anhtld commited on
Commit
ce20b52
·
verified ·
1 Parent(s): 6bb641f

Auto-sync: 2026-06-30 19:41:38

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": "2026-06-30T19:13:18+00:00",
4
  "mechanism_gap": {
5
  "best_clean_vs_direct_same_ckpt": 0.0776811594202898,
6
  "best_clean_vs_h16": 0.06318840579710144,
@@ -1918,9 +1918,101 @@
1918
  "std_success": 0.013130146843949137
1919
  },
1920
  "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_stacknowg": {
 
1921
  "label": "K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.01, no wrong-gripper component on Stack",
1922
- "missing": true,
1923
- "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_stacknowg_summary.json"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1924
  },
1925
  "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger002": {
1926
  "ci95_success": 0.03115620450679224,
 
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,
 
1918
  "std_success": 0.013130146843949137
1919
  },
1920
  "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_stacknowg": {
1921
+ "ci95_success": 0.0438491810213656,
1922
  "label": "K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.01, no wrong-gripper component on Stack",
1923
+ "mean_action_mse_to_best": 0.4006344801130826,
1924
+ "mean_progress": 0.571213552200006,
1925
+ "mean_success": 0.36000000000000004,
1926
+ "num_completed": 3,
1927
+ "per_task_success": {
1928
+ "LiftPegUpright-v1": {
1929
+ "mean_num_groups": 102.0,
1930
+ "mean_success": 0.26340865087329823,
1931
+ "std_success": 0.06057881385012102
1932
+ },
1933
+ "PickCube-v1": {
1934
+ "mean_num_groups": 196.66666666666666,
1935
+ "mean_success": 0.3272707566185827,
1936
+ "std_success": 0.03916710518683381
1937
+ },
1938
+ "PullCube-v1": {
1939
+ "mean_num_groups": 81.0,
1940
+ "mean_success": 0.22388547125389233,
1941
+ "std_success": 0.04131573991107309
1942
+ },
1943
+ "PushCube-v1": {
1944
+ "mean_num_groups": 102.0,
1945
+ "mean_success": 0.7604066042026859,
1946
+ "std_success": 0.08151850231627764
1947
+ },
1948
+ "StackCube-v1": {
1949
+ "mean_num_groups": 93.33333333333333,
1950
+ "mean_success": 0.20448070448070446,
1951
+ "std_success": 0.07686648155181856
1952
+ }
1953
+ },
1954
+ "seed_action_mse_to_best": {
1955
+ "0": 0.3848872706310257,
1956
+ "1": 0.39591813460478315,
1957
+ "2": 0.42109803510343896
1958
+ },
1959
+ "seed_progress": {
1960
+ "0": 0.5547143837315557,
1961
+ "1": 0.5689966015874044,
1962
+ "2": 0.5899296712810579
1963
+ },
1964
+ "seed_success": {
1965
+ "0": 0.3565217391304348,
1966
+ "1": 0.3443478260869565,
1967
+ "2": 0.3791304347826087
1968
+ },
1969
+ "selected_candidate_type_counts": {
1970
+ "retrieval_residual_policy_residual": 1550,
1971
+ "retrieval_residual_residual_near_miss": 102,
1972
+ "retrieval_residual_residual_near_miss+residual_wrong_gripper": 3,
1973
+ "retrieval_residual_residual_no_op": 40,
1974
+ "retrieval_residual_residual_no_op+residual_wrong_gripper": 16,
1975
+ "retrieval_residual_residual_wrong_gripper": 14
1976
+ },
1977
+ "selected_residual_scale_counts": {
1978
+ "0.35": 1585,
1979
+ "0.4": 17,
1980
+ "0.45": 123
1981
+ },
1982
+ "selected_type_outcomes": {
1983
+ "retrieval_residual_policy_residual": {
1984
+ "count": 1550.0,
1985
+ "mean_progress": 0.5643063750295269,
1986
+ "success_rate": 0.34903225806451615
1987
+ },
1988
+ "retrieval_residual_residual_near_miss": {
1989
+ "count": 102.0,
1990
+ "mean_progress": 0.5342439103177657,
1991
+ "success_rate": 0.3431372549019608
1992
+ },
1993
+ "retrieval_residual_residual_near_miss+residual_wrong_gripper": {
1994
+ "count": 3.0,
1995
+ "mean_progress": 1.0,
1996
+ "success_rate": 1.0
1997
+ },
1998
+ "retrieval_residual_residual_no_op": {
1999
+ "count": 40.0,
2000
+ "mean_progress": 0.7411756929010153,
2001
+ "success_rate": 0.5
2002
+ },
2003
+ "retrieval_residual_residual_no_op+residual_wrong_gripper": {
2004
+ "count": 16.0,
2005
+ "mean_progress": 0.8642677208408713,
2006
+ "success_rate": 0.8125
2007
+ },
2008
+ "retrieval_residual_residual_wrong_gripper": {
2009
+ "count": 14.0,
2010
+ "mean_progress": 0.6928790105240685,
2011
+ "success_rate": 0.6428571428571429
2012
+ }
2013
+ },
2014
+ "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_stacknowg_summary.json",
2015
+ "std_success": 0.017650246200160393
2016
  },
2017
  "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger002": {
2018
  "ci95_success": 0.03115620450679224,
results/paper_analysis.md CHANGED
@@ -1,6 +1,6 @@
1
  # Paper Analysis
2
 
