Auto-sync: 2026-06-30 02:40:05 (part 4)

results/paper_analysis.json

@@ -1,11 +1,11 @@
 {
-  "best_clean_key": "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8",
-  "generated_utc": "2026-06-29T23:07:43+00:00",
+  "best_clean_key": "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003",
+  "generated_utc": "2026-06-30T06:41:21+00:00",
   "mechanism_gap": {
-    "best_clean_vs_direct_same_ckpt": 0.07362318840579707,
-    "best_clean_vs_h16": 0.05913043478260871,
+    "best_clean_vs_direct_same_ckpt": 0.07304347826086954,
+    "best_clean_vs_h16": 0.058550724637681184,
     "same_state_full_vs_no_expert": 0.12347826086956515,
-    "same_state_no_expert_vs_best_clean": 0.21333333333333343,
+    "same_state_no_expert_vs_best_clean": 0.21391304347826096,
     "same_state_no_expert_vs_h16": 0.27246376811594214
   },
   "methods": {
@@ -1433,114 +1433,9 @@
       "std_success": 0.009889114266221456
     },
     "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8": {
-      "candidate_oracle_rollouts": 8,
-      "candidate_oracle_type_counts": {
-        "retrieval_residual_policy_residual": 945,
-        "retrieval_residual_residual_near_miss": 211,
-        "retrieval_residual_residual_near_miss+residual_wrong_gripper": 76,
-        "retrieval_residual_residual_no_op": 180,
-        "retrieval_residual_residual_no_op+residual_wrong_gripper": 77,
-        "retrieval_residual_residual_wrong_gripper": 236
-      },
-      "ci95_success": 0.03024411731871433,
-      "label": "K4 compatible tangents, no-op bonus 0.03, candidate-oracle prefix K=8 diagnostic",
-      "mean_action_mse_to_best": 0.4063519326314006,
-      "mean_candidate_oracle_improvement_rate": 0.5344927536231884,
-      "mean_candidate_oracle_progress": 0.63640114126646,
-      "mean_candidate_oracle_score_gain_over_selected": 0.14484467573895835,
-      "mean_candidate_oracle_success_rate": 0.4220289855072464,
-      "mean_progress": 0.571575072284851,
-      "mean_success": 0.3565217391304348,
-      "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": 0.3278169093386485,
-          "std_success": 0.02533117351097172
-        },
-        "PullCube-v1": {
-          "mean_num_groups": 81.0,
-          "mean_success": 0.21716032505506191,
-          "std_success": 0.0182535211270701
-        },
-        "PushCube-v1": {
-          "mean_num_groups": 102.0,
-          "mean_success": 0.7541548289389031,
-          "std_success": 0.08244271787813129
-        },
-        "StackCube-v1": {
-          "mean_num_groups": 93.33333333333333,
-          "mean_success": 0.19707329707329704,
-          "std_success": 0.0831467196651622
-        }
-      },
-      "seed_action_mse_to_best": {
-        "0": 0.3930192190424904,
-        "1": 0.39824521353549286,
-        "2": 0.42779136531621864
-      },
-      "seed_progress": {
-        "0": 0.5579317773906919,
-        "1": 0.571580135051323,
-        "2": 0.5852133044125383
-      },
-      "seed_success": {
-        "0": 0.35130434782608694,
-        "1": 0.34782608695652173,
-        "2": 0.37043478260869567
-      },
-      "selected_candidate_type_counts": {
-        "retrieval_residual_policy_residual": 1613,
-        "retrieval_residual_residual_near_miss": 8,
-        "retrieval_residual_residual_near_miss+residual_wrong_gripper": 5,
-        "retrieval_residual_residual_no_op": 56,
-        "retrieval_residual_residual_no_op+residual_wrong_gripper": 24,
-        "retrieval_residual_residual_wrong_gripper": 19
-      },
-      "selected_residual_scale_counts": {
-        "0.35": 1631,
-        "0.4": 11,
-        "0.45": 83
-      },
-      "selected_type_outcomes": {
-        "retrieval_residual_policy_residual": {
-          "count": 1613.0,
-          "mean_progress": 0.561925975703087,
-          "success_rate": 0.34593924364538126
-        },
