anhtld commited on
Commit
a6cc44f
·
verified ·
1 Parent(s): cd39378

Auto-sync: 2026-06-30 20:26:14

Browse files
results/paper_analysis.json CHANGED
@@ -1,6 +1,6 @@
1
  {
2
  "best_clean_key": "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001",
3
- "generated_utc": "2026-06-30T23:47:13+00:00",
4
  "mechanism_gap": {
5
  "best_clean_vs_direct_same_ckpt": 0.0776811594202898,
6
  "best_clean_vs_h16": 0.06318840579710144,
@@ -2305,6 +2305,16 @@
2305
  "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmglobal_wgpickpull_challenger0p01_norm_summary.json",
2306
  "std_success": 0.015780472566743406
2307
  },
 
 
 
 
 
 
 
 
 
 
2308
  "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgchallenger001": {
2309
  "ci95_success": 0.028111505659259583,
2310
  "label": "K4 compatible tangents, no-op bonus 0.03, near-miss/wrong-gripper challenger gate 0.01",
@@ -2499,6 +2509,200 @@
2499
  "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmwgchallenger0p01_pickpull_norm_summary.json",
2500
  "std_success": 0.01659024698116427
2501
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2502
  "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8": {
2503
  "candidate_oracle_rollouts": 8,
2504
  "candidate_oracle_type_counts": {
 
1
  {
2
  "best_clean_key": "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001",
3
+ "generated_utc": "2026-07-01T00:31:55+00:00",
4
  "mechanism_gap": {
5
  "best_clean_vs_direct_same_ckpt": 0.0776811594202898,
6
  "best_clean_vs_h16": 0.06318840579710144,
 
2305
  "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmglobal_wgpickpull_challenger0p01_norm_summary.json",
2306
  "std_success": 0.015780472566743406
2307
  },
2308
+ "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmglobal_wgpickpull_wgmargin003_challenger001": {
2309
+ "label": "K4 compatible tangents, no-op bonus 0.03, near-miss global + wrong-gripper Pick/Pull margin 0.03",
2310
+ "missing": true,
2311
+ "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmglobal_wgpickpull_wgmargin0p03_challenger0p01_summary.json"
2312
+ },
2313
+ "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmglobal_wgpickpull_wgmargin005_challenger001": {
2314
+ "label": "K4 compatible tangents, no-op bonus 0.03, near-miss global + wrong-gripper Pick/Pull margin 0.05",
2315
+ "missing": true,
2316
+ "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmglobal_wgpickpull_wgmargin0p05_challenger0p01_summary.json"
2317
+ },
2318
  "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgchallenger001": {
2319
  "ci95_success": 0.028111505659259583,
2320
  "label": "K4 compatible tangents, no-op bonus 0.03, near-miss/wrong-gripper challenger gate 0.01",
 