3
- Generated: `2026-06-30T19:13:18+00:00`
4
 
5
  ## Main Seed Statistics
6
 
@@ -54,7 +54,7 @@ Generated: `2026-06-30T19:13:18+00:00`
54
  | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8trace | K4 compatible tangents, no-op bonus 0.03, unique candidate-oracle prefix K=8 branch trace | 3 | 35.65% +/- 1.22 | +/- 3.02 | 57.16% | 0.406 | +5.91 pp |
55
  | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger002 | K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.02 | 3 | 36.00% +/- 1.25 | +/- 3.12 | 57.42% | 0.407 | +6.26 pp |
56
  | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001 | K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.01 | 3 | 36.06% +/- 1.41 | +/- 3.49 | 57.38% | 0.407 | +6.32 pp |
57
- | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_stacknowg | K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.01, no wrong-gripper component on Stack | 0 | missing | missing | missing | missing | missing |
58
  | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger0005 | K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.005 | 3 | 36.00% +/- 1.42 | +/- 3.54 | 57.34% | 0.407 | +6.26 pp |
59
  | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger0015 | K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.015 | 3 | 36.00% +/- 1.25 | +/- 3.12 | 57.38% | 0.407 | +6.26 pp |
60
  | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_scale035 | K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.01, scale-gated 0.35 | 3 | 36.00% +/- 1.14 | +/- 2.83 | 57.33% | 0.407 | +6.26 pp |
 
1
  # Paper Analysis
2
 
3
+ Generated: `2026-06-30T23:47:13+00:00`
4
 
5
  ## Main Seed Statistics
6
 
 
54
  | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8trace | K4 compatible tangents, no-op bonus 0.03, unique candidate-oracle prefix K=8 branch trace | 3 | 35.65% +/- 1.22 | +/- 3.02 | 57.16% | 0.406 | +5.91 pp |
55
  | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger002 | K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.02 | 3 | 36.00% +/- 1.25 | +/- 3.12 | 57.42% | 0.407 | +6.26 pp |
56
  | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001 | K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.01 | 3 | 36.06% +/- 1.41 | +/- 3.49 | 57.38% | 0.407 | +6.32 pp |
57
+ | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_stacknowg | K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.01, no wrong-gripper component on Stack | 3 | 36.00% +/- 1.77 | +/- 4.38 | 57.12% | 0.401 | +6.26 pp |
58
  | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger0005 | K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.005 | 3 | 36.00% +/- 1.42 | +/- 3.54 | 57.34% | 0.407 | +6.26 pp |
59
  | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger0015 | K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.015 | 3 | 36.00% +/- 1.25 | +/- 3.12 | 57.38% | 0.407 | +6.26 pp |
60
  | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_scale035 | K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.01, scale-gated 0.35 | 3 | 36.00% +/- 1.14 | +/- 2.83 | 57.33% | 0.407 | +6.26 pp |
results/paper_table_status.json CHANGED
@@ -1439,14 +1439,14 @@
1439
  "story_role": "task-local tangent compatibility mask for Stack where wrong-gripper residuals are an anti-affordance",
1440
  "fallback_success": null,
1441
  "pending_job": "14976136/14976144",
1442
- "path_exists": false,
1443
- "status": "pending",
1444
- "success": null,
1445
- "std_success": null,
1446
  "completed_seeds": null,
1447
- "num_completed": null,
1448
  "best_config": null,
1449
- "gain_vs_h16_policy": null
1450
  },
1451
  {
1452
  "key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger0005",
 
