden_at_8": 0.0,
875
+ "total_car_to_hidden_at_8": 0.0002920776605606079
876
+ },
877
+ "307068879": {
878
+ "any_generated_unsafe_known_at_8": 0.0,
879
+ "base_safety_label_known": 1.0,
880
+ "base_success": 0.0,
881
+ "base_unsafe_known": 0.0,
882
+ "base_utility": 0.20915235579013824,
883
+ "branch_car_at_8": 0.0018310099840164185,
884
+ "generated_safety_label_coverage_at_8": 1.0,
885
+ "generated_unsafe_rate_known_at_8": 0.0,
886
+ "hidden_chart_oracle_success_at_8": 1.0,
887
+ "hidden_chart_oracle_utility_at_8": 2.0,
888
+ "ncar_to_hidden_chart_oracle_at_8": 1.004161799291129,
889
+ "ncar_to_proposal_oracle_at_8": 0.32567740299816206,
890
+ "outcome_ptr_at_8": 0.0,
891
+ "pairwise_causal_calibration_ece": 0.07483884431566282,
892
+ "proposal_oracle_safety_label_known_at_8": 1.0,
893
+ "proposal_oracle_success_at_8": 0.0,
894
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
895
+ "proposal_oracle_unsafe_known_at_8": 0.0,
896
+ "proposal_oracle_utility_at_8": 0.2035302072763443,
897
+ "proposal_oracle_utility_gain_over_base_at_8": -0.005622148513793945,
898
+ "selected_safety_label_known_at_8": 1.0,
899
+ "selected_success_at_8": 0.0,
900
+ "selected_success_gain_over_base_at_8": 0.0,
901
+ "selected_unsafe_known_at_8": 0.0,
902
+ "selected_utility_at_8": 0.20169919729232788,
903
+ "selected_utility_gain_over_base_at_8": -0.007453158497810364,
904
+ "selector_gap_fraction_to_hidden_at_8": 0.0010224264414320273,
905
+ "selector_regret_at_8": 0.0018310099840164185,
906
+ "success_selector_gap_at_8": 0.0,
907
+ "success_support_gap_at_8": 1.0,
908
+ "success_total_car_to_hidden_at_8": 1.0,
909
+ "support_gap_at_8": 1.7964697927236557,
910
+ "support_gap_fraction_to_hidden_at_8": 1.003139378456884,
911
+ "total_car_to_hidden_at_8": 1.7983008027076721
912
+ },
913
+ "412102563": {
914
+ "any_generated_unsafe_known_at_8": 0.0,
915
+ "base_safety_label_known": 1.0,
916
+ "base_success": 1.0,
917
+ "base_unsafe_known": 0.0,
918
+ "base_utility": 2.0,
919
+ "branch_car_at_8": 0.6269922256469727,
920
+ "generated_safety_label_coverage_at_8": 1.0,
921
+ "generated_unsafe_rate_known_at_8": 0.0,
922
+ "hidden_chart_oracle_success_at_8": 1.0,
923
+ "hidden_chart_oracle_utility_at_8": 2.0,
924
+ "ncar_to_proposal_oracle_at_8": 0.5347509829497624,
925
+ "outcome_ptr_at_8": 0.0,
926
+ "pairwise_causal_calibration_ece": 0.2630363607222759,
927
+ "proposal_oracle_safety_label_known_at_8": 1.0,
928
+ "proposal_oracle_success_at_8": 0.0,
929
+ "proposal_oracle_success_gain_over_base_at_8": -1.0,
930
+ "proposal_oracle_unsafe_known_at_8": 0.0,
931
+ "proposal_oracle_utility_at_8": 0.8275061845779419,
932
+ "proposal_oracle_utility_gain_over_base_at_8": -1.172493815422058,
933
+ "selected_safety_label_known_at_8": 1.0,
934
+ "selected_success_at_8": 0.0,
935
+ "selected_success_gain_over_base_at_8": -1.0,
936
+ "selected_unsafe_known_at_8": 0.0,
937
+ "selected_utility_at_8": 0.20051395893096924,
938
+ "selected_utility_gain_over_base_at_8": -1.7994860410690308,
939
+ "selector_regret_at_8": 0.6269922256469727,
940
+ "success_selector_gap_at_8": 0.0,
941
+ "success_support_gap_at_8": 1.0,
942
+ "success_total_car_to_hidden_at_8": 1.0,
943
+ "support_gap_at_8": 1.172493815422058,
944
+ "total_car_to_hidden_at_8": 1.7994860410690308
945
+ },
946
+ "440577711": {
947
+ "any_generated_unsafe_known_at_8": 0.0,
948
+ "base_safety_label_known": 1.0,
949
+ "base_success": 1.0,
950
+ "base_unsafe_known": 0.0,
951
+ "base_utility": 2.0,
952
+ "branch_car_at_8": 0.0,
953
+ "generated_safety_label_coverage_at_8": 1.0,
954
+ "generated_unsafe_rate_known_at_8": 0.0,
955
+ "hidden_chart_oracle_success_at_8": 1.0,
956
+ "hidden_chart_oracle_utility_at_8": 2.0,
957
+ "outcome_ptr_at_8": 0.0,
958
+ "pairwise_causal_calibration_ece": 0.0,
959
+ "proposal_oracle_safety_label_known_at_8": 1.0,
960
+ "proposal_oracle_success_at_8": 1.0,
961
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
962
+ "proposal_oracle_unsafe_known_at_8": 0.0,
963
+ "proposal_oracle_utility_at_8": 2.0,
964
+ "proposal_oracle_utility_gain_over_base_at_8": 0.0,
965
+ "selected_safety_label_known_at_8": 1.0,
966
+ "selected_success_at_8": 1.0,
967
+ "selected_success_gain_over_base_at_8": 0.0,
968
+ "selected_unsafe_known_at_8": 0.0,
969
+ "selected_utility_at_8": 2.0,
970
+ "selected_utility_gain_over_base_at_8": 0.0,
971
+ "selector_regret_at_8": 0.0,
972
+ "success_selector_gap_at_8": 0.0,
973
+ "success_support_gap_at_8": 0.0,
974
+ "success_total_car_to_hidden_at_8": 0.0,
975
+ "support_gap_at_8": 0.0,
976
+ "total_car_to_hidden_at_8": 0.0
977
+ },
978
+ "441046872": {
979
+ "any_generated_unsafe_known_at_8": 0.0,
980
+ "base_safety_label_known": 1.0,
981
+ "base_success": 0.0,
982
+ "base_unsafe_known": 0.0,
983
+ "base_utility": 0.2097676396369934,
984
+ "branch_car_at_8": 0.000804021954536438,
985
+ "generated_safety_label_coverage_at_8": 1.0,
986
+ "generated_unsafe_rate_known_at_8": 0.0,
987
+ "hidden_chart_oracle_success_at_8": 1.0,
988
+ "hidden_chart_oracle_utility_at_8": 2.0,
989
+ "ncar_to_hidden_chart_oracle_at_8": 1.003523045662944,
990
+ "ncar_to_proposal_oracle_at_8": 0.14610429876487893,
991
+ "outcome_ptr_at_8": 0.0,
992
+ "pairwise_causal_calibration_ece": 0.1123926329678895,
993
+ "proposal_oracle_safety_label_known_at_8": 1.0,
994
+ "proposal_oracle_success_at_8": 0.0,
995
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
996
+ "proposal_oracle_unsafe_known_at_8": 0.0,
997
+ "proposal_oracle_utility_at_8": 0.2042645812034607,
998
+ "proposal_oracle_utility_gain_over_base_at_8": -0.005503058433532715,
999
+ "selected_safety_label_known_at_8": 1.0,
1000
+ "selected_success_at_8": 0.0,
1001
+ "selected_success_gain_over_base_at_8": 0.0,
1002
+ "selected_unsafe_known_at_8": 0.0,
1003
+ "selected_utility_at_8": 0.20346055924892426,
1004
+ "selected_utility_gain_over_base_at_8": -0.006307080388069153,
1005
+ "selector_gap_fraction_to_hidden_at_8": 0.00044911597641627144,
1006
+ "selector_regret_at_8": 0.000804021954536438,
1007
+ "success_selector_gap_at_8": 0.0,
1008
+ "success_support_gap_at_8": 1.0,
1009
+ "success_total_car_to_hidden_at_8": 1.0,
1010
+ "support_gap_at_8": 1.7957354187965393,
1011
+ "support_gap_fraction_to_hidden_at_8": 1.003073935292074,
1012
+ "total_car_to_hidden_at_8": 1.7965394407510757
1013
+ },
1014
+ "509430392": {
1015
+ "any_generated_unsafe_known_at_8": 0.0,
1016
+ "base_safety_label_known": 1.0,
1017
+ "base_success": 0.0,
1018
+ "base_unsafe_known": 0.0,
1019
+ "base_utility": 0.04073460400104523,
1020
+ "branch_car_at_8": 0.0,
1021
+ "generated_safety_label_coverage_at_8": 1.0,
1022
+ "generated_unsafe_rate_known_at_8": 0.0,
1023
+ "hidden_chart_oracle_success_at_8": 1.0,
1024
+ "hidden_chart_oracle_utility_at_8": 2.0,
1025
+ "ncar_to_hidden_chart_oracle_at_8": 0.0,
1026
+ "ncar_to_proposal_oracle_at_8": 0.0,
1027
+ "outcome_ptr_at_8": 1.0,
1028
+ "pairwise_causal_calibration_ece": 0.11637116801567593,
1029
+ "proposal_oracle_safety_label_known_at_8": 1.0,
1030
+ "proposal_oracle_success_at_8": 1.0,
1031
+ "proposal_oracle_success_gain_over_base_at_8": 1.0,
1032
+ "proposal_oracle_unsafe_known_at_8": 0.0,
1033
+ "proposal_oracle_utility_at_8": 2.0,
1034
+ "proposal_oracle_utility_gain_over_base_at_8": 1.9592653959989548,
1035
+ "selected_safety_label_known_at_8": 1.0,
1036
+ "selected_success_at_8": 1.0,
1037
+ "selected_success_gain_over_base_at_8": 1.0,
1038
+ "selected_unsafe_known_at_8": 0.0,
1039
+ "selected_utility_at_8": 2.0,
1040
+ "selected_utility_gain_over_base_at_8": 1.9592653959989548,
1041
+ "selector_gap_fraction_to_hidden_at_8": 0.0,
1042
+ "selector_regret_at_8": 0.0,
1043
+ "success_selector_gap_at_8": 0.0,
1044
+ "success_support_gap_at_8": 0.0,
1045
+ "success_total_car_to_hidden_at_8": 0.0,
1046
+ "support_gap_at_8": 0.0,
1047
+ "support_gap_fraction_to_hidden_at_8": 0.0,
1048
+ "total_car_to_hidden_at_8": 0.0
1049
+ },
1050
+ "585904130": {
1051
+ "any_generated_unsafe_known_at_8": 1.0,
1052
+ "base_safety_label_known": 1.0,
1053
+ "base_success": 0.0,
1054
+ "base_unsafe_known": 0.0,
1055
+ "base_utility": 0.07103653997182846,
1056
+ "branch_car_at_8": 0.16070519387722015,
1057
+ "generated_safety_label_coverage_at_8": 1.0,
1058
+ "generated_unsafe_rate_known_at_8": 0.875,
1059
+ "hidden_chart_oracle_success_at_8": 1.0,
1060
+ "hidden_chart_oracle_utility_at_8": 2.0,
1061
+ "ncar_to_hidden_chart_oracle_at_8": 0.9677841286713383,
1062
+ "ncar_to_proposal_oracle_at_8": 0.7211412341552217,
1063
+ "outcome_ptr_at_8": 1.0,
1064
+ "pairwise_causal_calibration_ece": 0.537235764832106,
1065
+ "proposal_oracle_safety_label_known_at_8": 1.0,
1066
+ "proposal_oracle_success_at_8": 0.0,
1067
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
1068
+ "proposal_oracle_unsafe_known_at_8": 1.0,
1069
+ "proposal_oracle_utility_at_8": 0.2938849627971649,
1070
+ "proposal_oracle_utility_gain_over_base_at_8": 0.22284842282533646,
1071
+ "selected_safety_label_known_at_8": 1.0,
1072
+ "selected_success_at_8": 0.0,
1073
+ "selected_success_gain_over_base_at_8": 0.0,
1074
+ "selected_unsafe_known_at_8": 0.0,
1075
+ "selected_utility_at_8": 0.13317976891994476,
1076
+ "selected_utility_gain_over_base_at_8": 0.0621432289481163,
1077
+ "selector_gap_fraction_to_hidden_at_8": 0.08331168381741826,
1078
+ "selector_regret_at_8": 0.16070519387722015,
1079
+ "success_selector_gap_at_8": 0.0,
1080
+ "success_support_gap_at_8": 1.0,
1081
+ "success_total_car_to_hidden_at_8": 1.0,
1082
+ "support_gap_at_8": 1.706115037202835,
1083
+ "support_gap_fraction_to_hidden_at_8": 0.8844724498710401,
1084
+ "total_car_to_hidden_at_8": 1.8668202310800552
1085
+ },
1086
+ "608016290": {
1087
+ "any_generated_unsafe_known_at_8": 0.0,
1088
+ "base_safety_label_known": 1.0,
1089
+ "base_success": 0.0,
1090
+ "base_unsafe_known": 0.0,
1091
+ "base_utility": 0.525360107421875,
1092
+ "branch_car_at_8": 0.1574377715587616,
1093
+ "generated_safety_label_coverage_at_8": 1.0,
1094
+ "generated_unsafe_rate_known_at_8": 0.0,
1095
+ "hidden_chart_oracle_success_at_8": 1.0,
1096
+ "hidden_chart_oracle_utility_at_8": 2.0,
1097
+ "ncar_to_hidden_chart_oracle_at_8": 1.076145372525531,
1098
+ "ncar_to_proposal_oracle_at_8": 3.4869346102323533,
1099
+ "outcome_ptr_at_8": 1.0,
1100
+ "pairwise_causal_calibration_ece": 0.098162172719824,
1101
+ "proposal_oracle_safety_label_known_at_8": 1.0,
1102
+ "proposal_oracle_success_at_8": 0.0,
1103
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
1104
+ "proposal_oracle_unsafe_known_at_8": 0.0,
1105
+ "proposal_oracle_utility_at_8": 0.5705108642578125,
1106
+ "proposal_oracle_utility_gain_over_base_at_8": 0.0451507568359375,
1107
+ "selected_safety_label_known_at_8": 1.0,
1108
+ "selected_success_at_8": 0.0,
1109
+ "selected_success_gain_over_base_at_8": 0.0,
1110
+ "selected_unsafe_known_at_8": 0.0,
1111
+ "selected_utility_at_8": 0.4130730926990509,
1112
+ "selected_utility_gain_over_base_at_8": -0.1122870147228241,
1113
+ "selector_gap_fraction_to_hidden_at_8": 0.10676353756001532,
1114
+ "selector_regret_at_8": 0.1574377715587616,
1115
+ "success_selector_gap_at_8": 0.0,
1116
+ "success_support_gap_at_8": 1.0,
1117
+ "success_total_car_to_hidden_at_8": 1.0,
1118
+ "support_gap_at_8": 1.4294891357421875,
1119
+ "support_gap_fraction_to_hidden_at_8": 0.9693818422631982,
1120
+ "total_car_to_hidden_at_8": 1.586926907300949
1121
+ },
1122
+ "626486813": {
1123
+ "any_generated_unsafe_known_at_8": 0.5,
1124
+ "base_safety_label_known": 1.0,
1125
+ "base_success": 0.0,
1126
+ "base_unsafe_known": 0.0,
1127
+ "base_utility": 0.04174089431762695,
1128
+ "branch_car_at_8": 0.0030257832258939743,
1129
+ "generated_safety_label_coverage_at_8": 1.0,
1130
+ "generated_unsafe_rate_known_at_8": 0.3125,
1131
+ "hidden_chart_oracle_success_at_8": 0.0,
1132
+ "hidden_chart_oracle_utility_at_8": 0.16558238118886948,
1133
+ "ncar_to_hidden_chart_oracle_at_8": 0.8734881395113818,
1134
+ "ncar_to_proposal_oracle_at_8": 1.0887571854184497,
1135
+ "outcome_ptr_at_8": 1.0,
1136
+ "pairwise_causal_calibration_ece": 0.22930626964762768,
1137
+ "proposal_oracle_safety_label_known_at_8": 1.0,
1138
+ "proposal_oracle_success_at_8": 0.0,
1139
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
1140
+ "proposal_oracle_unsafe_known_at_8": 0.0,
1141
+ "proposal_oracle_utility_at_8": 0.08006523177027702,
1142
+ "proposal_oracle_utility_gain_over_base_at_8": 0.03832433745265007,
1143
+ "selected_safety_label_known_at_8": 1.0,
1144
+ "selected_success_at_8": 0.0,
1145
+ "selected_success_gain_over_base_at_8": 0.0,
1146
+ "selected_unsafe_known_at_8": 0.0,
1147
+ "selected_utility_at_8": 0.07703944854438305,
1148
+ "selected_utility_gain_over_base_at_8": 0.035298554226756096,
1149
+ "selector_gap_fraction_to_hidden_at_8": 0.07621781757550655,
1150
+ "selector_regret_at_8": 0.0030257832258939743,
1151
+ "success_selector_gap_at_8": 0.0,
1152
+ "success_support_gap_at_8": 0.0,
1153
+ "success_total_car_to_hidden_at_8": 0.0,
1154
+ "support_gap_at_8": 0.08551714941859245,
1155
+ "support_gap_fraction_to_hidden_at_8": 0.797270536737948,
1156
+ "total_car_to_hidden_at_8": 0.08854293264448643
1157
+ },
1158
+ "636680137": {
1159
+ "any_generated_unsafe_known_at_8": 0.0,
1160
+ "base_safety_label_known": 1.0,
1161
+ "base_success": 0.0,
1162
+ "base_unsafe_known": 0.0,
1163
+ "base_utility": 0.11957359313964844,
1164
+ "branch_car_at_8": 0.04479973018169403,
1165
+ "generated_safety_label_coverage_at_8": 1.0,
1166
+ "generated_unsafe_rate_known_at_8": 0.0,
1167
+ "hidden_chart_oracle_success_at_8": 0.0,
1168
+ "hidden_chart_oracle_utility_at_8": 0.17056432366371155,
1169
+ "ncar_to_hidden_chart_oracle_at_8": 1.7778231310760573,
1170
+ "ncar_to_proposal_oracle_at_8": 8.71937375135354,
1171
+ "outcome_ptr_at_8": 1.0,
1172
+ "pairwise_causal_calibration_ece": 0.17961355132445345,
1173
+ "proposal_oracle_safety_label_known_at_8": 1.0,
1174
+ "proposal_oracle_success_at_8": 0.0,
1175
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
1176
+ "proposal_oracle_unsafe_known_at_8": 0.0,
1177
+ "proposal_oracle_utility_at_8": 0.1247115358710289,
1178
+ "proposal_oracle_utility_gain_over_base_at_8": 0.005137942731380463,
1179
+ "selected_safety_label_known_at_8": 1.0,
1180
+ "selected_success_at_8": 0.0,
1181
+ "selected_success_gain_over_base_at_8": 0.0,
1182
+ "selected_unsafe_known_at_8": 0.0,
1183
+ "selected_utility_at_8": 0.07991180568933487,
1184
+ "selected_utility_gain_over_base_at_8": -0.03966178745031357,
1185
+ "selector_gap_fraction_to_hidden_at_8": 0.8785857688497427,
1186
+ "selector_regret_at_8": 0.04479973018169403,
1187
+ "success_selector_gap_at_8": 0.0,
1188
+ "success_support_gap_at_8": 0.0,
1189
+ "success_total_car_to_hidden_at_8": 0.0,
1190
+ "support_gap_at_8": 0.04585278779268265,
1191
+ "support_gap_fraction_to_hidden_at_8": 0.8992377108824552,
1192
+ "total_car_to_hidden_at_8": 0.09065251797437668
1193
+ },
1194
+ "655884605": {
1195
+ "any_generated_unsafe_known_at_8": 0.0,
1196
+ "base_safety_label_known": 1.0,
1197
+ "base_success": 0.0,
1198
+ "base_unsafe_known": 0.0,
1199
+ "base_utility": 0.1440753936767578,
1200
+ "branch_car_at_8": 0.0026672184467315674,
1201
+ "generated_safety_label_coverage_at_8": 1.0,
1202
+ "generated_unsafe_rate_known_at_8": 0.0,
1203
+ "hidden_chart_oracle_success_at_8": 1.0,
1204
+ "hidden_chart_oracle_utility_at_8": 2.0,
1205
+ "ncar_to_hidden_chart_oracle_at_8": 0.9861618738608336,
1206
+ "ncar_to_proposal_oracle_at_8": 0.09408265091957171,
1207
+ "outcome_ptr_at_8": 1.0,
1208
+ "pairwise_causal_calibration_ece": 0.030740674705558885,
1209
+ "proposal_oracle_safety_label_known_at_8": 1.0,
1210
+ "proposal_oracle_success_at_8": 0.0,
1211
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
1212
+ "proposal_oracle_unsafe_known_at_8": 0.0,
1213
+ "proposal_oracle_utility_at_8": 0.17242512106895447,
1214
+ "proposal_oracle_utility_gain_over_base_at_8": 0.028349727392196655,
1215
+ "selected_safety_label_known_at_8": 1.0,
1216
+ "selected_success_at_8": 0.0,
1217
+ "selected_success_gain_over_base_at_8": 0.0,
1218
+ "selected_unsafe_known_at_8": 0.0,
1219
+ "selected_utility_at_8": 0.1697579026222229,
1220
+ "selected_utility_gain_over_base_at_8": 0.025682508945465088,
1221
+ "selector_gap_fraction_to_hidden_at_8": 0.0014371372832949142,
1222
+ "selector_regret_at_8": 0.0026672184467315674,
1223
+ "success_selector_gap_at_8": 0.0,
1224
+ "success_support_gap_at_8": 1.0,
1225
+ "success_total_car_to_hidden_at_8": 1.0,
1226
+ "support_gap_at_8": 1.8275748789310455,
1227
+ "support_gap_fraction_to_hidden_at_8": 0.9847247418911267,
1228
+ "total_car_to_hidden_at_8": 1.830242097377777
1229
+ },
1230
+ "672328262": {
1231
+ "any_generated_unsafe_known_at_8": 0.0,
1232
+ "base_safety_label_known": 1.0,
1233
+ "base_success": 0.0,
1234
+ "base_unsafe_known": 0.0,
1235
+ "base_utility": 0.10593423992395401,
1236
+ "branch_car_at_8": 0.003098629415035248,
1237
+ "generated_safety_label_coverage_at_8": 1.0,
1238
+ "generated_unsafe_rate_known_at_8": 0.0,
1239
+ "hidden_chart_oracle_success_at_8": 1.0,
1240
+ "hidden_chart_oracle_utility_at_8": 2.0,
1241
+ "ncar_to_hidden_chart_oracle_at_8": 1.019464283192897,
1242
+ "ncar_to_proposal_oracle_at_8": 0.09176223920635006,
1243
+ "outcome_ptr_at_8": 0.0,
1244