2509
  "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmwgchallenger0p01_pickpull_norm_summary.json",
2510
  "std_success": 0.01659024698116427
2511
  },
2512
+ "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgmargin003_challenger001": {
2513
+ "ci95_success": 0.03465460086179854,
2514
+ "label": "K4 compatible tangents, no-op bonus 0.03, near-miss challenger 0.01, wrong-gripper margin 0.03",
2515
+ "mean_action_mse_to_best": 0.42052025424397516,
2516
+ "mean_progress": 0.5725739098747895,
2517
+ "mean_success": 0.35768115942028983,
2518
+ "num_completed": 3,
2519
+ "per_task_success": {
2520
+ "LiftPegUpright-v1": {
2521
+ "mean_num_groups": 102.0,
2522
+ "mean_success": 0.2564642064288538,
2523
+ "std_success": 0.04855619832211069
2524
+ },
2525
+ "PickCube-v1": {
2526
+ "mean_num_groups": 196.66666666666666,
2527
+ "mean_success": 0.3289542583020844,
2528
+ "std_success": 0.04030135158178743
2529
+ },
2530
+ "PullCube-v1": {
2531
+ "mean_num_groups": 81.0,
2532
+ "mean_success": 0.228271436166173,
2533
+ "std_success": 0.03525140677829959
2534
+ },
2535
+ "PushCube-v1": {
2536
+ "mean_num_groups": 102.0,
2537
+ "mean_success": 0.757403601199683,
2538
+ "std_success": 0.07781968814460603
2539
+ },
2540
+ "StackCube-v1": {
2541
+ "mean_num_groups": 93.33333333333333,
2542
+ "mean_success": 0.19336959336959336,
2543
+ "std_success": 0.08682910965604235
2544
+ }
2545
+ },
2546
+ "seed_action_mse_to_best": {
2547
+ "0": 0.39712414266136675,
2548
+ "1": 0.4136190785184179,
2549
+ "2": 0.45081754155214065
2550
+ },
2551
+ "seed_progress": {
2552
+ "0": 0.5605675064440331,
2553
+ "1": 0.5702972661898188,
2554
+ "2": 0.5868569569905167
2555
+ },
2556
+ "seed_success": {
2557
+ "0": 0.3565217391304348,
2558
+ "1": 0.3443478260869565,
2559
+ "2": 0.37217391304347824
2560
+ },
2561
+ "selected_candidate_type_counts": {
2562
+ "retrieval_residual_policy_residual": 1400,
2563
+ "retrieval_residual_residual_near_miss": 91,
2564
+ "retrieval_residual_residual_near_miss+residual_wrong_gripper": 5,
2565
+ "retrieval_residual_residual_no_op": 38,
2566
+ "retrieval_residual_residual_no_op+residual_wrong_gripper": 27,
2567
+ "retrieval_residual_residual_wrong_gripper": 164
2568
+ },
2569
+ "selected_residual_scale_counts": {
2570
+ "0.35": 1480,
2571
+ "0.4": 44,
2572
+ "0.45": 201
2573
+ },
2574
+ "selected_type_outcomes": {
2575
+ "retrieval_residual_policy_residual": {
2576
+ "count": 1400.0,
2577
+ "mean_progress": 0.545508016923642,
2578
+ "success_rate": 0.3171428571428571
2579
+ },
2580
+ "retrieval_residual_residual_near_miss": {
2581
+ "count": 91.0,
2582
+ "mean_progress": 0.5661385127926594,
2583
+ "success_rate": 0.38461538461538464
2584
+ },
2585
+ "retrieval_residual_residual_near_miss+residual_wrong_gripper": {
2586
+ "count": 5.0,
2587
+ "mean_progress": 0.7557853102684021,
2588
+ "success_rate": 0.6
2589
+ },
2590
+ "retrieval_residual_residual_no_op": {
2591
+ "count": 38.0,
2592
+ "mean_progress": 0.7757954887653652,
2593
+ "success_rate": 0.5263157894736842
2594
+ },
2595
+ "retrieval_residual_residual_no_op+residual_wrong_gripper": {
2596
+ "count": 27.0,
2597
+ "mean_progress": 0.6916872307106301,
2598
+ "success_rate": 0.48148148148148145
2599
+ },
2600
+ "retrieval_residual_residual_wrong_gripper": {
2601
+ "count": 164.0,
2602
+ "mean_progress": 0.7349113159949278,
2603
+ "success_rate": 0.6219512195121951
2604
+ }
2605
+ },
2606
+ "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmwgmargin0p03_challenger0p01_summary.json",
2607
+ "std_success": 0.01394922830784459
2608
+ },
2609