-        "retrieval_residual_residual_near_miss": {
-          "count": 8.0,
-          "mean_progress": 0.8754628524184227,
-          "success_rate": 0.75
-        },
-        "retrieval_residual_residual_near_miss+residual_wrong_gripper": {
-          "count": 5.0,
-          "mean_progress": 0.7557869911193847,
-          "success_rate": 0.6
-        },
-        "retrieval_residual_residual_no_op": {
-          "count": 56.0,
-          "mean_progress": 0.7249722566921264,
-          "success_rate": 0.5
-        },
-        "retrieval_residual_residual_no_op+residual_wrong_gripper": {
-          "count": 24.0,
-          "mean_progress": 0.6711457151298722,
-          "success_rate": 0.4583333333333333
-        },
-        "retrieval_residual_residual_wrong_gripper": {
-          "count": 19.0,
-          "mean_progress": 0.6364115562878156,
-          "success_rate": 0.47368421052631576
-        }
-      },
-      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_oraclek8_summary.json",
-      "std_success": 0.012173913043478276
+      "label": "K4 compatible tangents, no-op bonus 0.03, unique candidate-oracle prefix K=8 diagnostic",
+      "missing": true,
+      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_oraclek8_summary.json"
     },
     "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_srcscorebonus002": {
       "ci95_success": 0.02131295542962397,
@@ -7202,38 +7097,38 @@
   },
   "paired_deltas": {
     "best_clean - canonical_h16": {
-      "ci95_delta": 0.04510826498554671,
-      "left": "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8",
-      "mean_delta": 0.05913043478260869,
+      "ci95_delta": 0.04370704391798567,
+      "left": "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003",
+      "mean_delta": 0.05855072463768116,
       "right": "h16_policy_canonical",
       "seed_deltas": {
-        "0": 0.0678260869565217,
+        "0": 0.06956521739130433,
         "1": 0.038260869565217404,
-        "2": 0.07130434782608697
+        "2": 0.06782608695652176
       },
       "seeds": [
         0,
         1,
         2
       ],
-      "std_delta": 0.018157054798105293
+      "std_delta": 0.01759303293390552
     },
     "best_clean - direct_same_ckpt": {
-      "ci95_delta": 0.045383317930687055,
-      "left": "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8",
-      "mean_delta": 0.07362318840579708,
+      "ci95_delta": 0.04121578446082111,
+      "left": "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003",
+      "mean_delta": 0.07304347826086956,
       "right": "near_miss_policy_bc5",
       "seed_deltas": {
-        "0": 0.07304347826086954,
+        "0": 0.07478260869565218,
         "1": 0.055652173913043446,
-        "2": 0.09217391304347827
+        "2": 0.08869565217391306
       },
       "seeds": [
         0,
         1,
         2
       ],
-      "std_delta": 0.018267769572856522
+      "std_delta": 0.0165902469811643
     },
     "full_lattice - no_expert_lattice": {
       "ci95_delta": 0.02628111194117426,
@@ -7291,16 +7186,16 @@
     "best_clean_vs_h16": {
       "LiftPegUpright-v1": 0.03278131303915899,
       "PickCube-v1": 0.1219595621769535,
-      "PullCube-v1": 0.019305840358471965,
-      "PushCube-v1": 0.003597657063003523,
-      "StackCube-v1": 0.05610685610685609
+      "PullCube-v1": 0.006887167413483192,
+      "PushCube-v1": 0.006846429323783432,
+      "StackCube-v1": 0.0594738594738595
     },
     "no_expert_vs_best_clean": {
       "LiftPegUpright-v1": 0.35374851747103375,
       "PickCube-v1": 0.29366195398804096,
-      "PullCube-v1": -0.018680539733171325,
-      "PushCube-v1": 0.04688815632569587,
-      "StackCube-v1": 0.2854441854441855
+      "PullCube-v1": -0.006261866788182552,
+      "PushCube-v1": 0.043639384064915965,
+      "StackCube-v1": 0.282077182077182
     }
   }
 }