+ "pairwise_causal_calibration_ece": 0.11992738854206779,
1245
+ "proposal_oracle_safety_label_known_at_8": 1.0,
1246
+ "proposal_oracle_success_at_8": 0.0,
1247
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
1248
+ "proposal_oracle_unsafe_known_at_8": 0.0,
1249
+ "proposal_oracle_utility_at_8": 0.07216622680425644,
1250
+ "proposal_oracle_utility_gain_over_base_at_8": -0.03376801311969757,
1251
+ "selected_safety_label_known_at_8": 1.0,
1252
+ "selected_success_at_8": 0.0,
1253
+ "selected_success_gain_over_base_at_8": 0.0,
1254
+ "selected_unsafe_known_at_8": 0.0,
1255
+ "selected_utility_at_8": 0.06906759738922119,
1256
+ "selected_utility_gain_over_base_at_8": -0.03686664253473282,
1257
+ "selector_gap_fraction_to_hidden_at_8": 0.001635967177248819,
1258
+ "selector_regret_at_8": 0.003098629415035248,
1259
+ "success_selector_gap_at_8": 0.0,
1260
+ "success_support_gap_at_8": 1.0,
1261
+ "success_total_car_to_hidden_at_8": 1.0,
1262
+ "support_gap_at_8": 1.9278337731957436,
1263
+ "support_gap_fraction_to_hidden_at_8": 1.0178283213980606,
1264
+ "total_car_to_hidden_at_8": 1.9309324026107788
1265
+ },
1266
+ "701667486": {
1267
+ "any_generated_unsafe_known_at_8": 0.0,
1268
+ "base_safety_label_known": 1.0,
1269
+ "base_success": 0.0,
1270
+ "base_unsafe_known": 0.0,
1271
+ "base_utility": 0.7912740707397461,
1272
+ "branch_car_at_8": 1.2084707021713257,
1273
+ "generated_safety_label_coverage_at_8": 1.0,
1274
+ "generated_unsafe_rate_known_at_8": 0.0,
1275
+ "hidden_chart_oracle_success_at_8": 1.0,
1276
+ "hidden_chart_oracle_utility_at_8": 2.0,
1277
+ "ncar_to_hidden_chart_oracle_at_8": 0.9997888379154961,
1278
+ "ncar_to_proposal_oracle_at_8": 0.9997888379154961,
1279
+ "outcome_ptr_at_8": 1.0,
1280
+ "pairwise_causal_calibration_ece": 0.19081260775351527,
1281
+ "proposal_oracle_safety_label_known_at_8": 1.0,
1282
+ "proposal_oracle_success_at_8": 1.0,
1283
+ "proposal_oracle_success_gain_over_base_at_8": 1.0,
1284
+ "proposal_oracle_unsafe_known_at_8": 0.0,
1285
+ "proposal_oracle_utility_at_8": 2.0,
1286
+ "proposal_oracle_utility_gain_over_base_at_8": 1.208725929260254,
1287
+ "selected_safety_label_known_at_8": 1.0,
1288
+ "selected_success_at_8": 0.0,
1289
+ "selected_success_gain_over_base_at_8": 0.0,
1290
+ "selected_unsafe_known_at_8": 0.0,
1291
+ "selected_utility_at_8": 0.7915292978286743,
1292
+ "selected_utility_gain_over_base_at_8": 0.00025522708892822266,
1293
+ "selector_gap_fraction_to_hidden_at_8": 0.9997888461869231,
1294
+ "selector_regret_at_8": 1.2084707021713257,
1295
+ "success_selector_gap_at_8": 1.0,
1296
+ "success_support_gap_at_8": 0.0,
1297
+ "success_total_car_to_hidden_at_8": 1.0,
1298
+ "support_gap_at_8": 0.0,
1299
+ "support_gap_fraction_to_hidden_at_8": 0.0,
1300
+ "total_car_to_hidden_at_8": 1.2084707021713257
1301
+ },
1302
+ "747260437": {
1303
+ "any_generated_unsafe_known_at_8": 1.0,
1304
+ "base_safety_label_known": 1.0,
1305
+ "base_success": 1.0,
1306
+ "base_unsafe_known": 0.0,
1307
+ "base_utility": 2.0,
1308
+ "branch_car_at_8": 0.02974964678287506,
1309
+ "generated_safety_label_coverage_at_8": 1.0,
1310
+ "generated_unsafe_rate_known_at_8": 1.0,
1311
+ "hidden_chart_oracle_success_at_8": 0.0,
1312
+ "hidden_chart_oracle_utility_at_8": 0.2786632478237152,
1313
+ "ncar_to_hidden_chart_oracle_at_8": 0.07153763725445794,
1314
+ "ncar_to_proposal_oracle_at_8": 0.016393451416906714,
1315
+ "outcome_ptr_at_8": 0.0,
1316
+ "pairwise_causal_calibration_ece": 0.30546034643967757,
1317
+ "proposal_oracle_safety_label_known_at_8": 1.0,
1318
+ "proposal_oracle_success_at_8": 0.0,
1319
+ "proposal_oracle_success_gain_over_base_at_8": -1.0,
1320
+ "proposal_oracle_unsafe_known_at_8": 1.0,
1321
+ "proposal_oracle_utility_at_8": 0.18527252972126007,
1322
+ "proposal_oracle_utility_gain_over_base_at_8": -1.81472747027874,
1323
+ "selected_safety_label_known_at_8": 1.0,
1324
+ "selected_success_at_8": 0.0,
1325
+ "selected_success_gain_over_base_at_8": -1.0,
1326
+ "selected_unsafe_known_at_8": 1.0,
1327
+ "selected_utility_at_8": 0.155522882938385,
1328
+ "selected_utility_gain_over_base_at_8": -1.844477117061615,
1329
+ "selector_gap_fraction_to_hidden_at_8": 0.017282874338947683,
1330
+ "selector_regret_at_8": 0.02974964678287506,
1331
+ "success_selector_gap_at_8": 0.0,
1332
+ "success_support_gap_at_8": 0.0,
1333
+ "success_total_car_to_hidden_at_8": 0.0,
1334
+ "support_gap_at_8": 0.09339071810245514,
1335
+ "support_gap_fraction_to_hidden_at_8": 0.05425476333110376,
1336
+ "total_car_to_hidden_at_8": 0.1231403648853302
1337
+ },
1338
+ "750448296": {
1339
+ "any_generated_unsafe_known_at_8": 1.0,
1340
+ "base_safety_label_known": 1.0,
1341
+ "base_success": 0.3333333333333333,
1342
+ "base_unsafe_known": 0.0,
1343
+ "base_utility": 0.7605033268531164,
1344
+ "branch_car_at_8": 0.05147687097390493,
1345
+ "generated_safety_label_coverage_at_8": 1.0,
1346
+ "generated_unsafe_rate_known_at_8": 0.9583333333333334,
1347
+ "hidden_chart_oracle_success_at_8": 0.6666666666666666,
1348
+ "hidden_chart_oracle_utility_at_8": 1.4079964955647786,
1349
+ "ncar_to_hidden_chart_oracle_at_8": 3.6663220458717465,
1350
+ "ncar_to_proposal_oracle_at_8": 0.22983336179896974,
1351
+ "outcome_ptr_at_8": 0.3333333333333333,
1352
+ "pairwise_causal_calibration_ece": 0.22679962996282277,
1353
+ "proposal_oracle_safety_label_known_at_8": 1.0,
1354
+ "proposal_oracle_success_at_8": 0.0,
1355
+ "proposal_oracle_success_gain_over_base_at_8": -0.3333333333333333,
1356
+ "proposal_oracle_unsafe_known_at_8": 0.6666666666666666,
1357
+ "proposal_oracle_utility_at_8": 0.2481476068496704,
1358
+ "proposal_oracle_utility_gain_over_base_at_8": -0.512355720003446,
1359
+ "selected_safety_label_known_at_8": 1.0,
1360
+ "selected_success_at_8": 0.0,
1361
+ "selected_success_gain_over_base_at_8": -0.3333333333333333,
1362
+ "selected_unsafe_known_at_8": 1.0,
1363
+ "selected_utility_at_8": 0.19667073587576547,
1364
+ "selected_utility_gain_over_base_at_8": -0.5638325909773508,
1365
+ "selector_gap_fraction_to_hidden_at_8": 0.7611347320927621,
1366
+ "selector_regret_at_8": 0.05147687097390493,
1367
+ "success_selector_gap_at_8": 0.0,
1368
+ "success_support_gap_at_8": 0.6666666666666666,
1369
+ "success_total_car_to_hidden_at_8": 0.6666666666666666,
1370
+ "support_gap_at_8": 1.1598488887151082,
1371
+ "support_gap_fraction_to_hidden_at_8": 2.905190365647023,
1372
+ "total_car_to_hidden_at_8": 1.2113257596890132
1373
+ },
1374
+ "767747424": {
1375
+ "any_generated_unsafe_known_at_8": 0.75,
1376
+ "base_safety_label_known": 1.0,
1377
+ "base_success": 0.5,
1378
+ "base_unsafe_known": 0.0,
1379
+ "base_utility": 1.0747087560594082,
1380
+ "branch_car_at_8": 0.45911610033363104,
1381
+ "generated_safety_label_coverage_at_8": 1.0,
1382
+ "generated_unsafe_rate_known_at_8": 0.75,
1383
+ "hidden_chart_oracle_success_at_8": 0.25,
1384
+ "hidden_chart_oracle_utility_at_8": 0.6737305596470833,
1385
+ "ncar_to_hidden_chart_oracle_at_8": 1.15659739162032,
1386
+ "ncar_to_proposal_oracle_at_8": 0.5383285029875181,
1387
+ "outcome_ptr_at_8": 0.25,
1388
+ "pairwise_causal_calibration_ece": 0.14681410264927175,
1389
+ "proposal_oracle_safety_label_known_at_8": 1.0,
1390
+ "proposal_oracle_success_at_8": 0.25,
1391
+ "proposal_oracle_success_gain_over_base_at_8": -0.25,
1392
+ "proposal_oracle_unsafe_known_at_8": 0.75,
1393
+ "proposal_oracle_utility_at_8": 0.5809952197596431,
1394
+ "proposal_oracle_utility_gain_over_base_at_8": -0.4937135362997651,
1395
+ "selected_safety_label_known_at_8": 1.0,
1396
+ "selected_success_at_8": 0.0,
1397
+ "selected_success_gain_over_base_at_8": -0.5,
1398
+ "selected_unsafe_known_at_8": 0.75,
1399
+ "selected_utility_at_8": 0.12187911942601204,
1400
+ "selected_utility_gain_over_base_at_8": -0.9528296366333961,
1401
+ "selector_gap_fraction_to_hidden_at_8": 0.7611605820457173,
1402
+ "selector_regret_at_8": 0.45911610033363104,
1403
+ "success_selector_gap_at_8": 0.25,
1404
+ "success_support_gap_at_8": 0.0,
1405
+ "success_total_car_to_hidden_at_8": 0.25,
1406
+ "support_gap_at_8": 0.09709237236529589,
1407
+ "support_gap_fraction_to_hidden_at_8": 0.5478765791717298,
1408
+ "total_car_to_hidden_at_8": 0.5518514402210712
1409
+ }
1410
+ }
workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/metrics_by_task.json ADDED
@@ -0,0 +1,178 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "LiftPegUpright-v1": {
3
+ "any_generated_unsafe_known_at_8": 0.0,
4
+ "base_safety_label_known": 1.0,
5
+ "base_success": 0.0,
6
+ "base_unsafe_known": 0.0,
7
+ "base_utility": 0.4796089559793472,
8
+ "branch_car_at_8": 0.2874205827713013,
9
+ "generated_safety_label_coverage_at_8": 1.0,
10
+ "generated_unsafe_rate_known_at_8": 0.0,
11
+ "hidden_chart_oracle_success_at_8": 1.0,
12
+ "hidden_chart_oracle_utility_at_8": 2.0,
13
+ "ncar_to_hidden_chart_oracle_at_8": 1.0173553456526272,
14
+ "ncar_to_proposal_oracle_at_8": 1.0338735248844815,
15
+ "outcome_ptr_at_8": 0.4,
16
+ "pairwise_causal_calibration_ece": 0.1245853527683611,
17
+ "proposal_oracle_safety_label_known_at_8": 1.0,
18
+ "proposal_oracle_success_at_8": 0.2,
19
+ "proposal_oracle_success_gain_over_base_at_8": 0.2,
20
+ "proposal_oracle_unsafe_known_at_8": 0.0,
21
+ "proposal_oracle_utility_at_8": 0.7410696536302567,
22
+ "proposal_oracle_utility_gain_over_base_at_8": 0.2614606976509094,
23
+ "selected_safety_label_known_at_8": 1.0,
24
+ "selected_success_at_8": 0.0,
25
+ "selected_success_gain_over_base_at_8": 0.0,
26
+ "selected_unsafe_known_at_8": 0.0,
27
+ "selected_utility_at_8": 0.4536490708589554,
28
+ "selected_utility_gain_over_base_at_8": -0.025959885120391844,
29
+ "selector_gap_fraction_to_hidden_at_8": 0.22204696485387942,
30
+ "selector_regret_at_8": 0.2874205827713013,
31
+ "success_selector_gap_at_8": 0.2,
32
+ "success_support_gap_at_8": 0.8,
33
+ "success_total_car_to_hidden_at_8": 1.0,
34
+ "support_gap_at_8": 1.2589303463697434,
35
+ "support_gap_fraction_to_hidden_at_8": 0.7953083876473557,
36
+ "total_car_to_hidden_at_8": 1.5463509291410447
37
+ },
38
+ "PickCube-v1": {
39
+ "any_generated_unsafe_known_at_8": 0.0,
40
+ "base_safety_label_known": 1.0,
41
+ "base_success": 1.0,
42
+ "base_unsafe_known": 0.0,
43
+ "base_utility": 2.0,
44
+ "branch_car_at_8": 1.4563714861869812,
45
+ "generated_safety_label_coverage_at_8": 1.0,
46
+ "generated_unsafe_rate_known_at_8": 0.0,
47
+ "hidden_chart_oracle_success_at_8": 1.0,
48
+ "hidden_chart_oracle_utility_at_8": 2.0,
49
+ "outcome_ptr_at_8": 0.0,
50
+ "pairwise_causal_calibration_ece": 0.12267671398942098,
51
+ "proposal_oracle_safety_label_known_at_8": 1.0,
52
+ "proposal_oracle_success_at_8": 1.0,
53
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
54
+ "proposal_oracle_unsafe_known_at_8": 0.0,
55
+ "proposal_oracle_utility_at_8": 2.0,
56
+ "proposal_oracle_utility_gain_over_base_at_8": 0.0,
57
+ "selected_safety_label_known_at_8": 1.0,
58
+ "selected_success_at_8": 0.0,
59
+ "selected_success_gain_over_base_at_8": -1.0,
60
+ "selected_unsafe_known_at_8": 0.0,
61
+ "selected_utility_at_8": 0.5436285138130188,
62
+ "selected_utility_gain_over_base_at_8": -1.4563714861869812,
63
+ "selector_regret_at_8": 1.4563714861869812,
64
+ "success_selector_gap_at_8": 1.0,
65
+ "success_support_gap_at_8": 0.0,
66
+ "success_total_car_to_hidden_at_8": 1.0,
67
+ "support_gap_at_8": 0.0,
68
+ "total_car_to_hidden_at_8": 1.4563714861869812
69
+ },
70
+ "PullCube-v1": {
71
+ "any_generated_unsafe_known_at_8": 0.72,
72
+ "base_safety_label_known": 1.0,
73
+ "base_success": 0.36,
74
+ "base_unsafe_known": 0.0,
75
+ "base_utility": 0.8032506120204925,
76
+ "branch_car_at_8": 0.10624422326683998,
77
+ "generated_safety_label_coverage_at_8": 1.0,
78
+ "generated_unsafe_rate_known_at_8": 0.625,
79
+ "hidden_chart_oracle_success_at_8": 0.52,
80
+ "hidden_chart_oracle_utility_at_8": 1.1467878860235214,
81
+ "ncar_to_hidden_chart_oracle_at_8": 1.3318783387964193,
82
+ "ncar_to_proposal_oracle_at_8": 0.9248858515826692,
83
+ "outcome_ptr_at_8": 0.36,
84
+ "pairwise_causal_calibration_ece": 0.20159853773134337,
85
+ "proposal_oracle_safety_label_known_at_8": 1.0,
86
+ "proposal_oracle_success_at_8": 0.16,
87
+ "proposal_oracle_success_gain_over_base_at_8": -0.2,
88
+ "proposal_oracle_unsafe_known_at_8": 0.6,
89
+ "proposal_oracle_utility_at_8": 0.45270782738924026,
90
+ "proposal_oracle_utility_gain_over_base_at_8": -0.35054278463125227,
91
+ "selected_safety_label_known_at_8": 1.0,
92
+ "selected_success_at_8": 0.12,
93
+ "selected_success_gain_over_base_at_8": -0.24,
94
+ "selected_unsafe_known_at_8": 0.56,
95
+ "selected_utility_at_8": 0.3464636041224003,
96
+ "selected_utility_gain_over_base_at_8": -0.4567870078980923,
97
+ "selector_gap_fraction_to_hidden_at_8": 0.5617890754661544,
98
+ "selector_regret_at_8": 0.10624422326683998,
99
+ "success_selector_gap_at_8": 0.04,
100
+ "success_support_gap_at_8": 0.36,
101
+ "success_total_car_to_hidden_at_8": 0.4,
102
+ "support_gap_at_8": 0.6982270714640617,
103
+ "support_gap_fraction_to_hidden_at_8": 1.011357898635456,
104
+ "total_car_to_hidden_at_8": 0.8003242819011211
105
+ },
106
+ "PushCube-v1": {
107
+ "any_generated_unsafe_known_at_8": 0.0,
108
+ "base_safety_label_known": 1.0,
109
+ "base_success": 0.8333333333333334,
110
+ "base_unsafe_known": 0.0,
111
+ "base_utility": 1.6734557673335075,
112
+ "branch_car_at_8": 0.31880776087443036,
113
+ "generated_safety_label_coverage_at_8": 1.0,
114
+ "generated_unsafe_rate_known_at_8": 0.0,
115
+ "hidden_chart_oracle_success_at_8": 0.8333333333333334,
116
+ "hidden_chart_oracle_utility_at_8": 1.7015800600250561,
117
+ "ncar_to_hidden_chart_oracle_at_8": 0.015806854319084362,
118
+ "ncar_to_proposal_oracle_at_8": 0.010005592343544699,
119
+ "outcome_ptr_at_8": 0.16666666666666666,
120
+ "pairwise_causal_calibration_ece": 0.20497329769741546,
121
+ "proposal_oracle_safety_label_known_at_8": 1.0,
122
+ "proposal_oracle_success_at_8": 0.6666666666666666,
123
+ "proposal_oracle_success_gain_over_base_at_8": -0.16666666666666666,
124
+ "proposal_oracle_unsafe_known_at_8": 0.0,
125
+ "proposal_oracle_utility_at_8": 1.3802802537878354,
126
+ "proposal_oracle_utility_gain_over_base_at_8": -0.2931755135456721,
127
+ "selected_safety_label_known_at_8": 1.0,
128
+ "selected_success_at_8": 0.5,
129
+ "selected_success_gain_over_base_at_8": -0.3333333333333333,
130
+ "selected_unsafe_known_at_8": 0.0,
131
+ "selected_utility_at_8": 1.0614724929134052,
132
+ "selected_utility_gain_over_base_at_8": -0.6119832744201025,
133
+ "selector_gap_fraction_to_hidden_at_8": 0.012105901794610257,
134
+ "selector_regret_at_8": 0.31880776087443036,
135
+ "success_selector_gap_at_8": 0.16666666666666666,
136
+ "success_support_gap_at_8": 0.16666666666666666,
137
+ "success_total_car_to_hidden_at_8": 0.3333333333333333,
138
+ "support_gap_at_8": 0.32129980623722076,
139
+ "support_gap_fraction_to_hidden_at_8": 0.003700952612754926,
140
+ "total_car_to_hidden_at_8": 0.6401075671116511
141
+ },
142
+ "StackCube-v1": {
143
+ "any_generated_unsafe_known_at_8": 0.0,
144
+ "base_safety_label_known": 1.0,
145
+ "base_success": 0.09090909090909091,
146
+ "base_unsafe_known": 0.0,
147
+ "base_utility": 0.5003613992170854,
148
+ "branch_car_at_8": 0.26306972991336475,
149
+ "generated_safety_label_coverage_at_8": 1.0,
150
+ "generated_unsafe_rate_known_at_8": 0.0,
151
+ "hidden_chart_oracle_success_at_8": 1.0,
152
+ "hidden_chart_oracle_utility_at_8": 2.0,
153
+ "ncar_to_hidden_chart_oracle_at_8": 1.0574720335451175,
154
+ "ncar_to_proposal_oracle_at_8": 2.6639490050538304,
155
+ "outcome_ptr_at_8": 0.6363636363636364,
156
+ "pairwise_causal_calibration_ece": 0.2021201140493732,
157
+ "proposal_oracle_safety_label_known_at_8": 1.0,
158
+ "proposal_oracle_success_at_8": 0.09090909090909091,
159
+ "proposal_oracle_success_gain_over_base_at_8": 0.0,
160
+ "proposal_oracle_unsafe_known_at_8": 0.0,
161
+ "proposal_oracle_utility_at_8": 0.5384775440801274,
162
+ "proposal_oracle_utility_gain_over_base_at_8": 0.03811614486304196,
163
+ "selected_safety_label_known_at_8": 1.0,
164
+ "selected_success_at_8": 0.0,
165
+ "selected_success_gain_over_base_at_8": -0.09090909090909091,
166
+ "selected_unsafe_known_at_8": 0.0,
167
+ "selected_utility_at_8": 0.27540781416676263,
168
+ "selected_utility_gain_over_base_at_8": -0.22495358505032279,
169
+ "selector_gap_fraction_to_hidden_at_8": 0.1782854308460104,
170
+ "selector_regret_at_8": 0.26306972991336475,
171
+ "success_selector_gap_at_8": 0.09090909090909091,
172
+ "success_support_gap_at_8": 0.9090909090909091,
173
+ "success_total_car_to_hidden_at_8": 1.0,
174
+ "support_gap_at_8": 1.4615224559198727,
175
+ "support_gap_fraction_to_hidden_at_8": 0.8791866095136187,
176
+ "total_car_to_hidden_at_8": 1.7245921858332374
177
+ }
178
+ }
workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/report.md ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # CTT Generated Measured Rollout
2
+
3
+ Rows: `48`
4
+ K: `8`
5
+ Checkpoint: `runs/ctt_residual_base_context_obs_seed2/model.pt`
6
+
7
+ | Metric | Mean |
8
+ | --- | ---: |
9
+ | OutcomePTR@K | 0.3958 |
10
+ | SelectorRegret@K | 0.2158 |
11
+ | SupportGap@K | 0.8677 |
12
+
13
+ These are measured simulator rollouts of decoded CTT candidates, not PPTC proxies.
workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/run.log ADDED
@@ -0,0 +1,535 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ start
2
+ importing gymnasium/mani_skill
3
+ imported gymnasium/mani_skill
4
+ loaded checkpoint=runs/ctt_residual_base_context_obs_seed2/model.pt
5
+ loaded charts source=301 target=48
6
+ generated proposals rows=48 total=384
7
+ env init sequential task=LiftPegUpright-v1 batch=1 sim=physx_cpu render=cpu
8
+ execute sequential task=LiftPegUpright-v1 chart=ms-5799584ed40e211c50a5 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
9
+ execute sequential task=LiftPegUpright-v1 chart=ms-5799584ed40e211c50a5 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
10
+ execute sequential task=LiftPegUpright-v1 chart=ms-5799584ed40e211c50a5 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
11
+ execute sequential task=LiftPegUpright-v1 chart=ms-5799584ed40e211c50a5 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
12
+ execute sequential task=LiftPegUpright-v1 chart=ms-5799584ed40e211c50a5 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
13
+ execute sequential task=LiftPegUpright-v1 chart=ms-5799584ed40e211c50a5 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
14
+ execute sequential task=LiftPegUpright-v1 chart=ms-5799584ed40e211c50a5 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
15
+ execute sequential task=LiftPegUpright-v1 chart=ms-5799584ed40e211c50a5 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
16
+ execute sequential task=LiftPegUpright-v1 chart=ms-5799584ed40e211c50a5 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
17
+ batch done sequential task=LiftPegUpright-v1
18
+ env init sequential task=LiftPegUpright-v1 batch=1 sim=physx_cpu render=cpu
19
+ execute sequential task=LiftPegUpright-v1 chart=ms-715db95b13ebc6c977c8 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
20
+ execute sequential task=LiftPegUpright-v1 chart=ms-715db95b13ebc6c977c8 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
21
+ execute sequential task=LiftPegUpright-v1 chart=ms-715db95b13ebc6c977c8 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
22
+ execute sequential task=LiftPegUpright-v1 chart=ms-715db95b13ebc6c977c8 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
23
+ execute sequential task=LiftPegUpright-v1 chart=ms-715db95b13ebc6c977c8 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
24
+ execute sequential task=LiftPegUpright-v1 chart=ms-715db95b13ebc6c977c8 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
25
+ execute sequential task=LiftPegUpright-v1 chart=ms-715db95b13ebc6c977c8 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
26
+ execute sequential task=LiftPegUpright-v1 chart=ms-715db95b13ebc6c977c8 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
27
+ execute sequential task=LiftPegUpright-v1 chart=ms-715db95b13ebc6c977c8 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
28
+ batch done sequential task=LiftPegUpright-v1
29
+ env init sequential task=LiftPegUpright-v1 batch=1 sim=physx_cpu render=cpu
30
+ execute sequential task=LiftPegUpright-v1 chart=ms-7b229b60d7f64fb624cf candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
31
+ execute sequential task=LiftPegUpright-v1 chart=ms-7b229b60d7f64fb624cf candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
32
+ execute sequential task=LiftPegUpright-v1 chart=ms-7b229b60d7f64fb624cf candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
33
+ execute sequential task=LiftPegUpright-v1 chart=ms-7b229b60d7f64fb624cf candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
34
+ execute sequential task=LiftPegUpright-v1 chart=ms-7b229b60d7f64fb624cf candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
35
+ execute sequential task=LiftPegUpright-v1 chart=ms-7b229b60d7f64fb624cf candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
36
+ execute sequential task=LiftPegUpright-v1 chart=ms-7b229b60d7f64fb624cf candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
37
+ execute sequential task=LiftPegUpright-v1 chart=ms-7b229b60d7f64fb624cf candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
38
+ execute sequential task=LiftPegUpright-v1 chart=ms-7b229b60d7f64fb624cf candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
39
+ batch done sequential task=LiftPegUpright-v1
40
+ env init sequential task=LiftPegUpright-v1 batch=1 sim=physx_cpu render=cpu
41
+ execute sequential task=LiftPegUpright-v1 chart=ms-99f263cb3f383b1f6243 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
42
+ execute sequential task=LiftPegUpright-v1 chart=ms-99f263cb3f383b1f6243 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
43
+ execute sequential task=LiftPegUpright-v1 chart=ms-99f263cb3f383b1f6243 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
44
+ execute sequential task=LiftPegUpright-v1 chart=ms-99f263cb3f383b1f6243 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
45
+ execute sequential task=LiftPegUpright-v1 chart=ms-99f263cb3f383b1f6243 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
46
+ execute sequential task=LiftPegUpright-v1 chart=ms-99f263cb3f383b1f6243 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
47
+ execute sequential task=LiftPegUpright-v1 chart=ms-99f263cb3f383b1f6243 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
48
+ execute sequential task=LiftPegUpright-v1 chart=ms-99f263cb3f383b1f6243 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
49
+ execute sequential task=LiftPegUpright-v1 chart=ms-99f263cb3f383b1f6243 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
50
+ batch done sequential task=LiftPegUpright-v1
51
+ env init sequential task=LiftPegUpright-v1 batch=1 sim=physx_cpu render=cpu
52
+ execute sequential task=LiftPegUpright-v1 chart=ms-ec2e4afb34b439731136 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
53
+ execute sequential task=LiftPegUpright-v1 chart=ms-ec2e4afb34b439731136 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
54
+ execute sequential task=LiftPegUpright-v1 chart=ms-ec2e4afb34b439731136 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
55
+ execute sequential task=LiftPegUpright-v1 chart=ms-ec2e4afb34b439731136 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
56
+ execute sequential task=LiftPegUpright-v1 chart=ms-ec2e4afb34b439731136 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
57
+ execute sequential task=LiftPegUpright-v1 chart=ms-ec2e4afb34b439731136 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
58
+ execute sequential task=LiftPegUpright-v1 chart=ms-ec2e4afb34b439731136 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
59
+ execute sequential task=LiftPegUpright-v1 chart=ms-ec2e4afb34b439731136 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
60
+ execute sequential task=LiftPegUpright-v1 chart=ms-ec2e4afb34b439731136 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
61
+ batch done sequential task=LiftPegUpright-v1
62
+ env init sequential task=PickCube-v1 batch=1 sim=physx_cpu render=cpu
63
+ execute sequential task=PickCube-v1 chart=ms-a9ac13384fe5afc835f1 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
64
+ execute sequential task=PickCube-v1 chart=ms-a9ac13384fe5afc835f1 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
65
+ execute sequential task=PickCube-v1 chart=ms-a9ac13384fe5afc835f1 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
66
+ execute sequential task=PickCube-v1 chart=ms-a9ac13384fe5afc835f1 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
67
+ execute sequential task=PickCube-v1 chart=ms-a9ac13384fe5afc835f1 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
68
+ execute sequential task=PickCube-v1 chart=ms-a9ac13384fe5afc835f1 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
69
+ execute sequential task=PickCube-v1 chart=ms-a9ac13384fe5afc835f1 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
70
+ execute sequential task=PickCube-v1 chart=ms-a9ac13384fe5afc835f1 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
71
+ execute sequential task=PickCube-v1 chart=ms-a9ac13384fe5afc835f1 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
72
+ batch done sequential task=PickCube-v1
73
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
74
+ execute sequential task=PullCube-v1 chart=ms-06babf453b22ba8f2407 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
75
+ execute sequential task=PullCube-v1 chart=ms-06babf453b22ba8f2407 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
76
+ execute sequential task=PullCube-v1 chart=ms-06babf453b22ba8f2407 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
77
+ execute sequential task=PullCube-v1 chart=ms-06babf453b22ba8f2407 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
78
+ execute sequential task=PullCube-v1 chart=ms-06babf453b22ba8f2407 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
79
+ execute sequential task=PullCube-v1 chart=ms-06babf453b22ba8f2407 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
80
+ execute sequential task=PullCube-v1 chart=ms-06babf453b22ba8f2407 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
81
+ execute sequential task=PullCube-v1 chart=ms-06babf453b22ba8f2407 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
82
+ execute sequential task=PullCube-v1 chart=ms-06babf453b22ba8f2407 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
83
+ batch done sequential task=PullCube-v1
84
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
85
+ execute sequential task=PullCube-v1 chart=ms-0e1ff8d16372796dac97 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
86
+ execute sequential task=PullCube-v1 chart=ms-0e1ff8d16372796dac97 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
87
+ execute sequential task=PullCube-v1 chart=ms-0e1ff8d16372796dac97 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
88
+ execute sequential task=PullCube-v1 chart=ms-0e1ff8d16372796dac97 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
89
+ execute sequential task=PullCube-v1 chart=ms-0e1ff8d16372796dac97 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
90
+ execute sequential task=PullCube-v1 chart=ms-0e1ff8d16372796dac97 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
91
+ execute sequential task=PullCube-v1 chart=ms-0e1ff8d16372796dac97 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
92
+ execute sequential task=PullCube-v1 chart=ms-0e1ff8d16372796dac97 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
93
+ execute sequential task=PullCube-v1 chart=ms-0e1ff8d16372796dac97 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
94
+ batch done sequential task=PullCube-v1
95
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
96
+ execute sequential task=PullCube-v1 chart=ms-223b0a8319beb051a1ee candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
97
+ execute sequential task=PullCube-v1 chart=ms-223b0a8319beb051a1ee candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
98
+ execute sequential task=PullCube-v1 chart=ms-223b0a8319beb051a1ee candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
99
+ execute sequential task=PullCube-v1 chart=ms-223b0a8319beb051a1ee candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
100
+ execute sequential task=PullCube-v1 chart=ms-223b0a8319beb051a1ee candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
101
+ execute sequential task=PullCube-v1 chart=ms-223b0a8319beb051a1ee candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
102
+ execute sequential task=PullCube-v1 chart=ms-223b0a8319beb051a1ee candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
103
+ execute sequential task=PullCube-v1 chart=ms-223b0a8319beb051a1ee candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
104
+ execute sequential task=PullCube-v1 chart=ms-223b0a8319beb051a1ee candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
105
+ batch done sequential task=PullCube-v1
106
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
107
+ execute sequential task=PullCube-v1 chart=ms-34726202b475d29cdf38 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
108
+ execute sequential task=PullCube-v1 chart=ms-34726202b475d29cdf38 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
109
+ execute sequential task=PullCube-v1 chart=ms-34726202b475d29cdf38 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
110
+ execute sequential task=PullCube-v1 chart=ms-34726202b475d29cdf38 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
111
+ execute sequential task=PullCube-v1 chart=ms-34726202b475d29cdf38 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
112
+ execute sequential task=PullCube-v1 chart=ms-34726202b475d29cdf38 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
113
+ execute sequential task=PullCube-v1 chart=ms-34726202b475d29cdf38 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
114
+ execute sequential task=PullCube-v1 chart=ms-34726202b475d29cdf38 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
115
+ execute sequential task=PullCube-v1 chart=ms-34726202b475d29cdf38 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
116
+ batch done sequential task=PullCube-v1
117
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
118
+ execute sequential task=PullCube-v1 chart=ms-4cc40587ec490257b1d2 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
119
+ execute sequential task=PullCube-v1 chart=ms-4cc40587ec490257b1d2 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
120
+ execute sequential task=PullCube-v1 chart=ms-4cc40587ec490257b1d2 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
121
+ execute sequential task=PullCube-v1 chart=ms-4cc40587ec490257b1d2 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
122
+ execute sequential task=PullCube-v1 chart=ms-4cc40587ec490257b1d2 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
123
+ execute sequential task=PullCube-v1 chart=ms-4cc40587ec490257b1d2 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
124
+ execute sequential task=PullCube-v1 chart=ms-4cc40587ec490257b1d2 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
125
+ execute sequential task=PullCube-v1 chart=ms-4cc40587ec490257b1d2 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
126
+ execute sequential task=PullCube-v1 chart=ms-4cc40587ec490257b1d2 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
127
+ batch done sequential task=PullCube-v1
128
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
129
+ execute sequential task=PullCube-v1 chart=ms-62f8c18a15c566ec4b00 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
130
+ execute sequential task=PullCube-v1 chart=ms-62f8c18a15c566ec4b00 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
131
+ execute sequential task=PullCube-v1 chart=ms-62f8c18a15c566ec4b00 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
132
+ execute sequential task=PullCube-v1 chart=ms-62f8c18a15c566ec4b00 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
133
+ execute sequential task=PullCube-v1 chart=ms-62f8c18a15c566ec4b00 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
134
+ execute sequential task=PullCube-v1 chart=ms-62f8c18a15c566ec4b00 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
135
+ execute sequential task=PullCube-v1 chart=ms-62f8c18a15c566ec4b00 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
136
+ execute sequential task=PullCube-v1 chart=ms-62f8c18a15c566ec4b00 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
137
+ execute sequential task=PullCube-v1 chart=ms-62f8c18a15c566ec4b00 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
138
+ batch done sequential task=PullCube-v1
139
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
140
+ execute sequential task=PullCube-v1 chart=ms-6661b157c41ad4d6ab85 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
141
+ execute sequential task=PullCube-v1 chart=ms-6661b157c41ad4d6ab85 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
142
+ execute sequential task=PullCube-v1 chart=ms-6661b157c41ad4d6ab85 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
143
+ execute sequential task=PullCube-v1 chart=ms-6661b157c41ad4d6ab85 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
144
+ execute sequential task=PullCube-v1 chart=ms-6661b157c41ad4d6ab85 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
145
+ execute sequential task=PullCube-v1 chart=ms-6661b157c41ad4d6ab85 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
146
+ execute sequential task=PullCube-v1 chart=ms-6661b157c41ad4d6ab85 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
147
+ execute sequential task=PullCube-v1 chart=ms-6661b157c41ad4d6ab85 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
148
+ execute sequential task=PullCube-v1 chart=ms-6661b157c41ad4d6ab85 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
149
+ batch done sequential task=PullCube-v1
150
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
151
+ execute sequential task=PullCube-v1 chart=ms-6b789356d65b3312fbae candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
152
+ execute sequential task=PullCube-v1 chart=ms-6b789356d65b3312fbae candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
153
+ execute sequential task=PullCube-v1 chart=ms-6b789356d65b3312fbae candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
154
+ execute sequential task=PullCube-v1 chart=ms-6b789356d65b3312fbae candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
155
+ execute sequential task=PullCube-v1 chart=ms-6b789356d65b3312fbae candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
156
+ execute sequential task=PullCube-v1 chart=ms-6b789356d65b3312fbae candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
157
+ execute sequential task=PullCube-v1 chart=ms-6b789356d65b3312fbae candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
158
+ execute sequential task=PullCube-v1 chart=ms-6b789356d65b3312fbae candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
159
+ execute sequential task=PullCube-v1 chart=ms-6b789356d65b3312fbae candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
160
+ batch done sequential task=PullCube-v1
161
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
162
+ execute sequential task=PullCube-v1 chart=ms-7923b8196b397863cf4f candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
163
+ execute sequential task=PullCube-v1 chart=ms-7923b8196b397863cf4f candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
164
+ execute sequential task=PullCube-v1 chart=ms-7923b8196b397863cf4f candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
165
+ execute sequential task=PullCube-v1 chart=ms-7923b8196b397863cf4f candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
166
+ execute sequential task=PullCube-v1 chart=ms-7923b8196b397863cf4f candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
167
+ execute sequential task=PullCube-v1 chart=ms-7923b8196b397863cf4f candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
168
+ execute sequential task=PullCube-v1 chart=ms-7923b8196b397863cf4f candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
169
+ execute sequential task=PullCube-v1 chart=ms-7923b8196b397863cf4f candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
170
+ execute sequential task=PullCube-v1 chart=ms-7923b8196b397863cf4f candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
171
+ batch done sequential task=PullCube-v1
172
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
173
+ execute sequential task=PullCube-v1 chart=ms-7df455efd413d77e6ce3 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
174
+ execute sequential task=PullCube-v1 chart=ms-7df455efd413d77e6ce3 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
175
+ execute sequential task=PullCube-v1 chart=ms-7df455efd413d77e6ce3 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
176
+ execute sequential task=PullCube-v1 chart=ms-7df455efd413d77e6ce3 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
177
+ execute sequential task=PullCube-v1 chart=ms-7df455efd413d77e6ce3 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
178
+ execute sequential task=PullCube-v1 chart=ms-7df455efd413d77e6ce3 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
179
+ execute sequential task=PullCube-v1 chart=ms-7df455efd413d77e6ce3 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
180
+ execute sequential task=PullCube-v1 chart=ms-7df455efd413d77e6ce3 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
181
+ execute sequential task=PullCube-v1 chart=ms-7df455efd413d77e6ce3 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
182
+ batch done sequential task=PullCube-v1
183
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
184
+ execute sequential task=PullCube-v1 chart=ms-7e295b5859c07c90487f candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
185
+ execute sequential task=PullCube-v1 chart=ms-7e295b5859c07c90487f candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
186
+ execute sequential task=PullCube-v1 chart=ms-7e295b5859c07c90487f candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
187
+ execute sequential task=PullCube-v1 chart=ms-7e295b5859c07c90487f candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
188
+ execute sequential task=PullCube-v1 chart=ms-7e295b5859c07c90487f candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
189
+ execute sequential task=PullCube-v1 chart=ms-7e295b5859c07c90487f candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
