Auto-sync: 2026-06-30 15:00:21 (part 3)

results/paper_analysis.json

@@ -1,6 +1,6 @@
 {
   "best_clean_key": "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001",
-  "generated_utc": "2026-06-30T18:49:56+00:00",
+  "generated_utc": "2026-06-30T19:13:18+00:00",
   "mechanism_gap": {
     "best_clean_vs_direct_same_ckpt": 0.0776811594202898,
     "best_clean_vs_h16": 0.06318840579710144,
@@ -1917,6 +1917,11 @@
       "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_scales035040_summary.json",
       "std_success": 0.013130146843949137
     },
+    "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_stacknowg": {
+      "label": "K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.01, no wrong-gripper component on Stack",
+      "missing": true,
+      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_stacknowg_summary.json"
+    },
     "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger002": {
       "ci95_success": 0.03115620450679224,
       "label": "K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.02",
@@ -2111,6 +2116,103 @@
       "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p03_summary.json",
       "std_success": 0.012819330079707819
     },
+    "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmglobal_wgpickpull_challenger001": {
+      "ci95_success": 0.03920403094295293,
+      "label": "K4 compatible tangents, no-op bonus 0.03, near-miss global + wrong-gripper challenger on Pick/Pull, gate 0.01",
+      "mean_action_mse_to_best": 0.41923150534080206,
+      "mean_progress": 0.574088785921579,
+      "mean_success": 0.36057971014492757,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.26340865087329823,
+          "std_success": 0.06057881385012102
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.3305568224046485,
+          "std_success": 0.04196214617114575
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.228271436166173,
+          "std_success": 0.03525140677829959
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.7604066042026859,
+          "std_success": 0.08151850231627764
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.19707329707329704,
+          "std_success": 0.0831467196651622
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.39629442195691494,
+        "1": 0.41279180422911177,
+        "2": 0.4486082898363795
+      },
+      "seed_progress": {
+        "0": 0.559952266274722,
+        "1": 0.5719372708739145,
+        "2": 0.5903768206161002
+      },
+      "seed_success": {
+        "0": 0.3582608695652174,
+        "1": 0.34608695652173915,
+        "2": 0.3773913043478261
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 1444,
+        "retrieval_residual_residual_near_miss": 91,
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": 5,
+        "retrieval_residual_residual_no_op": 38,
+        "retrieval_residual_residual_no_op+residual_wrong_gripper": 27,
+        "retrieval_residual_residual_wrong_gripper": 120
+      },
+      "selected_residual_scale_counts": {
+        "0.35": 1506,
+        "0.4": 36,
+        "0.45": 183
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 1444.0,
+          "mean_progress": 0.5575818072795194,
+          "success_rate": 0.3337950138504155
+        },
+        "retrieval_residual_residual_near_miss": {
+          "count": 91.0,
+          "mean_progress": 0.5661385127926594,
+          "success_rate": 0.38461538461538464
+        },
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": {
+          "count": 5.0,
+          "mean_progress": 0.7557853102684021,
+          "success_rate": 0.6
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 38.0,
+          "mean_progress": 0.7757954887653652,
+          "success_rate": 0.5263157894736842
+        },
+        "retrieval_residual_residual_no_op+residual_wrong_gripper": {
+          "count": 27.0,
+          "mean_progress": 0.6916872307106301,
+          "success_rate": 0.48148148148148145
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 120.0,
+          "mean_progress": 0.6808475915446053,
+          "success_rate": 0.575
+        }
+      },
+      "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",
+      "std_success": 0.015780472566743406
+    },
     "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgchallenger001": {
       "ci95_success": 0.028111505659259583,
       "label": "K4 compatible tangents, no-op bonus 0.03, near-miss/wrong-gripper challenger gate 0.01",
@@ -2209,9 +2311,101 @@
       "std_success": 0.011315490606343858
     },
     "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgchallenger001_pickpull": {
+      "ci95_success": 0.04121578446082103,
       "label": "K4 compatible tangents, no-op bonus 0.03, near-miss/wrong-gripper challenger gate 0.01 on Pick/Pull only",
-      "missing": true,
-      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmwgchallenger0p01_pickpull_summary.json"
+      "mean_action_mse_to_best": 0.41794628433532255,
+      "mean_progress": 0.5735963408033935,
+      "mean_success": 0.36000000000000004,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.26340865087329823,
+          "std_success": 0.06057881385012102
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.3305568224046485,
+          "std_success": 0.04196214617114575
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.228271436166173,
+          "std_success": 0.03525140677829959
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.757403601199683,
+          "std_success": 0.07781968814460603
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.19707329707329704,
+          "std_success": 0.0831467196651622
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.3939786813845453,
+        "1": 0.41151866878018434,
+        "2": 0.4483415028412381
+      },
+      "seed_progress": {
+        "0": 0.559967816375007,
+        "1": 0.5704443854190733,
+        "2": 0.5903768206161002
+      },
+      "seed_success": {
+        "0": 0.3582608695652174,
+        "1": 0.3443478260869565,
+        "2": 0.3773913043478261
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 1463,
+        "retrieval_residual_residual_near_miss": 72,
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": 5,
+        "retrieval_residual_residual_no_op": 38,
+        "retrieval_residual_residual_no_op+residual_wrong_gripper": 27,
+        "retrieval_residual_residual_wrong_gripper": 120
+      },
+      "selected_residual_scale_counts": {
+        "0.35": 1516,
+        "0.4": 35,
+        "0.45": 174
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 1463.0,
+          "mean_progress": 0.5617103827593661,
+          "success_rate": 0.3390293916609706
+        },
+        "retrieval_residual_residual_near_miss": {
+          "count": 72.0,
+          "mean_progress": 0.47270800791577333,
+          "success_rate": 0.2777777777777778
+        },
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": {
+          "count": 5.0,
+          "mean_progress": 0.7557853102684021,
+          "success_rate": 0.6
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 38.0,
+          "mean_progress": 0.7757954887653652,
+          "success_rate": 0.5263157894736842
+        },
+        "retrieval_residual_residual_no_op+residual_wrong_gripper": {
+          "count": 27.0,
+          "mean_progress": 0.6916872307106301,
+          "success_rate": 0.48148148148148145
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 120.0,
+          "mean_progress": 0.6808475915446053,
+          "success_rate": 0.575
+        }
+      },
+      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmwgchallenger0p01_pickpull_norm_summary.json",
+      "std_success": 0.01659024698116427
     },
     "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8": {
       "candidate_oracle_rollouts": 8,

