Auto-sync: 2026-06-30 05:12:29 (part 3)
Browse files- results/paper_analysis.json +195 -1
- results/paper_analysis.md +3 -1
- results/paper_core_results.md +1 -1
- results/paper_story_memo.md +6 -3
- results/paper_table_status.json +12 -12
- results/paper_table_status.md +2 -2
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-
|
| 4 |
"mechanism_gap": {
|
| 5 |
"best_clean_vs_direct_same_ckpt": 0.0776811594202898,
|
| 6 |
"best_clean_vs_h16": 0.06318840579710144,
|
|
@@ -1723,6 +1723,200 @@
|
|
| 1723 |
"source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p015_summary.json",
|
| 1724 |
"std_success": 0.012541047914657353
|
| 1725 |
},
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1726 |
"residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger002": {
|
| 1727 |
"ci95_success": 0.03115620450679224,
|
| 1728 |
"label": "K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.02",
|
|
|
|
| 1 |
{
|
| 2 |
"best_clean_key": "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001",
|
| 3 |
+
"generated_utc": "2026-06-30T09:13:25+00:00",
|
| 4 |
"mechanism_gap": {
|
| 5 |
"best_clean_vs_direct_same_ckpt": 0.0776811594202898,
|
| 6 |
"best_clean_vs_h16": 0.06318840579710144,
|
|
|
|
| 1723 |
"source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p015_summary.json",
|
| 1724 |
"std_success": 0.012541047914657353
|
| 1725 |
},
|
| 1726 |
+
"residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_scale035": {
|
| 1727 |
+
"ci95_success": 0.02833199143417802,
|
| 1728 |
+
"label": "K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.01, scale-gated 0.35",
|
| 1729 |
+
"mean_action_mse_to_best": 0.40711666752759745,
|
| 1730 |
+
"mean_progress": 0.5732656509485569,
|
| 1731 |
+
"mean_success": 0.36000000000000004,
|
| 1732 |
+
"num_completed": 3,
|
| 1733 |
+
"per_task_success": {
|
| 1734 |
+
"LiftPegUpright-v1": {
|
| 1735 |
+
"mean_num_groups": 102.0,
|
| 1736 |
+
"mean_success": 0.26340865087329823,
|
| 1737 |
+
"std_success": 0.06057881385012102
|
| 1738 |
+
},
|
| 1739 |
+
"PickCube-v1": {
|
| 1740 |
+
"mean_num_groups": 196.66666666666666,
|
| 1741 |
+
"mean_success": 0.33270553922727836,
|
| 1742 |
+
"std_success": 0.029907974490119102
|
| 1743 |
+
},
|
| 1744 |
+
"PullCube-v1": {
|
| 1745 |
+
"mean_num_groups": 81.0,
|
| 1746 |
+
"mean_success": 0.2214893293840662,
|
| 1747 |
+
"std_success": 0.02089044075040798
|
| 1748 |
+
},
|
| 1749 |
+
"PushCube-v1": {
|
| 1750 |
+
"mean_num_groups": 102.0,
|
| 1751 |
+
"mean_success": 0.7571578319419062,
|
| 1752 |
+
"std_success": 0.0856403530080274
|
| 1753 |
+
},
|
| 1754 |
+
"StackCube-v1": {
|
| 1755 |
+
"mean_num_groups": 93.33333333333333,
|
| 1756 |
+
"mean_success": 0.20044030044030045,
|
| 1757 |
+
"std_success": 0.08896982358351117
|
| 1758 |
+
}
|
| 1759 |
+
},
|
| 1760 |
+
"seed_action_mse_to_best": {
|
| 1761 |
+
"0": 0.39446062079104394,
|
| 1762 |
+
"1": 0.3992000118400092,
|
| 1763 |
+
"2": 0.4276893699517393
|
| 1764 |
+
},
|
| 1765 |
+
"seed_progress": {
|
| 1766 |
+
"0": 0.5603290924210222,
|
| 1767 |
+
"1": 0.5727110972935738,
|
| 1768 |
+
"2": 0.5867567631310743
|
| 1769 |
+
},
|
| 1770 |
+
"seed_success": {
|
| 1771 |
+
"0": 0.3582608695652174,
|
| 1772 |
+
"1": 0.34956521739130436,
|
| 1773 |
+
"2": 0.37217391304347824
|
| 1774 |
+
},
|
| 1775 |
+
"selected_candidate_type_counts": {
|
| 1776 |
+
"retrieval_residual_policy_residual": 1531,
|
| 1777 |
+
"retrieval_residual_residual_near_miss": 90,
|
| 1778 |
+
"retrieval_residual_residual_near_miss+residual_wrong_gripper": 5,
|
| 1779 |
+
"retrieval_residual_residual_no_op": 56,