+ "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgmargin005_challenger001": {
2610
+ "ci95_success": 0.03242840579710147,
2611
+ "label": "K4 compatible tangents, no-op bonus 0.03, near-miss challenger 0.01, wrong-gripper margin 0.05",
2612
+ "mean_action_mse_to_best": 0.41712138921886255,
2613
+ "mean_progress": 0.5713764009092921,
2614
+ "mean_success": 0.35594202898550725,
2615
+ "num_completed": 3,
2616
+ "per_task_success": {
2617
+ "LiftPegUpright-v1": {
2618
+ "mean_num_groups": 102.0,
2619
+ "mean_success": 0.2564642064288538,
2620
+ "std_success": 0.04855619832211069
2621
+ },
2622
+ "PickCube-v1": {
2623
+ "mean_num_groups": 196.66666666666666,
2624
+ "mean_success": 0.3272707566185827,
2625
+ "std_success": 0.03916710518683381
2626
+ },
2627
+ "PullCube-v1": {
2628
+ "mean_num_groups": 81.0,
2629
+ "mean_success": 0.22388547125389233,
2630
+ "std_success": 0.04131573991107309
2631
+ },
2632
+ "PushCube-v1": {
2633
+ "mean_num_groups": 102.0,
2634
+ "mean_success": 0.7541032711666796,
2635
+ "std_success": 0.07644139692615382
2636
+ },
2637
+ "StackCube-v1": {
2638
+ "mean_num_groups": 93.33333333333333,
2639
+ "mean_success": 0.19336959336959336,
2640
+ "std_success": 0.08682910965604235
2641
+ }
2642
+ },
2643
+ "seed_action_mse_to_best": {
2644
+ "0": 0.3951253530658458,
2645
+ "1": 0.411393547800734,
2646
+ "2": 0.44484526679000774
2647
+ },
2648
+ "seed_progress": {
2649
+ "0": 0.5606453491382949,
2650
+ "1": 0.5687696125099192,
2651
+ "2": 0.5847142410796622
2652
+ },
2653
+ "seed_success": {
2654
+ "0": 0.3565217391304348,
2655
+ "1": 0.3426086956521739,
2656
+ "2": 0.36869565217391304
2657
+ },
2658
+ "selected_candidate_type_counts": {
2659
+ "retrieval_residual_policy_residual": 1447,
2660
+ "retrieval_residual_residual_near_miss": 91,
2661
+ "retrieval_residual_residual_near_miss+residual_wrong_gripper": 5,
2662
+ "retrieval_residual_residual_no_op": 38,
2663
+ "retrieval_residual_residual_no_op+residual_wrong_gripper": 27,
2664
+ "retrieval_residual_residual_wrong_gripper": 117
2665
+ },
2666
+ "selected_residual_scale_counts": {
2667
+ "0.35": 1507,
2668
+ "0.4": 39,
2669
+ "0.45": 179
2670
+ },
2671
+ "selected_type_outcomes": {
2672
+ "retrieval_residual_policy_residual": {
2673
+ "count": 1447.0,
2674
+ "mean_progress": 0.5467335695066794,
2675
+ "success_rate": 0.3213545266067726
2676
+ },
2677
+ "retrieval_residual_residual_near_miss": {
2678
+ "count": 91.0,
2679
+ "mean_progress": 0.5661385127926594,
2680
+ "success_rate": 0.38461538461538464
2681
+ },
2682
+ "retrieval_residual_residual_near_miss+residual_wrong_gripper": {
2683
+ "count": 5.0,
2684
+ "mean_progress": 0.7557853102684021,
2685
+ "success_rate": 0.6
2686
+ },
2687
+ "retrieval_residual_residual_no_op": {
2688
+ "count": 38.0,
2689
+ "mean_progress": 0.7757954887653652,
2690
+ "success_rate": 0.5263157894736842
2691
+ },
2692
+ "retrieval_residual_residual_no_op+residual_wrong_gripper": {
2693
+ "count": 27.0,
2694
+ "mean_progress": 0.6916872307106301,
2695
+ "success_rate": 0.48148148148148145
2696
+ },
2697
+ "retrieval_residual_residual_wrong_gripper": {
2698
+ "count": 117.0,
2699
+ "mean_progress": 0.7781837732873411,
2700
+ "success_rate": 0.6666666666666666
2701
+ }
2702
+ },
2703
+ "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmwgmargin0p05_challenger0p01_summary.json",
2704
+ "std_success": 0.01305313652080894
2705
+ },
2706
  "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8": {
2707
  "candidate_oracle_rollouts": 8,
2708
  "candidate_oracle_type_counts": {
results/paper_analysis.md CHANGED
@@ -1,6 +1,6 @@
1
  # Paper Analysis
2
 