results/paper_analysis.md

@@ -1,6 +1,6 @@
 # Paper Analysis
 
-Generated: `2026-06-30T18:49:56+00:00`
+Generated: `2026-06-30T19:13:18+00:00`
 
 ## Main Seed Statistics
 
@@ -54,13 +54,15 @@ Generated: `2026-06-30T18:49:56+00:00`
 | 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 |
 | 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 |
 | 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 |
+| 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 |
 | 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 |
 | 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 |
 | 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 |
 | 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 |
 | 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 |
 | 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 |
-| 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 | 0 | missing | missing | missing | missing | missing |
+| 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 |
+| 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 |
 | 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 |
 | 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 |
 | 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

@@ -1429,6 +1429,25 @@
       "best_config": null,
       "gain_vs_h16_policy": 0.06318840579710144
     },
+    {
+      "key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_stacknowg",
+      "label": "K4 compatible residual retrieval, near-miss challenger gate 0.01, no wrong-gripper component on Stack",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_stacknowg_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "task-local tangent compatibility mask for Stack where wrong-gripper residuals are an anti-affordance",
+      "fallback_success": null,
+      "pending_job": "14976136/14976144",
+      "path_exists": false,
+      "status": "pending",
+      "success": null,
+      "std_success": null,
+      "completed_seeds": null,
+      "num_completed": null,
+      "best_config": null,
+      "gain_vs_h16_policy": null
+    },
     {
       "key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger0005",
       "label": "K4 compatible residual retrieval, near-miss challenger gate 0.005",
@@ -1546,21 +1565,40 @@
     {
       "key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgchallenger001_pickpull",
       "label": "K4 compatible residual retrieval, near-miss/wrong-gripper challenger gate 0.01 on Pick/Pull only",
-      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmwgchallenger0p01_pickpull_summary.json",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmwgchallenger0p01_pickpull_norm_summary.json",
       "clean_deployment": "yes",
       "same_state_proposals": "no",
       "expert_proposal": "no",
       "story_role": "task-conditional challenger gate using branch trace evidence that wrong-gripper tangents help Pick/Pull but harm other tasks",
       "fallback_success": null,
-      "pending_job": "14974294/14974296",
-      "path_exists": false,
-      "status": "pending",
-      "success": null,
-      "std_success": null,
+      "pending_job": "14974602/14974608",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.36,
+      "std_success": 0.01659024698116427,
       "completed_seeds": null,
-      "num_completed": null,
+      "num_completed": 3,
       "best_config": null,
-      "gain_vs_h16_policy": null
+      "gain_vs_h16_policy": 0.06260869565217392
+    },
+    {
+      "key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmglobal_wgpickpull_challenger001",
+      "label": "K4 compatible residual retrieval, near-miss global + wrong-gripper challenger on Pick/Pull, gate 0.01",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmglobal_wgpickpull_challenger0p01_norm_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "type-specific affordance gate: keep near-miss global while restricting wrong-gripper challenger to acquisition/contact tasks",
+      "fallback_success": null,
+      "pending_job": "14974610/14974611",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.36057971014492757,
+      "std_success": 0.015780472566743406,
+      "completed_seeds": null,
+      "num_completed": 3,
+      "best_config": null,
+      "gain_vs_h16_policy": 0.0631884057971015
     },
     {
       "key": "typed_proposal_lattice_types6_prepend_margin000",
@@ -2951,23 +2989,23 @@
     }
   ],
   "best_clean": {
-    "key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001",