results/paper_analysis.md

@@ -1,6 +1,6 @@
 # Paper Analysis
 
-Generated: `2026-06-29T23:07:43+00:00`
+Generated: `2026-06-30T06:41:21+00:00`
 
 ## Main Seed Statistics
 
@@ -50,7 +50,7 @@ Generated: `2026-06-29T23:07:43+00:00`
 | residual_k4_composemasked_dropnmnoop_grid035040045 | K4 composed type-consensus tangents, masked, drop near-miss+no-op composite | 3 | 35.48% +/- 1.25 | +/- 3.12 | 57.00% | 0.406 | +5.74 pp |
 | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003 | K4 composed type-consensus tangents, masked, drop near-miss+no-op composite | 3 | 35.59% +/- 0.99 | +/- 2.46 | 57.07% | 0.406 | +5.86 pp |
 | residual_k4_composemasked_dropnmnoop_grid035040045_margin010_noopbonus003 | K4 compatible tangents, margin 0.10, no-op bonus 0.03 | 3 | 34.67% +/- 0.96 | +/- 2.38 | 56.45% | 0.418 | +4.93 pp |
-| residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8 | K4 compatible tangents, no-op bonus 0.03, candidate-oracle prefix K=8 diagnostic | 3 | 35.65% +/- 1.22 | +/- 3.02 | 57.16% | 0.406 | +5.91 pp |
+| residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8 | K4 compatible tangents, no-op bonus 0.03, unique candidate-oracle prefix K=8 diagnostic | 0 | missing | missing | missing | missing | missing |
 | 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 |
 | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmbonus002 | K4 composed compatible tangents, no-op bonus 0.03, singleton near-miss bonus 0.02 | 3 | 35.59% +/- 0.99 | +/- 2.46 | 57.10% | 0.406 | +5.86 pp |
 | residual_k4_composemasked_dropnmnoop_grid035040045_srcscorebonus002 | K4 composed type-consensus tangents, masked, drop near-miss+no-op composite, source-score bonus 0.02 | 3 | 35.48% +/- 1.22 | +/- 3.02 | 57.02% | 0.408 | +5.74 pp |
@@ -98,8 +98,8 @@ Generated: `2026-06-29T23:07:43+00:00`
 
 | comparison | seeds | mean delta | 95% CI | seed deltas |
 |---|---:|---:|---:|---|
-| best_clean - canonical_h16 | 3 | +5.91 pp | +/- 4.51 | 0:+6.78, 1:+3.83, 2:+7.13 |
-| best_clean - direct_same_ckpt | 3 | +7.36 pp | +/- 4.54 | 0:+7.30, 1:+5.57, 2:+9.22 |
+| best_clean - canonical_h16 | 3 | +5.86 pp | +/- 4.37 | 0:+6.96, 1:+3.83, 2:+6.78 |
+| best_clean - direct_same_ckpt | 3 | +7.30 pp | +/- 4.12 | 0:+7.48, 1:+5.57, 2:+8.87 |
 | no_expert_lattice - canonical_h16 | 3 | +27.25 pp | +/- 8.58 | 0:+23.30, 1:+28.70, 2:+29.74 |
 | full_lattice - no_expert_lattice | 3 | +12.35 pp | +/- 2.63 | 0:+13.57, 1:+11.83, 2:+11.65 |
 | policy_candidate_lattice - no_expert_lattice | 3 | -16.29 pp | +/- 7.55 | 0:-15.48, 1:-13.74, 2:-19.65 |
@@ -110,31 +110,25 @@ Generated: `2026-06-29T23:07:43+00:00`
 |---|---:|---:|---:|---:|---:|---:|---:|
 | LiftPegUpright-v1 | 23.06% | 26.34% | 62.90% | 61.72% | 73.19% | +3.28 pp | +35.37 pp |
 | PickCube-v1 | 20.59% | 32.78% | 58.13% | 62.15% | 84.19% | +12.20 pp | +29.37 pp |
-| PullCube-v1 | 19.79% | 21.72% | 15.02% | 19.85% | 22.41% | +1.93 pp | -1.87 pp |
-| PushCube-v1 | 75.06% | 75.42% | 82.54% | 80.10% | 81.92% | +0.36 pp | +4.69 pp |
-| StackCube-v1 | 14.10% | 19.71% | 50.41% | 48.25% | 60.83% | +5.61 pp | +28.54 pp |
+| PullCube-v1 | 19.79% | 20.47% | 15.02% | 19.85% | 22.41% | +0.69 pp | -0.63 pp |
+| PushCube-v1 | 75.06% | 75.74% | 82.54% | 80.10% | 81.92% | +0.68 pp | +4.36 pp |
+| StackCube-v1 | 14.10% | 20.04% | 50.41% | 48.25% | 60.83% | +5.95 pp | +28.21 pp |
 
 ## Mechanism Gap
 
-- Best clean residual transport improves over canonical h16 by +5.91 pp.
+- Best clean residual transport improves over canonical h16 by +5.86 pp.
 - Same-state no-expert lattice improves over canonical h16 by +27.25 pp.
-- Remaining clean-to-same-state proposal gap is +21.33 pp.
+- Remaining clean-to-same-state proposal gap is +21.39 pp.
 - Full lattice adds expert proposals and reaches 69.33%, a +12.35 pp gain over no-expert.
 