|
| 1780 |
+
"retrieval_residual_residual_no_op+residual_wrong_gripper": 24,
|
| 1781 |
+
"retrieval_residual_residual_wrong_gripper": 19
|
| 1782 |
+
},
|
| 1783 |
+
"selected_residual_scale_counts": {
|
| 1784 |
+
"0.35": 1631,
|
| 1785 |
+
"0.4": 11,
|
| 1786 |
+
"0.45": 83
|
| 1787 |
+
},
|
| 1788 |
+
"selected_type_outcomes": {
|
| 1789 |
+
"retrieval_residual_policy_residual": {
|
| 1790 |
+
"count": 1531.0,
|
| 1791 |
+
"mean_progress": 0.5667326277051034,
|
| 1792 |
+
"success_rate": 0.35075114304376226
|
| 1793 |
+
},
|
| 1794 |
+
"retrieval_residual_residual_near_miss": {
|
| 1795 |
+
"count": 90.0,
|
| 1796 |
+
"mean_progress": 0.5404321867422873,
|
| 1797 |
+
"success_rate": 0.36666666666666664
|
| 1798 |
+
},
|
| 1799 |
+
"retrieval_residual_residual_near_miss+residual_wrong_gripper": {
|
| 1800 |
+
"count": 5.0,
|
| 1801 |
+
"mean_progress": 0.7557869911193847,
|
| 1802 |
+
"success_rate": 0.6
|
| 1803 |
+
},
|
| 1804 |
+
"retrieval_residual_residual_no_op": {
|
| 1805 |
+
"count": 56.0,
|
| 1806 |
+
"mean_progress": 0.7249722566921264,
|
| 1807 |
+
"success_rate": 0.5
|
| 1808 |
+
},
|
| 1809 |
+
"retrieval_residual_residual_no_op+residual_wrong_gripper": {
|
| 1810 |
+
"count": 24.0,
|
| 1811 |
+
"mean_progress": 0.6711457151298722,
|
| 1812 |
+
"success_rate": 0.4583333333333333
|
| 1813 |
+
},
|
| 1814 |
+
"retrieval_residual_residual_wrong_gripper": {
|
| 1815 |
+
"count": 19.0,
|
| 1816 |
+
"mean_progress": 0.6364115562878156,
|
| 1817 |
+
"success_rate": 0.47368421052631576
|
| 1818 |
+
}
|
| 1819 |
+
},
|
| 1820 |
+
"source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_scale035_summary.json",
|
| 1821 |
+
"std_success": 0.011404240911829549
|
| 1822 |
+
},
|
| 1823 |
+
"residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_scales035040": {
|
| 1824 |
+
"ci95_success": 0.032619725485313965,
|
| 1825 |
+
"label": "K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.01, scale-gated 0.35/0.40",
|
| 1826 |
+
"mean_action_mse_to_best": 0.40716694236098644,
|
| 1827 |
+
"mean_progress": 0.5736821288862468,
|
| 1828 |
+
"mean_success": 0.36000000000000004,
|
| 1829 |
+
"num_completed": 3,
|
| 1830 |
+
"per_task_success": {
|
| 1831 |
+
"LiftPegUpright-v1": {
|
| 1832 |
+
"mean_num_groups": 102.0,
|
| 1833 |
+
"mean_success": 0.26340865087329823,
|
| 1834 |
+
"std_success": 0.06057881385012102
|
| 1835 |
+
},
|
| 1836 |
+
"PickCube-v1": {
|
| 1837 |
+
"mean_num_groups": 196.66666666666666,
|
| 1838 |
+
"mean_success": 0.3308939450243798,
|
| 1839 |
+
"std_success": 0.032984654416421025
|
| 1840 |
+
},
|
| 1841 |
+
"PullCube-v1": {
|
| 1842 |
+
"mean_num_groups": 81.0,
|
| 1843 |
+
"mean_success": 0.22519303308776994,
|
| 1844 |
+
"std_success": 0.02468122407826519
|
| 1845 |
+
},
|
| 1846 |
+
"PushCube-v1": {
|
| 1847 |
+
"mean_num_groups": 102.0,
|
| 1848 |
+
"mean_success": 0.7571578319419062,
|
| 1849 |
+
"std_success": 0.0856403530080274
|
| 1850 |
+
},
|
| 1851 |
+
"StackCube-v1": {
|
| 1852 |
+
"mean_num_groups": 93.33333333333333,
|
| 1853 |
+
"mean_success": 0.20044030044030045,
|
| 1854 |
+
"std_success": 0.08896982358351117
|
| 1855 |
+
}
|
| 1856 |
+
},
|
| 1857 |
+
"seed_action_mse_to_best": {
|
| 1858 |
+
"0": 0.39465496900450925,
|
| 1859 |
+
"1": 0.3990837439878479,
|
| 1860 |
+
"2": 0.42776211409060205
|
| 1861 |
+
},
|
| 1862 |
+
"seed_progress": {
|
| 1863 |
+
"0": 0.5607444118218415,
|
| 1864 |
+
"1": 0.5718394296454347,
|
| 1865 |
+
"2": 0.5884625451914642
|
| 1866 |
+
},
|
| 1867 |
+
"seed_success": {
|
| 1868 |
+
"0": 0.3582608695652174,
|
| 1869 |
+