1439
  "story_role": "task-local tangent compatibility mask for Stack where wrong-gripper residuals are an anti-affordance",
1440
  "fallback_success": null,
1441
  "pending_job": "14976136/14976144",
1442
+ "path_exists": true,
1443
+ "status": "complete",
1444
+ "success": 0.36,
1445
+ "std_success": 0.017650246200160393,
1446
  "completed_seeds": null,
1447
+ "num_completed": 3,
1448
  "best_config": null,
1449
+ "gain_vs_h16_policy": 0.06260869565217392
1450
  },
1451
  {
1452
  "key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger0005",
results/paper_table_status.md CHANGED
@@ -78,7 +78,7 @@ Baseline h=16 policy: 29.74%
78
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8trace | K4 compatible residual retrieval, unique candidate-oracle prefix K=8 branch trace | complete | 43.07% | +13.33 pp | diagnostic | no | no | diagnostic branch trace for deployable selector calibration over the local tangent chart |
79
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger002 | K4 compatible residual retrieval, near-miss challenger gate 0.02 | complete | 36.00% | +6.26 pp | yes | no | no | trace-motivated two-stage selector calibration over the compatible local tangent chart |
80
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001 | K4 compatible residual retrieval, near-miss challenger gate 0.01 | complete | 36.06% | +6.32 pp | yes | no | no | lower-margin sensitivity for trace-motivated singleton near-miss challenger calibration |
81
- | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_stacknowg | K4 compatible residual retrieval, near-miss challenger gate 0.01, no wrong-gripper component on Stack | pending 14976136/14976144 | pending | pending | yes | no | no | task-local tangent compatibility mask for Stack where wrong-gripper residuals are an anti-affordance |
82
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger0005 | K4 compatible residual retrieval, near-miss challenger gate 0.005 | complete | 36.00% | +6.26 pp | yes | no | no | fine lower-margin calibration for the trace-motivated singleton near-miss challenger gate |
83
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger0015 | K4 compatible residual retrieval, near-miss challenger gate 0.015 | complete | 36.00% | +6.26 pp | yes | no | no | fine upper-margin calibration for the trace-motivated singleton near-miss challenger gate |
84
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_scale035 | K4 compatible residual retrieval, near-miss challenger gate 0.01, scale-gated 0.35 | complete | 36.00% | +6.26 pp | yes | no | no | trace-motivated near-miss challenger with tangent-length reliability gating at the shortest scale |
 
78
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8trace | K4 compatible residual retrieval, unique candidate-oracle prefix K=8 branch trace | complete | 43.07% | +13.33 pp | diagnostic | no | no | diagnostic branch trace for deployable selector calibration over the local tangent chart |
79
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger002 | K4 compatible residual retrieval, near-miss challenger gate 0.02 | complete | 36.00% | +6.26 pp | yes | no | no | trace-motivated two-stage selector calibration over the compatible local tangent chart |
80
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001 | K4 compatible residual retrieval, near-miss challenger gate 0.01 | complete | 36.06% | +6.32 pp | yes | no | no | lower-margin sensitivity for trace-motivated singleton near-miss challenger calibration |
81
+ | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_stacknowg | K4 compatible residual retrieval, near-miss challenger gate 0.01, no wrong-gripper component on Stack | complete | 36.00% | +6.26 pp | yes | no | no | task-local tangent compatibility mask for Stack where wrong-gripper residuals are an anti-affordance |
82
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger0005 | K4 compatible residual retrieval, near-miss challenger gate 0.005 | complete | 36.00% | +6.26 pp | yes | no | no | fine lower-margin calibration for the trace-motivated singleton near-miss challenger gate |
83
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger0015 | K4 compatible residual retrieval, near-miss challenger gate 0.015 | complete | 36.00% | +6.26 pp | yes | no | no | fine upper-margin calibration for the trace-motivated singleton near-miss challenger gate |
84
  | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_scale035 | K4 compatible residual retrieval, near-miss challenger gate 0.01, scale-gated 0.35 | complete | 36.00% | +6.26 pp | yes | no | no | trace-motivated near-miss challenger with tangent-length reliability gating at the shortest scale |