3
- Generated: `2026-06-30T23:47:13+00:00`
4
 
5
  ## Main Seed Statistics
6
 
@@ -61,8 +61,12 @@ Generated: `2026-06-30T23:47:13+00:00`
61
  | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_scales035040 | K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.01, scale-gated 0.35/0.40 | 3 | 36.00% +/- 1.31 | +/- 3.26 | 57.37% | 0.407 | +6.26 pp |
62
  | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger003 | K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.03 | 3 | 35.94% +/- 1.28 | +/- 3.18 | 57.36% | 0.407 | +6.20 pp |
63
  | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgchallenger001 | K4 compatible tangents, no-op bonus 0.03, near-miss/wrong-gripper challenger gate 0.01 | 3 | 35.94% +/- 1.13 | +/- 2.81 | 57.40% | 0.424 | +6.20 pp |
 
 
64
  | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgchallenger001_pickpull | K4 compatible tangents, no-op bonus 0.03, near-miss/wrong-gripper challenger gate 0.01 on Pick/Pull only | 3 | 36.00% +/- 1.66 | +/- 4.12 | 57.36% | 0.418 | +6.26 pp |
65
  | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmglobal_wgpickpull_challenger001 | K4 compatible tangents, no-op bonus 0.03, near-miss global + wrong-gripper challenger on Pick/Pull, gate 0.01 | 3 | 36.06% +/- 1.58 | +/- 3.92 | 57.41% | 0.419 | +6.32 pp |
 
 
66
  | typed_proposal_lattice_types6_prepend_margin000 | Typed proposal lattice head, six families, policy-prepended margin 0.00 | 3 | 31.30% +/- 3.81 | +/- 9.48 | 54.03% | 0.480 | +1.57 pp |
67
  | typed_proposal_lattice_types6_prepend_margin005 | Typed proposal lattice head, six families, policy-prepended margin 0.05 | 3 | 32.17% +/- 3.05 | +/- 7.57 | 54.74% | 0.467 | +2.43 pp |
68
  | typed_proposal_lattice_types4safe_prepend_margin005 | Typed proposal lattice head, safe four families, policy-prepended margin 0.05 | 3 | 33.45% +/- 3.39 | +/- 8.43 | 55.47% | 0.456 | +3.71 pp |
 
1
  # Paper Analysis
2
 
3
+ Generated: `2026-07-01T00:31:55+00:00`
4
 
5
  ## Main Seed Statistics
6
 
 
61
  | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_scales035040 | K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.01, scale-gated 0.35/0.40 | 3 | 36.00% +/- 1.31 | +/- 3.26 | 57.37% | 0.407 | +6.26 pp |
62
  | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger003 | K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.03 | 3 | 35.94% +/- 1.28 | +/- 3.18 | 57.36% | 0.407 | +6.20 pp |
63
  | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgchallenger001 | K4 compatible tangents, no-op bonus 0.03, near-miss/wrong-gripper challenger gate 0.01 | 3 | 35.94% +/- 1.13 | +/- 2.81 | 57.40% | 0.424 | +6.20 pp |
64
+ | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgmargin003_challenger001 | K4 compatible tangents, no-op bonus 0.03, near-miss challenger 0.01, wrong-gripper margin 0.03 | 3 | 35.77% +/- 1.39 | +/- 3.47 | 57.26% | 0.421 | +6.03 pp |
65
+ | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgmargin005_challenger001 | K4 compatible tangents, no-op bonus 0.03, near-miss challenger 0.01, wrong-gripper margin 0.05 | 3 | 35.59% +/- 1.31 | +/- 3.24 | 57.14% | 0.417 | +5.86 pp |
66
  | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgchallenger001_pickpull | K4 compatible tangents, no-op bonus 0.03, near-miss/wrong-gripper challenger gate 0.01 on Pick/Pull only | 3 | 36.00% +/- 1.66 | +/- 4.12 | 57.36% | 0.418 | +6.26 pp |
67
  | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmglobal_wgpickpull_challenger001 | K4 compatible tangents, no-op bonus 0.03, near-miss global + wrong-gripper challenger on Pick/Pull, gate 0.01 | 3 | 36.06% +/- 1.58 | +/- 3.92 | 57.41% | 0.419 | +6.32 pp |
68
+ | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmglobal_wgpickpull_wgmargin003_challenger001 | K4 compatible tangents, no-op bonus 0.03, near-miss global + wrong-gripper Pick/Pull margin 0.03 | 0 | missing | missing | missing | missing | missing |
69
+ | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmglobal_wgpickpull_wgmargin005_challenger001 | K4 compatible tangents, no-op bonus 0.03, near-miss global + wrong-gripper Pick/Pull margin 0.05 | 0 | missing | missing | missing | missing | missing |
70
  | typed_proposal_lattice_types6_prepend_margin000 | Typed proposal lattice head, six families, policy-prepended margin 0.00 | 3 | 31.30% +/- 3.81 | +/- 9.48 | 54.03% | 0.480 | +1.57 pp |
71
  | typed_proposal_lattice_types6_prepend_margin005 | Typed proposal lattice head, six families, policy-prepended margin 0.05 | 3 | 32.17% +/- 3.05 | +/- 7.57 | 54.74% | 0.467 | +2.43 pp |
72
  | typed_proposal_lattice_types4safe_prepend_margin005 | Typed proposal lattice head, safe four families, policy-prepended margin 0.05 | 3 | 33.45% +/- 3.39 | +/- 8.43 | 55.47% | 0.456 | +3.71 pp |
results/paper_table_status.json CHANGED
@@ -1562,6 +1562,44 @@
1562
  "best_config": null,
1563
  "gain_vs_h16_policy": 0.06202898550724639
1564
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1565
  {
1566
  "key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgchallenger001_pickpull",
1567
  "label": "K4 compatible residual retrieval, near-miss/wrong-gripper challenger gate 0.01 on Pick/Pull only",
@@ -1600,6 +1638,44 @@
1600
  "best_config": null,
1601
  "gain_vs_h16_policy": 0.0631884057971015
1602
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1603
  {
1604
  "key": "typed_proposal_lattice_types6_prepend_margin000",
1605
  "label": "Typed proposal lattice head, six families, policy-prepended margin 0.00",
 