-    "label": "K4 compatible residual retrieval, near-miss challenger gate 0.01",
-    "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_summary.json",
+    "key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmglobal_wgpickpull_challenger001",
+    "label": "K4 compatible residual retrieval, near-miss global + wrong-gripper challenger on Pick/Pull, gate 0.01",
+    "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmglobal_wgpickpull_challenger0p01_norm_summary.json",
     "clean_deployment": "yes",
     "same_state_proposals": "no",
     "expert_proposal": "no",
-    "story_role": "lower-margin sensitivity for trace-motivated singleton near-miss challenger calibration",
+    "story_role": "type-specific affordance gate: keep near-miss global while restricting wrong-gripper challenger to acquisition/contact tasks",
     "fallback_success": null,
-    "pending_job": "14954526/14954527",
+    "pending_job": "14974610/14974611",
     "path_exists": true,
     "status": "complete",
-    "success": 0.3605797101449275,
-    "std_success": 0.01405722394548653,
+    "success": 0.36057971014492757,
+    "std_success": 0.015780472566743406,
     "completed_seeds": null,
     "num_completed": 3,
     "best_config": null,
-    "gain_vs_h16_policy": 0.06318840579710144
+    "gain_vs_h16_policy": 0.0631884057971015
   },
   "best_mechanism_no_expert": {
     "key": "no_expert_lattice",
@@ -2992,7 +3030,7 @@
     "Use no-expert same-state lattice as the conservative mechanism result, not as deployment-clean inference.",
     "Use full lattice only as an upper result because it includes expert proposals.",
     "Do not claim external SOTA from this table alone; add current external baselines separately.",
-    "Current best clean deployment row is K4 compatible residual retrieval, near-miss challenger gate 0.01 at 36.06%.",
+    "Current best clean deployment row is K4 compatible residual retrieval, near-miss global + wrong-gripper challenger on Pick/Pull, gate 0.01 at 36.06%.",
     "Trust-region field optimization should be framed as a negative/diagnostic ablation.",
     "Train-state counterfactual residual retrieval is a positive clean bridge but remains below the current clean best.",
     "KNN counterfactual residual retrieval is a positive clean bridge but remains below the current clean best.",

results/paper_table_status.md

@@ -78,13 +78,15 @@ Baseline h=16 policy: 29.74%
 | 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 |
 | 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 |
 | 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 |
+| 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 |
 | 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 |
 | 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 |
 | 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 |
 | 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 |
 | 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 |
 | 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 |
-| 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 | pending 14974294/14974296 | pending | pending | yes | no | no | task-conditional challenger gate using branch trace evidence that wrong-gripper tangents help Pick/Pull but harm other tasks |
+| 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 |
+| 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 |
 | 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 |
 | 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 |
 | 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 |
@@ -164,7 +166,7 @@ Baseline h=16 policy: 29.74%
 - Use no-expert same-state lattice as the conservative mechanism result, not as deployment-clean inference.
 - Use full lattice only as an upper result because it includes expert proposals.
 - Do not claim external SOTA from this table alone; add current external baselines separately.
-- Current best clean deployment row is K4 compatible residual retrieval, near-miss challenger gate 0.01 at 36.06%.
+- Current best clean deployment row is K4 compatible residual retrieval, near-miss global + wrong-gripper challenger on Pick/Pull, gate 0.01 at 36.06%.
 - Trust-region field optimization should be framed as a negative/diagnostic ablation.
 - Train-state counterfactual residual retrieval is a positive clean bridge but remains below the current clean best.
 - KNN counterfactual residual retrieval is a positive clean bridge but remains below the current clean best.