"1": 0.34782608695652173,
|
| 1870 |
+
"2": 0.3739130434782609
|
| 1871 |
+
},
|
| 1872 |
+
"selected_candidate_type_counts": {
|
| 1873 |
+
"retrieval_residual_policy_residual": 1528,
|
| 1874 |
+
"retrieval_residual_residual_near_miss": 93,
|
| 1875 |
+
"retrieval_residual_residual_near_miss+residual_wrong_gripper": 5,
|
| 1876 |
+
"retrieval_residual_residual_no_op": 56,
|
| 1877 |
+
"retrieval_residual_residual_no_op+residual_wrong_gripper": 24,
|
| 1878 |
+
"retrieval_residual_residual_wrong_gripper": 19
|
| 1879 |
+
},
|
| 1880 |
+
"selected_residual_scale_counts": {
|
| 1881 |
+
"0.35": 1567,
|
| 1882 |
+
"0.4": 75,
|
| 1883 |
+
"0.45": 83
|
| 1884 |
+
},
|
| 1885 |
+
"selected_type_outcomes": {
|
| 1886 |
+
"retrieval_residual_policy_residual": {
|
| 1887 |
+
"count": 1528.0,
|
| 1888 |
+
"mean_progress": 0.5667018804414848,
|
| 1889 |
+
"success_rate": 0.3507853403141361
|
| 1890 |
+
},
|
| 1891 |
+
"retrieval_residual_residual_near_miss": {
|
| 1892 |
+
"count": 93.0,
|
| 1893 |
+
"mean_progress": 0.5495107629166175,
|
| 1894 |
+
"success_rate": 0.3655913978494624
|
| 1895 |
+
},
|
| 1896 |
+
"retrieval_residual_residual_near_miss+residual_wrong_gripper": {
|
| 1897 |
+
"count": 5.0,
|
| 1898 |
+
"mean_progress": 0.7557869911193847,
|
| 1899 |
+
"success_rate": 0.6
|
| 1900 |
+
},
|
| 1901 |
+
"retrieval_residual_residual_no_op": {
|
| 1902 |
+
"count": 56.0,
|
| 1903 |
+
"mean_progress": 0.7249722566921264,
|
| 1904 |
+
"success_rate": 0.5
|
| 1905 |
+
},
|
| 1906 |
+
"retrieval_residual_residual_no_op+residual_wrong_gripper": {
|
| 1907 |
+
"count": 24.0,
|
| 1908 |
+
"mean_progress": 0.6711457151298722,
|
| 1909 |
+
"success_rate": 0.4583333333333333
|
| 1910 |
+
},
|
| 1911 |
+
"retrieval_residual_residual_wrong_gripper": {
|
| 1912 |
+
"count": 19.0,
|
| 1913 |
+
"mean_progress": 0.6364115562878156,
|
| 1914 |
+
"success_rate": 0.47368421052631576
|
| 1915 |
+
}
|
| 1916 |
+
},
|
| 1917 |
+
"source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_scales035040_summary.json",
|
| 1918 |
+
"std_success": 0.013130146843949137
|
| 1919 |
+
},
|
| 1920 |
"residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger002": {
|
| 1921 |
"ci95_success": 0.03115620450679224,
|
| 1922 |
"label": "K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.02",
|
results/paper_analysis.md
CHANGED
|
@@ -1,6 +1,6 @@
|
|
| 1 |
# Paper Analysis
|
| 2 |
|
| 3 |
-
Generated: `2026-06-
|
| 4 |
|
| 5 |
## Main Seed Statistics
|
| 6 |
|
|
@@ -56,6 +56,8 @@ Generated: `2026-06-30T08:40:51+00:00`
|
|
| 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_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 |
|
| 58 |
| 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 |
|
|
|
|
|
|
|
| 59 |
| 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 |
|
| 60 |
| 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 |
|
| 61 |
| residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmbonus001 | K4 composed compatible tangents, no-op bonus 0.03, singleton near-miss bonus 0.01 | 3 | 35.59% +/- 0.99 | +/- 2.46 | 57.10% | 0.406 | +5.86 pp |
|
|
|
|
| 1 |
# Paper Analysis
|
| 2 |
|
| 3 |
+
Generated: `2026-06-30T09:13:25+00:00`
|
| 4 |
|
| 5 |
## Main Seed Statistics
|
| 6 |
|
|
|
|
| 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_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 |
|
| 58 |
| 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 |
|
| 59 |
+
| 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 |
|
| 60 |
+
| 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 |
|
| 61 |