190
+ execute sequential task=PullCube-v1 chart=ms-7e295b5859c07c90487f candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
191
+ execute sequential task=PullCube-v1 chart=ms-7e295b5859c07c90487f candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
192
+ execute sequential task=PullCube-v1 chart=ms-7e295b5859c07c90487f candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
193
+ batch done sequential task=PullCube-v1
194
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
195
+ execute sequential task=PullCube-v1 chart=ms-8577816da2917a2bb63d candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
196
+ execute sequential task=PullCube-v1 chart=ms-8577816da2917a2bb63d candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
197
+ execute sequential task=PullCube-v1 chart=ms-8577816da2917a2bb63d candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
198
+ execute sequential task=PullCube-v1 chart=ms-8577816da2917a2bb63d candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
199
+ execute sequential task=PullCube-v1 chart=ms-8577816da2917a2bb63d candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
200
+ execute sequential task=PullCube-v1 chart=ms-8577816da2917a2bb63d candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
201
+ execute sequential task=PullCube-v1 chart=ms-8577816da2917a2bb63d candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
202
+ execute sequential task=PullCube-v1 chart=ms-8577816da2917a2bb63d candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
203
+ execute sequential task=PullCube-v1 chart=ms-8577816da2917a2bb63d candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
204
+ batch done sequential task=PullCube-v1
205
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
206
+ execute sequential task=PullCube-v1 chart=ms-8e56b0eaa2604db6749e candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
207
+ execute sequential task=PullCube-v1 chart=ms-8e56b0eaa2604db6749e candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
208
+ execute sequential task=PullCube-v1 chart=ms-8e56b0eaa2604db6749e candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
209
+ execute sequential task=PullCube-v1 chart=ms-8e56b0eaa2604db6749e candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
210
+ execute sequential task=PullCube-v1 chart=ms-8e56b0eaa2604db6749e candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
211
+ execute sequential task=PullCube-v1 chart=ms-8e56b0eaa2604db6749e candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
212
+ execute sequential task=PullCube-v1 chart=ms-8e56b0eaa2604db6749e candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
213
+ execute sequential task=PullCube-v1 chart=ms-8e56b0eaa2604db6749e candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
214
+ execute sequential task=PullCube-v1 chart=ms-8e56b0eaa2604db6749e candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
215
+ batch done sequential task=PullCube-v1
216
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
217
+ execute sequential task=PullCube-v1 chart=ms-972c5be72d0b24be7cd5 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
218
+ execute sequential task=PullCube-v1 chart=ms-972c5be72d0b24be7cd5 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
219
+ execute sequential task=PullCube-v1 chart=ms-972c5be72d0b24be7cd5 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
220
+ execute sequential task=PullCube-v1 chart=ms-972c5be72d0b24be7cd5 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
221
+ execute sequential task=PullCube-v1 chart=ms-972c5be72d0b24be7cd5 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
222
+ execute sequential task=PullCube-v1 chart=ms-972c5be72d0b24be7cd5 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
223
+ execute sequential task=PullCube-v1 chart=ms-972c5be72d0b24be7cd5 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
224
+ execute sequential task=PullCube-v1 chart=ms-972c5be72d0b24be7cd5 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
225
+ execute sequential task=PullCube-v1 chart=ms-972c5be72d0b24be7cd5 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
226
+ batch done sequential task=PullCube-v1
227
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
228
+ execute sequential task=PullCube-v1 chart=ms-9757fc494e79b5543bb6 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
229
+ execute sequential task=PullCube-v1 chart=ms-9757fc494e79b5543bb6 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
230
+ execute sequential task=PullCube-v1 chart=ms-9757fc494e79b5543bb6 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
231
+ execute sequential task=PullCube-v1 chart=ms-9757fc494e79b5543bb6 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
232
+ execute sequential task=PullCube-v1 chart=ms-9757fc494e79b5543bb6 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
233
+ execute sequential task=PullCube-v1 chart=ms-9757fc494e79b5543bb6 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
234
+ execute sequential task=PullCube-v1 chart=ms-9757fc494e79b5543bb6 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
235
+ execute sequential task=PullCube-v1 chart=ms-9757fc494e79b5543bb6 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
236
+ execute sequential task=PullCube-v1 chart=ms-9757fc494e79b5543bb6 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
237
+ batch done sequential task=PullCube-v1
238
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
239
+ execute sequential task=PullCube-v1 chart=ms-a8731059713697ce5e2a candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
240
+ execute sequential task=PullCube-v1 chart=ms-a8731059713697ce5e2a candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
241
+ execute sequential task=PullCube-v1 chart=ms-a8731059713697ce5e2a candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
242
+ execute sequential task=PullCube-v1 chart=ms-a8731059713697ce5e2a candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
243
+ execute sequential task=PullCube-v1 chart=ms-a8731059713697ce5e2a candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
244
+ execute sequential task=PullCube-v1 chart=ms-a8731059713697ce5e2a candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
245
+ execute sequential task=PullCube-v1 chart=ms-a8731059713697ce5e2a candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
246
+ execute sequential task=PullCube-v1 chart=ms-a8731059713697ce5e2a candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
247
+ execute sequential task=PullCube-v1 chart=ms-a8731059713697ce5e2a candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
248
+ batch done sequential task=PullCube-v1
249
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
250
+ execute sequential task=PullCube-v1 chart=ms-ac94b48febe523da15a5 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
251
+ execute sequential task=PullCube-v1 chart=ms-ac94b48febe523da15a5 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
252
+ execute sequential task=PullCube-v1 chart=ms-ac94b48febe523da15a5 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
253
+ execute sequential task=PullCube-v1 chart=ms-ac94b48febe523da15a5 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
254
+ execute sequential task=PullCube-v1 chart=ms-ac94b48febe523da15a5 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
255
+ execute sequential task=PullCube-v1 chart=ms-ac94b48febe523da15a5 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
256
+ execute sequential task=PullCube-v1 chart=ms-ac94b48febe523da15a5 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
257
+ execute sequential task=PullCube-v1 chart=ms-ac94b48febe523da15a5 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
258
+ execute sequential task=PullCube-v1 chart=ms-ac94b48febe523da15a5 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
259
+ batch done sequential task=PullCube-v1
260
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
261
+ execute sequential task=PullCube-v1 chart=ms-b49043f15f4846de5b13 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
262
+ execute sequential task=PullCube-v1 chart=ms-b49043f15f4846de5b13 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
263
+ execute sequential task=PullCube-v1 chart=ms-b49043f15f4846de5b13 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
264
+ execute sequential task=PullCube-v1 chart=ms-b49043f15f4846de5b13 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
265
+ execute sequential task=PullCube-v1 chart=ms-b49043f15f4846de5b13 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
266
+ execute sequential task=PullCube-v1 chart=ms-b49043f15f4846de5b13 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
267
+ execute sequential task=PullCube-v1 chart=ms-b49043f15f4846de5b13 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
268
+ execute sequential task=PullCube-v1 chart=ms-b49043f15f4846de5b13 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
269
+ execute sequential task=PullCube-v1 chart=ms-b49043f15f4846de5b13 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
270
+ batch done sequential task=PullCube-v1
271
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
272
+ execute sequential task=PullCube-v1 chart=ms-b7d7021b782f73816083 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
273
+ execute sequential task=PullCube-v1 chart=ms-b7d7021b782f73816083 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
274
+ execute sequential task=PullCube-v1 chart=ms-b7d7021b782f73816083 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
275
+ execute sequential task=PullCube-v1 chart=ms-b7d7021b782f73816083 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
276
+ execute sequential task=PullCube-v1 chart=ms-b7d7021b782f73816083 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
277
+ execute sequential task=PullCube-v1 chart=ms-b7d7021b782f73816083 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
278
+ execute sequential task=PullCube-v1 chart=ms-b7d7021b782f73816083 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
279
+ execute sequential task=PullCube-v1 chart=ms-b7d7021b782f73816083 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
280
+ execute sequential task=PullCube-v1 chart=ms-b7d7021b782f73816083 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
281
+ batch done sequential task=PullCube-v1
282
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
283
+ execute sequential task=PullCube-v1 chart=ms-b8c3ec3ea915c5d35b3e candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
284
+ execute sequential task=PullCube-v1 chart=ms-b8c3ec3ea915c5d35b3e candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
285
+ execute sequential task=PullCube-v1 chart=ms-b8c3ec3ea915c5d35b3e candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
286
+ execute sequential task=PullCube-v1 chart=ms-b8c3ec3ea915c5d35b3e candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
287
+ execute sequential task=PullCube-v1 chart=ms-b8c3ec3ea915c5d35b3e candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
288
+ execute sequential task=PullCube-v1 chart=ms-b8c3ec3ea915c5d35b3e candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
289
+ execute sequential task=PullCube-v1 chart=ms-b8c3ec3ea915c5d35b3e candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
290
+ execute sequential task=PullCube-v1 chart=ms-b8c3ec3ea915c5d35b3e candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
291
+ execute sequential task=PullCube-v1 chart=ms-b8c3ec3ea915c5d35b3e candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
292
+ batch done sequential task=PullCube-v1
293
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
294
+ execute sequential task=PullCube-v1 chart=ms-b8fc1568ce66c45c8d8b candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
295
+ execute sequential task=PullCube-v1 chart=ms-b8fc1568ce66c45c8d8b candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
296
+ execute sequential task=PullCube-v1 chart=ms-b8fc1568ce66c45c8d8b candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
297
+ execute sequential task=PullCube-v1 chart=ms-b8fc1568ce66c45c8d8b candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
298
+ execute sequential task=PullCube-v1 chart=ms-b8fc1568ce66c45c8d8b candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
299
+ execute sequential task=PullCube-v1 chart=ms-b8fc1568ce66c45c8d8b candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
300
+ execute sequential task=PullCube-v1 chart=ms-b8fc1568ce66c45c8d8b candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
301
+ execute sequential task=PullCube-v1 chart=ms-b8fc1568ce66c45c8d8b candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
302
+ execute sequential task=PullCube-v1 chart=ms-b8fc1568ce66c45c8d8b candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
303
+ batch done sequential task=PullCube-v1
304
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
305
+ execute sequential task=PullCube-v1 chart=ms-da555a03dd382d93e119 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
306
+ execute sequential task=PullCube-v1 chart=ms-da555a03dd382d93e119 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
307
+ execute sequential task=PullCube-v1 chart=ms-da555a03dd382d93e119 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
308
+ execute sequential task=PullCube-v1 chart=ms-da555a03dd382d93e119 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
309
+ execute sequential task=PullCube-v1 chart=ms-da555a03dd382d93e119 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
310
+ execute sequential task=PullCube-v1 chart=ms-da555a03dd382d93e119 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
311
+ execute sequential task=PullCube-v1 chart=ms-da555a03dd382d93e119 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
312
+ execute sequential task=PullCube-v1 chart=ms-da555a03dd382d93e119 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
313
+ execute sequential task=PullCube-v1 chart=ms-da555a03dd382d93e119 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
314
+ batch done sequential task=PullCube-v1
315
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
316
+ execute sequential task=PullCube-v1 chart=ms-f5dd1c43ebc4c0cfc3f1 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
317
+ execute sequential task=PullCube-v1 chart=ms-f5dd1c43ebc4c0cfc3f1 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
318
+ execute sequential task=PullCube-v1 chart=ms-f5dd1c43ebc4c0cfc3f1 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
319
+ execute sequential task=PullCube-v1 chart=ms-f5dd1c43ebc4c0cfc3f1 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
320
+ execute sequential task=PullCube-v1 chart=ms-f5dd1c43ebc4c0cfc3f1 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
321
+ execute sequential task=PullCube-v1 chart=ms-f5dd1c43ebc4c0cfc3f1 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
322
+ execute sequential task=PullCube-v1 chart=ms-f5dd1c43ebc4c0cfc3f1 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
323
+ execute sequential task=PullCube-v1 chart=ms-f5dd1c43ebc4c0cfc3f1 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
324
+ execute sequential task=PullCube-v1 chart=ms-f5dd1c43ebc4c0cfc3f1 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
325
+ batch done sequential task=PullCube-v1
326
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
327
+ execute sequential task=PullCube-v1 chart=ms-f9740a08a98cb251b761 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
328
+ execute sequential task=PullCube-v1 chart=ms-f9740a08a98cb251b761 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
329
+ execute sequential task=PullCube-v1 chart=ms-f9740a08a98cb251b761 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
330
+ execute sequential task=PullCube-v1 chart=ms-f9740a08a98cb251b761 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
331
+ execute sequential task=PullCube-v1 chart=ms-f9740a08a98cb251b761 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
332
+ execute sequential task=PullCube-v1 chart=ms-f9740a08a98cb251b761 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
333
+ execute sequential task=PullCube-v1 chart=ms-f9740a08a98cb251b761 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
334
+ execute sequential task=PullCube-v1 chart=ms-f9740a08a98cb251b761 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
335
+ execute sequential task=PullCube-v1 chart=ms-f9740a08a98cb251b761 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
336
+ batch done sequential task=PullCube-v1
337
+ env init sequential task=PullCube-v1 batch=1 sim=physx_cpu render=cpu
338
+ execute sequential task=PullCube-v1 chart=ms-fa8b4ba41ecac8030271 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
339
+ execute sequential task=PullCube-v1 chart=ms-fa8b4ba41ecac8030271 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
340
+ execute sequential task=PullCube-v1 chart=ms-fa8b4ba41ecac8030271 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
341
+ execute sequential task=PullCube-v1 chart=ms-fa8b4ba41ecac8030271 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
342
+ execute sequential task=PullCube-v1 chart=ms-fa8b4ba41ecac8030271 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
343
+ execute sequential task=PullCube-v1 chart=ms-fa8b4ba41ecac8030271 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
344
+ execute sequential task=PullCube-v1 chart=ms-fa8b4ba41ecac8030271 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
345
+ execute sequential task=PullCube-v1 chart=ms-fa8b4ba41ecac8030271 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
346
+ execute sequential task=PullCube-v1 chart=ms-fa8b4ba41ecac8030271 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
347
+ batch done sequential task=PullCube-v1
348
+ env init sequential task=PushCube-v1 batch=1 sim=physx_cpu render=cpu
349
+ execute sequential task=PushCube-v1 chart=ms-05f34a90fa7355262b8e candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
350
+ execute sequential task=PushCube-v1 chart=ms-05f34a90fa7355262b8e candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
351
+ execute sequential task=PushCube-v1 chart=ms-05f34a90fa7355262b8e candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
352
+ execute sequential task=PushCube-v1 chart=ms-05f34a90fa7355262b8e candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
353
+ execute sequential task=PushCube-v1 chart=ms-05f34a90fa7355262b8e candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
354
+ execute sequential task=PushCube-v1 chart=ms-05f34a90fa7355262b8e candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
355
+ execute sequential task=PushCube-v1 chart=ms-05f34a90fa7355262b8e candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
356
+ execute sequential task=PushCube-v1 chart=ms-05f34a90fa7355262b8e candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
357
+ execute sequential task=PushCube-v1 chart=ms-05f34a90fa7355262b8e candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
358
+ batch done sequential task=PushCube-v1
359
+ env init sequential task=PushCube-v1 batch=1 sim=physx_cpu render=cpu
360
+ execute sequential task=PushCube-v1 chart=ms-2413d1155c87f0442485 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
361
+ execute sequential task=PushCube-v1 chart=ms-2413d1155c87f0442485 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
362
+ execute sequential task=PushCube-v1 chart=ms-2413d1155c87f0442485 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
363
+ execute sequential task=PushCube-v1 chart=ms-2413d1155c87f0442485 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
364
+ execute sequential task=PushCube-v1 chart=ms-2413d1155c87f0442485 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
365
+ execute sequential task=PushCube-v1 chart=ms-2413d1155c87f0442485 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
366
+ execute sequential task=PushCube-v1 chart=ms-2413d1155c87f0442485 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
367
+ execute sequential task=PushCube-v1 chart=ms-2413d1155c87f0442485 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
368
+ execute sequential task=PushCube-v1 chart=ms-2413d1155c87f0442485 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
369
+ batch done sequential task=PushCube-v1
370
+ env init sequential task=PushCube-v1 batch=1 sim=physx_cpu render=cpu
371
+ execute sequential task=PushCube-v1 chart=ms-54346a8f79bd9dca1dfe candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
372
+ execute sequential task=PushCube-v1 chart=ms-54346a8f79bd9dca1dfe candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
373
+ execute sequential task=PushCube-v1 chart=ms-54346a8f79bd9dca1dfe candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
374
+ execute sequential task=PushCube-v1 chart=ms-54346a8f79bd9dca1dfe candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
375
+ execute sequential task=PushCube-v1 chart=ms-54346a8f79bd9dca1dfe candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
376
+ execute sequential task=PushCube-v1 chart=ms-54346a8f79bd9dca1dfe candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
377