1562
  "best_config": null,
1563
  "gain_vs_h16_policy": 0.06202898550724639
1564
  },
1565
+ {
1566
+ "key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgmargin003_challenger001",
1567
+ "label": "K4 compatible residual retrieval, near-miss challenger 0.01, wrong-gripper margin 0.03",
1568
+ "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmwgmargin0p03_challenger0p01_summary.json",
1569
+ "clean_deployment": "yes",
1570
+ "same_state_proposals": "no",
1571
+ "expert_proposal": "no",
1572
+ "story_role": "family-specific reliability margin: keep near-miss sensitive while requiring stronger wrong-gripper evidence",
1573
+ "fallback_success": null,
1574
+ "pending_job": "14989648/14989649",
1575
+ "path_exists": true,
1576
+ "status": "complete",
1577
+ "success": 0.35768115942028983,
1578
+ "std_success": 0.013949228307844588,
1579
+ "completed_seeds": null,
1580
+ "num_completed": 3,
1581
+ "best_config": null,
1582
+ "gain_vs_h16_policy": 0.06028985507246376
1583
+ },
1584
+ {
1585
+ "key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgmargin005_challenger001",
1586
+ "label": "K4 compatible residual retrieval, near-miss challenger 0.01, wrong-gripper margin 0.05",
1587
+ "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmwgmargin0p05_challenger0p01_summary.json",
1588
+ "clean_deployment": "yes",
1589
+ "same_state_proposals": "no",
1590
+ "expert_proposal": "no",
1591
+ "story_role": "family-specific reliability margin upper setting for wrong-gripper challenger evidence",
1592
+ "fallback_success": null,
1593
+ "pending_job": "14989651/14989659",
1594
+ "path_exists": true,
1595
+ "status": "complete",
1596
+ "success": 0.35594202898550725,
1597
+ "std_success": 0.01305313652080894,
1598
+ "completed_seeds": null,
1599
+ "num_completed": 3,
1600
+ "best_config": null,
1601
+ "gain_vs_h16_policy": 0.058550724637681184
1602
+ },
1603
  {
1604
  "key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgchallenger001_pickpull",
1605
  "label": "K4 compatible residual retrieval, near-miss/wrong-gripper challenger gate 0.01 on Pick/Pull only",
 