| 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 |
|
| 62 |
| 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 |
|
| 63 |
| residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmbonus001 | K4 composed compatible tangents, no-op bonus 0.03, singleton near-miss bonus 0.01 | 3 | 35.59% +/- 0.99 | +/- 2.46 | 57.10% | 0.406 | +5.86 pp |
|
results/paper_core_results.md
CHANGED
|
@@ -60,7 +60,7 @@ lattice is `+27.25 pp`, and the remaining clean-to-same-state proposal gap is
|
|
| 60 |
| K4 composed type-consensus residual retrieval, exact compatibility mask + no-op/near-miss singleton priors | No | No | 35.59% | +5.86 pp | Exact singleton near-miss priors 0.01/0.02 tie the top row and raise progress slightly; transferred near-miss tangents are high-precision but too sparse to close the proposal gap |
|
| 61 |
| K4 compatible residual retrieval, unique candidate-oracle prefix K=8 | No | No | 43.07% diagnostic | +13.33 pp diagnostic | Diagnostic-only measured oracle over generated clean candidate prefix; unique-action trace shows real selector headroom, not duplicate-action inflation |
|
| 62 |
| K4 compatible residual retrieval + near-miss challenger gate 0.01 | No | No | 36.06% | +6.32 pp | Current best clean deployment row; a two-stage selector keeps the robust compatible-chart anchor and only lets singleton near-miss tangents override under a tightly calibrated positive field margin |
|
| 63 |
-
| K4 compatible residual retrieval + near-miss challenger fine calibration | No | No | 36.00-36.00% | +6.26 pp | Margins 0.005/0.015/0.02 form a near-tie plateau around the 0.01 optimum; 0.03 and near-miss+wrong-gripper are lower or higher-MSE, so the effect is calibrated singleton near-miss geometry rather than broad residual mixing |
|
| 64 |
| K4 compatible residual retrieval, margin 0.10 | No | No | 34.67% | +4.93 pp | Naively lowering abstention selects too many bad nonzero tangents and falls below the 36.06% top row |
|
| 65 |
| K4 composed type-consensus residual retrieval, exact compatibility mask + source-score prior | No | No | 35.48% | +5.74 pp | Measured train-source reward confidence does not replace the sparse typed prior on the compatible chart; useful negative calibration |
|
| 66 |
| K4 composed type-consensus residual retrieval, exact compatibility mask + no-op/source-score priors | No | No | 35.54% | +5.80 pp | Adding measured train-source reward confidence to the current typed-prior chart ties near-best but does not beat exact compatibility masking alone |
|
|
|
|
| 60 |
| K4 composed type-consensus residual retrieval, exact compatibility mask + no-op/near-miss singleton priors | No | No | 35.59% | +5.86 pp | Exact singleton near-miss priors 0.01/0.02 tie the top row and raise progress slightly; transferred near-miss tangents are high-precision but too sparse to close the proposal gap |
|
| 61 |
| K4 compatible residual retrieval, unique candidate-oracle prefix K=8 | No | No | 43.07% diagnostic | +13.33 pp diagnostic | Diagnostic-only measured oracle over generated clean candidate prefix; unique-action trace shows real selector headroom, not duplicate-action inflation |
|
| 62 |
| K4 compatible residual retrieval + near-miss challenger gate 0.01 | No | No | 36.06% | +6.32 pp | Current best clean deployment row; a two-stage selector keeps the robust compatible-chart anchor and only lets singleton near-miss tangents override under a tightly calibrated positive field margin |
|
| 63 |
+