-## Candidate-Oracle Diagnostic
-
-- Oracle over the deployed candidate prefix reaches 42.20% with mean progress 63.64%; this is diagnostic-only because it uses measured rollout outcomes after generating candidates.
-- Mean oracle-prefix score gain over the selected branch is +0.145, which isolates ranking/abstention headroom inside the clean proposal set.
-- Candidate-oracle best type counts: {'retrieval_residual_policy_residual': 945, 'retrieval_residual_residual_near_miss': 211, 'retrieval_residual_residual_near_miss+residual_wrong_gripper': 76, 'retrieval_residual_residual_no_op': 180, 'retrieval_residual_residual_no_op+residual_wrong_gripper': 77, 'retrieval_residual_residual_wrong_gripper': 236}.
-
 ## Selection Histograms
 
 - `same_state_near_miss`: lattice_near_miss=1725 (100.0%)
 - `same_state_no_expert`: lattice_near_miss=1263 (73.2%), lattice_no_op=222 (12.9%), lattice_random_negative=144 (8.3%), lattice_wrong_gripper=62 (3.6%), lattice_wrong_direction=34 (2.0%)
 - `same_state_policy_baseline`: policy_continuous=1022 (59.2%), lattice_near_miss=448 (26.0%), lattice_no_op=119 (6.9%), lattice_random_negative=75 (4.3%), lattice_wrong_gripper=45 (2.6%), lattice_wrong_direction=16 (0.9%)
 - `same_state_full`: lattice_expert=977 (56.6%), lattice_near_miss=348 (20.2%), lattice_no_op=177 (10.3%), lattice_random_negative=138 (8.0%), lattice_wrong_gripper=55 (3.2%), lattice_wrong_direction=30 (1.7%)
-- `residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8`: retrieval_residual_policy_residual=1613 (93.5%), retrieval_residual_residual_no_op=56 (3.2%), retrieval_residual_residual_no_op+residual_wrong_gripper=24 (1.4%), retrieval_residual_residual_wrong_gripper=19 (1.1%), retrieval_residual_residual_near_miss=8 (0.5%), retrieval_residual_residual_near_miss+residual_wrong_gripper=5 (0.3%)
-- `residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8` residual scale counts: {'0.35': 1631, '0.4': 11, '0.45': 83}
+- `residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003`: retrieval_residual_policy_residual=1613 (93.5%), retrieval_residual_residual_no_op=56 (3.2%), retrieval_residual_residual_no_op+residual_wrong_gripper=24 (1.4%), retrieval_residual_residual_wrong_gripper=19 (1.1%), retrieval_residual_residual_near_miss=8 (0.5%), retrieval_residual_residual_near_miss+residual_wrong_gripper=5 (0.3%)
+- `residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003` residual scale counts: {'0.35': 1631, '0.4': 11, '0.45': 83}
 
 ## Selected-Type Outcomes
 

results/paper_core_results.md

@@ -58,7 +58,8 @@ lattice is `+27.25 pp`, and the remaining clean-to-same-state proposal gap is
 | K4 composed type-consensus residual retrieval, exact compatibility mask without typed prior | No | No | 35.48% | +5.74 pp | Isolates the algebraic compatibility mask: it improves pure masked composition but does not fully replace the sparse typed no-op prior |
 | K4 composed type-consensus residual retrieval, no-op prior + exact near-miss/no-op incompatibility mask | No | No | 35.59% | +5.86 pp | Current best clean diagnostic; dropping only the weak `near_miss+no_op` composite preserves useful singleton/composite tangents and lowers action MSE |
 | 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 |
-| K4 compatible residual retrieval, candidate-oracle prefix K=8 | No | No | 41.04% seed-0 diagnostic | +11.30 pp seed-0 | Diagnostic-only measured oracle over generated clean candidate prefix; seed 0 shows ranking/abstention headroom inside the proposal set, remaining seeds pending |
+| K4 compatible residual retrieval, unique candidate-oracle prefix K=8 | No | No | pending | pending | Diagnostic-only measured oracle over generated clean candidate prefix; first non-unique run was archived because duplicate policy branches inflated headroom |
+| 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 35.59% top row |
 | 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 |
 | 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 |
 | K4 composed type-consensus residual retrieval, masked + component-wise no-op prior | No | No | 35.36% | +5.62 pp | Propagating the no-op prior into every composite over-selects composed tangents and drops below the exact-prior row; useful negative calibration |