1638
  "best_config": null,
1639
  "gain_vs_h16_policy": 0.0631884057971015
1640
  },
1641
+ {
1642
+ "key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmglobal_wgpickpull_wgmargin003_challenger001",
1643
+ "label": "K4 compatible residual retrieval, near-miss global + wrong-gripper Pick/Pull margin 0.03",
1644
+ "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmglobal_wgpickpull_wgmargin0p03_challenger0p01_summary.json",
1645
+ "clean_deployment": "yes",
1646
+ "same_state_proposals": "no",
1647
+ "expert_proposal": "no",
1648
+ "story_role": "type-specific affordance margin: keep near-miss globally sensitive while requiring stronger wrong-gripper evidence only on Pick/Pull",
1649
+ "fallback_success": null,
1650
+ "pending_job": "14989757/14989775",
1651
+ "path_exists": false,
1652
+ "status": "pending",
1653
+ "success": null,
1654
+ "std_success": null,
1655
+ "completed_seeds": null,
1656
+ "num_completed": null,
1657
+ "best_config": null,
1658
+ "gain_vs_h16_policy": null
1659
+ },
1660
+ {
1661
+ "key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmglobal_wgpickpull_wgmargin005_challenger001",
1662
+ "label": "K4 compatible residual retrieval, near-miss global + wrong-gripper Pick/Pull margin 0.05",
1663
+ "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmglobal_wgpickpull_wgmargin0p05_challenger0p01_summary.json",
1664
+ "clean_deployment": "yes",
1665
+ "same_state_proposals": "no",
1666
+ "expert_proposal": "no",
1667
+ "story_role": "upper setting for type-specific wrong-gripper evidence on Pick/Pull while preserving global near-miss calibration",
1668
+ "fallback_success": null,
1669
+ "pending_job": "14989760/14989780",
1670
+ "path_exists": false,
1671
+ "status": "pending",
1672
+ "success": null,
1673
+ "std_success": null,
1674
+ "completed_seeds": null,
1675
+ "num_completed": null,
1676
+ "best_config": null,
1677
+ "gain_vs_h16_policy": null
1678
+ },
1679
  {
1680
  "key": "typed_proposal_lattice_types6_prepend_margin000",
1681
  "label": "Typed proposal lattice head, six families, policy-prepended margin 0.00",
results/paper_table_status.md CHANGED
@@ -85,8 +85,12 @@ Baseline h=16 policy: 29.74%
85
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_scales035040 | K4 compatible residual retrieval, near-miss challenger gate 0.01, scale-gated 0.35/0.40 | complete | 36.00% | +6.26 pp | yes | no | no | trace-motivated near-miss challenger with tangent-length reliability gating over the two shortest scales |
86
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger003 | K4 compatible residual retrieval, near-miss challenger gate 0.03 | complete | 35.94% | +6.20 pp | yes | no | no | upper-margin sensitivity for trace-motivated singleton near-miss challenger calibration |
87
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgchallenger001 | K4 compatible residual retrieval, near-miss/wrong-gripper challenger gate 0.01 | complete | 35.94% | +6.20 pp | yes | no | no | trace-motivated challenger calibration test for whether wrong-gripper residuals carry conditional selector headroom |
 
 
88
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgchallenger001_pickpull | K4 compatible residual retrieval, near-miss/wrong-gripper challenger gate 0.01 on Pick/Pull only | complete | 36.00% | +6.26 pp | yes | no | no | task-conditional challenger gate using branch trace evidence that wrong-gripper tangents help Pick/Pull but harm other tasks |
89
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmglobal_wgpickpull_challenger001 | K4 compatible residual retrieval, near-miss global + wrong-gripper challenger on Pick/Pull, gate 0.01 | complete | 36.06% | +6.32 pp | yes | no | no | type-specific affordance gate: keep near-miss global while restricting wrong-gripper challenger to acquisition/contact tasks |
 
 
90
  | typed_proposal_lattice_types6_prepend_margin000 | Typed proposal lattice head, six families, policy-prepended margin 0.00 | complete | 31.30% | +1.57 pp | yes | no | no | model-generated counterfactual proposal support test; field scores typed proposals rather than retrieved train-state residuals |
91
  | typed_proposal_lattice_types6_prepend_margin005 | Typed proposal lattice head, six families, policy-prepended margin 0.05 | complete | 32.17% | +2.43 pp | yes | no | no | model-generated counterfactual proposal support test with a conservative policy-abstention margin |
92
  | typed_proposal_lattice_types4safe_prepend_margin005 | Typed proposal lattice head, safe four families, policy-prepended margin 0.05 | complete | 33.45% | +3.71 pp | yes | no | no | anti-goal proposal-family mask test for the model-generated typed lattice |
 