| K4 compatible residual retrieval + near-miss challenger fine calibration | No | No | 36.00-36.00% | +6.26 pp | Margins 0.005/0.015/0.02 form a near-tie plateau around the 0.01 optimum; scale-gating the challenger to 0.35 or 0.35/0.40 lowers MSE slightly but does not recover the lost success; 0.03 and near-miss+wrong-gripper are lower or higher-MSE, so the effect is calibrated singleton near-miss geometry rather than broad residual mixing |
|
| 64 |
| K4 compatible residual retrieval, margin 0.10 | No | No | 34.67% | +4.93 pp | Naively lowering abstention selects too many bad nonzero tangents and falls below the 36.06% top row |
|
| 65 |
| K4 composed type-consensus residual retrieval, exact compatibility mask + source-score prior | No | No | 35.48% | +5.74 pp | Measured train-source reward confidence does not replace the sparse typed prior on the compatible chart; useful negative calibration |
|
| 66 |
| K4 composed type-consensus residual retrieval, exact compatibility mask + no-op/source-score priors | No | No | 35.54% | +5.80 pp | Adding measured train-source reward confidence to the current typed-prior chart ties near-best but does not beat exact compatibility masking alone |
|
results/paper_story_memo.md
CHANGED
|
@@ -33,7 +33,7 @@ when queried on proposal geometry that matches those local counterfactuals.
|
|
| 33 |
| Train-source score priors do not explain the compatible tangent-chart gain | replacing the typed no-op prior with a measured train-source reward-score prior on the exact compatibility chart reaches 35.48%; adding it to the typed no-op row reaches 35.54%, still below the previous 35.59% compatibility row | Negative diagnostic: compatibility masking and sparse typed abstention matter more than continuous train-source confidence on this chart |
|
| 34 |
| Singleton near-miss priors do not open a new clean route | adding a small exact `residual_near_miss` singleton prior 0.01 or 0.02 on the compatible chart ties the previous 35.59% compatibility row and slightly raises progress to 57.10% | Tie diagnostic: revived near-miss tangents are high-precision but too sparse to close the clean-to-same-state proposal gap |
|
| 35 |
| Candidate-prefix oracle needs unique-action hygiene | the first K=8 prefix diagnostic was archived as `_nonunique`; the deduplicated unique-action trace reaches 43.07% candidate-oracle success with mean best branch rank 2.85 | Supported diagnostic: proposal headroom is real, but branch success falls with rank, so selector calibration must be conditional |
|
| 36 |
-
| Near-miss challenger calibration improves clean deployment | the trace-motivated two-stage selector keeps the compatible-chart anchor and lets singleton near-miss tangents override only under a tightly calibrated margin; margin 0.01 reaches 36.06%, while 0.005/0.015/0.02 tie at 36.00% | Current best clean row; small but story-aligned selector-calibration gain, with wrong-gripper
|
| 37 |
| Component-wise composite priors do not add the gain | propagating the no-op bonus into composite types reaches 35.36%, below the exact-prior masked composition row at 35.54% and the exact compatibility row at 35.59% | Negative/near-tie diagnostic: sparse exact priors are better than broadly rewarding every no-op-containing composite |
|
| 38 |
| Composite trust-radius penalty explains but does not improve the success top line | composite-only L2 penalty 0.02 ties 35.54% while lowering action MSE from 0.4106 to 0.4079; with the exact compatibility mask it lowers MSE further to 0.4048 but drops back to 35.54% | Tie/negative diagnostic: composed tangents need a local trust radius, but compatibility masking gives the success optimum |
|
| 39 |