results/paper_story_memo.md

@@ -32,7 +32,7 @@ when queried on proposal geometry that matches those local counterfactuals.
 | Masked tangent composition gives a small clean lift | K4 composed type-consensus transport is 35.30% after anti-goal composite masking; exact `near_miss+no_op` masking alone reaches 35.48%; typed no-op prior reaches 35.54%; combining the exact compatibility mask with the typed prior reaches 35.59% | Current best clean result; controlled compatible tangent-chart composition, not unmasked proposal accumulation |
 | 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 35.59% top row | Negative diagnostic: compatibility masking and sparse typed abstention matter more than continuous train-source confidence on this chart |
 | 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 35.59% top 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 |
-| Candidate-prefix oracle exposes ranking/abstention headroom | seed 0 of the K=8 prefix diagnostic reaches 41.04% candidate-oracle success vs 35.13% selected rollout, with +0.139 mean score gain and 53.9% states improved by some generated prefix candidate | Partial diagnostic: the compatible proposal prefix contains useful actions; field ranking/margin, not only proposal existence, is now a measurable bottleneck |
+| Candidate-prefix oracle needs unique-action hygiene | the first K=8 prefix diagnostic exposed apparent headroom, but raw rows showed duplicate zero-residual/policy branches across scale slots; summaries were archived as `_nonunique` and a unique-action rerun is pending | In progress: use the deduplicated diagnostic before making ranking/abstention claims |
 | 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 |
 | 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 |
 | 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 |
@@ -81,8 +81,8 @@ clean proposal result, the intended main rows are:
 15. K4 mean-by-type tangent consensus + typed no-op prior 0.025-0.035: 35.25%
 16. K4 mean-by-type tangent consensus + scale-grid typed no-op prior: 35.42%
 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%
-18. K4 compatible tangent candidate-oracle prefix K=8 diagnostic: seed 0 candidate-oracle reaches 41.04% vs 35.13% selected; remaining seeds pending
-19. K4 compatible tangent abstention margin 0.10 diagnostic: submitted after oracle seed 0 exposed prefix headroom; pending
+18. K4 compatible tangent unique-action candidate-oracle prefix K=8 diagnostic: pending rerun after archiving the non-unique oracle summary
+19. K4 compatible tangent abstention margin 0.10 diagnostic: 34.67%; naive lower abstention hurts despite apparent non-unique oracle headroom
 20. K4 masked composed type-consensus tangent transport + component-wise no-op prior: 35.36%
 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
 22. K4 mean-by-type tangent consensus + upper/wide scale diagnostics: 35.36% for 0.40/0.45/0.50; 35.13% for 0.35/0.45/0.55
@@ -139,7 +139,7 @@ test-time search. The cleaner novelty is:
 
 ## Job Status
 
-Last checked: `2026-06-29 22:34 UTC`. The K4 masked composed type-consensus
+Last checked: `2026-06-30 06:34 UTC`. The K4 masked composed type-consensus
 sweep completed and produced a new clean best, 35.59%, when the exact
 `residual_near_miss+residual_no_op` composite is additionally masked. Pure
 masked composition reached 35.30%, exact compatibility masking without typed
@@ -158,13 +158,15 @@ below the top row.
 The singleton near-miss revival diagnostic completed: exact `residual_near_miss`
 bonuses 0.01 and 0.02 both tie the 35.59% top row and slightly raise mean
 progress to 57.10%, but do not improve success.
-The candidate-oracle prefix diagnostic is in flight: seed 0 of job `14935471`
-completed on the correct h16 dataset and shows K=8 candidate-oracle success
-41.04% vs 35.13% selected rollout, mean progress 62.48%, score gain +0.139,
-and 53.9% of states improved by some generated prefix candidate. Seeds 1-2
-remain pending; summary `14935502` and rebuild `14935530` are dependency-gated.
-The oracle-motivated clean abstention test at margin 0.10 is also pending as
-job `14936131`, with summary `14936132` and rebuild `14936133`.
+The original candidate-oracle prefix diagnostic (`14935471`) completed but is
+now treated as non-unique because the prefix contained duplicate policy/zero
+residual actions across residual-scale slots; its summaries were archived with
+the `_nonunique` suffix rather than used as current evidence. The evaluator now
+deduplicates candidate actions before oracle selection and blocks invalid padding
+branches from winning. The unique-action rerun is pending as GPU array
+`14953513`, with summary `14953522` and rebuild `14953524`. The clean
+compatible-chart margin `0.10` test completed at 34.67%, below the 35.59% top
+row, so simple lower abstention is not the answer.
 