85
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_scales035040 | K4 compatible residual retrieval, near-miss challenger gate 0.01, scale-gated 0.35/0.40 | complete | 36.00% | +6.26 pp | yes | no | no | trace-motivated near-miss challenger with tangent-length reliability gating over the two shortest scales |
86
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger003 | K4 compatible residual retrieval, near-miss challenger gate 0.03 | complete | 35.94% | +6.20 pp | yes | no | no | upper-margin sensitivity for trace-motivated singleton near-miss challenger calibration |
87
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgchallenger001 | K4 compatible residual retrieval, near-miss/wrong-gripper challenger gate 0.01 | complete | 35.94% | +6.20 pp | yes | no | no | trace-motivated challenger calibration test for whether wrong-gripper residuals carry conditional selector headroom |
88
+ | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgmargin003_challenger001 | K4 compatible residual retrieval, near-miss challenger 0.01, wrong-gripper margin 0.03 | complete | 35.77% | +6.03 pp | yes | no | no | family-specific reliability margin: keep near-miss sensitive while requiring stronger wrong-gripper evidence |
89
+ | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgmargin005_challenger001 | K4 compatible residual retrieval, near-miss challenger 0.01, wrong-gripper margin 0.05 | complete | 35.59% | +5.86 pp | yes | no | no | family-specific reliability margin upper setting for wrong-gripper challenger evidence |
90
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgchallenger001_pickpull | K4 compatible residual retrieval, near-miss/wrong-gripper challenger gate 0.01 on Pick/Pull only | complete | 36.00% | +6.26 pp | yes | no | no | task-conditional challenger gate using branch trace evidence that wrong-gripper tangents help Pick/Pull but harm other tasks |
91
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmglobal_wgpickpull_challenger001 | K4 compatible residual retrieval, near-miss global + wrong-gripper challenger on Pick/Pull, gate 0.01 | complete | 36.06% | +6.32 pp | yes | no | no | type-specific affordance gate: keep near-miss global while restricting wrong-gripper challenger to acquisition/contact tasks |
92
+ | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmglobal_wgpickpull_wgmargin003_challenger001 | K4 compatible residual retrieval, near-miss global + wrong-gripper Pick/Pull margin 0.03 | pending 14989757/14989775 | pending | pending | yes | no | no | type-specific affordance margin: keep near-miss globally sensitive while requiring stronger wrong-gripper evidence only on Pick/Pull |
93
+ | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmglobal_wgpickpull_wgmargin005_challenger001 | K4 compatible residual retrieval, near-miss global + wrong-gripper Pick/Pull margin 0.05 | pending 14989760/14989780 | pending | pending | yes | no | no | upper setting for type-specific wrong-gripper evidence on Pick/Pull while preserving global near-miss calibration |
94
  | typed_proposal_lattice_types6_prepend_margin000 | Typed proposal lattice head, six families, policy-prepended margin 0.00 | complete | 31.30% | +1.57 pp | yes | no | no | model-generated counterfactual proposal support test; field scores typed proposals rather than retrieved train-state residuals |
95
  | typed_proposal_lattice_types6_prepend_margin005 | Typed proposal lattice head, six families, policy-prepended margin 0.05 | complete | 32.17% | +2.43 pp | yes | no | no | model-generated counterfactual proposal support test with a conservative policy-abstention margin |
96
  | typed_proposal_lattice_types4safe_prepend_margin005 | Typed proposal lattice head, safe four families, policy-prepended margin 0.05 | complete | 33.45% | +3.71 pp | yes | no | no | anti-goal proposal-family mask test for the model-generated typed lattice |
scripts/build_paper_analysis.py CHANGED
@@ -497,6 +497,22 @@ METHODS = [
497
  "k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmglobal_wgpickpull_challenger0p01_norm_summary.json"
498
  ),
499
  ),
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
500
  MethodSpec(
501
  key="typed_proposal_lattice_types6_prepend_margin000",
502
  label="Typed proposal lattice head, six families, policy-prepended margin 0.00",
 