+ execute sequential task=PushCube-v1 chart=ms-54346a8f79bd9dca1dfe candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
378
+ execute sequential task=PushCube-v1 chart=ms-54346a8f79bd9dca1dfe candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
379
+ execute sequential task=PushCube-v1 chart=ms-54346a8f79bd9dca1dfe candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
380
+ batch done sequential task=PushCube-v1
381
+ env init sequential task=PushCube-v1 batch=1 sim=physx_cpu render=cpu
382
+ execute sequential task=PushCube-v1 chart=ms-64f98d163b6bff29332b candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
383
+ execute sequential task=PushCube-v1 chart=ms-64f98d163b6bff29332b candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
384
+ execute sequential task=PushCube-v1 chart=ms-64f98d163b6bff29332b candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
385
+ execute sequential task=PushCube-v1 chart=ms-64f98d163b6bff29332b candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
386
+ execute sequential task=PushCube-v1 chart=ms-64f98d163b6bff29332b candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
387
+ execute sequential task=PushCube-v1 chart=ms-64f98d163b6bff29332b candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
388
+ execute sequential task=PushCube-v1 chart=ms-64f98d163b6bff29332b candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
389
+ execute sequential task=PushCube-v1 chart=ms-64f98d163b6bff29332b candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
390
+ execute sequential task=PushCube-v1 chart=ms-64f98d163b6bff29332b candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
391
+ batch done sequential task=PushCube-v1
392
+ env init sequential task=PushCube-v1 batch=1 sim=physx_cpu render=cpu
393
+ execute sequential task=PushCube-v1 chart=ms-c2ae2b78ab7891e74c80 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
394
+ execute sequential task=PushCube-v1 chart=ms-c2ae2b78ab7891e74c80 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
395
+ execute sequential task=PushCube-v1 chart=ms-c2ae2b78ab7891e74c80 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
396
+ execute sequential task=PushCube-v1 chart=ms-c2ae2b78ab7891e74c80 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
397
+ execute sequential task=PushCube-v1 chart=ms-c2ae2b78ab7891e74c80 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
398
+ execute sequential task=PushCube-v1 chart=ms-c2ae2b78ab7891e74c80 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
399
+ execute sequential task=PushCube-v1 chart=ms-c2ae2b78ab7891e74c80 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
400
+ execute sequential task=PushCube-v1 chart=ms-c2ae2b78ab7891e74c80 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
401
+ execute sequential task=PushCube-v1 chart=ms-c2ae2b78ab7891e74c80 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
402
+ batch done sequential task=PushCube-v1
403
+ env init sequential task=PushCube-v1 batch=1 sim=physx_cpu render=cpu
404
+ execute sequential task=PushCube-v1 chart=ms-d4c37ed44fb338c1d551 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
405
+ execute sequential task=PushCube-v1 chart=ms-d4c37ed44fb338c1d551 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
406
+ execute sequential task=PushCube-v1 chart=ms-d4c37ed44fb338c1d551 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
407
+ execute sequential task=PushCube-v1 chart=ms-d4c37ed44fb338c1d551 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
408
+ execute sequential task=PushCube-v1 chart=ms-d4c37ed44fb338c1d551 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
409
+ execute sequential task=PushCube-v1 chart=ms-d4c37ed44fb338c1d551 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
410
+ execute sequential task=PushCube-v1 chart=ms-d4c37ed44fb338c1d551 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
411
+ execute sequential task=PushCube-v1 chart=ms-d4c37ed44fb338c1d551 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
412
+ execute sequential task=PushCube-v1 chart=ms-d4c37ed44fb338c1d551 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
413
+ batch done sequential task=PushCube-v1
414
+ env init sequential task=StackCube-v1 batch=1 sim=physx_cpu render=cpu
415
+ execute sequential task=StackCube-v1 chart=ms-146ff6c3f22b3a49af3a candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
416
+ execute sequential task=StackCube-v1 chart=ms-146ff6c3f22b3a49af3a candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
417
+ execute sequential task=StackCube-v1 chart=ms-146ff6c3f22b3a49af3a candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
418
+ execute sequential task=StackCube-v1 chart=ms-146ff6c3f22b3a49af3a candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
419
+ execute sequential task=StackCube-v1 chart=ms-146ff6c3f22b3a49af3a candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
420
+ execute sequential task=StackCube-v1 chart=ms-146ff6c3f22b3a49af3a candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
421
+ execute sequential task=StackCube-v1 chart=ms-146ff6c3f22b3a49af3a candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
422
+ execute sequential task=StackCube-v1 chart=ms-146ff6c3f22b3a49af3a candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
423
+ execute sequential task=StackCube-v1 chart=ms-146ff6c3f22b3a49af3a candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
424
+ batch done sequential task=StackCube-v1
425
+ env init sequential task=StackCube-v1 batch=1 sim=physx_cpu render=cpu
426
+ execute sequential task=StackCube-v1 chart=ms-15fc8cc6f74cf589056f candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
427
+ execute sequential task=StackCube-v1 chart=ms-15fc8cc6f74cf589056f candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
428
+ execute sequential task=StackCube-v1 chart=ms-15fc8cc6f74cf589056f candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
429
+ execute sequential task=StackCube-v1 chart=ms-15fc8cc6f74cf589056f candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
430
+ execute sequential task=StackCube-v1 chart=ms-15fc8cc6f74cf589056f candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
431
+ execute sequential task=StackCube-v1 chart=ms-15fc8cc6f74cf589056f candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
432
+ execute sequential task=StackCube-v1 chart=ms-15fc8cc6f74cf589056f candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
433
+ execute sequential task=StackCube-v1 chart=ms-15fc8cc6f74cf589056f candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
434
+ execute sequential task=StackCube-v1 chart=ms-15fc8cc6f74cf589056f candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
435
+ batch done sequential task=StackCube-v1
436
+ env init sequential task=StackCube-v1 batch=1 sim=physx_cpu render=cpu
437
+ execute sequential task=StackCube-v1 chart=ms-36bc1364d2439099e9ed candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
438
+ execute sequential task=StackCube-v1 chart=ms-36bc1364d2439099e9ed candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
439
+ execute sequential task=StackCube-v1 chart=ms-36bc1364d2439099e9ed candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
440
+ execute sequential task=StackCube-v1 chart=ms-36bc1364d2439099e9ed candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
441
+ execute sequential task=StackCube-v1 chart=ms-36bc1364d2439099e9ed candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
442
+ execute sequential task=StackCube-v1 chart=ms-36bc1364d2439099e9ed candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
443
+ execute sequential task=StackCube-v1 chart=ms-36bc1364d2439099e9ed candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
444
+ execute sequential task=StackCube-v1 chart=ms-36bc1364d2439099e9ed candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
445
+ execute sequential task=StackCube-v1 chart=ms-36bc1364d2439099e9ed candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
446
+ batch done sequential task=StackCube-v1
447
+ env init sequential task=StackCube-v1 batch=1 sim=physx_cpu render=cpu
448
+ execute sequential task=StackCube-v1 chart=ms-420280f15a522ce21c72 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
449
+ execute sequential task=StackCube-v1 chart=ms-420280f15a522ce21c72 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
450
+ execute sequential task=StackCube-v1 chart=ms-420280f15a522ce21c72 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
451
+ execute sequential task=StackCube-v1 chart=ms-420280f15a522ce21c72 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
452
+ execute sequential task=StackCube-v1 chart=ms-420280f15a522ce21c72 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
453
+ execute sequential task=StackCube-v1 chart=ms-420280f15a522ce21c72 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
454
+ execute sequential task=StackCube-v1 chart=ms-420280f15a522ce21c72 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
455
+ execute sequential task=StackCube-v1 chart=ms-420280f15a522ce21c72 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
456
+ execute sequential task=StackCube-v1 chart=ms-420280f15a522ce21c72 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
457
+ batch done sequential task=StackCube-v1
458
+ env init sequential task=StackCube-v1 batch=1 sim=physx_cpu render=cpu
459
+ execute sequential task=StackCube-v1 chart=ms-4aed1197d0906ae51c34 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
460
+ execute sequential task=StackCube-v1 chart=ms-4aed1197d0906ae51c34 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
461
+ execute sequential task=StackCube-v1 chart=ms-4aed1197d0906ae51c34 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
462
+ execute sequential task=StackCube-v1 chart=ms-4aed1197d0906ae51c34 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
463
+ execute sequential task=StackCube-v1 chart=ms-4aed1197d0906ae51c34 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
464
+ execute sequential task=StackCube-v1 chart=ms-4aed1197d0906ae51c34 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
465
+ execute sequential task=StackCube-v1 chart=ms-4aed1197d0906ae51c34 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
466
+ execute sequential task=StackCube-v1 chart=ms-4aed1197d0906ae51c34 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
467
+ execute sequential task=StackCube-v1 chart=ms-4aed1197d0906ae51c34 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
468
+ batch done sequential task=StackCube-v1
469
+ env init sequential task=StackCube-v1 batch=1 sim=physx_cpu render=cpu
470
+ execute sequential task=StackCube-v1 chart=ms-55e78c04643bd06fd5f8 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
471
+ execute sequential task=StackCube-v1 chart=ms-55e78c04643bd06fd5f8 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
472
+ execute sequential task=StackCube-v1 chart=ms-55e78c04643bd06fd5f8 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
473
+ execute sequential task=StackCube-v1 chart=ms-55e78c04643bd06fd5f8 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
474
+ execute sequential task=StackCube-v1 chart=ms-55e78c04643bd06fd5f8 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
475
+ execute sequential task=StackCube-v1 chart=ms-55e78c04643bd06fd5f8 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
476
+ execute sequential task=StackCube-v1 chart=ms-55e78c04643bd06fd5f8 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
477
+ execute sequential task=StackCube-v1 chart=ms-55e78c04643bd06fd5f8 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
478
+ execute sequential task=StackCube-v1 chart=ms-55e78c04643bd06fd5f8 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
479
+ batch done sequential task=StackCube-v1
480
+ env init sequential task=StackCube-v1 batch=1 sim=physx_cpu render=cpu
481
+ execute sequential task=StackCube-v1 chart=ms-77536801924a2a78688e candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
482
+ execute sequential task=StackCube-v1 chart=ms-77536801924a2a78688e candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
483
+ execute sequential task=StackCube-v1 chart=ms-77536801924a2a78688e candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
484
+ execute sequential task=StackCube-v1 chart=ms-77536801924a2a78688e candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
485
+ execute sequential task=StackCube-v1 chart=ms-77536801924a2a78688e candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
486
+ execute sequential task=StackCube-v1 chart=ms-77536801924a2a78688e candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
487
+ execute sequential task=StackCube-v1 chart=ms-77536801924a2a78688e candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
488
+ execute sequential task=StackCube-v1 chart=ms-77536801924a2a78688e candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
489
+ execute sequential task=StackCube-v1 chart=ms-77536801924a2a78688e candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
490
+ batch done sequential task=StackCube-v1
491
+ env init sequential task=StackCube-v1 batch=1 sim=physx_cpu render=cpu
492
+ execute sequential task=StackCube-v1 chart=ms-9dd1ff4c297a81b13071 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
493
+ execute sequential task=StackCube-v1 chart=ms-9dd1ff4c297a81b13071 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
494
+ execute sequential task=StackCube-v1 chart=ms-9dd1ff4c297a81b13071 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
495
+ execute sequential task=StackCube-v1 chart=ms-9dd1ff4c297a81b13071 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
496
+ execute sequential task=StackCube-v1 chart=ms-9dd1ff4c297a81b13071 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
497
+ execute sequential task=StackCube-v1 chart=ms-9dd1ff4c297a81b13071 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
498
+ execute sequential task=StackCube-v1 chart=ms-9dd1ff4c297a81b13071 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
499
+ execute sequential task=StackCube-v1 chart=ms-9dd1ff4c297a81b13071 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
500
+ execute sequential task=StackCube-v1 chart=ms-9dd1ff4c297a81b13071 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
501
+ batch done sequential task=StackCube-v1
502
+ env init sequential task=StackCube-v1 batch=1 sim=physx_cpu render=cpu
503
+ execute sequential task=StackCube-v1 chart=ms-d7254aa722729aa1a031 candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
504
+ execute sequential task=StackCube-v1 chart=ms-d7254aa722729aa1a031 candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
505
+ execute sequential task=StackCube-v1 chart=ms-d7254aa722729aa1a031 candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
506
+ execute sequential task=StackCube-v1 chart=ms-d7254aa722729aa1a031 candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
507
+ execute sequential task=StackCube-v1 chart=ms-d7254aa722729aa1a031 candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
508
+ execute sequential task=StackCube-v1 chart=ms-d7254aa722729aa1a031 candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
509
+ execute sequential task=StackCube-v1 chart=ms-d7254aa722729aa1a031 candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
510
+ execute sequential task=StackCube-v1 chart=ms-d7254aa722729aa1a031 candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
511
+ execute sequential task=StackCube-v1 chart=ms-d7254aa722729aa1a031 candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
512
+ batch done sequential task=StackCube-v1
513
+ env init sequential task=StackCube-v1 batch=1 sim=physx_cpu render=cpu
514
+ execute sequential task=StackCube-v1 chart=ms-d776afa14e6322cadedd candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
515
+ execute sequential task=StackCube-v1 chart=ms-d776afa14e6322cadedd candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
516
+ execute sequential task=StackCube-v1 chart=ms-d776afa14e6322cadedd candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
517
+ execute sequential task=StackCube-v1 chart=ms-d776afa14e6322cadedd candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
518
+ execute sequential task=StackCube-v1 chart=ms-d776afa14e6322cadedd candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
519
+ execute sequential task=StackCube-v1 chart=ms-d776afa14e6322cadedd candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
520
+ execute sequential task=StackCube-v1 chart=ms-d776afa14e6322cadedd candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
521
+ execute sequential task=StackCube-v1 chart=ms-d776afa14e6322cadedd candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
522
+ execute sequential task=StackCube-v1 chart=ms-d776afa14e6322cadedd candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
523
+ batch done sequential task=StackCube-v1
524
+ env init sequential task=StackCube-v1 batch=1 sim=physx_cpu render=cpu
525
+ execute sequential task=StackCube-v1 chart=ms-e887dbfa0133af46829d candidate=0 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
526
+ execute sequential task=StackCube-v1 chart=ms-e887dbfa0133af46829d candidate=1 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
527
+ execute sequential task=StackCube-v1 chart=ms-e887dbfa0133af46829d candidate=2 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
528
+ execute sequential task=StackCube-v1 chart=ms-e887dbfa0133af46829d candidate=3 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
529
+ execute sequential task=StackCube-v1 chart=ms-e887dbfa0133af46829d candidate=4 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
530
+ execute sequential task=StackCube-v1 chart=ms-e887dbfa0133af46829d candidate=5 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
531
+ execute sequential task=StackCube-v1 chart=ms-e887dbfa0133af46829d candidate=6 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
532
+ execute sequential task=StackCube-v1 chart=ms-e887dbfa0133af46829d candidate=7 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
533
+ execute sequential task=StackCube-v1 chart=ms-e887dbfa0133af46829d candidate=8 shape=(1, 1, 16, 7) clip_actions=False execution_action_scale=1.0 execution_action_transform=identity
534
+ batch done sequential task=StackCube-v1
535
+ rollout complete rows=48
workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/source_data_hash.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ 67f7b4a692f7c7e71da24378dc71e3399c5d35afc14c494869eba6087b765c42
workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/source_index.json ADDED
The diff for this file is too large to render. See raw diff
 
workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/source_split_hash.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ 867a94a1c67686c082765671a7bf8d74a9d633085e1b2619c83197c60bf58372
workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/split_hash.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ 84151df9e88c219b76751677558cc5bbbc19082bf358cea7330158e9e403e25f
workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/table.tex ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ % Auto-generated by scripts/eval_metrics.py
2
+ \begin{tabular}{lrrrr}
3
+ \toprule
4
+ Metric & N & Micro mean & CI low & CI high \\
5
+ \midrule
6
+ any\_generated\_unsafe\_known\_at\_8 & 48 & 0.3750 & 0.2500 & 0.5208 \\
7
+ base\_safety\_label\_known & 48 & 1.0000 & 1.0000 & 1.0000 \\
8
+ base\_success & 48 & 0.3333 & 0.2083 & 0.4583 \\
9
+ base\_unsafe\_known & 48 & 0.0000 & 0.0000 & 0.0000 \\
10
+ base\_utility & 48 & 0.8338 & 0.5920 & 1.0708 \\
11
+ branch\_car\_at\_8 & 48 & 0.2158 & 0.1011 & 0.3661 \\
12
+ generated\_safety\_label\_coverage\_at\_8 & 48 & 1.0000 & 1.0000 & 1.0000 \\
13
+ generated\_unsafe\_rate\_known\_at\_8 & 48 & 0.3255 & 0.2214 & 0.4505 \\
14
+ hidden\_chart\_oracle\_success\_at\_8 & 48 & 0.7292 & 0.6042 & 0.8333 \\
15
+ hidden\_chart\_oracle\_utility\_at\_8 & 48 & 1.5183 & 1.2906 & 1.7056 \\
16
+ ncar\_to\_hidden\_chart\_oracle\_at\_8 & 35 & 1.1333 & 0.8365 & 1.5441 \\
17
+ ncar\_to\_proposal\_oracle\_at\_8 & 40 & 1.3481 & 0.5300 & 2.5628 \\
18
+ outcome\_ptr\_at\_8 & 48 & 0.3958 & 0.2708 & 0.5417 \\
19
+ pairwise\_causal\_calibration\_ece & 48 & 0.1925 & 0.1548 & 0.2249 \\
20
+ proposal\_oracle\_safety\_label\_known\_at\_8 & 48 & 1.0000 & 1.0000 & 1.0000 \\
21
+ proposal\_oracle\_success\_at\_8 & 48 & 0.2292 & 0.1042 & 0.3333 \\
22
+ proposal\_oracle\_success\_gain\_over\_base\_at\_8 & 48 & -0.1042 & -0.2292 & 0.0208 \\
23
+ proposal\_oracle\_unsafe\_known\_at\_8 & 48 & 0.3125 & 0.1875 & 0.4583 \\
24
+ proposal\_oracle\_utility\_at\_8 & 48 & 0.6506 & 0.4362 & 0.8420 \\
25
+ proposal\_oracle\_utility\_gain\_over\_base\_at\_8 & 48 & -0.1833 & -0.4133 & 0.0378 \\
26
+ selected\_safety\_label\_known\_at\_8 & 48 & 1.0000 & 1.0000 & 1.0000 \\
27
+ selected\_success\_at\_8 & 48 & 0.1250 & 0.0417 & 0.2292 \\
28
+ selected\_success\_gain\_over\_base\_at\_8 & 48 & -0.2083 & -0.3125 & -0.0833 \\
29
+ selected\_unsafe\_known\_at\_8 & 48 & 0.2917 & 0.1875 & 0.4375 \\
30
+ selected\_utility\_at\_8 & 48 & 0.4348 & 0.2639 & 0.6220 \\
31
+ selected\_utility\_gain\_over\_base\_at\_8 & 48 & -0.3990 & -0.5913 & -0.1621 \\
32