 - `14858328`-`14858333`: completed train-split `field_selected_noexpert_bc5`;
   direct rollout is 26.84%, field-guided best is 27.65%.

results/paper_table_status.json

@@ -1355,22 +1355,22 @@
     },
     {
       "key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8",
-      "label": "K4 compatible residual retrieval, candidate-oracle prefix K=8 diagnostic",
+      "label": "K4 compatible residual retrieval, unique candidate-oracle prefix K=8 diagnostic",
       "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_oraclek8_summary.json",
       "clean_deployment": "diagnostic",
       "same_state_proposals": "no",
       "expert_proposal": "no",
       "story_role": "diagnostic oracle over generated clean candidate prefix to separate proposal headroom from field ranking",
       "fallback_success": null,
-      "pending_job": "14935471/14935502",
-      "path_exists": true,
-      "status": "complete",
-      "success": 0.4220289855072464,
-      "std_success": 0.012173913043478276,
+      "pending_job": "14953513/14953522",
+      "path_exists": false,
+      "status": "pending",
+      "success": null,
+      "std_success": null,
       "completed_seeds": null,
-      "num_completed": 3,
+      "num_completed": null,
       "best_config": null,
-      "gain_vs_h16_policy": 0.12463768115942031
+      "gain_vs_h16_policy": null
     },
     {
       "key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmbonus001",
@@ -2618,7 +2618,6 @@
     "KNN counterfactual residual retrieval is a positive clean bridge but remains below the current clean best.",
     "K4 train-state residual retrieval, scale 0.40, safe residuals, mean-by-type tangent consensus improves the clean bridge but is not yet the main result.",
     "Policy-baseline abstention inside the same-state no-expert lattice drops from 56.99% to 40.70%, so the mechanism result should emphasize counterfactual proposal geometry rather than policy fallback.",
-    "Singleton near-miss priors 0.01/0.02 tie the clean best rather than improving it; the compatible chart preserves high-precision near-miss tangents, but they remain sparse.",
-    "Candidate-oracle prefix K=8 is diagnostic-only; compare it to the clean selected row to decide whether the next paper move should be proposal generation or field ranking."
+    "Singleton near-miss priors 0.01/0.02 tie the clean best rather than improving it; the compatible chart preserves high-precision near-miss tangents, but they remain sparse."
   ]
 }

results/paper_table_status.md

@@ -74,7 +74,7 @@ Baseline h=16 policy: 29.74%
 | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045 | K4 composed type-consensus residual retrieval, masked, drop near-miss+no-op composite | complete | 35.48% | +5.74 pp | yes | no | no | exact compatibility mask without typed no-op prior |
 | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, masked, drop near-miss+no-op composite | complete | 35.59% | +5.86 pp | yes | no | no | exact incompatibility mask for one weak composed tangent pair |
 | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_margin010_noopbonus003 | K4 compatible residual retrieval, margin 0.10, no-op bonus 0.03 | complete | 34.67% | +4.93 pp | yes | no | no | oracle-motivated abstention sensitivity on the compatible local tangent chart |
-| retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8 | K4 compatible residual retrieval, candidate-oracle prefix K=8 diagnostic | complete | 42.20% | +12.46 pp | diagnostic | no | no | diagnostic oracle over generated clean candidate prefix to separate proposal headroom from field ranking |
+| retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8 | K4 compatible residual retrieval, unique candidate-oracle prefix K=8 diagnostic | pending 14953513/14953522 | pending | pending | diagnostic | no | no | diagnostic oracle over generated clean candidate prefix to separate proposal headroom from field ranking |
 | 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 |
 | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmbonus002 | K4 composed compatible residual retrieval, no-op bonus 0.03, singleton near-miss bonus 0.02 | complete | 35.59% | +5.86 pp | yes | no | no | stronger singleton near-miss revival prior on the compatible local tangent chart |
 | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_srcscorebonus002 | K4 composed type-consensus residual retrieval, masked, drop near-miss+no-op composite, source-score bonus 0.02 | complete | 35.48% | +5.74 pp | yes | no | no | train-measured source-score prior on the compatible local tangent chart |
@@ -151,4 +151,3 @@ Baseline h=16 policy: 29.74%
 - K4 train-state residual retrieval, scale 0.40, safe residuals, mean-by-type tangent consensus improves the clean bridge but is not yet the main result.
 - Policy-baseline abstention inside the same-state no-expert lattice drops from 56.99% to 40.70%, so the mechanism result should emphasize counterfactual proposal geometry rather than policy fallback.
 - Singleton near-miss priors 0.01/0.02 tie the clean best rather than improving it; the compatible chart preserves high-precision near-miss tangents, but they remain sparse.
-- Candidate-oracle prefix K=8 is diagnostic-only; compare it to the clean selected row to decide whether the next paper move should be proposal generation or field ranking.