scripts/build_paper_analysis.py

@@ -409,6 +409,14 @@ METHODS = [
             "k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_summary.json"
         ),
     ),
+    MethodSpec(
+        key="residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_stacknowg",
+        label="K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.01, no wrong-gripper component on Stack",
+        summary_path=(
+            "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_"
+            "k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_stacknowg_summary.json"
+        ),
+    ),
     MethodSpec(
         key="residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger0005",
         label="K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.005",
@@ -462,7 +470,15 @@ METHODS = [
         label="K4 compatible tangents, no-op bonus 0.03, near-miss/wrong-gripper challenger gate 0.01 on Pick/Pull only",
         summary_path=(
             "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_"
-            "k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmwgchallenger0p01_pickpull_summary.json"
+            "k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmwgchallenger0p01_pickpull_norm_summary.json"
+        ),
+    ),
+    MethodSpec(
+        key="residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmglobal_wgpickpull_challenger001",
+        label="K4 compatible tangents, no-op bonus 0.03, near-miss global + wrong-gripper challenger on Pick/Pull, gate 0.01",
+        summary_path=(
+            "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_"
+            "k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmglobal_wgpickpull_challenger0p01_norm_summary.json"
         ),
     ),
     MethodSpec(

scripts/build_paper_table_status.py

@@ -765,6 +765,16 @@ SPECS = [
         story_role="lower-margin sensitivity for trace-motivated singleton near-miss challenger calibration",
         pending_job="14954526/14954527",
     ),
+    ResultSpec(
+        key="retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger001_stacknowg",
+        label="K4 compatible residual retrieval, near-miss challenger gate 0.01, no wrong-gripper component on Stack",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p01_stacknowg_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="task-local tangent compatibility mask for Stack where wrong-gripper residuals are an anti-affordance",
+        pending_job="14976136/14976144",
+    ),
     ResultSpec(
         key="retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger0005",
         label="K4 compatible residual retrieval, near-miss challenger gate 0.005",
@@ -828,12 +838,22 @@ SPECS = [
     ResultSpec(
         key="retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmwgchallenger001_pickpull",
         label="K4 compatible residual retrieval, near-miss/wrong-gripper challenger gate 0.01 on Pick/Pull only",
-        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmwgchallenger0p01_pickpull_summary.json",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmwgchallenger0p01_pickpull_norm_summary.json",
         clean_deployment="yes",
         same_state_proposals="no",
         expert_proposal="no",
         story_role="task-conditional challenger gate using branch trace evidence that wrong-gripper tangents help Pick/Pull but harm other tasks",
-        pending_job="14974294/14974296",
+        pending_job="14974602/14974608",
+    ),
+    ResultSpec(
+        key="retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmglobal_wgpickpull_challenger001",
+        label="K4 compatible residual retrieval, near-miss global + wrong-gripper challenger on Pick/Pull, gate 0.01",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmglobal_wgpickpull_challenger0p01_norm_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="type-specific affordance gate: keep near-miss global while restricting wrong-gripper challenger to acquisition/contact tasks",
+        pending_job="14974610/14974611",
     ),
     ResultSpec(
         key="typed_proposal_lattice_types6_prepend_margin000",