+ selector\_gap\_fraction\_to\_hidden\_at\_8 & 35 & 0.3723 & 0.1653 & 0.6711 \\
33
+ selector\_regret\_at\_8 & 48 & 0.2158 & 0.1011 & 0.3661 \\
34
+ success\_selector\_gap\_at\_8 & 48 & 0.1042 & 0.0208 & 0.2083 \\
35
+ success\_support\_gap\_at\_8 & 48 & 0.5000 & 0.3333 & 0.6667 \\
36
+ success\_total\_car\_to\_hidden\_at\_8 & 48 & 0.6042 & 0.4583 & 0.7292 \\
37
+ support\_gap\_at\_8 & 48 & 0.8699 & 0.5969 & 1.1405 \\
38
+ support\_gap\_fraction\_to\_hidden\_at\_8 & 35 & 0.8852 & 0.6005 & 1.2127 \\
39
+ total\_car\_to\_hidden\_at\_8 & 48 & 1.0835 & 0.8389 & 1.2861 \\
40
+ \bottomrule
41
+ \end{tabular}
workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/target_index.json ADDED
@@ -0,0 +1,911 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "audience": "evaluator_only",
3
+ "base_candidate_types": [
4
+ "policy",
5
+ "policy_residual",
6
+ "anchor",
7
+ "base",
8
+ "expert"
9
+ ],
10
+ "candidate_type_counts": {
11
+ "expert": 410,
12
+ "near_miss": 820,
13
+ "no_op": 410,
14
+ "random_negative": 4100,
15
+ "wrong_direction": 410,
16
+ "wrong_gripper": 410
17
+ },
18
+ "content_hash": "0c38d73dc9bc4ab756e12006671b5d12f1a23358b281e54b3f6a9ae3cbf6e9af",
19
+ "dataset": "/scratch/knguy52/dovla/experiments/six_task_h16_collection",
20
+ "deployment_candidate_excludes_expert": true,
21
+ "deployment_clean": false,
22
+ "embedding_content_hash": "981d3b1b0aa9dc962e5a22fd695fa9befd21e1334f23b8c3843591e8be89ef83",
23
+ "epsilon": 0.05,
24
+ "format": "cil_charts_npz",
25
+ "group_hashes": [
26
+ "0007bdeb2cd982d3d00c7e6ff04b6a464135679254e223a5a10392734b8bc6d8",
27
+ "001c01ea87ba60e3ec7040ae6af17bfe4fef64fbf40ad906fe9258c9f85c5670",
28
+ "006cfb2673b00a6eb8b01af6cb8db204644de9fe5e9bf901e615fa9681cf67bd",
29
+ "00d1fad89bb7e1cc01e432e0a8c1dc39edc7045d2d25ba7c52256142f9bc65ce",
30
+ "011a665d2baee1b90a589a205463340204b7876a99c942f742523ab7a8ba3ca3",
31
+ "0155f9f15f95b3ad606169d1a4b59acbdc87a773f813376518f26c5d9803ea46",
32
+ "017ebedb06ec921569a797619e60f77551213b496a8ea3c2090c6646422ee6b5",
33
+ "0192ac22f91a294bbe7f46d861b1ccae131ce03bf2ebb4ba697c662b381c5663",
34
+ "01c2280c26df28f19e73270020c6cbf35a1b739c75d2feab218b5d40a9dad833",
35
+ "03bc15045e86dd5d67b89da015844ec52d13ba891ae0f4b688d917a4b8fbd83e",
36
+ "041d9ba8e26c507546ef7a2f66cb7a86668d7c787109f1e7eba310a5b43c7e31",
37
+ "04e1992a37d6b1709dac4e28293d6c0d72820804fac25484e00bbc1e4e85475b",
38
+ "0539ee8d5137076cadeae47404a0abe44d481204301deb270e6195549dbb1269",
39
+ "060e1b400d7abb4299e11bd36999d4bce7619d78e1efa1cd6bb32448d18b82c6",
40
+ "060e87c11d7e86ba8ea45b2b29af89e13a7361f760086655fdc31264d3b04aa4",
41
+ "07816113197976ecaefd996167fb5e52546b19a8cabf7039977b7783d0ccb2bf",
42
+ "07fc1d31e75fcf0eb9e0fd9a98ba44f6a098e3133ad6d0445a1ad311cb266506",
43
+ "08a53f6c8bfde25489a92225280c06265510ad0ae4ec4198855d05bf04c7634b",
44
+ "08db98843e2d5fb7ee992c3e24b9fd9e2fd8a19e6f40121773cc2d5f20fb6af9",
45
+ "09f1c778bdddceb7899e203f6c753e6d0253336c8d75af278461fee391558b61",
46
+ "0a63c4a73b38e0d4a76c8d29bee9854e127e2031428c90e386efee75af003605",
47
+ "0ab666a8ae83fd476443f8ebc8f0b2691ad50e2eb1f0758f87794a3462923793",
48
+ "0b95650f3d54877b77a9233592c962c57f8bad12fbdaf921e71a68e7c23f51a9",
49
+ "0ca0c6ad41d25d93fa19e41c544c322fbd0b28cf0b0e958e58a0cc4372e4ea3b",
50
+ "0ca567a3525a032b0ebe77f43dfe52fae4cf7d5927d93665f91c8df38b012e34",
51
+ "0dc1912c8531fcf9d63e6749a869ee71f42a50b5611546e6ed444984c196c145",
52
+ "0ffe9443040477642faee0dea246f46552cb396f9c79dfeeb68c9df76ab3c748",
53
+ "10bfbae4e9ed03cafce9b24b296005666d7d59b41ada06e8f5ccb036a6026c01",
54
+ "110094e9c49ec0207ec52f6b8811e69e78eb2640ff324b8a3df0a051089749bf",
55
+ "1130fedbc94b55e0bc812730a8e868ac57f1caf9c47d29ee76ee5e30990dd5a0",
56
+ "118ae5475b33c80dbe0d15afd837cad81fefd30882ff5dd7a4f983a5f096ec1e",
57
+ "11a49c253e136f618265b8cc3a298c9330eaaaf27b90cddc81956e1be00e8e9e",
58
+ "11db66da8a58abd540863fb58670a0c872fecf04a0d4605ecc5aaf202ba9b44f",
59
+ "11edbe43f364dfd31a5828ad1b80e2098516e263ac48f7cfd5503aa3ad240dba",
60
+ "124680a15078f0d17c61f90627f3090e05506597713ffac03335db6b158c3e5d",
61
+ "1286eb0f819a1528450a601f86862e7da0902610afcd4234b6b6aacce65a7672",
62
+ "12ea115a8e2ff6f16d0c3204d3f509da16cdf4637e2fd7f88d0e59849af4f397",
63
+ "1332761b5e3a7caf73b98206b79c354732f061bd0e87f9c05ca54ec435948c65",
64
+ "1338f4e10decd4b2ef526d6a56e52f6bdd6c1edff755f821d3ed7011c8bb7710",
65
+ "13ba81ea4a28629132b3e2f6f79918644e50ff998c5f4ea420517486664046d0",
66
+ "13bd8f20d6c071f69abc0b50596a8b36eb937c81d5a068e0c75b58f9420e712c",
67
+ "156cb442d6089cc43d86909848c269035cc01d61656e0c60ab049b1278e9c129",
68
+ "15a4586f96ad878eeacd79cebaba12826404f415bb66c501e90dc057b0687d64",
69
+ "15dd893e5e2312081f4a9417ac36d979625f59fa34d06d38e3464be81efe87a4",
70
+ "16e933e25ec817fd4f7003d5a7c41fae6e5b7453df3b1ef44e909f918a51b97d",
71
+ "17498920ffddbc4da2157a94486acb91dfc38e4c35db9d7aa798733022c0029d",
72
+ "17648e6070d70caf099513f8b22bd5e05d6469c693735f91ecb9c14fd2d9ea90",
73
+ "17de3e35901cbbb48f0b82afb84ff8e458fb16f1b5d95cdbd4fda0b057607a11",
74
+ "18f8644cba3b47cdcb40de42a4311045f636ffa35409667d5271c9371d6f4a4f",
75
+ "19ba2745ea505a71af1411ae666c3d76691fea459f0c9fa2b303a317bb0477ec",
76
+ "19febae2111ae96f694bca67a33585e53e9cf746f2791302b821e9c8fbdd013c",
77
+ "1a07326a6db79bca54149427b5372a83b2b59e6f6871661531b53b3bf34891ee",
78
+ "1b7c2dd4731b1b141fb0a8cdde0b10ed8e5452cf1902de623eda00506ad2dd3c",
79
+ "1e671df1d262128a0fd7fe67827f41a77324121b2b589e9dae109ebd9df4b1f8",
80
+ "20b6fe19c6bb23ad1e4537441c01ac152df2dbd9f778162da30f65d86e6bce9a",
81
+ "20cab2c31288a131d80ea3f10ac0e38cee76b70a9fc9f87a34990e08787c3d9e",
82
+ "21dee04c93f399c0f8b577ad691edc1df072d20667b41dd0f668c6bdd20afb2c",
83
+ "25426f80d4feac9877cfed3648cceb7f183c8873494b897c7b7baee2d03f5f4e",
84
+ "26f527bcc8937fb979a2c57829d7f7f74c781c952131827b753b98e46a39340c",
85
+ "27330aad72bdf87bbe7895be6cc76bd79ed905b35bbdcb8abc4e88569fe2c3b4",
86
+ "27e60a9d68cb7e0fb5a2bc33baab8f60bfaed45f28ebc2b62c370fcc13e4c2f5",
87
+ "289a46819f19ddbd713b024c7adcc4876d131fadcec221b3afb19fba6f51e0b1",
88
+ "2927736a2ee4ac6fe22321499437215277bb7f4bc7d667ea05fb07168f8000c6",
89
+ "296f4972aaa4d080779c61f6397712c8e2a6ec17334177fea2f800275fe8a103",
90
+ "29ae3eb9b205af85419065034b255129b4a9b4aea5bfe88eb58ed2ca54a396ac",
91
+ "2a9aa97145b3978bab7cae14241a509bd7ac0001e852599f67c7ab0abbdf0294",
92
+ "2c2dd96ec73997f143a8f2a2a95e72cf81d434ea35d8545257d171b47b124d15",
93
+ "2d11bf97e91ac12402d0473647389db88906a432c45ca478ed91e120a521ade8",
94
+ "2d7cb0415ccc1d977c0d895ae130ee2aa6e1fd5929965ae56ca3ad5308949cc6",
95
+ "2e4ee75075d62464482af570e7f3b49996f5fff4bcc59f3f8d9f3ac7d3b83260",
96
+ "2e926dbc2b276c0faf482cbd4e1bc71fdb16060082f56f03ac414501870fa3ff",
97
+ "2f4dadc2c85fe72fc4dbd65c7c56b8b3575e8f1214e1bc33cbb845c9399c54d2",
98
+ "2f575f5fa8b95c7aaa7acb325dce6ae9aace80a7cd583560263607763a5f3c7e",
99
+ "2fc5d24ebb0a14f83b82e2b1e152953b8ecc751227bce3f542522e2e951c812e",
100
+ "302cf9c2c3f4b0b32800f27a93e3a72e8af78b0468fed7510a6b6d78b54f0444",
101
+ "30fd16ab2320464cc117580159e21267cdc800c367d0a414d4a14201d66acc6a",
102
+ "314101e4b9327d7a93fdde8db8691efe2e4b3115fdfb2e8e50b07db65fc3f7d8",
103
+ "31c82d6ce0b5154e8a3521472a93875659c6c196f3a0913f01c3f50018f3adc0",
104
+ "32a4c067144fedae644310fd76b2d31b0ea4033124d5aa4f58eadda0e01897ec",
105
+ "35615503a8aabbea5fca0b60bc48a5ee2c895614cf624b53f1659e86bcdead9b",
106
+ "3741c12184df2ff5f5684e1dbe2f5b42793596d06f5765c52cc3523f8f51aa9f",
107
+ "390a604b7a108529e8b5cf2c05d98ddd5f48a26bb303a51314e6e29cc58626e4",
108
+ "390c50924276ff6f5710a185124f893e10f3ca8fe3c3ff062ff8bd29f428f063",
109
+ "39307306d7ea31f45d96ea06077a7480c4f12eb2c0df1b6773d9a26a08441c7f",
110
+ "3939b5cfd9535e1ead171b2e7a4926a67c7b22ba851bb522042ea966eea689d3",
111
+ "396e4e2cb8b6c382148822ad790ce798a2d89e1b31541a4d7937d59ca17b4624",
112
+ "39daabe27804e1a04bfd80272eee8d0faaf3d829d5917e1682ee697677f619b9",
113
+ "3a4318a4127672d9e627d5a90932efd77a523904ca5bb09ef1aa5856db198378",
114
+ "3a95110704b90641fd930392654967270afff99064cb506d80f4b65f9fea5870",
115
+ "3ad13fb77cf2201159e82f6bc4a31d9d0432c7befad4aa0faeb46e5bc2b77939",
116
+ "3b66af1f06aa0dbff3bedbd570ef7da51e94a4905c194925c21e2e30e254c73a",
117
+ "3b8857c883ccff813f632384ec0bf8903402aba8d457651e3ae0dea5d301816f",
118
+ "3b9ab98960b6fcfbd4fc1d44d7fceb1d3092d58f16c5a22e27d09f671fde85ef",
119
+ "3bdcae16a99d8be20e38c268f7978794c8be3d2440b03f15b0cc547ba01a190f",
120
+ "3d25bc64c127f6e9e8705ed0752427fcde08986f9b9df5f0973be25f0ebde5e5",
121
+ "3f87b13daa1dee73f5ebc4fac31c3a21cf5ffdc8a0f49d6ea892f22e51ffe861",
122
+ "40d23f5141ec7975458fc3b927a2f8a2db1000386e1d57126a44bb857b953adc",
123
+ "426ce6db08ba8a44ae1aef32bca73631c9b1cc8d8fdd05d78753473e342aa138",
124
+ "430bf806bca6f4d3a98b04d24c674025db1104f21f08e54495025acec7b06c9d",
125
+ "44226a0fb144c7aa91efeebe7396bd161662f555a4463ff99412b6c0f3a3347c",
126
+ "45f55fff3194d5f7835b0256fa70bab97b7e98804fc527cbe1dfd76293546d84",
127
+ "4606724b3a2fbc92a849fee2055b74f2259fd0a3b3260f5bd6ce2a747f77a03b",
128
+ "4653d720c6c3b9f3f2a261d0d497b690aa8c775035d9fd59cc46b95aa076d8b4",
129
+ "46f1cd713f5787b4f716536dc34b130412cb1f7dbf8976efce364997c0af9508",
130
+ "47c955f7b5b743a82e6cdcfa317911df4c62f5b3d515878c5ac577de9085fc7c",
131
+ "4813b938c6cbb66bfad1f5f304e5855675ba2a3b2c6a7c7d0171b474f5fc92a5",
132
+ "488677a09ee03b123e04b10abaf22dbe0048e5e9250f5b319537c61f5bf9fd85",
133
+ "491ccbc633f3baaecfd5c75c87f310e9f6c6aa200ad3081c33d7968b606b1ef9",
134
+ "49c4707f784561b72f55a09f737461ff99cd6686c14876adb6b89306a4e28eb3",
135
+ "4a8f69b72cc5a99c2c6f67d6d764ad302d03b00171b48c7c003c8363d272a501",
136
+ "4afff14d190bc0d789f2c3d25468e9595c4fcd7257cbee03d9c31862e5adb33b",
137
+ "4bcbf4fe55bc9d53c0222262a4fc0eec440925c9b2fa09116a3416c778ec0dea",
138
+ "4d0892179c6b597c2cc5198362d7307f64456f1c52783e6601797bfeac0109ef",
139
+ "4d2e20812db7c117c936bdde771c8662ba2ea9c2c53126587233b1071432a1e3",
140
+ "4e8e264c27bdb9e0ffa85d4bbb8e902ebd0d425e28140ed85532949dbbc06cf3",
141
+ "53319010c3e47da4f287cb993637477450c08dbf2f805bbdb04d51839c100125",
142
+ "54a66bca37c53a4ca495f1d6689c40237e96507eb32ba5255fd5bed21bd649bd",
143
+ "54af21b2db061a9acbd755fc11eabd879388eb6352d7dc271b3bdfcef2d665ed",
144
+ "54b53edb39204026ffd6d52e586e41eea40baeb706fc106d70bbf86f5df11ffe",
145
+ "5521e3454bb71307ac62ed9e9977d0714ab97c277d9f8141b81471c87cccb677",
146
+ "5559effcfb0080cc08b73b8dda196c70db04db6a1c5280d1746f08ad750b203e",
147
+ "5615a1c3e44a77826bfa1c8a2b87b1cfe2d1184ff5d69a3d8f683417bcf26878",
148
+ "56347a884046431a2a42fe99aff7f11ab7137f82f71e5c7ea68647c700faf55e",
149
+ "5718cd0f31018f14c665af23dfcfa6b09a54942784840a8813979452f9293aeb",
150
+ "572b0440ebe785a3792ddd86fcf327172e208fc40ed8cfa5b0cf192e2656128b",
151
+ "5790177d4a2a26bdd20e020d4ae51a6030f8a62be923eca0fbe1aa1c3c6d3101",
152
+ "57bd10c2d8b2c64f3ad25dc0ad76373cfa7966bc3aba69e9b45cd838e27c216f",
153
+ "57f3a07ac9c4781586ae1f3d449a6a75cee9769838203d09fe08f5d9f41efd6e",
154
+ "58361c52234f8b7203632e636e5f293eafedbf76036b9f0bc0450ddcd9092cc0",
155
+ "58798920cd5a91cadacc57fc6bfe08363d3474a4c56fbd37fd47702a06deb56d",
156
+ "58c02a6d10b19dbb08be5b48bf0f0e2970d4887c885ee8bcb2d59072f0b3c021",
157
+ "5985eac9aafb36ad2e5b08e5cd9b82c739e3d7e4af4b3dd05b68544b35ee88f5",
158
+ "5a706a95969a9afb3d65fbf4685d4582298d4ccef54228c9347748efa93fa5f1",
159
+ "5d7c7a165f74e9c88f12324cdbc1002c76afaa60d4a4c1f4c606232566c041f5",
160
+ "602c17dcdfea1ac52c0c79522adf4fde16dad0c420bc2e0d98d99e026a1580a7",
161
+ "6131f664c32a5c9885d259e22ab51b879a0f3229b27d6377429a64a22f0b7866",
162
+ "61ad6d99513b83bd00f84969872c02bd52c65d967988aa473c1c65eefd676b19",
163
+ "6210f692303bdb56b4b73d612a4ec399d88dc618028d8310105fdf36f5f494f3",
164
+ "6310027f48e518cc010476dd0e73806f4e56d2dc532b4573a267f731d3d41414",
165
+ "63ff4c6a3242883ad009a47609d15ea0dc04505a0a127eba84908f9349b1ccec",
166
+ "6425d424ac4f051c0e98574978f16af52be2edab0ff253971d403046dc1822fc",
167
+ "654680e590d30b7d29662d3989e91ad89865fcef62f711d16e2e54cdef061702",
168
+ "6555addfd8e7249db007b2265be1eda1348095f6ac1ef8bc1ca4d2c377a61eca",
169
+ "65f42165abed0efb45ae2178cfa3e497e669433563390e046de233109e2973f7",
170
+ "6731583ab3fa78827bf3a348596ab585d1cb9bfbd483c6c3fb5be4fa78c1dffb",
171
+ "67d6104d2642f4d213c4b90c58441aa30f1fe59aed66ce213a41aa5293ac8b0d",
172
+ "68788bcba35ac159814128f8f58a288eee4aad3b5224422e19e90c3c4fb854e7",
173
+ "69366a4250958c6bb863eeff854d1196b293f366c610435d34474e5b852935c7",
174
+ "6ab049c8a12bc666a3bc45954c81d044be3467db09d9cba50cf2ffe161e8286b",
175
+ "6aef49775887d2ee69ac038f4f5e4200423b05ca8596ef63f9607e45e8df3ac1",
176
+ "6b17ce5b388838e500322d13a0949950c33b35d30c4f52e4e4caa66634ccf3ce",
177
+ "6b368523bafeeddfc6513acc429687b726c48c24ded69b93b08a7e5c39d56c94",
178
+ "6b96045a98825402848f23f009cae525b6ee776bc142438b127d62581609bd82",
179
+ "6cc7f10311bacf627324494096f25e2ba2590a46d7674e98d239761f8ce6db78",
180
+ "6cdf5a6b67a31c75d95ed9d727ccbfd12eb7b5b8d2cf4b28e088f94ac73c20a0",
181
+ "6cdffacb4cfb379d127f8d67e10ccedf20fab3eaefa8f67e76ee1dcab6b8c4d6",
182
+ "6d3cf3c690d1e5d386abcb9d2f8eb6131242cc99bf8ec9de652d35318b133302",
183
+ "6d4b5efbeee4bbb85e9eecd3d2146c5c55dc09db9004de08b942ae783f90e69d",
184
+ "6dc5a4f395fae529157472d50a877fedc68291881ae443925b79e94addf5ba75",
185
+ "6e1c0d9ae3e2b4e506f4ad74ae1701e83826ff41c972a03294f096a306512420",
186
+ "6e3c8c35af08870d44640081ac7bc508f7baf402e85e00079d4788c70df494ba",
187
+ "6e633d849696f95b9fca05ea76d75278fca6e751061402fe01a45f66eaecc53c",
188
+ "6e65747b0aa8f15a9aebae69609dfe679d7df160af7c6ec1f1bb7f4fbe77e11e",
189
+ "6e6fe0b9c29af9252a35479a24f8df339830ecd240d13e156cf14b614478e23b",
190
+ "6ed932592463c7c0ee3080ff49fbd3256eb376d9fbea4f5a1d87a580cf0f9dda",
191
+ "6f328699052eecafc730cff3cfb923ea06180f7817272f881f296f7710867fe3",
192
+ "6f8eed619bd516d812be34fc41c2d3344b1ebe04dd9f9ffaf0dd0d4253292800",
193
+ "6fd537d3a74739c21a4d95e906e35bc9f8ac1d53a863ceefbdac5a6ad01abe1d",
194
+ "6ff807542ca183bed4a6bc219660e8b3524c8e2490c6999bcaaef7194ab78e75",
195
+ "70cff96648b78a021ae4ae0a3ae4bbd152b6d426ffb68a2c7b3f68af5602c0d0",
196
+ "70d0d3d9515ec960eb34915c9f32b77b6bced2a8905be335ff334f798f1166ce",
197
+ "710e3e86832665cc2074333617d3fbd3d8bf5811fee593d059f31f396d7f45c5",
198
+ "71109eede5531750b395203b228e8aa9803373b1254d6a6b009dac051c8af5d8",
199
+ "717c86ddddac537d41f88b7c8228005f20ae3b4e5392725af2ced0c9b47d9313",
200
+ "7182b822c86e7c2b152072bb8b3b8c0adb77de730cf2b3baca138060f00eb802",
201
+ "71a939bb2fe3fe621a38aa2814300823a716a74bc4b351585ea0a9b4c083af71",
202
+ "71ef53fd5d733e59abd86c7df4b5ce62ecbf738fed4715ba552a3aae2cf82b47",
203
+ "72d461aed408dc559a73710d45414f0643ca50cd5da610b3fb0129dc52d15a77",
204
+ "74071642727c9ed6ccb11bcd048db68fad618d722c4714d42034fa378ae706fc",
205
+ "747bb5a792e01b6644c91a98c94264a270c7ca36d6d536d163d94f3d3b84f64a",
206
+ "74dabef6f2d19582dae9e8df33c2cdad9e8756edbf8486325843c86d7b6f0b92",
207
+ "74ea51a817d117bc7c0ddb7a50770bad7c9b886520ab14a39f541d78cc4d36b3",
208
+ "769a1243de2280c5949abf534eedb5006633c3f60e6aaf414b0f305a07621cae",
209
+ "780b8bc74f7babe33315a1906cc6dbe0d9676079d9fd3fd303d5050cf6fa6adb",
210
+ "789467c49d28ac5926ca74c0dcc6f87e3e7ab1594f4dcd5809ca813f3b7a08ba",
211
+ "78da1ea6dd550cde2487750afac571d0cadb81071c2d183336b0919933de30f9",
212
+ "78f0984ef9c77e6ffba37b4260c3148c649339e761bed5b465537daf2370c1c7",
213
+ "79913de53c1646ad7cfe7622a3a84331e0cf60560679d6e4273d47ee5eaa6f72",
214
+ "799144770a2b4e07cf01947ea8b77a5839807ddd355170b87ba06d92866c47e6",
215
+ "7b03f51b82cb4798ba44b348c8842a1aafe968e70b111066740cbd3e48da93b5",
216
+ "7b5ad9f84a572208a6e92c733b19c068d3fe74dc5972dc7fde17d99c44488540",
217
+ "7c7651c266b7ab2504195b2d0fefd9ec3683df12bc20328007641a032f9f7f53",
218
+ "7d17467b25d44d0b1195dadbec457a87ba304966bbd8b1370047ae7de2ed2a05",
219
+ "7e884e9e13768e48225aa52c48b97e5c3252a649c90f3cc8010bfe9d5acf929f",
220
+ "7ed28f6688ca2c3971200d3ea0d000688e89f520ef4b6dde6701cff6d93379cc",
221
+ "7edc32f4e2da616996e003993e5d5b26ceefc4520692074b60aecac240798107",
222
+ "7eeb7fbdf38454f42aba21f2a65eb7f5cef9b83ae061ba05911959505cb91ba8",
223
+ "80425f425d21a43c3a24d0cbec2e1cf2da3c84b49ca37d17b3cfd87bd3c064fd",
224
+ "814fff4d3ddebdbcf578f3ecd2ddb15bd73f13c5dc97645f825dbec1229d769a",
225
+ "81dc75b3c4c5278f77872721fd7025351b3b6ee35b3c6c717382933a12846b54",
226
+ "81f2a221e07b8b87c8a6af2b8ae92daa1894ce793038b89fac66661c516a13fd",
227
+ "81f5585e360289f12de31c6ca3c5a19d35b36706dc818c1bbe84c52100ad5a13",
228
+ "8324086d105ca455bb60adb5bf2ef75570cb6bf61b3707a7b257e2d83bc69e4c",
229
+ "836794cd8ef10d37d1054cf7b79817cda0f0e4844bd6bd0ab0b0868f6d68a542",
230
+ "8411461a46bd466e5747381bc5c7bd43f492093add98cdac5797dbcf6f1e4660",
231
+ "84f58f57f6c72ec586fc22b38330eaef5bd838bc1c8fad56c29efa0be7eccdab",
232
+ "853523426e3dc330502ecba19f3a9a7628a12f7e8359b2c70eda0102c3e35cbe",
233
+ "855e6d393936d0aaad4a90d2d40b5c4c64187e11eaa8afffb8ece7639c14b5e0",
234
+ "86605eb145ee658ad400ea7473b837813106c904a4a958cc26b7ca43d3bbc6bd",
235
+ "866b5df371574f0c85056b58126624b34a949227402540918ec1d02bc129e272",
236
+ "870233bf592af1c5e6a70cdb7610b7fbf59102b2f94055aebc8e09e2e8abc066",
237
+ "872c4c55617194769ba5d872580f47a75a5553e70437e32d24560ac206156dc7",
238
+ "87be64c00fc483b0db889d7dd12e643222ade5c822c7da2b07c187aaa88d0304",
239
+ "8832815aaf4d20db32b8caa39dcc740a02e1cfda52854f00ba80336119562bd7",
240
+ "88b818d5b25b12992343aa730f98dbd25cdf13e183a93821ce3cef54976759fd",
241
+ "88ef007b6f8a349b461726d0ef1a8a5c3bd62e30f3991e195d5e2f04776936b1",
242
+ "88f52aef40dcab119d399e383866b95f979bb7d4371f9a8eb5b6d4946eec3f98",
243
+ "8937ec18e5415087ed1ebe671504dfe7d5dfd912bcb98bf64f57e0e10608eac8",
244
+ "8991f2d1867e1f9e9fe7e2256e7618d4c9462140ad82ba83fa96b7f8fb5e8276",
245
+ "8c3845902baa2a56d30004689da7ecfde2e18ff94b438f9e998b4f9f916fb8ab",
246
+ "8d68bc7e841c900e0d54a943f421aed61e30d4e2825f206290e5db04be19fd0d",
247
+ "8e847b6c90e52cdd68637c92a3f32650d4a7efdda87157c979666cc5f4fb3e5b",
248
+ "8ec13ad05622d6aff476a4c6bdfc7f33377157eb0dea77be7eca5031e0d3a3d5",
249
+ "8ece48d24519b345a1ec5c273690fe8220026167b26d5b3784c81200a5a268be",
250
+ "8f5666d6f88d29c0588631652e9d376f2982e19c855afbd67f6aa4f238961dbe",
251
+ "8f69f26a86bb62dc8b45b1cf30d53b0aa16f67ef2bbe7a01907ef6fa8c717c16",
252
+ "9113117e178903566bfc2a2e7c42b158479eefce407e38c9471535d54410139f",
253
+ "912a29f600f58a940b4136ec5c3f33d838a5cf45d1892a36c32c30a20fc56d90",
254
+ "91a57d162d23cc952f493e1199ccd4df33082ae5ec3ccc03efb6464dfd80befe",
255
+ "9329a96a4393cf87af20d2ecc4a8f2ad75b48b4927a13adb9c4191e7fbcf0aa6",
256
+ "94c63f1068cab884a46fb92b43196c046699581be0000d27c39c5e6bf6be5cc8",
257
+ "94cf3f3f5169047786afa7adfd68388bc79fe520513c78ac932ba27c94587ac9",
258
+ "954728266ac0b749dabafa2bb3086106ffcdbd7292e3aa76f44d146caeaafacb",
259
+ "95522b12879b74fe21b4f599bb83862a37cc15f9e84d42b6aab7e504f8a1fa89",
260
+ "9555a76e17a8b21212bf28ef681236caae41eeca7234bfa6b0df3e73c392ccf6",
261
+ "95b1f08688d5fab30d3aca9a2a2f85791c5a02177d91528220a2949b7dce7a1a",
262
+ "95b56f347bf4ac20fb9a084a4bb6edb9dfd2b34f9e4038a1b585824cff694b3b",
263
+ "95cdab0e351436c33ced6d47089d6ea983ed92337d64b30d411a34aa86fd585b",
264
+ "96e4d7865577d8da6547b55e2c33ca0ac0ee5230481d3bd46d8821c87fdcdfc9",
265
+ "973af4a9a83fcc5878392f67c590cf33e0186ac03379a3212158d8a53e3a87f5",
266
+ "982e2c67ab8c0989552f2a3f55886ea7910f94ac5ec1387c06e7e25d2405fc6f",
267
+ "988234ed7064ca88bdc403d8f3efbeeffc9a6140a3582ab6188f3a25cc43b8b8",
268
+ "98d242f623064f662fdb46a6618a85c21e3008128d17d7a335fe4b4c0da53878",
269
+ "99a43eccbd1f875bb1829289d71802ee71fbc78fc5dd3b65c6a183298a618709",
270
+ "9a8bc8321d59c58993ce3b72c915471a8451dd849e01687572558908cb6e26aa",
271
+ "9ab1290957acf6d87b204c4e97a0244c387bfe772de321090254f20df1910ba7",
272
+ "9b95e8ee0fa27fcab0bae060223433ff78dbe902cc3c0eb0d7549f800d54f103",
273
+ "9c964184dfef49dfeabac2ca70841d826ec042b80b5dba4b58d04cb53285af78",
274
+ "9deab55a3c9634b580db728cf42c9dba00767f8f37775d9ef76d87bb003336cf",
275
+ "9e21e4e980e2779799b01befcdb8102810411029d89a297d9ecc949de0f1dbfd",
276
+ "9e69dbf0aca54bb0d954a1143e42665cb9305eb8ef652ba2f477a65caeed5442",
277
+ "9e9f8296409fa67996c815e91feed202f598167eaf5b3f0350204007c0d1ed51",
278
+ "a03e1c1fa4e17c3344c98a3a5ee8cd8d3f37f2909cddc21e54822be80de5670f",
279
+ "a05f8cbfbb664035c98ab714a65313ddb72b573a5b4c3e4c1996a8113bb22681",
280
+ "a0a7649bedd18507db87f80a6e26a9ee942b4bb231afeab0dcc767b21db8608a",
281
+ "a0c28dac590a141cc127b54db3207f36026eca5cb620bfd2f844c82b67167890",
282
+ "a1edc1e88a06025bac194b9de08c8b710f8349b6ee2f18cd555a63f32541709f",
283
+ "a25d21d56eace92ddfdfacfe5321db18a80cb5191bbeb689e8e5f6c0dc598963",
284
+ "a282bba95a8a4542a289106bc37d27244dc98b311f3ac7acb7eaf5cafd02ba26",
285
+ "a46235803d16cf09a827c39210dc98dcbc1cade1fac99080c14ba14859bbbfa2",
286
+ "a49c8b8a5645e0a29fbfd20224c970fad477d04cb71a206fd29ee922fca57092",
287
+ "a4c799cc8b693658c4249c3b7c0fd69c45a1d41b7be3b38068e962d789f939da",
288
+ "a627e1c3f3b15e4faca6e36a48a0daf6af290f4587658b40c99b8d1bac26c1f1",
289
+ "a6dc7c54ebce2d5033b2e32245e52be15832ad9c5a62add5710c35b00fcab220",
290
+ "a7fd4637538900354d849d161c465112206803a259690ecb85a335217dcbc299",
291
+ "a7ffc8ec75bc9251416105d48a95eb466cde6aab8ba4be79dedbdd60a6bf80a5",
292
+ "a896bf4376349911127945d17ebfc10ec457401ed5d2e3b06e558ba74164f2e0",
293
+ "a90f3b019efb29fb36dd037c3d043d49c93440e6cefdcab2b213c0cdf73c4677",
294
+ "a916116487ef88d1762152ffd77390eefe84c69a954cf6071ec48f48599183f4",
295
+ "ab9b70ab4f40fa4d69438dcdc65e19212f8b1ded5130a89a4289934cf35b0ef9",
296