497
  "k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmglobal_wgpickpull_challenger0p01_norm_summary.json"
498
  ),
499
  ),
500
+ MethodSpec(
501
+ key="residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmglobal_wgpickpull_wgmargin003_challenger001",
502
+ label="K4 compatible tangents, no-op bonus 0.03, near-miss global + wrong-gripper Pick/Pull margin 0.03",
503
+ summary_path=(
504
+ "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_"
505
+ "k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmglobal_wgpickpull_wgmargin0p03_challenger0p01_summary.json"
506
+ ),
507
+ ),
508
+ MethodSpec(
509
+ key="residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmglobal_wgpickpull_wgmargin005_challenger001",
510
+ label="K4 compatible tangents, no-op bonus 0.03, near-miss global + wrong-gripper Pick/Pull margin 0.05",
511
+ summary_path=(
512
+ "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_"
513
+ "k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmglobal_wgpickpull_wgmargin0p05_challenger0p01_summary.json"
514
+ ),
515
+ ),
516
  MethodSpec(
517
  key="typed_proposal_lattice_types6_prepend_margin000",
518
  label="Typed proposal lattice head, six families, policy-prepended margin 0.00",
scripts/build_paper_table_status.py CHANGED
@@ -843,7 +843,7 @@ SPECS = [
843
  same_state_proposals="no",
844
  expert_proposal="no",
845
  story_role="family-specific reliability margin: keep near-miss sensitive while requiring stronger wrong-gripper evidence",
846
- pending_job="pending",
847
  ),
848
  ResultSpec(
849
  key="retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgmargin005_challenger001",
@@ -853,7 +853,7 @@ SPECS = [
853
  same_state_proposals="no",
854
  expert_proposal="no",
855
  story_role="family-specific reliability margin upper setting for wrong-gripper challenger evidence",
856
- pending_job="pending",
857
  ),
858
  ResultSpec(
859
  key="retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgchallenger001_pickpull",
@@ -875,6 +875,26 @@ SPECS = [
875
  story_role="type-specific affordance gate: keep near-miss global while restricting wrong-gripper challenger to acquisition/contact tasks",
876
  pending_job="14974610/14974611",
877
  ),
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
878
  ResultSpec(
879
  key="typed_proposal_lattice_types6_prepend_margin000",
880
  label="Typed proposal lattice head, six families, policy-prepended margin 0.00",
 
843
  same_state_proposals="no",
844
  expert_proposal="no",
845
  story_role="family-specific reliability margin: keep near-miss sensitive while requiring stronger wrong-gripper evidence",
846
+ pending_job="14989648/14989649",
847
  ),
848
  ResultSpec(
849
  key="retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgmargin005_challenger001",
 
853
  same_state_proposals="no",
854
  expert_proposal="no",
855
  story_role="family-specific reliability margin upper setting for wrong-gripper challenger evidence",
856
+ pending_job="14989651/14989659",
857
  ),
858
  ResultSpec(
859
  key="retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgchallenger001_pickpull",
 
875
  story_role="type-specific affordance gate: keep near-miss global while restricting wrong-gripper challenger to acquisition/contact tasks",
876
  pending_job="14974610/14974611",
877
  ),
878
+ ResultSpec(
879
+ key="retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmglobal_wgpickpull_wgmargin003_challenger001",
880
+ label="K4 compatible residual retrieval, near-miss global + wrong-gripper Pick/Pull margin 0.03",
881
+ path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmglobal_wgpickpull_wgmargin0p03_challenger0p01_summary.json",
882
+ clean_deployment="yes",
883
+ same_state_proposals="no",
884
+ expert_proposal="no",
885
+ story_role="type-specific affordance margin: keep near-miss globally sensitive while requiring stronger wrong-gripper evidence only on Pick/Pull",
886
+ pending_job="14989757/14989775",
887
+ ),
888
+ ResultSpec(
889
+ key="retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmglobal_wgpickpull_wgmargin005_challenger001",
890
+ label="K4 compatible residual retrieval, near-miss global + wrong-gripper Pick/Pull margin 0.05",
891
+ path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmglobal_wgpickpull_wgmargin0p05_challenger0p01_summary.json",
892
+ clean_deployment="yes",
893
+ same_state_proposals="no",
894
+ expert_proposal="no",
895
+ story_role="upper setting for type-specific wrong-gripper evidence on Pick/Pull while preserving global near-miss calibration",
896
+ pending_job="14989760/14989780",
897
+ ),
898
  ResultSpec(
899
  key="typed_proposal_lattice_types6_prepend_margin000",
900
  label="Typed proposal lattice head, six families, policy-prepended margin 0.00",