| Minimum-energy residual regularization does not add the gain | action L2 penalty 0.05 ties the previous 35.42% scale-grid row, while 0.10/0.20 reach 35.36% | Negative/tie diagnostic: the clean bridge is not explained by shortest-action bias |
|
|
@@ -83,7 +83,7 @@ clean proposal result, the intended main rows are:
|
|
| 83 |
16. K4 mean-by-type tangent consensus + scale-grid typed no-op prior: 35.42%
|
| 84 |
17. K4 masked composed type-consensus tangent transport: 35.30%; with exact near-miss/no-op compatibility mask only: 35.48%; with typed no-op prior: 35.54%; with exact compatibility mask + typed no-op prior: 35.59%; adding singleton near-miss prior 0.01/0.02 ties 35.59%; with source-score prior on that exact chart: 35.48%; with typed no-op + source-score: 35.54%
|
| 85 |
18. K4 compatible tangent unique-action candidate-oracle prefix K=8 diagnostic: 43.07%; branch trace shows mean best branch rank 2.85, unique count 8.00, and real but conditional selector headroom
|
| 86 |
-
19. K4 compatible tangent near-miss challenger gate 0.01: 36.06%; current best clean deployment row from trace-motivated two-stage selector calibration. Fine margins 0.005/0.015/0.02 tie at 36.00%
|
| 87 |
20. K4 compatible tangent abstention margin 0.10 diagnostic: 34.67%; naive lower abstention hurts despite candidate-prefix oracle headroom
|
| 88 |
20. K4 masked composed type-consensus tangent transport + component-wise no-op prior: 35.36%
|
| 89 |
21. K4 masked composed type-consensus tangent transport + composite L2 trust penalty: 35.54% at 0.02; 35.48% at 0.05; 35.54% with exact compatibility mask + L2 0.02
|
|
@@ -172,7 +172,10 @@ completed at 36.00% for margin `0.02` (`14954280`/`14954281`), then improved to
|
|
| 172 |
36.06% for margin `0.01` (`14954526`/`14954527`), the current clean best. Fine
|
| 173 |
calibration jobs for margins `0.005` and `0.015` (`14955089`/`14955090` and
|
| 174 |
`14955093`/`14955094`) tie 36.00%, confirming a tight selector-calibration
|
| 175 |
-
plateau around the 0.01 optimum.
|
|
|
|
|
|
|
|
|
|
| 176 |
completed at 34.67%, so simple lower abstention is not the answer.
|
| 177 |
|
| 178 |
- `14858328`-`14858333`: completed train-split `field_selected_noexpert_bc5`;
|
|
|
|
| 33 |
| Train-source score priors do not explain the compatible tangent-chart gain | replacing the typed no-op prior with a measured train-source reward-score prior on the exact compatibility chart reaches 35.48%; adding it to the typed no-op row reaches 35.54%, still below the previous 35.59% compatibility row | Negative diagnostic: compatibility masking and sparse typed abstention matter more than continuous train-source confidence on this chart |
|
| 34 |
| Singleton near-miss priors do not open a new clean route | adding a small exact `residual_near_miss` singleton prior 0.01 or 0.02 on the compatible chart ties the previous 35.59% compatibility row and slightly raises progress to 57.10% | Tie diagnostic: revived near-miss tangents are high-precision but too sparse to close the clean-to-same-state proposal gap |
|
| 35 |
| Candidate-prefix oracle needs unique-action hygiene | the first K=8 prefix diagnostic was archived as `_nonunique`; the deduplicated unique-action trace reaches 43.07% candidate-oracle success with mean best branch rank 2.85 | Supported diagnostic: proposal headroom is real, but branch success falls with rank, so selector calibration must be conditional |
|
| 36 |
+