+ "aba716b3a181e91cff5f83f1dd4d4286fd7e99f8eed0ef876484bc8b8aba2693",
297
+ "abab0537479480471fdbce7a9bcf265bbdfd65f1ff3d4ddba0d9e5b4d9ced9c7",
298
+ "abd2d9b083620a1d2f671211125c53dae979b997c88f67d45807fde6ba372408",
299
+ "ac262a059fca131e899b1bdfc77dd816d399c94616d1a11263671b33316c9599",
300
+ "ac82dc75ef05680a107045d99f26a35dc300d55a7118e422be93355ae5fac944",
301
+ "ac9a9ccae1b5a7809b3d1ba1be44a71c7169e5b56fde550852a96d7d1b2b5f11",
302
+ "acadc12353fd67a26dea25cb68d37740bcff84752f1b00af344301299cee8ab3",
303
+ "adef0f0e85961cabe53053836f2638e83459f78e3ad8b1206b1e22a0e64891c9",
304
+ "af6cfab458e6bc9e744c73a8f478747bf77f1c1b56ad30d6945a264c42846a76",
305
+ "b0a4b3c916589cbbf3e48b7820e66f919131ddf5330b0ec58872adfabdfb2e05",
306
+ "b16c55c36f3685acf7707bc575ad1069ae9c058d8dac86f3ae2fb73a7ea8e4a7",
307
+ "b1b630785cf6bcb5200be4fbe6b9cc5c681f9385f3bb7f677bcd7e852b26ee91",
308
+ "b1d040fcfd9ccc5f178579bbbf895ee2e385f427ec30f8e4786c6161fffb4fb5",
309
+ "b2ac951600008b0fea2ba3cbc5c474594519c9b2a70edf9ad163abccdba1c660",
310
+ "b2d67426cdddc00acbd06b2ea113db9f9c441162ce9d022ec8510efb48146e46",
311
+ "b3e016b64cd0cda66aea9e3f44eabb40943e1556d32bf2d6518745d9e427b42d",
312
+ "b3e07c07f4c4b008fed527f167c7f11f17ed4d1e4163cdb0ba2e1ba93751cef3",
313
+ "b4a6b9d0180583975911ea25c8f7ff0107152d92d6f948141112b83ae2f2febb",
314
+ "b4c5ade87ab82415346618aaae5728f15104ce3b60b11bfdddc2159c6a9ce481",
315
+ "b760bb947e6d5b6459228799cb111df0a4b41b67b3da2089bbaf4de6e4841ba0",
316
+ "b78983bf3b1bd83a3100707a4c76d66dc1153abd74008158a7eae0b34c7c625a",
317
+ "b7f5d3a5b89c5ef68650155cec7493fe609f21b5512adb4cb5c08e2bc5c95e82",
318
+ "b857988895276b50f5533c84b38013e0642ffde0591f66678448db56a4d53698",
319
+ "b85ce13d03e3e70bb8eae0a820bce4ce7767d89c703c131e882ab916db55a5b9",
320
+ "b8ae3abe48f5037066867e455a75f4dc2e501a75429cbadfa9a3e278d9fa3491",
321
+ "b8d79019da05f874f40725409b294289a4276b49881cca8108b173e706bd7acc",
322
+ "b916ffa0d5a66bf3293242e6a86b3e8ce4247761990159b3a157d23d9051a9d6",
323
+ "b92b5690c983699be359b27f42314e1781ae82e010ceaa53d87e248c713cc6bc",
324
+ "bac079989a07132080b3cbf18c956f43eeb90ac243320d36f316db69e2445685",
325
+ "bb47e3ab316278d9b61cd0ba233e0d1ac96702fc09b54daf21c66e6e5b0b6b93",
326
+ "bb834499aae77dc571eabca63b060e2d3af4fbd15e6ff08f31920d56ea69e512",
327
+ "bbb0f2d656da337147899b331f4eaf2d42d27f2e5f433a0f89f7a47d94e00630",
328
+ "bbbfae82c10d69302fadbdb59607d17542b7c36eec05c410bc3d19bc59f9c606",
329
+ "bbe394b68e620d4f2a298e9cc932cf544c2ed1976881670a5252e80034a31635",
330
+ "bc2ad162fee094abc641053665817bfac9c6d8571a8c981ff7467ecef07197f5",
331
+ "bcb3abf244ba1c9e1185b4d5b82978e3eb525b215b76e9e612cc4ea5946c27dd",
332
+ "bce02811d92bf052116e829fb9514b73ee7d3ce06179145ea87933ebb43864c1",
333
+ "bf3222ed5d399bc0b39e8f850bc7179267c74589ee334893c821256c09dfabf4",
334
+ "bf85b54e54f81929282389b1f6bd71e1500e906a8770aec49d5177f9afc401a0",
335
+ "bf9919db71fcbe389329b0a80a25e2f32e92c242de4367040031b773d1289e6e",
336
+ "c03eae7c3673288b28e3cba77f65491cb09ddc36f5fdcdb230595ab59ba00ba1",
337
+ "c1b6ffdac70e3fd228a441606b57a74a87e52ce8db51bf777d574c9c7074afe4",
338
+ "c2324ff1a84058708a778b0dbc4b62f7906541001319661e7705704bab5215ea",
339
+ "c398fa7831e97ed2d41cebafda84cdca1fb256ff8c423cf29cd61be3d2323e29",
340
+ "c3ee012f79a5e9a579c87d0ad6bd09b69e83911acc2f9469b9acf2d4ed020c5d",
341
+ "c3f9fd85aa2ffd14ad57d3d1d4641cab1b3307a16c469f2ffb17cfb7e38d89e8",
342
+ "c41423d563f1efefe224467078c3953d4362fbbe5403605bbe47c348648de873",
343
+ "c5f435d77f24b0a361e07a4fc92f2656d7f3a16678319c32b0fd53c5344cfee0",
344
+ "c6d7ba85d51e713f466fe5874a814a5fd4fafd841b780a3c69a5bafa4b181641",
345
+ "c79a1fdd63925c9369e8215bc03205b5440f525d2475760b35e65d128471bb33",
346
+ "c8263aa33563f118a277f94880e01ab9d85d4354a82f6f7e67ffc3acfc368199",
347
+ "c9014998c9c1a3ddacaefe61c79cd910019cc7c85e0ad6764bcd96ef646602a7",
348
+ "ca1064517da75cd2d73d187c83f59e29253e92a902468e8516555cf39ff531a0",
349
+ "ca106e061bffbf113ab31a4e8ca571ca19b684c4ef22397bb2e0d26f9cd9b4b3",
350
+ "ca22d2c910d5fb8a7cd6fe1e0e6038c21c9016158da0453c288f4beff35e704f",
351
+ "cbceacfd6887672ae36da858fe55fa1eb1248a461fc918e0622988e80f5a5f1e",
352
+ "cd2f4d7a45b33110be96a05b5200017d318e687c2c442003197bc078dd0e1c56",
353
+ "cd2f7561761d5260d96e29be54ec5289642c93b0341a343eb3b71743ba5de1b7",
354
+ "cd6a7a408ac632e85ba353dd2b3bb53543ba25005bb4348f2173439583b7bad5",
355
+ "cf56c541150872145af25c650445aa02562aded8c2b651ec81114e7e47a57ed5",
356
+ "d0137dfc0e9dff20e467a5f183dc18b2d1c3050db62c7199f0e2cfad24ea871d",
357
+ "d09894ecf5a8fa94f7f80a75c597cf2bffa38d6cb843ae78c60d1251cd7d9f91",
358
+ "d2323e66b21bd254d3edacbbe48e7fb41082dab2a47186fbeaaf447e41e847b8",
359
+ "d4febb5402691fe81541c78897f18467b8b19f52824417b71cb52fe042ab3297",
360
+ "d54d9e3ce3b3bf3a2e721539b4526f27c785ac8ffb273ad8d635d429d91de2e6",
361
+ "d5622631d3095d08a0ab948349dddf22ac6d309b9d6349f62e2ae741eb6c81f0",
362
+ "d6054aa280062a63064d8c2f4388169a0d99bb46a4f09429470e58e34b720885",
363
+ "d63ceb0c6327e4e658901e155923c96ada2ae0f7948f260dfde05808c10410e6",
364
+ "d6457fe91b57d3f6407ab52c979c1b168dbbd4d345f6811b7fa4a33ecf15fe16",
365
+ "d6c5114dd7d96b6d349244caea137f41896d0d65803f0aeac524b8e0205750e6",
366
+ "d94be29e8f3f5fce3d572a23f82a9facc3d547e1a20c9a894e4ecf3a505bb7de",
367
+ "da1f46665f95caf29e033238a60bc885a1edf30918538ba53f3e0e0f3ba758d8",
368
+ "daa82746340669b4b9dd39be180a98893ecab290c5646775f778ac810e68f3ad",
369
+ "dabc9c6421417336d9a4e9edb999c62918760f81f9242600163c27387d2f9f4c",
370
+ "dac9b0b33161d877114e396f23173a296b826315bb7b848af25049d417224541",
371
+ "db8e6d2930c641154a58cd407e957fd411456a1c36ba3fe1c09ecbfb597da737",
372
+ "db94722d3c9d3930d765dbab1cf5fdc3fe350540f62b8dc51fc61b50d5629c98",
373
+ "dc6b01f69ef6d3e1b599273193d300428662385a90a8629e4f123271f0bf24d2",
374
+ "dcbf7ef1a04611b24582d73025ca744e409e3520c59d5d9cc9b6723a6be2afe0",
375
+ "dcecd123fe1669da91e710e444a07c091072a11e64fc84446fb3bc81ea9e3ce6",
376
+ "dd81493e182cf772ba90bf45866646a9a8180a9548cc340ab07affd73a32d646",
377
+ "dd970f73df9b378c8f4c0088c022f03554adbe682beebbecd7b3188a26fd654e",
378
+ "de5c405220e11f90a248d6d56eb9a5c255a3b1a5a442b960cc4ef528e0e2cbbd",
379
+ "de845c1c7260617e6c524215ac2bc50ab2a43cb03a503dc094191d9e177c82ef",
380
+ "de995837785e14496378f2052cafec8afedb2b7d37d1346cd3e2ea8692b1a7f3",
381
+ "deb729f802c6de2e79c3fd458dca6850809ab1e9182087b3e52321e0a6c04ab2",
382
+ "dec23fbc274e9e3db3bc7de70755489d25f997578849bd4c77e8d5af7294b2ac",
383
+ "df756c0684905a591127492bff47368f36f108205b68be5cc043c4b1c9970ca9",
384
+ "dfb562f0d7787ee13a030d751845ce49fad81d82b842b7109910d8f690d874ea",
385
+ "e0e3e3192a8b85362e8ff1adef5eab8d0697d0bd7f332811cf7aaf294972adf4",
386
+ "e211a10f5efe01ebfd0cb01d9e5eb79029c055a5581ce243ac0105114fbb8bd6",
387
+ "e27411a52403ae9c96c241f02eaa738c6c672d120bb7c2ecb116ef6cd54923cd",
388
+ "e402472d711aa8265d42ee16c79e379d1100395fc1b82794271be38c777ab591",
389
+ "e40581f959f5ab0a8dbce6d1bc17c8cad4670eef1b510d3afa33252813c4823f",
390
+ "e4c2b23fda6ec02c51518fc851e448d8152aeda7b093dd2d15bc1db0db380566",
391
+ "e4ff400de7b7014c6b0734e759450d155ea061938988ed24378f40e554d6cca8",
392
+ "e62a79147d1f8f603ded27c8d52b08d118f5a8fc184adc48b5d5571096c45758",
393
+ "e6563012f3543c9f23ebbe8027b50f3d8e5acfd007bc7ff934baf8d2589cbb3c",
394
+ "e680cab2209055baf3effdba0f1d71cd7da1deece72fb624134f9aee2c51e127",
395
+ "e71666a8b3de77d8df6e1b1a26963014b3ba3b61853de1875c1336ae57c8a168",
396
+ "e788f248a6745006fb0ad0bf27262261b2ebea392f69ab028e954707371fbf84",
397
+ "ea25d7b2fa5d864a13bfd22c91e08e6599feb32ec362771b074c4ee41855ec95",
398
+ "eb096a05c4bf4447357dbc3a226cd702332dcebcbcee6b64999711e7b6457bd3",
399
+ "eba83fa004dfdd799a0c3fce1ad49b604bea8ce6f5dc845ea8324420c5e014bf",
400
+ "ec8965c52e6bff3ceef758795b77cda27ae291980a34c71f6f35070c23b9eb86",
401
+ "ecee380c7161fc1d9ce94ee1bc93a54592a7de47e8cbec8f0498cd202efd1cb5",
402
+ "ed13af6ff41fa3a204559b6ee6ccaf19314f18ae79cc7e0c6f82ddd42d579524",
403
+ "edc41a9635ad187d84d07a131e2bf5f76832890dde4fb554815c434f32a9fb48",
404
+ "ef2be2eb1e6f2503105b625e92af4afa7257b73606c056a2b9067349289123c1",
405
+ "f0163a9dcccb7b915942a0e573c636679992e901a39f17d813e02e2c4cacf47c",
406
+ "f03b66c738c1a74db7ae30b9e2d78d30ce7691d71a8127b8e33c83deab349931",
407
+ "f194e7426583ffc175688919aa806c891d58311d42dd922bd3cf0f22ecc58c78",
408
+ "f2145b0764b9c49093beb8acd3879bcbece4c86bc2545ad30b3547998db5d4d3",
409
+ "f24f90f3764d940281cf638876c99ff9ec19c46a579399823ac6553a50d44c07",
410
+ "f25c17c7b1a9160f302e78521b4fa9db1d1ed431d2a6029a4e66e7c0bd78ca69",
411
+ "f3262cc7ae7b8d652129e3b19be2e71b5393fefd194cdb4f18fecced393fa692",
412
+ "f3e221015253c624bd18e1a1ec6d6cdf855b179a528aae544a28c431850c8137",
413
+ "f420bb272354584966ad45fab45db2e96c03bd92ab85b739fbe8701a4f795ca7",
414
+ "f440c37361c65c981cce45f160116a897ece4e18d4b6c1fd8b7251ae4ea014a6",
415
+ "f57eae02d4855b2505566f705543467f3758edb3639dcf63428b3d655d6900ca",
416
+ "f5888d79c11cd5039cbffac6bd92e13cacdee543a08ceda6cb12136cd3e64514",
417
+ "f5df5241fb7af7d427738394acfeed4ef40025f3524f2184b21b12dc7f3277d7",
418
+ "f6764cf8c80ebce0a10960fcf97f1d2f868d201bb1421ec2f1bcc635f2a905b1",
419
+ "f6fea0b0cf86027b13ad2b74423087279ecb1e27d32a81f90e3aba7affc1bc02",
420
+ "f76737636cc2793fa94a81b48ebb62e6a293a1258721bfc2c3a2df3d524b918b",
421
+ "f76fc301ed789c51fb593b92c333b4c0f0341732501af0964feee339c9b750e5",
422
+ "f882570ee3e809b38602c4a7e97f2bb7510d2851392d45c9be98d38366e837ed",
423
+ "f942783262e9e60ee5ee39b71aaf23f5868d7ea0248d437aaee1219675d8ad40",
424
+ "f9db175886ea8c5ca4046ddec0176b9897516872b3a1a0848d3be424b95087ad",
425
+ "f9e667dc320b33690cd860b7e66cd789630658bbcc3af8c16cf978412f98f3b0",
426
+ "fa169af3e954857659521ebdb22e95188911b7c5926adbbf493f38e05db0825e",
427
+ "fa67862c4434071236731974e363313151f32440f87e6f2cfbf0eca76509ca83",
428
+ "fa8aca36011d8a7acb85b0f053082ab1f42a48f40fd19f7300260f8eac252254",
429
+ "fad035c7def0678143fb088abe594306a81cf92a616e9a4bc3615ec385d87956",
430
+ "fbaaab822f820adebe7e5c56f4daa8356cb50e0869cff2b9d569596dbcb6bbb1",
431
+ "fbf9101b8d684ee753ce6d5f8280413b68f418fd59da9772221682f8280b9d4f",
432
+ "fe06e5a7372a1b9e026043e3bc89063f91dbb1ac1c6a4c2d34b93f295f3e51a1",
433
+ "fe37bf2cf2b401d02ba34ab0c77d52b89fd07224bea409bbc911a391430c8ded",
434
+ "fe9cc6e4934a28231d7c864c377b2a1b72e4da2b05bf70064db4c6199fe432b9",
435
+ "ff354ad90b2deec831a173c06965f2dffc0ea3f17b056e3e5ad55763a2c52d93"
436
+ ],
437
+ "include_outcomes": true,
438
+ "label_counts": {
439
+ "negative": 5179,
440
+ "neutral": 1222,
441
+ "positive": 159
442
+ },
443
+ "leakage_contract": {
444
+ "deployment_must_not_read_outcomes": true,
445
+ "nontrain_outcomes_are_evaluator_only": true,
446
+ "train_split_only_for_retrieval": true
447
+ },
448
+ "max_flat_action_dim": 112,
449
+ "num_groups_exported": 410,
450
+ "num_groups_scanned": 2873,
451
+ "num_groups_skipped_by_split": 2463,
452
+ "num_rows": 6560,
453
+ "object_embedding_content_hash": "4e8773e46aef035fe87b3a6b631ae0bd9c82ba2c7310c62545a6b8c8a86373d5",
454
+ "object_embedding_manifest": {
455
+ "dataset": "embeddings",
456
+ "dim": 64,
457
+ "extractor": "rgb_object_layout_v1",
458
+ "num_embeddings": 410,
459
+ "path": "object_embeddings_rgb_layout.npz",
460
+ "reads_outcomes": false
461
+ },
462
+ "observation_embedding_manifest": {
463
+ "dataset": "embeddings",
464
+ "dim": 32,
465
+ "extractor": "rgb_jpeg_stats_v1",
466
+ "num_embeddings": 410,
467
+ "path": "observation_embeddings_rgb_stats.npz",
468
+ "reads_outcomes": false
469
+ },
470
+ "retrieval_index_allowed": false,
471
+ "schema_version": 1,
472
+ "shard_content_hashes": {
473
+ "charts_00000.npz": "42b54c3fe5a42b22a225fc233171d098875d16595a7754b59a2c4efeb49f8ca1"
474
+ },
475
+ "shards": [
476
+ {
477
+ "num_rows": 6560,
478
+ "path": "charts_00000.npz",
479
+ "sha256": "e1dab883a6d5ccf9fab9111d291de115a0fc4de77487bc7a740c207adb62e74e"
480
+ }
481
+ ],
482
+ "skip_counts": {},
483
+ "split": "test",
484
+ "split_fractions": {
485
+ "test": 0.15,
486
+ "train": 0.7,
487
+ "val": 0.15
488
+ },
489
+ "split_hash": "84151df9e88c219b76751677558cc5bbbc19082bf358cea7330158e9e403e25f",
490
+ "split_policy": "stable-hash",
491
+ "split_seed": 0,
492
+ "state_hashes": [
493
+ "004c8ac0df224f86988595e4f552c47d31562c42302aa5e5b9f120d8e8fa65a4",
494
+ "0080acc64d7fb2f9154ab9883e2ec006f0bcbdf4d05e9de28a69b8637047074a",
495
+ "009dbd5ee8d586cbaed373d3aef6402664edf778a9cba1ddf011d37e365ac87a",
496
+ "00bb526c8634dedb0fdfb5d5160db2bd76ce136b8dde4a99387834fc490b08cd",
497
+ "01a4efa12c9caf9b24c5f12b2d6a3effd752762c4600eed12ca5f58e21bc138e",
498
+ "01dbcbab515cea225b9d1159da4a9ae93c8396d2737ffaf0e0fb44d6cdef329b",
499
+ "03c28c71f2a5c79aeb63a10fc4d39169536bbae8755b8c49427b8e4a48da595a",
500
+ "03d19f01e4757695bc7fbe0353e9d20b344bf375e396036255ceb13922409821",
501
+ "04513e79db3e0f6e1aaf91d54465e1a63985974c670008fdcd8bcc54571e25c3",
502
+ "04672832863fb8c9eb5422426815f398a4ac41f10f3a1985f382f771359417e6",
503
+ "0551ca60a30e84d93604c0e342308647e32350ab89152aab60acc95b77917df2",
504
+ "05d47d5040c2c080ff7315348e0edc79482271c1058d079d5aae6886b213d93b",
505
+ "061d1b5c2864f8f74c82a62e35804bd3f5c960ad4032a9507a4ecbd957b6e8cd",
506
+ "0621a9d94a00c788ab283cf27213fec18be5b8dc4dc2ab1ede7fb6bcdf24f9fe",
507
+ "06a625946bca0cb354a0db0e144972691fe3fcc172e9f5718cf1e3e6edb65695",
508
+ "06d116fa981309af8f407def79c5b227cffa4d2fd31b90a727b1c3b6d61b65e1",
509
+ "08bbc76d14a4d92a2d94a0c80aaf9c96244c992862820eff10ad593023ae0604",
510
+ "0a5e1fbbd0fa9543c54f068a799b1eb48b78219dd146bb0e11f8f4423db7b02a",
511
+ "0af43195d1134606a56243bd158f50390371cde4e065c1c7dabc81b23a8c48a5",
512
+ "0c4c8b752dfa7f773514cbdb30f2bf7c53ab452cd92dada5e78d197b090addcb",
513
+ "0c58ceafb2828682cd2912eec6950d52762ffd0688ece98dbbdac6f6332d312c",
514
+ "0cca086201e7689d11f790b065798d7412ea5fbcaa22ad0940f84ead8be24e40",
515
+ "0d84e8b2f6d6efab89062a5efde3fafd7b8c2cdce1e6b7431821c9a7db1c3c4c",
516
+ "101de590da3b55f23abbf9ec98af6d0a734e4a0553fb957d9b9d650a4b1478de",
517
+ "1025bfbc572dd9623c6437a0229cde7d21e9e9619d6fcb7b6e78d57136fa93c2",
518
+ "105f9d47a5ba7e9de7277b3acdc95f540e794c5079f96b10b7d22e158b5f1d3a",
519
+ "10b35b7e101af08ab05deccfece53119f6aa8ffec4f8eadd097c67bb4c2e693f",
520
+ "1106abc92d2ab317b3a2dd77c69dafbf6305c822e070962f6f4e9ef97a8d9048",
521
+ "110f5c52d9c59ff4973970df78b74791be7398e8def0e185fe2603bb579d941f",
522
+ "113ace08ebcc14c1eb1e3e36c94802b5743dfa5f86f84c7bf0a1233ee4b71069",
523
+ "11e9d80be620a82f0f4d35f32f34104d819c14c3a3fe91f6d36891216d61c08a",
524
+ "12d681493bce24cc2dcbd133ff88a2e966d9c2ad735da586c6b7326cf9f76b32",
525
+ "12fbf78fad03688ba083cf5fd1f230726abb50770c0178351ccd4ef8ab830eda",
526
+ "13d6ac7b78af7527b4cbad8162a99a176e2aab866a7e4c4f8f87977d6f69c568",
527
+ "1496bccafdba868e26fec89c211b482ab2de808587e4e3f9581c0dd41e43a6ce",
528
+ "166f0fda1b9dd55677371d0eec1d2ee86f83d90c38271147b317df1dfe130692",
529
+ "17e60393d3f9877967b76a87ce36a1dcadd8a32911e1d3c13b6fb4c2f5e240af",
530
+ "184aa9501e510d89392eba64da1a3ba528632244212ee6c51e6c01b56c51f68f",
531
+ "184e008447d983793dbe18e5eda7b4e4343b3e7468eb91d47138c9b769f7d60a",
532
+ "19c23bc68b3795084dacd86e7719257fe73fc252cf422971851b5a500784f52d",
533
+ "1a21e0b5713c3c77ff1624f68452413a6622858f35026b4fdcf53934aa8fd3f2",
534
+ "1a8477054d7eaf84012250057e2a8f3a2be85d33ee68b2c07bc44120096802db",
535
+ "1abf1ed7e7b56348168c666a020ed1f04c92d7e0b3c92c951a98c4bac472ee89",
536
+ "1bfb07cdb48308cc49cebadfbe47b6985d0978657c5f55b62994fc08d892de8c",
537
+ "1c08830110844a563176b393ff600fb718a717b7dbff65d1ca79cf2734127d16",
538
+ "1c17e9a25aef52bbae4e31ff916b44a4c9d2e48d34f717551041852bc0efa920",
539
+ "1cd587a2f725692688dd2f896c2d3dd331975425a355e119f3fa0ffc4ecc2fc3",
540
+ "1d370ef467069ad818738972fd37aa221b2b1ea703a52e3819834deaac481be3",
541
+ "1d484a5c583a8b80af98481df8d2288078e0ca01eb1485fc7bffde3db7bf5673",
542
+ "1dfbd15fc68a9af8be56276e9f73652b081baff65a231b6d9168c1f22d73569d",
543
+ "1ea9a7ebc086ac2722fe0054a89af39a8972c2b78bef52a88d982649e800674e",
544
+ "1f0156c63e661221dae2b2794349306f1a999c3af2b95dabef534c3f42ff47bf",
545
+ "1f8067c0c0e82f005e7c3e256bdb713707aa0c428bd09ad637e69e3e363417ed",
546
+ "1fdb9db401793099759ed7596ab5495f7ec5c3dabca2521894574c4dbf265615",
547
+ "203596e5c1acc347af07d6369911526437a3a56635d02abba76718bdd2ee5063",
548
+ "204a6b75a085b3d884fd5e1607a2104d8d71e1228cf1bbe01017a73b166a9ffa",
549
+ "212133c8ba1f27881c61acf0ef2690910b8608569a2e00a40e1a007a61009804",
550
+ "217b3a0cd9bc0931c3e84ccc56c6e074340df9a77623c0eebbf14829bf830596",
551
+ "2524ac64cca15c8962b381b1fe2517105b5cc936813c40db59a1545d76b3bd69",
552
+ "252511afe52c3f9f509bd7c96dfe0938c89fb072523ae09b2a30948044a973c6",
553
+ "2630c85c8000a2dc9bd3777fa33951d602f2aa6b209832f346d76e8d922e7092",
554
+ "27100aca96c8e5f7009eb06e931e2531cbe51e89453e82c5357e441120fd92a3",
555
+ "2770f5bcbf2fb7972bb0fdf00f6309d80a438a356af61e44f6e0a87d9e9e9749",
556
+ "27a0f083a047b857fb190b8ce227af6470dbc9877e12ddec667ccf092bb38aec",
557
+ "287ee74ce4a16b40805867a6a84bd0ce31a8170bb87b024b7150cf6286225239",
558
+ "290915eb95c98623bede55f945feea19b31b9e0b55d498087586eb78ab673124",
559
+ "296d2d3d87283de2989d7feda5c1964bf0ca5f465cb1eb35a099058310bbccb0",
560
+ "298f0b6e389c94bad5eeb13a7bb276714273bee1a2355aa9a9ed2a3df00e230a",
561
+ "29cdfbc917293161550d97ea67ee5954df89f926f032460834ca3b42c3b0c0b4",
562
+ "29dec43762af9d614595a41f2684991e44ed10ae23dc53695d97267c85180d44",
563
+ "2a45934c4f237e7c55633a9340de671f1cc832add504b384487aad994940ebbd",
564
+ "2b56257937c61e54482358d290c70c7dd15533527c7a5c67304515f78ed16b8d",
565
+ "2bb9f77251763945536e30206e23872a28f964e93483bbb8af48f890d4997b5e",
566
+ "2bbf500cedf656b58907e06d4debf3ca6db083799a6f8e1b8dfe29d1b5ebde8b",
567
+ "2bcdc0a2add7d61b5d8bb831139d7ee1ba000f9ebd8f53610d24c83e0d3bfeec",
568
+ "2c8114ad5af7d1ed8952927e827195dc61069af77148844b8daa64d9b5681823",
569
+ "2d843ce2bf559df3e6e3d82f042bc26bf0a136135ce1d9d1593cf6243c64d035",
570
+ "2ddb82f8e201cc38d0cfcb6ab09d47204cfeb2e34f593b7004246cc461de2b8f",
571
+ "2df7da74da815e2293af7467feb178731a363c43bd04efb4fb90e749ea15dc5e",
572
+ "2e41728c1feb0baebd448e13a996a81a5a8966a98d399747910f1ca4a303ef70",
573
+ "306593faf7ac803611b57887a3d49cf391cda0afc34ae9717556467e2738697a",
574
+ "313a6956aa6fea5af537785643e7c18cc1f49a20ab356139294ff9836cd0041b",
575
+ "3192576ecac24034f3313854514349784b72cf1ccb3f15f780f09805b3c9049b",
576
+ "31ad8725309a36c83919b4e22219f6acc28ab74eacc1a21b6ce25a5ebe5c9a3e",
577
+ "31b01d9176f54313c51fb2b38cf190dec881b3f764ecf6deb9c5e0cc6dd39273",
578
+ "320b235b4b99a4c37396625aaeea37928b009140100d0816877947ec7fc4d5ea",
579
+ "3284ef95cf53b90e6245fea8805e438913bb7e2e64ff30a4165176c8f20a31e4",
580
+ "3392a6b9cfe682e3a90d835d4839fe4a9f905b691d4926c47260f7c8ba1b4a85",
581
+ "33c7dd7da206cb007ddfa4493161c2c693a42115ceb4e5f6a7f3cfd902345782",
582
+ "33faabea19353c19d4553ccdf76da78922ab20629399b3732ce4ae61907ac65f",
583
+ "34b5d85b732d320d0452c8df51466aff4fef9e0652eb24bb1bc66ff71ec6581c",
584
+ "35b21fc47388d8aae43bcd09e67381efd5aaa4d9bd21e5fb141a97d04cf81309",
585
+ "361c89942e74f15b1a64cab7fb3f05f16b25fd5b7eaeaaa79f47d2d6f8125a4e",
586
+ "366c6de86f1c9584274f3bc105667febc2bfb9bb718424768f668379bb93db08",
587
+ "369716a89b5a2a776551105e267ea8b2e1b4fc1ce2ca204b9c7ac23d1f4ee4a4",
588
+ "37107fa603a746d3044f70b0776bdeac5e4b62238f7af96d29a6ecdee8a842d7",
589
+ "380cfbaec777fd891c1597a776a2c240e6dc769b8906eed47a8ccb95bbaa78cc",
590
+ "383201ea110c3d298bf177866f91b326429a6f3a5d783d55e37c65f9a4d953c0",
591
+ "39247017df9f8e2f6b38525b4b8f539fdee0705d84c2d712645dafa8782ab94b",
592
+ "39a632a0dd972df1c6dd5b17ad64c4d685e856d636c08c058e54649fdb6f19c5",
593
+ "3a2244fff97f5e28b0893251fa2f63de51f9a17f1444cd9c61e1881183d65292",
594
+ "3ac68a2f2e6a24ff85471e36d646ee3d679168bf3e47dcbad5b881778e00b80d",
595
+ "3b739c0fe2bfe255cd7fb936e9f6e1845f71bd8c743a1fc8fb433296d95983cd",
596
+ "3be79bc31455e70d3ee04d04661e6d0b9fa99490ec0cccf42c16181c5861d8ae",
597
+ "3c18c8d7856f3b7ef04eac25dcbbf89fd5a1fe4843b539c7cb7882166dbcb2d8",
598
+ "3c8fdca1da5807ad4d0d5b5268370d5ffbb381eddb06dc8913187cbce8845887",
599
+ "3d353295a9cf0f3044dce3a7738138da3284c52ffdb32c5739f23644685c673e",
600
+ "3fce9b41bb738817096475807f2c5f81bc80ae97dab2169ad97014f6f0207c88",
601
+ "40fc07492f163d29cb6b2060b9b6bf4e5a560fd464bb30bc1abb8178cc5cb939",
602
+ "4142225cb1fc7c8d3b9c9f862c7ace7615be14132c42ad12cda618b9fed8995a",
603
+ "415941b56aeb483dc738f2fcb7a3ba5cfb2d06a29f588ae81513ec7ab266c2c1",
604
+ "4170672fc2ac340d5028f3a26a4048c04081eaeff1169e3bef194e10700b4f2f",
605
+ "41ad134a4e792921e2f152c35b002d24ba3e27978e3b7cea2e0512785f99958f",
606
+ "436872d61296107eb46ced244e1d4f49a60f41216b525d10b96fa7539e69636c",
607
+ "44cef6c21d800d28c49026c8244c298f3f35909ded05a7a44a6c482d70ffc87c",
608
+ "451f523412ce5db8060f637f418611a2d16ca92309e3a94eaa06c0d1e3bd494e",
609
+ "451f5bc41b81488efecf18360f35b71854d9f3e65791f439ef0e9dab0d5b37dd",
610
+ "45571b7a74ffc59271a2476ff48c2c826d616f3c37a4e16ba4d4a92d91b6a8af",
611
+ "459dd06d08f6d9b60ba6e4de2c1cdcd3a3a91e1eb7304d60266ed8e8f2130b91",
612
+ "45bcc6d17d9f2c38c31a1be7da5510e6b6ea61fdc3f448689e7bf557b917b69c",
613
+ "4641eae58a63093a0c751e22e7b86bc9b700f0d4c36acab080b7b6cf6c2dfb1e",
614
+ "469c1a5dc65e24634bf784a133c9fea1a62498a700cd990756b9b027fc41d172",
615
+ "4804a2cfc0023e955e593c78381d3b61286259d442c14c89d1d1e92407953a92",
616
+ "48f7954a804296dbce08b9c95d314387500bb222018cae413205cf937b672257",
617
+ "4981fdf0cff45e9cfd46e9af6198b260a55a671e7c6efddbf0164e1aef93afd1",
618
+ "498457d95c5a3f99cecd457405f7d21120b9a08e4c5ed64780e9be6bf751f3ce",