scripts/build_paper_analysis.py

@@ -379,7 +379,7 @@ METHODS = [
     ),
     MethodSpec(
         key="residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8",
-        label="K4 compatible tangents, no-op bonus 0.03, candidate-oracle prefix K=8 diagnostic",
+        label="K4 compatible tangents, no-op bonus 0.03, unique candidate-oracle prefix K=8 diagnostic",
         summary_path=(
             "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_"
             "k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_oraclek8_summary.json"
@@ -846,6 +846,9 @@ def _normalize_summary(data: dict[str, Any], rows: list[dict[str, Any]], *, sour
         output.update(
             {
                 "candidate_oracle_rollouts": int(data.get("candidate_oracle_rollouts") or 0),
+                "candidate_oracle_unique_tolerance": data.get(
+                    "candidate_oracle_unique_tolerance"
+                ),
                 "mean_candidate_oracle_success_rate": _mean(candidate_oracle_success),
                 "mean_candidate_oracle_progress": _mean(
                     [
@@ -862,6 +865,13 @@ def _normalize_summary(data: dict[str, Any], rows: list[dict[str, Any]], *, sour
                         is not None
                     ]
                 ),
+                "mean_candidate_oracle_unique_count": _mean(
+                    [
+                        float(row.get("candidate_oracle_unique_count") or 0.0)
+                        for row in rows
+                        if row.get("candidate_oracle_unique_count") is not None
+                    ]
+                ),
                 "mean_candidate_oracle_improvement_rate": _mean(
                     [
                         float(row["candidate_oracle_improvement_rate"])
@@ -1111,6 +1121,10 @@ def _render_markdown(report: dict[str, Any]) -> str:
                     f"{oracle.get('mean_candidate_oracle_score_gain_over_selected', 0.0):+.3f}, "
                     "which isolates ranking/abstention headroom inside the clean proposal set."
                 ),
+                (
+                    "- Mean unique candidates in the prefix: "
+                    f"{oracle.get('mean_candidate_oracle_unique_count', 0.0):.2f}."
+                ),
                 (
                     "- Candidate-oracle best type counts: "
                     f"{oracle.get('candidate_oracle_type_counts', {})}."

scripts/build_paper_table_status.py

@@ -727,13 +727,13 @@ SPECS = [
     ),
     ResultSpec(
         key="retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8",
-        label="K4 compatible residual retrieval, candidate-oracle prefix K=8 diagnostic",
+        label="K4 compatible residual retrieval, unique candidate-oracle prefix K=8 diagnostic",
         path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_oraclek8_summary.json",
         clean_deployment="diagnostic",
         same_state_proposals="no",
         expert_proposal="no",
         story_role="diagnostic oracle over generated clean candidate prefix to separate proposal headroom from field ranking",
-        pending_job="14935471/14935502",
+        pending_job="14953513/14953522",
     ),
     ResultSpec(
         key="retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmbonus001",

scripts/eval_maniskill_policy_rollout.py

@@ -260,6 +260,14 @@ def main(argv: list[str] | None = None) -> int:
         "residual candidates and report the best measured outcome in that prefix. "
         "Zero disables the diagnostic and preserves deployment evaluation.",
     )
+    parser.add_argument(
+        "--candidate-oracle-unique-tolerance",
+        type=float,
+        default=1.0e-6,
+        help="Max absolute action difference for treating two diagnostic oracle candidates "
+        "as duplicates. Duplicate candidates are rolled only as invalid padding and cannot "
+        "win the oracle choice.",
+    )
     args = parser.parse_args(argv)
     lattice_exclude_types = tuple(
         item.strip() for item in args.lattice_exclude_types.split(",") if item.strip()
@@ -335,6 +343,7 @@ def main(argv: list[str] | None = None) -> int:
         candidate_type_bonuses=candidate_type_bonuses,
         candidate_type_bonus_components=args.candidate_type_bonus_components,
         candidate_oracle_rollouts=args.candidate_oracle_rollouts,
+        candidate_oracle_unique_tolerance=args.candidate_oracle_unique_tolerance,
     )
     print(json.dumps({key: value for key, value in result.items() if key != "rows"}, indent=2))
     return 0