| Near-miss challenger calibration improves clean deployment | the trace-motivated two-stage selector keeps the compatible-chart anchor and lets singleton near-miss tangents override only under a tightly calibrated margin; margin 0.01 reaches 36.06%, while 0.005/0.015/0.02 and scale-gated 0.35 or 0.35/0.40 variants tie at 36.00% | Current best clean row; small but story-aligned selector-calibration gain, with scale-gating explaining MSE but not improving success and wrong-gripper challengers rejected by lower success or higher MSE |
|
| 37 |
| Component-wise composite priors do not add the gain | propagating the no-op bonus into composite types reaches 35.36%, below the exact-prior masked composition row at 35.54% and the exact compatibility row at 35.59% | Negative/near-tie diagnostic: sparse exact priors are better than broadly rewarding every no-op-containing composite |
|
| 38 |
| Composite trust-radius penalty explains but does not improve the success top line | composite-only L2 penalty 0.02 ties 35.54% while lowering action MSE from 0.4106 to 0.4079; with the exact compatibility mask it lowers MSE further to 0.4048 but drops back to 35.54% | Tie/negative diagnostic: composed tangents need a local trust radius, but compatibility masking gives the success optimum |
|
| 39 |
| Minimum-energy residual regularization does not add the gain | action L2 penalty 0.05 ties the previous 35.42% scale-grid row, while 0.10/0.20 reach 35.36% | Negative/tie diagnostic: the clean bridge is not explained by shortest-action bias |
|
|
|
|
| 83 |
16. K4 mean-by-type tangent consensus + scale-grid typed no-op prior: 35.42%
|
| 84 |
17. K4 masked composed type-consensus tangent transport: 35.30%; with exact near-miss/no-op compatibility mask only: 35.48%; with typed no-op prior: 35.54%; with exact compatibility mask + typed no-op prior: 35.59%; adding singleton near-miss prior 0.01/0.02 ties 35.59%; with source-score prior on that exact chart: 35.48%; with typed no-op + source-score: 35.54%
|
| 85 |
18. K4 compatible tangent unique-action candidate-oracle prefix K=8 diagnostic: 43.07%; branch trace shows mean best branch rank 2.85, unique count 8.00, and real but conditional selector headroom
|
| 86 |
+
19. K4 compatible tangent near-miss challenger gate 0.01: 36.06%; current best clean deployment row from trace-motivated two-stage selector calibration. Fine margins 0.005/0.015/0.02 tie at 36.00%; scale-gating the challenger to 0.35 or 0.35/0.40 also ties 36.00% with slightly lower MSE; 0.03 and near-miss+wrong-gripper variants are lower or higher-MSE.
|
| 87 |
20. K4 compatible tangent abstention margin 0.10 diagnostic: 34.67%; naive lower abstention hurts despite candidate-prefix oracle headroom
|
| 88 |
20. K4 masked composed type-consensus tangent transport + component-wise no-op prior: 35.36%
|
| 89 |
21. K4 masked composed type-consensus tangent transport + composite L2 trust penalty: 35.54% at 0.02; 35.48% at 0.05; 35.54% with exact compatibility mask + L2 0.02
|
|
|
|
| 172 |
36.06% for margin `0.01` (`14954526`/`14954527`), the current clean best. Fine
|
| 173 |
calibration jobs for margins `0.005` and `0.015` (`14955089`/`14955090` and
|
| 174 |
`14955093`/`14955094`) tie 36.00%, confirming a tight selector-calibration
|
| 175 |
+
plateau around the 0.01 optimum. Scale-gated challenger jobs at `0.35` and
|
| 176 |
+
`0.35/0.40` (`14955663`/`14955664` and `14955665`/`14955666`) also tie 36.00%
|
| 177 |
+
with slightly lower MSE, so tangent-length reliability explains action error but
|
| 178 |
+
does not replace the all-scale success optimum. The clean compatible-chart margin `0.10` test
|
| 179 |
completed at 34.67%, so simple lower abstention is not the answer.