619
+ "49d01844c161be1a3d2e570af94081a9527688b303ff92f984a752dd7dc87be2",
620
+ "4a73a9bf644e075b82f021ea56d9a66ffcd2ae6eede7be08b7db5a61473bc880",
621
+ "4ad614012d8df8ff7af9ba31d7a63b365d7992e11244b6a05f9bf24d92fbaee7",
622
+ "4b69e82727dbc355de5bd1921ea3124b831d4ced84915a8ea17cb1f666b72272",
623
+ "4b718fc596fa2894b76c4c3388eaa3cafa24a3af8a21a0f43510acab83df8f9f",
624
+ "4bab1aff689033aa4c5ca9ad5b166c0cd307e959e387e0654cb5eed13954fc89",
625
+ "4d1b2b46737f492f016c7dc383e3c7f669c785b3dd5240ac3721b358acbd8210",
626
+ "4d9004e2f0824a680dafdd2d3e869569f5ea3223dfd53857e67e90b452f37ec7",
627
+ "4d9b125e84a3703202d74eb7f29a681fa27e540a947b33b760a539a566dbfe03",
628
+ "4e7e45bb17bc9bd419ff97fee6db24b20ffb4cda75234ea855dc5666b6ac48ff",
629
+ "4edfb7c99fb79709890e6c45b460e44d5137860108f9dda149ffd7843eef1029",
630
+ "4eee787b9350fc39a739c41725af1a82512bae12c02713dcb64838730c98f876",
631
+ "4f01d0243893a64e9b0690140be0a84485018cba8c65d95baf195a77e01eb9a3",
632
+ "4f4d34c6580fbf468a3fc98965c01a35bf44997246e9871b664b1e19d91672ae",
633
+ "5018bb862f81847720619f8f0faf45793f51bafd9f02598bb3413d039e1ab16c",
634
+ "50cba560b4f304187a517a419477bedfc0c351246253780d3ca27f1935089076",
635
+ "51831cae9b2c0df8f6b4ba9e837bdf36fe82bcbd3211a421f31d99e95b98f7dc",
636
+ "525e8f64554faae4e6c637c78347bd2d133858ce5eedc1be0424aab9c618db14",
637
+ "52d082bc104632d9938fe8dfc7ece166c2904d572f64a45bd2ea620c32472417",
638
+ "52dc9d114df63ecad5379ac4d8a5b08e86d96d991d02b80861701cfe841342b5",
639
+ "533f765b91e8a44d0458df84c596a73394224692a5581ca91ebb30cde24e642d",
640
+ "53edbf5e4b0447cebdce43a796490ae0e070759f15c2e4d59dca2b93e42165a8",
641
+ "540e5b5c8295af60000acf168a53e7961db4644654d1b601d8aae8e0bda8acc3",
642
+ "541a78131848d13136d40bcf2334ca0c86868c3086b4360a34b2b29ed97fcb5f",
643
+ "542cc1af0c8e1265cbac8ceda239074b8d0f0309c06e5060131d228318b73650",
644
+ "5443fe36272511f859309a002380c67a01a450dc75b3e5b60fd4cd5d59da0ed7",
645
+ "54c4fb2d3e171d1c0f230ed8f9340280988aa897b23b90e3a93a276d22b76025",
646
+ "5519a34de4d24b8bba071f9b8117857e6e65941dc1258a6d4eff403442e70b72",
647
+ "55b7248a422de620d58db64541a22f179cfa3cca0313ded1cabf0a5e9ff35dae",
648
+ "580cc59e7949ec5c4ba6fb09ac960eb34c0035f5b4290317db2f4eab7f14e27a",
649
+ "58a330fb02ee4f7a3f5adb72db91181dc23bb1320e7019b4f5558f44b5d2391a",
650
+ "5a1041b583b67b6b9346d9abc55560b53e3517805648f43a2669b2b8bffeb5d5",
651
+ "5a6a0132c83f62a6cba965d40358ab914c16fc6bcd2831dd3b3db26572c90a1a",
652
+ "5ab3935d6a5fdb5903b53fab49f4d289b48fa1b120bd62db92fd3d5f632d7f83",
653
+ "5c2eb08e8e176b6c3b4634919de6c8c8efeb454b315fbf3c774760c66a301655",
654
+ "5ce2b13654f31d24873de055c34629f000c246ed4397ecc2062f5cad378c3111",
655
+ "5dd091b805a96934a5b52bf4aba8ed9ded2b5bc7ee913dfa714fec9fb4dc76c4",
656
+ "5dfbaffaa9eb3d8a67ea9eb6c41885b2f3c45786f3ce69af01f8c0f30fd7977f",
657
+ "5e02ce65c5893406958f95b7f69cf802e042c60b08221b8727fa5d5016c8d0a7",
658
+ "5e22c92b199860d5664399375d86ebe84f1ede853411ff31a20208d9f8ee6a4d",
659
+ "5f33bc4ce5537fcb41f8aa551245f8e14f27d37045edb29c2a0bd8bd65e6f69e",
660
+ "5f4b2c0a277a52ecd448643366ec97cab00a082e455713a067c3eecc708b8b4e",
661
+ "5fca01db1ee71e3462ea89895a1813ccbffe8441ae9f399cd47ef2c82563a705",
662
+ "609d3b921f29270929930ca2794207cc6dbe1b51aaa0b2f024ef3c60154057a2",
663
+ "60c4aa72fc44b13fd9de803623fc04e1744910226d2a32e7675457172f1e3b7f",
664
+ "6133e8cd1edc38c7c1c9d4b50b506985f5b3df8259cae0bcdd0c2e065de20d24",
665
+ "618d53f71a4fb3768c995518218aec746c3fbef2f2890dc9eee4ae6c250671f6",
666
+ "61dc8dc952c410ae48a4de75e006138717f8e3fbb1b120a74834023df56da812",
667
+ "62afee6acafef492a58695ae10b9aedccbb2acac25c8ded11c5fb04807b22c0f",
668
+ "630fca898e849462716b2727150d539f480951c2a6c37ed073048e5fb4de5b18",
669
+ "65573e39f2f6860c70d6b314fe203f46c6ba472871e4721cec880898c78a4c46",
670
+ "66990e5666d927abe3213ae2dd12b4bedd73a278757255cdb9a11fb546391d55",
671
+ "66de57d988227dfd54fb3f690ec5af4ed71193fd4f61cdcc1aca010245a452d0",
672
+ "68b57c72ec6aee1c6890ef817b1260a7a0a856c58768d54e656882cc85dbc970",
673
+ "68c584d3957300195152c69e1948188ce43d86fa1b0d4a8968e7d691078c8852",
674
+ "68d6aca0ea953877a2168d497f30b8e229688c61a3b59953ac01afb43c967184",
675
+ "69ceb0c1bbde81e023627bab10dff9e478e537ca180c123d0121f126f5d5a238",
676
+ "6a635ff442f0c208a7b1d825475664e95d509bcfa4f778eafbeb53b4d8f060bb",
677
+ "6b2f36a818a271d6e6cbd61d5b781d546ae232c8b6a36f4b7c5f46af3ed69d31",
678
+ "6b65bff555349fccdbf699bf95f33340bb33aa1e97f013b8fe4b4430143bcdd3",
679
+ "6bc4432668aff0b13a2ecd06031e03c0ec27716ff87160f2210efac773fee27a",
680
+ "6c37f0f057cb86f846f323998c192dd8cf0125775bbbbe1922aeb0f417ca4bf2",
681
+ "6c4e8f674283aa3bb9912d8bdbfe80445128011c1ac2c0aa40d0cebaaf149bb0",
682
+ "6ceab1a3fe900414b3efc02ccbfc7733367f4a2a307aa10a1e6b520c9c40dc3a",
683
+ "6d7647910e0eeac9e7ad78604e39ef4206b3e0ff2674f26f87728f3d5895b988",
684
+ "6e2f8cd9e87038683721f1c73ea42c66818715dbc90b5efb91712e2372d64c53",
685
+ "7047d073070dea7c149f283594ccb01b9d15823a32d1d2ee5cdcf14ba98f9cf4",
686
+ "71838e854303096f4fb33060cfad1239512f66037e7d095730160c705b321ef0",
687
+ "722a603518ac43d933180723219e85970b11435f7fe6ebcbf95b5970e1a66c3d",
688
+ "72916d366fca7fd42d84d1991dce5aa7ca0cc15dde6d7901bf497469084bced8",
689
+ "73a3da42a96f0b11d3c388986c79599f864bb73b8469ade8ff54a3f0d2420ef8",
690
+ "745e713464767e651b7aac34ee800172fa3c5c6b0ba9e17e89382e4e5d532dbf",
691
+ "758bb504a252912208617fbfda26b0ba282d611a05489aa347d4b76ea784246f",
692
+ "75ffe6c452283ea1d91964876699acea114cb569458a6db255d9132b78642af4",
693
+ "7684f854663ef8b0a8a060f1ee08b373bbffee93a958b44c2899d6ea0f0093a4",
694
+ "7748451d5342802025c0f777ab931f2b6305636f0ae150c512c02e67f9b7b1fd",
695
+ "7771858a975f19dd2a5d2a987380a89c1daaea46e2269dbf9cdb0bdf223e9a90",
696
+ "77f4a3433bbca295de58539f3ca0ba94e6d085b038cd4a17f3cfe0902ace83ec",
697
+ "7813c7d55bdf518a2df2e9833493cfaeffd5e33f94b6865fe6df60b63f22d5e1",
698
+ "79d8413ff55b1ead8d9e7ba6c4b7d87143e16db9eeded02c6c89deca1155caf3",
699
+ "7a22a3d76b498054510260a2de6d56eba2d91ee40cb97b95231f10bf5ee48fc3",
700
+ "7ba9c357b43342420854faa64a1a18919b9f2116be800bcf16eeaa34eac92dd4",
701
+ "7bfae4c9ce7ed3f9c163d0cd1278b9975e11ed69cea67cb756b13dd05c7dfff3",
702
+ "7c50cd496d0231841656bd1b7f05909b3cf6ef4b690a21479840abdf69238015",
703
+ "7cdac0e35b745afd541a0d39587041db27c3edcefde7d0707afc2240e46fae7f",
704
+ "7cf49e44d1ad79dd40eb976f3cf877c1c2274ae1d1c7c95a72cbb6fef1ad2705",
705
+ "7cff5f5043122dcac5d4bfa75fec4b58f31d10a93667245e47df355ec9e41c8b",
706
+ "7d2dfeb31ab2b882590e91b09c40467085557dbfa32e3b9d82555090f7a7680a",
707
+ "7dc774183005bbc53c1c05bf6553bc96f66af5a75840ba036a17a57c7eefe50a",
708
+ "7f17ae8f11690a38cdb78ae290f55e434ec12879163774b086d9b218b6b00eed",
709
+ "7f6aac77c62d80535ded477fefe1aecb148015a4661d460e9c4aeefb60b787a1",
710
+ "7f9d0e1e7f5333329d16bada92ab9f0dc535bb5c6ea1cbeda1785fe5ce277e6c",
711
+ "8012580d27745c8f642ba233fc3e9f8534c428bf93b2683d737a406980a211d0",
712
+ "8087f093dc46ed0cd75381c3fb368e7eef340e4cf4d17f2bd5f98d501a04bcb3",
713
+ "80e66d45f829bc00ef1bab55b7a1118cc3de88dad03890515669a30d28a07f3d",
714
+ "817194c76893fcd404b84d28c81f5c585528dd364adccdb0020099ea2a45ef61",
715
+ "81dc06c678b17a10a5328067c6a555d140a4a8cea4a740f989cad1c5eba1bde5",
716
+ "8263af2d38b886b758dee8743e5380d470847a9587bedcdcf156741f7fb7b7ce",
717
+ "829e4c245f07ae3bfc0f84dd04e18c20a12b3bc6fb6ca6398b9a883b4c78ba05",
718
+ "8302f54e4192d55bd528f65a89b85730bae2159d5f595b71ecc573d1f724a5c8",
719
+ "832eb6c1219ab8c9f21b0862eace2813c9cc258152e43be5430ed7e38a00d3a4",
720
+ "841704d928cc50b4f815e0d2aca87bfb7d5544ef898c0b075f6da841da49bc97",
721
+ "8420245211e57e268a80dd17eb943380955e2f35802389221ec3756c6ec650a7",
722
+ "84f5fda143bfe216c4a329b777baaef820c950492b5d37758de5de6c40757231",
723
+ "85b4e8c8eaa65df6e928b1f2e179221685d4666fedd66f61b6e2ddb9f93b4a27",
724
+ "86024a863eb5d7b5e07ee37c6f26c1f4e42b2382c620f7d276d6224c0e84e8e4",
725
+ "86a37906b8f2764056054d05e3bfe1717445b7bfbb9d80949e6a5d3cf8bab93e",
726
+ "86b09b8fc5929eaf4bd5b3a643da815fcd8808fd58874ff4985e83f36ac8fc7d",
727
+ "86cb44b722a497cd7695eaf6e55eb819814a3d0143da512f7196c991e2d992bf",
728
+ "86d9cfbfb911259b06d55c4caed3c5168b16456ba7a203955dfb1be1368c255f",
729
+ "87cb23c1f7d34522de599ba212cd9c98f0700b35028c980fb9743b6c7b1fdfcc",
730
+ "8847011462d8bb9b32da2cb2f323e4637255d08b1568f0acd53037386a9ccec3",
731
+ "887b8f5a5bd1e0af38fc3c3a3b4f1c4cd3978c62722ed115ef8714bbc3bbd63c",
732
+ "88bfaad3f1a51f584fe3ec523b26d87b69106aab68209e92bdf8225325402df9",
733
+ "895df5db13b1825e92afb3683df71a311d6342bbd5b4717e1743ed232c228c8a",
734
+ "8b4efbe4bf29c6e6f317dc7b1acaeca3484f57adeae10261ef30407067ae5183",
735
+ "8bbf5589e3e95684873774134a0295889ab56e7fa3d8194e25961abd57aebac8",
736
+ "8c0b28b33462657c0fe32ddae2385e8282595012a68f38a6324f432f285678fe",
737
+ "8c73e1f98107d56cef9cb4f8fbd770c8ce1bfc50a7e7714d9baf19da133a5473",
738
+ "8c7414c1a2171f34029e543278a252ea3ec4a329b305a21d6cf260908c54dbc1",
739
+ "8d1a19151db3e8c51bd49a0afa2c26b2bc9dbc5733e98b85d1f3044e4451565c",
740
+ "8e6dbd3ceaa4bda1238eeb367c6a1394429476ec29f9febde591076bca8753d2",
741
+ "8ee259732a2f9df72cbf32ef940b0bb9d57081162b499546415f2a7e1edc697a",
742
+ "8fa7ce14d6aae5441afb5a92ae8a21ddbff16df27b9e4416239efff10dcfa7d6",
743
+ "910a59f3b14d6705740884aa44a918e05b085f2549101ec3f2078a0d435171d9",
744
+ "9204c03c4d9e8f503a9ba0404e7d6c93d8ddda8f2e00a8d5af52f3872023afa8",
745
+ "9218b4a66e4c1f7f96a48208da6921421ca4b88b0f4cd0d9176cfe9d66ffa648",
746
+ "925e4d065431537778b7aa6abd97dba382c8cd36e5cd6c6c9cd1e80df7c43d2a",
747
+ "928f77d740903d209ccc848190516bdca7d2cb4a484a7f64affdb0321dd986b6",
748
+ "93073054dc6e1c09b5395a8a0e7a0c24d32b41a5cc610ee08b1b06aaab15ee4e",
749
+ "9336967bedb64864799ab11c5dbe9952acdc679f041362bad35a44cde48db893",
750
+ "938638b3d83e2c4ce8e55523172d3e5c40d6e0ad6befe77889a0afbd5a849e06",
751
+ "94218311ff38c818716db2d747e481493b3931fc0695120760970c4a0480e458",
752
+ "9482adc1860f04494332ec88e6152471b6f5d4d9f125c0c3c12b0ac549d316d3",
753
+ "96e2e52c28c9e4680a18626c0140830389dafc6c7d1d6188f4c65a2ba9a8336f",
754
+ "971c548a21973ba775d26a2352808f7e2ed2e589dfbac75b3f9f3de62b1b4159",
755
+ "972daee2426e4aa9de3349121d502546f9245d59879def1dc544903348eb8f40",
756
+ "97ff36496e76c5cf2f19a554d1200ed33907c21a94148c93f068b641300afc61",
757
+ "994f3c33f66415d1c656e5999ee93123a8efbde509ddebaf67f9cf4c36fab195",
758
+ "99f254b4352512081a8988206af3659313bf7e094f7dbc934f326e27f4d7f441",
759
+ "9a44dc704a815d1aa0282d892aced643751494d9615a2ff64ec9787fab28b4af",
760
+ "9ae8ddf4f81e72c61a6110836deef4b05c1a5e61ad05d59f13c98a1c0fdc4b5d",
761
+ "9b3df01e01f3f6ad44012553e9158f31dfbb8751f2641ff32d89cbe1aaac4e90",
762
+ "9f666af66871af874000eb3547499cdd497980e5f624e62891d131aa219edc13",
763
+ "9f66fe744971c2223c34a29582dfeb7c4eeb41b7c0491d971cfb3ddb222ce936",
764
+ "9f9392c9a1e3239f1f067430efee9e0d1aec04ab8e1f69219ba6b296830bcf77",
765
+ "9fdefef352160c118ddeaa2c65497f31c869153f49a6be9dbcc6a07ee323a088",
766
+ "a0318f5a6454cf76f9b1050c3f3cfd33e65f873f9c1a9e21a65464016ce793dc",
767
+ "a14fde8ad0141e8b0e4c90d266146502861ea4c2be9a6cc300c7873342782449",
768
+ "a26a933196c6e8d9a888064bbe08f292e2e94bc045de8065084f274545399f5d",
769
+ "a391b9801bd7b38f17800db9946ba707cf9ec8ac80886d41b996cd395b132aa6",
770
+ "a3fffd039b0ab8b4ddc1528dbf442024639a31d61a114e5c6e48e0eb83dd5c2e",
771
+ "a41c1b36c5bebec916f2705919989410973d668f02747d79163784723878a52f",
772
+ "a49b957ee6e9c2a8ce5c4b95a5e4b2733a1f769523a8aeca1e889063798b2abb",
773
+ "a4d4fe2f70675daaf6963acabd0be398e7c0691c72b2c53e8ba7449a2542a3ac",
774
+ "a53c4f46e164ffbbef3a1ea4404e6b0ba5b7585e50ca2e5b31cd14b0175e9f83",
775
+ "a55011f4e9452242ea1e75a6156924b3038f46d3834d72b32425f197fcdc2aa2",
776
+ "a5670c4656c2fc83ec0835d2cd1178d1f1dff5a6c6e650848166850894ec07e6",
777
+ "a5717ca8b1c11e62d65ea76e663c34be028ac408eb6347e0c7612c3a1d6bae8c",
778
+ "a58b4667aa301fcbc2989d1b985a508498e1e06bfbb1ae12688743238492c3f2",
779
+ "a627ed3f0eb883a5734065199140c0ff4a700edc3daf9418bfe37d24388684d9",
780
+ "a7108a97f142fd2c64524d835bfdf3b2508c1d28823a476b2f55aeb211e37914",
781
+ "a8253532fbeeb9dc1ce83943582abcdb22b8f79b95d344c0a313cf5642c23c7d",
782
+ "a9438a6c670c9bfcdb791f2b5e0eff0e9d95f20503d579712edc98a7357b5317",
783
+ "aa918da55ac73194b1790aab33a89fa6c909714ecbd059e7faf8219913d1a61c",
784
+ "adeb52eddd4b76ebeabb5c01a1e758ed4a1eb69630a7b88069b3db4c101d0362",
785
+ "afa82734c0d1064f2fbb3d1c3148a2c79aa527bc85c7f24ffc9b004b7a834cd4",
786
+ "b092057e67c260ac6b71c234a623580868023dad0c210c335b596cb9415027f5",
787
+ "b0a637b42be52cd5c8141a7d809af9b9585ad553128af789e376795758690d77",
788
+ "b17e096f91b76956dca8fbf7b4a9e29b25c5054093485cde96a544a93674055c",
789
+ "b1d96ef4ffcdaecd193f824207cb9a41ed13365a66d5babc1d094e1e388b329e",
790
+ "b2425eff321e4e14d28f4b15e1a3a08ac098f81df6f875ba674e22ec51f5c0ba",
791
+ "b2543c77717a17f4ffc5a305118cbf02340a8cf1822d3106c9f62f18d0d120e6",
792
+ "b2fc04aefa3015b01c19efc2d6f5ce4fd54a1efd4f781659fb192c9216c0fc62",
793
+ "b335da7048bddcd73da25488b069038b40e03540be5f0d3e7701c9f5679d87ea",
794
+ "b497fb7a9f160cde00734f5c6463ced31d45620deb393511e7320e1a617fd2b5",
795
+ "b6387c9cad97b6ad7d78aefe24da89842e0b66ae6f7b7f5c7430fc1932bab8f7",
796
+ "b6613d86b06870aabd3e77cdc40e07add0b7cd39ab1597a0cca9100c7623b08c",
797
+ "b704b3daf899a707d6ac953a73cd4b4fa2be25c28a42e68ca522b936e24e6780",
798
+ "b75598647ad0876b0fd0c948d3ffe274fda5836dd09a8082d38583c240795850",
799
+ "b7858986f41da3681f1bcb05e4ee0246aaf211d39ecad3c82169dcc29c0c129d",
800
+ "b8210bbbb17d9b8f8012aa590f342c48852129402dcc0f698cb9d56a4206b004",
801
+ "b82e15a939ed6645ad5e9df36d8d160b30573b3985e4358509d66e336f15ef59",
802
+ "b8dbea530bbc958ea037ba2e06253f7298aef387c07c0de1b24245ad46531d8d",
803
+ "b927fc91034dded9b5fa01550fbecec586c085e4c774cd2edfa046a277faf9b7",
804
+ "b988a2792940ca8d17ec2ce543eebe57e66249aa8b2cbc1c70402799b3be0125",
805
+ "b98d3244c9c96739b2e1cfd1b97f9da4b9c68a7e3dc3f4f3076319d5e80fec2a",
806
+ "b99e7af54c13259b55f224b8818d9528423c345030a6f3ffd7ad7b3f22f6db0d",
807
+ "bb89f6bffb0936dc5931158a9362efeb008d937305c9da483301c88c6cfac999",
808
+ "bd4e2f910d6fd6379c069dcd4d0e731e8ff3e0d2bf3f024fc06b81df2fc51328",
809
+ "bd6447a23cea8dad5bc2e1809b758c0a3e274c4de21241f223eb378e0b99882a",
810
+ "bd7e598dc4357d7bd8a1d3f694ef0df9ac3995a3adb03dc6dada27148e3f7492",
811
+ "bd81669bb70aebf49a09579ab6b4552f883549524b3d7f4037c7fa92014e2fef",
812
+ "bd9b87bb81182b11eca9a81ad4d641ad3c5f9ee5c0ac37588c87a80bbab817c8",
813
+ "be3755ec87ca16137465fd498097e6bbab4be6389fda0800cad627b88474635e",
814
+ "be932b39171e6de32382027b2a706e3eea3aee791f5f1f90f254785ee32a8e70",
815
+ "be93706fbf393d4acc924253ce386c1bb5b8219c7f7dbd89ec6f0cbe06c99421",
816
+ "bf20906f1df86186fc89379fb3d8d5f1094d59b37bf28d44745fd6e41c418df6",
817
+ "c0573c1117981e9dc1822f5049602f3ad6ee4d9e6035a78a08b1edb61d43cc31",
818
+ "c1572bfce8b029273db50114aa8f945f8782abb4fa00e75fb075cd088eaa2636",
819
+ "c2a78cb7ec81ecc9bec383ab129cfce90d41f9296463b8b13eeef2884b007ad3",
820
+ "c2c4b79ab1489bc009bee083761079395f48f9d226be87b448a8ac7677452aeb",
821
+ "c37dbc80b577307cd0167b0ac8b82fe8f9fe2312c8f7c0552869e0c8a67e4866",
822
+ "c38279a86508c9ccb3d1cce1c09cdb2d9045397b8c244b5d83528d2538ffb32e",
823
+ "c5a2768beaeaa416cb5363a838278aa53e384f6b0b4e6f11ea06c8aaa5bcf591",
824
+ "c6e4353914d70195a7f3f4d97879d8fa17fe2d1ed2c23fabe14fb2f47ba28d7f",
825
+ "c8ee886768ecf9a263bcdb7282a1973dbcf61d19330d3d4abb8b99f1e4d0918b",
826
+ "c90391506e9de36d761156577b5e8d052a20096078c670027e0215abfaea32d9",
827
+ "c96d9dad9475e57944682bdb32ddc29b749431038e75c7a6d0594098487b4f14",
828
+ "ca5f1d4a686c704495ef54c4a7d7484e3f47c1f38a361e2db68aff9d72a772a4",
829
+ "ccebed809dd4feb69b4c3b3b3d4ae939e3fde214427ea99cd97f55e49ad5b42d",
830
+ "cd0cbdef471e564daa3377e022e622206bf384eb799a3caf052027da379e6bfe",
831
+ "cde05777ebf519d9d5cbc32c70d627fbcfa09a20102c054f4abd1577bb4c3fa6",
832
+ "cf5ad1f54bff9d9bf70b2e6d6510c2b4470c794f4568e809c65a18648a592451",
833
+ "cfe85387971de2625083a3f9deaacf348c759306b9a761e17874413ed47c4600",
834
+ "d044049ee253eeba638fcfe3f60bd58be42a0f49f4915b3b97e44803f34f1d95",
835
+ "d1099b0413b1a443c49d3e1ea31f6990d75e225453b33be430c8062caca79338",
836
+ "d11033de27092c7639922e586ac7022b8cb55dbcc6f08e393913be04e7f7205d",
837
+ "d20f64af91a020cd60de869540dbbfc8ac32d4faf450480c13eb1f3eb7fa3e52",
838
+ "d2bbd3c5548fdd09f813c2490d153c78a8af74c755c02cf3497588ac321476af",
839
+ "d3c38cd61a49db7b8aa336013e687a396c892c9a0cc4702d46911747632f948c",
840
+ "d415e0b8c0093a1ba14258e40e1aabb6905d2c98a6abc8f08a82ac9431eeaf43",
841
+ "d508361dcf52ae785901f36dcd855764730b11ab400fbb1a4c7c636f6738d992",
842
+ "d543e02b29981c2ecde24881f79b62ad7d65cbd99cacd4e640b6c410f46a9be5",
843
+ "d73c8ee431ade6e1b2e7d74facb8e1f8149eea7f5424edf80ae17ab6ba42d275",
844
+ "d74a31a15bf6d956e45b667a347932b4412571ebf44bd68d86b62ac45383c4c5",
845
+ "d81e0524a2810d7fa47cc5b5b66da3e1275dafffe084e34f2fbe7557049528f4",
846
+ "d83a6a42b0753e8485ab9264e7030e516f48dd44ee6ebe7860afeda9acfd7a0e",
847
+ "d84bc4b541486ecf463b0997e370a76c3fa182761fb345b5f083fd7512038f17",
848
+ "d875623a552ebe002143094173871690a95e4ee54140b6d74253633465298a79",
849
+ "d8aa9109d8e7c1a9b15ebeb5368eabae6eb9d5aa3923637e8996421092190a74",
850
+ "d9791761a59227220ece34dcbb5fb3d3d9903a0b5dbd85c0d01c5214b69dde0b",
851
+ "db7ceb27359a5b9813a451b448aabaec30c1f132501cc65986389544983bc8ca",
852
+ "dc26688fd3f3bf8de98b25462d69a60ce7f466033f919553c3c6872ac74faafc",
853
+ "dd6e4c3998f535b5b66ef4799c105bcf571c1763e11727c695c73b78e73d2a38",
854
+ "dd9c0a9f9a6d0a920109059ae2a43d8b433d033743a279059bde17398d503004",
855
+ "debdf6dcff2f2e957651abf47190dcaa4cb956b8aa9cfc0fecb9e768edcf3b98",
856
+ "dee19db3e3450e84e3f99706ffec37d698cb13f7cd4ac76729265dcac2b5d44a",
857
+ "df30dc755cc38e920ae8d9323a6f8efd3ac976f752462d95f2bae16453034c29",
858
+ "e05ae3f86a0fc0d1406d0ff5948773844f2f1741aa2485f1fba4e0edd4e81db4",
859
+ "e0cbcefc194c72bf1cff83654a299ccdbdec23fefcaf2ff45f8b2f0e1cf0e437",
860
+ "e112f179406d415cabe26b68440af8e51c816d156ed85696828a6c5159346fce",
861
+ "e16894303620c815a4f49c8586d737c13fa0cd6d054f40aaf9063c027f5fdcad",
862
+ "e28785317eeff5ecdd6145b6140d781c770449a4762922c3b6d397d19498c57d",
863
+ "e589aa96273ccb4165d74fdb416e368bc41355cf1ee6c177e790a5d4fcee58b4",
864
+ "e690a21a340ca630f26db2cca31587c42c453874419e4c04a741887c54ffc80d",
865
+ "e7002e6dbbff27e85564f42aef85fad9a707d1815ccb6c3f3e47740c85908aa3",
866
+ "e739d56f8c7a1568e17edeac600fed4f40c525e12c35210580205cff932fcfa7",
867
+ "e7e9a266f6f7a82554e3b14befb797bf4a41984ba7e2663c93be7f036c78f946",
868
+ "e8fe9a92835f0ef53330f5a4b771384380cdebd5d52bb5e86679c4dc1734ae7e",
869
+ "e96888d7d3f71f62047afcfdee212cd651206b6669c981a79fc818b87919828c",
870
+ "eab57568b298ef1cd4b0dccda61e2298241c4cd12e93b24cdc49f2c92b25cc04",
871
+ "eb5950bff96fac945a9dc33dd00739bb6d95503a14f6dd40afa3ade6cd3a0d90",
872
+ "eb7ce7ee77e377f4fa97ac4c426749c83fe7390cd8b7097e3a2aee692599962f",
873
+ "ec283bbf9c03dc73b62e0c0f00a178a23e21b90e62f654cb3d466e08e395d28a",
874
+ "ed8e402bd1a016413be08f0e9cb1dc26b15d976e8759930198c0304431f07da5",
875
+ "edb5574d009ee18eb19563f09587481989006ac02d572258ac7d0a704afec484",
876
+ "edd6b6c2ccd3ecb13e3012ff88852a7fd88c78f27e5597c4b438505d3608e84c",
877
+ "ee12e16c71c41dac8dc947c5423bf17c0cc9b5aa5deec8f55057a83bc3727d16",
878
+ "eeac3052aba13021d1be5b7b22a9aa2115f4d96c2612a0138e5f7d976ad26874",
879
+ "f02d0492dbabd9d62040a2988045eb1087fcad2a13d2cf1b04e131a27bacabd1",
880
+ "f0d8e74afb68279e6115cc369cadb222f6341efd28548430abc988f740ed13de",
881
+ "f229a18ffd0abf4772fe87b68a7ce6180b9109fd50d2bcf13da322766b2dad8c",
882
+ "f240d197b98753c2837c574f02af5dad6602bb02d2d1dde5c4401b50d4aa5706",
883
+ "f30fc00d0a59c04c029946514b3accc81c0c96b96402d67ab19b7cf4ab2b9ed0",
884
+ "f4b0a16d17858978ad779ed1161d562e7353d7007ea77c0f93c9489ce7ae128f",
885
+ "f531c44b42abd4e5690d70a90dc0f51f4fce56b65bbe89c769188fcab1af7417",
886
+ "f64570e9631e9bafcffb6fa049db5cbdc97ae7437f632af52674ea46447b307a",
887
+ "f694e411986769400a4bd755f4e6f4ae50794804ca363e2820b35e77ab4bf516",
888
+ "f7bda20eb265cef604dfa3fa03a0abecc1f6ad204710288c654b690370311cd7",
889
+ "f7c750e5b29465e56a8bbf1fa5046013b9c1ce4342f3f792dde4b80d0a0b2e83",
890
+ "f90b5f66967564ea762b74465943b69cf96fc71f926f926a591fe6ef05545805",
891
+ "f9f06da14bf1d57386851690bb94288cd3b1a368917765b46d11e3f5e2802052",
892
+ "f9f33b2057db3c7d9726109d9617bd37f8ef089b6e1583377d08385d52e02e8a",
893
+ "fa032edcd86968b3d1117c5978936894c733e21bcbacf7c98b9b521e8ee87dad",
894
+ "fa1ae0da300cdc977927390f502bc32ed7910a0752dfa98a4a2465ad918e6aee",
895
+ "fb23fde25430e3a53a7d807e41969b14e8dfc53885d86dcff62f082e8794ad23",
896
+ "fb597f4c310e4a3284c3000d0f9f528a0e9bfe1decfdab629820b72eb6f14555",
897
+ "fc0d39918a7b6774005db3d0c2cf333ee63ba116e2108f400a5b2fdcc7ab53ba",
898
+ "fc87e2c106a635f1ed1011b83f3ee89297d53c893493f49342b18dd4a8fae6c5",
899
+ "fcbf462ada431ea4e42e9190f3a8e717842696f4ba1ea470300e0e16f9bdf0fc",
900
+ "fd265b85a51de10189dfc2af020f3210ff668958fa854c0af4764307b38f9c61",
901
+ "fe27613ddfeb61f58d93ffd4cb8ac1348f41c747f7912e0cf55c20b97365b92e",
902
+ "fefcb03d77be6e00c50fd6d97c393c82f31a5cdf8d41ec6135c98edc6ddda4e5"
903
+ ],
904
+ "task_counts": {
905
+ "LiftPegUpright-v1": 1280,
906
+ "PickCube-v1": 2352,
907
+ "PullCube-v1": 800,
908
+ "PushCube-v1": 1120,
909
+ "StackCube-v1": 1008
910
+ }
911
+ }
workspace/runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed2/train.log ADDED
@@ -0,0 +1 @@
 
 
1
+ CTT checkpoint trained separately; rollout used checkpoint=runs/ctt_residual_base_context_obs_seed2/model.pt