scripts/slurm/eval_maniskill_policy_rollout.sbatch

@@ -80,6 +80,7 @@ if [[ -n "${CANDIDATE_TYPE_BONUSES_COLON:-}" ]]; then
 fi
 CANDIDATE_TYPE_BONUS_COMPONENTS="${CANDIDATE_TYPE_BONUS_COMPONENTS:-0}"
 CANDIDATE_ORACLE_ROLLOUTS="${CANDIDATE_ORACLE_ROLLOUTS:-0}"
+CANDIDATE_ORACLE_UNIQUE_TOLERANCE="${CANDIDATE_ORACLE_UNIQUE_TOLERANCE:-1e-6}"
 
 module load StdEnv/2023 apptainer/1.4.5
 cd "$PROJECT_DIR"
@@ -150,4 +151,5 @@ apptainer exec --nv \
     --lattice-exclude-types "$LATTICE_EXCLUDE_TYPES" \
     --candidate-type-bonuses "$CANDIDATE_TYPE_BONUSES" \
     --candidate-oracle-rollouts "$CANDIDATE_ORACLE_ROLLOUTS" \
+    --candidate-oracle-unique-tolerance "$CANDIDATE_ORACLE_UNIQUE_TOLERANCE" \
     "${EXTRA_ARGS[@]}"

scripts/slurm/summarize_h16_policy_ckpt.sbatch

@@ -107,6 +107,9 @@ for result_path in sorted(base_dir.glob(f"seed_*/{out_name}")):
                 "candidate_type_bonus_components", False
             ),
             "candidate_oracle_rollouts": data.get("candidate_oracle_rollouts", 0),
+            "candidate_oracle_unique_tolerance": data.get(
+                "candidate_oracle_unique_tolerance"
+            ),
             "candidate_oracle_success_rate": data.get(
                 "candidate_oracle_success_rate"
             ),
@@ -116,6 +119,7 @@ for result_path in sorted(base_dir.glob(f"seed_*/{out_name}")):
             "candidate_oracle_score_gain_over_selected": data.get(
                 "candidate_oracle_score_gain_over_selected"
             ),
+            "candidate_oracle_unique_count": data.get("candidate_oracle_unique_count"),
             "candidate_oracle_improvement_rate": data.get(
                 "candidate_oracle_improvement_rate"
             ),
@@ -170,6 +174,14 @@ if candidate_oracle_rows:
                 int(row.get("candidate_oracle_rollouts") or 0)
                 for row in candidate_oracle_rows
             ),
+            "candidate_oracle_unique_tolerance": next(
+                (
+                    row.get("candidate_oracle_unique_tolerance")
+                    for row in candidate_oracle_rows
+                    if row.get("candidate_oracle_unique_tolerance") is not None
+                ),
+                None,
+            ),
             "mean_candidate_oracle_success_rate": statistics.mean(
                 row["candidate_oracle_success_rate"] for row in candidate_oracle_rows
             ),
@@ -186,6 +198,12 @@ if candidate_oracle_rows:
                 row["candidate_oracle_score_gain_over_selected"]
                 for row in candidate_oracle_rows
             ),
+            "mean_candidate_oracle_unique_count": statistics.mean(
+                row.get("candidate_oracle_unique_count")
+                or row.get("candidate_oracle_rollouts")
+                or 0
+                for row in candidate_oracle_rows
+            ),
             "mean_candidate_oracle_improvement_rate": statistics.mean(
                 row["candidate_oracle_improvement_rate"]
                 for row in candidate_oracle_rows
@@ -242,6 +260,10 @@ if candidate_oracle_rows:
     lines.extend(
         [
             f"Candidate-oracle prefix: top {summary['candidate_oracle_rollouts']} candidates",
+            (
+                "Candidate-oracle unique tolerance: "
+                f"{summary.get('candidate_oracle_unique_tolerance')}"
+            ),
             (
                 "Candidate-oracle success: "
                 f"{summary['mean_candidate_oracle_success_rate']:.2%}"
@@ -254,6 +276,10 @@ if candidate_oracle_rows:
                 "Candidate-oracle gain over selected branch: "
                 f"{summary['mean_candidate_oracle_score_gain_over_selected']:+.3f}"
             ),
+            (
+                "Candidate-oracle unique count: "
+                f"{summary['mean_candidate_oracle_unique_count']:.2f}"
+            ),
             (
                 "Candidate-oracle improvement rate: "
                 f"{summary['mean_candidate_oracle_improvement_rate']:.2%}"