scripts/eval_maniskill_policy_rollout.py

@@ -267,12 +267,24 @@ def main(argv: list[str] | None = None) -> int:
         help="Optional comma-separated task IDs where challenger candidates may override "
         "the primary retrieval-residual selection. Empty enables challengers for all tasks.",
     )
+    parser.add_argument(
+        "--retrieval-residual-challenger-type-tasks",
+        default="",
+        help="Optional semicolon-separated candidate_type=TaskA|TaskB limits for individual "
+        "challenger types. Types without an entry keep the global challenger task behavior.",
+    )
     parser.add_argument(
         "--lattice-exclude-types",
         default="",
         help="Comma-separated candidate_type values to mask in selection-mode=lattice, "
         "for example 'expert'.",
     )
+    parser.add_argument(
+        "--lattice-exclude-type-tasks",
+        default="",
+        help="Optional semicolon-separated candidate_type=TaskA|TaskB exclusions applied "
+        "only on listed tasks, e.g. 'wrong_gripper=StackCube-v1'.",
+    )
     parser.add_argument(
         "--candidate-type-bonuses",
         default="",
@@ -305,6 +317,21 @@ def main(argv: list[str] | None = None) -> int:
     lattice_exclude_types = tuple(
         item.strip() for item in args.lattice_exclude_types.split(",") if item.strip()
     )
+    lattice_exclude_type_tasks: dict[str, tuple[str, ...]] = {}
+    try:
+        for item in args.lattice_exclude_type_tasks.split(";"):
+            item = item.strip()
+            if not item:
+                continue
+            key, value = item.split("=", 1)
+            lattice_exclude_type_tasks[key.strip()] = tuple(
+                task.strip() for task in value.split("|") if task.strip()
+            )
+    except ValueError as exc:
+        raise SystemExit(
+            "--lattice-exclude-type-tasks must contain semicolon-separated "
+            "candidate_type=TaskA|TaskB entries"
+        ) from exc
     retrieval_residual_challenger_types = tuple(
         item.strip()
         for item in args.retrieval_residual_challenger_types.split(",")
@@ -315,6 +342,21 @@ def main(argv: list[str] | None = None) -> int:
         for item in args.retrieval_residual_challenger_tasks.split(",")
         if item.strip()
     )
+    retrieval_residual_challenger_type_tasks: dict[str, tuple[str, ...]] = {}
+    try:
+        for item in args.retrieval_residual_challenger_type_tasks.split(";"):
+            item = item.strip()
+            if not item:
+                continue
+            key, value = item.split("=", 1)
+            retrieval_residual_challenger_type_tasks[key.strip()] = tuple(
+                task.strip() for task in value.split("|") if task.strip()
+            )
+    except ValueError as exc:
+        raise SystemExit(
+            "--retrieval-residual-challenger-type-tasks must contain "
+            "semicolon-separated candidate_type=TaskA|TaskB entries"
+        ) from exc
     proposal_lattice_types = tuple(
         item.strip()
         for item in args.proposal_lattice_types.split(",")
@@ -402,7 +444,11 @@ def main(argv: list[str] | None = None) -> int:
         retrieval_residual_challenger_scales=retrieval_residual_challenger_scales,
         retrieval_residual_challenger_margin=args.retrieval_residual_challenger_margin,
         retrieval_residual_challenger_tasks=retrieval_residual_challenger_tasks,
+        retrieval_residual_challenger_type_tasks=(
+            retrieval_residual_challenger_type_tasks
+        ),
         lattice_exclude_types=lattice_exclude_types,
+        lattice_exclude_type_tasks=lattice_exclude_type_tasks,
         candidate_type_bonuses=candidate_type_bonuses,
         candidate_type_bonus_components=args.candidate_type_bonus_components,
         candidate_oracle_rollouts=args.candidate_oracle_rollouts,

scripts/slurm/eval_maniskill_policy_rollout.sbatch

@@ -87,10 +87,12 @@ RETRIEVAL_RESIDUAL_CHALLENGER_TASKS="${RETRIEVAL_RESIDUAL_CHALLENGER_TASKS:-}"
 if [[ -n "${RETRIEVAL_RESIDUAL_CHALLENGER_TASKS_COLON:-}" ]]; then
   RETRIEVAL_RESIDUAL_CHALLENGER_TASKS="${RETRIEVAL_RESIDUAL_CHALLENGER_TASKS_COLON//:/,}"
 fi
+RETRIEVAL_RESIDUAL_CHALLENGER_TYPE_TASKS="${RETRIEVAL_RESIDUAL_CHALLENGER_TYPE_TASKS:-}"
 LATTICE_EXCLUDE_TYPES="${LATTICE_EXCLUDE_TYPES:-}"
 if [[ -n "${LATTICE_EXCLUDE_TYPES_COLON:-}" ]]; then
   LATTICE_EXCLUDE_TYPES="${LATTICE_EXCLUDE_TYPES_COLON//:/,}"
 fi
+LATTICE_EXCLUDE_TYPE_TASKS="${LATTICE_EXCLUDE_TYPE_TASKS:-}"
 CANDIDATE_TYPE_BONUSES="${CANDIDATE_TYPE_BONUSES:-}"
 if [[ -n "${CANDIDATE_TYPE_BONUSES_COLON:-}" ]]; then
   CANDIDATE_TYPE_BONUSES="${CANDIDATE_TYPE_BONUSES_COLON//:/,}"
@@ -170,7 +172,9 @@ apptainer exec --nv \
     --retrieval-residual-challenger-scales "$RETRIEVAL_RESIDUAL_CHALLENGER_SCALES" \
     --retrieval-residual-challenger-margin "$RETRIEVAL_RESIDUAL_CHALLENGER_MARGIN" \
     --retrieval-residual-challenger-tasks "$RETRIEVAL_RESIDUAL_CHALLENGER_TASKS" \
+    --retrieval-residual-challenger-type-tasks "$RETRIEVAL_RESIDUAL_CHALLENGER_TYPE_TASKS" \
     --lattice-exclude-types "$LATTICE_EXCLUDE_TYPES" \
+    --lattice-exclude-type-tasks "$LATTICE_EXCLUDE_TYPE_TASKS" \
     --candidate-type-bonuses "$CANDIDATE_TYPE_BONUSES" \
     --candidate-oracle-rollouts "$CANDIDATE_ORACLE_ROLLOUTS" \
     --candidate-oracle-unique-tolerance "$CANDIDATE_ORACLE_UNIQUE_TOLERANCE" \