|
| 180 |
|
| 181 |
- `14858328`-`14858333`: completed train-split `field_selected_noexpert_bc5`;
|
results/paper_table_status.json
CHANGED
|
@@ -1477,14 +1477,14 @@
|
|
| 1477 |
"story_role": "trace-motivated near-miss challenger with tangent-length reliability gating at the shortest scale",
|
| 1478 |
"fallback_success": null,
|
| 1479 |
"pending_job": "14955663/14955664",
|
| 1480 |
-
"path_exists":
|
| 1481 |
-
"status": "
|
| 1482 |
-
"success":
|
| 1483 |
-
"std_success":
|
| 1484 |
"completed_seeds": null,
|
| 1485 |
-
"num_completed":
|
| 1486 |
"best_config": null,
|
| 1487 |
-
"gain_vs_h16_policy":
|
| 1488 |
},
|
| 1489 |
{
|
| 1490 |
"key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_scales035040",
|
|
@@ -1496,14 +1496,14 @@
|
|
| 1496 |
"story_role": "trace-motivated near-miss challenger with tangent-length reliability gating over the two shortest scales",
|
| 1497 |
"fallback_success": null,
|
| 1498 |
"pending_job": "14955665/14955666",
|
| 1499 |
-
"path_exists":
|
| 1500 |
-
"status": "
|
| 1501 |
-
"success":
|
| 1502 |
-
"std_success":
|
| 1503 |
"completed_seeds": null,
|
| 1504 |
-
"num_completed":
|
| 1505 |
"best_config": null,
|
| 1506 |
-
"gain_vs_h16_policy":
|
| 1507 |
},
|
| 1508 |
{
|
| 1509 |
"key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger003",
|
|
|
|
| 1477 |
"story_role": "trace-motivated near-miss challenger with tangent-length reliability gating at the shortest scale",
|
| 1478 |
"fallback_success": null,
|
| 1479 |
"pending_job": "14955663/14955664",
|
| 1480 |
+
"path_exists": true,
|
| 1481 |
+
"status": "complete",
|
| 1482 |
+
"success": 0.36,
|
| 1483 |
+
"std_success": 0.011404240911829549,
|
| 1484 |
"completed_seeds": null,
|
| 1485 |
+
"num_completed": 3,
|
| 1486 |
"best_config": null,
|
| 1487 |
+
"gain_vs_h16_policy": 0.06260869565217392
|
| 1488 |
},
|
| 1489 |
{
|
| 1490 |
"key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_scales035040",
|
|
|
|
| 1496 |
"story_role": "trace-motivated near-miss challenger with tangent-length reliability gating over the two shortest scales",
|
| 1497 |
"fallback_success": null,
|
| 1498 |
"pending_job": "14955665/14955666",
|
| 1499 |
+
"path_exists": true,
|
| 1500 |
+
"status": "complete",
|
| 1501 |
+
"success": 0.36,
|
| 1502 |
+
"std_success": 0.013130146843949137,
|
| 1503 |
"completed_seeds": null,
|
| 1504 |
+
"num_completed": 3,
|
| 1505 |
"best_config": null,
|
| 1506 |
+
"gain_vs_h16_policy": 0.06260869565217392
|
| 1507 |
},
|
| 1508 |
{
|
| 1509 |
"key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger003",
|
results/paper_table_status.md
CHANGED
|
@@ -80,8 +80,8 @@ Baseline h=16 policy: 29.74%
|
|
| 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_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 |
|
| 82 |
| 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 |
|
| 83 |
-
| retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_scale035 | K4 compatible residual retrieval, near-miss challenger gate 0.01, scale-gated 0.35 |
|
| 84 |
-
| 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 |
|
| 85 |
| 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 |
|
| 86 |
| 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 |
|
| 87 |
| retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmbonus001 | K4 composed compatible residual retrieval, no-op bonus 0.03, singleton near-miss bonus 0.01 | complete | 35.59% | +5.86 pp | yes | no | no | revive high-precision singleton near-miss tangents without boosting toxic near-miss+no-op composites |
|
|
|
|
| 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_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 |
|
| 82 |
| 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 |
|
| 83 |
+
| 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 |
|
| 84 |
+
| 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 |
|
| 85 |
| 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 |
|
| 86 |
| 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 |
|
| 87 |
| retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmbonus001 | K4 composed compatible residual retrieval, no-op bonus 0.03, singleton near-miss bonus 0.01 | complete | 35.59% | +5.86 pp | yes | no | no | revive high-precision singleton near-miss tangents without boosting toxic near-miss+no-op composites |
|