scripts/slurm/summarize_h16_policy_ckpt.sbatch

@@ -115,7 +115,11 @@ for result_path in sorted(base_dir.glob(f"seed_*/{out_name}")):
             "retrieval_residual_challenger_tasks": data.get(
                 "retrieval_residual_challenger_tasks", []
             ),
+            "retrieval_residual_challenger_type_tasks": data.get(
+                "retrieval_residual_challenger_type_tasks", {}
+            ),
             "lattice_exclude_types": data.get("lattice_exclude_types", []),
+            "lattice_exclude_type_tasks": data.get("lattice_exclude_type_tasks", {}),
             "candidate_type_bonuses": data.get("candidate_type_bonuses", {}),
             "candidate_type_bonus_components": data.get(
                 "candidate_type_bonus_components", False
@@ -401,6 +405,26 @@ for row in rows:
         if challenger_tasks
         else ""
     )
+    challenger_type_tasks = row.get("retrieval_residual_challenger_type_tasks") or {}
+    challenger_type_task_suffix = (
+        " with "
+        + ";".join(
+            f"{candidate_type}=" + ",".join(tasks)
+            for candidate_type, tasks in sorted(challenger_type_tasks.items())
+        )
+        if challenger_type_tasks
+        else ""
+    )
+    exclude_type_tasks = row.get("lattice_exclude_type_tasks") or {}
+    exclude_type_task_suffix = (
+        " mask "
+        + ";".join(
+            f"{candidate_type}=" + ",".join(tasks)
+            for candidate_type, tasks in sorted(exclude_type_tasks.items())
+        )
+        if exclude_type_tasks
+        else ""
+    )
     lines.append(
         "| {seed} | {mode} | {k} | {policy_cand} | {retrieval} | {metric} | {anchor} | {direction} | {reduce}{challenger} | {min_success:.2f} | {type_success_bonus:.3f} | {consensus_penalty:.3f} | {min_source_progress:.2f} | {source_progress_bonus:.3f} | {source_score_bonus:.3f} | {scale:.2f} | {scale_grid} | {margin:.3f} | {sigma:.2f} | {steps} | {trust:.2f} | "
         "{success:.2%} | {progress:.2%} | {oracle:.2%} | {candidate_oracle} | {oracle_gain} | {mse:.3f} |".format(
@@ -413,7 +437,8 @@ for row in rows:
             anchor=row.get("retrieval_residual_anchor") or "none",
             direction=row.get("retrieval_residual_direction")
             or "candidate_minus_anchor",
-            reduce=row.get("retrieval_residual_reduce") or "none",
+            reduce=(row.get("retrieval_residual_reduce") or "none")
+            + exclude_type_task_suffix,
             challenger=(
                 " + challenger "
                 + ",".join(row.get("retrieval_residual_challenger_types") or [])
@@ -428,6 +453,7 @@ for row in rows:
                     else ""
                 )
                 + challenger_task_suffix
+                + challenger_type_task_suffix
                 + f"@{row.get('retrieval_residual_challenger_margin') or 0.0:.2f}"
                 if row.get("retrieval_residual_challenger_types")
                 else ""