Auto-sync: 2026-06-28 22:48:37 (part 3)

results/paper_core_results.md

@@ -38,8 +38,10 @@ and the remaining clean-to-same-state proposal gap is `+21.74 pp`.
 | Train-state residual retrieval, policy/no-op/wrong-gripper, scale 0.35 | No | No | 33.74% | +4.00 pp | Typed tangent transport before abstention |
 | Train-state residual retrieval, safe residuals + advantage margin 0.20 | No | No | 34.84% | +5.10 pp | Abstains unless field advantage beats policy |
 | K2 train-state residual retrieval, safe residuals + advantage margin 0.20 | No | No | 35.01% | +5.28 pp | Previous best clean diagnostic; abstention makes a small train-neighborhood useful |
+| K2 task-relative residual retrieval, safe residuals + advantage margin 0.20 | No | No | 34.26% | +4.52 pp | Actor-pose-only retrieval is too lossy; raw full-state similarity is better for residual transfer |
 | K4 train-state residual retrieval, safe residuals + mean-by-type tangent consensus | No | No | 34.96% | +5.22 pp | Near-tie clean diagnostic; consensus alone does not beat raw K2 residuals |
 | K4 mean-by-type residual retrieval + no-op prior 0.03 | No | No | 35.25% | +5.51 pp | Current best clean diagnostic; 0.025-0.035 forms a small plateau that nudges high-value no-op residuals without changing the core proposal family |
+| K4 task-relative mean-by-type residual retrieval + no-op prior 0.03 | No | No | 34.43% | +4.70 pp | Task-relative target/reference pose retrieval underperforms the raw-metric no-op plateau |
 | K4 kernel-weighted residual consensus + no-op prior 0.03 | No | No | 35.13-35.19% | +5.39-5.45 pp | Distance-weighted tangent interpolation is plausible but does not beat equal mean-consensus no-op plateau |
 | K4 field-softmax residual barycenter + no-op prior 0.03 | No | No | 34.84-35.19% | +5.10-5.45 pp | Field-conditioned aggregation finds high-value sparse corrections, but lower margins over-select them; it does not beat the equal mean-consensus no-op plateau |
 | K4 mean-by-type residual retrieval + wrong-gripper typed prior | No | No | 35.19-35.25% | +5.45-5.51 pp | Wrong-gripper-only is lower and two-family priors only tie the no-op plateau; useful negative/tie diagnostic |
@@ -72,14 +74,14 @@ Suggested main-table rows:
 4. Near-miss proposal + field, BC x5 field checkpoint
 5. Trust-region field optimization
 6. Best non-expert proposal + field
-7. Field-selected no-expert policy + field, seed-0 train map
-8. Field-selected no-expert policy + field, aligned allmap
-9. Train-state residual retrieval, scale 0.50
-10. Train-state residual retrieval, typed safe families at scale 0.35
-11. Train-state residual retrieval, typed safe families + advantage margin 0.20
-12. K2 train-state residual retrieval, typed safe families + advantage margin 0.20
-13. K4 train-state residual retrieval, mean-by-type tangent consensus
-14. K4 mean-by-type residual retrieval + no-op prior plateau, canonical 0.03
+7. Field-selected no-expert policy + field, aligned allmap
+8. Train-state residual retrieval, scale 0.50
+9. Train-state residual retrieval, typed safe families at scale 0.35
+10. Train-state residual retrieval, typed safe families + advantage margin 0.20
+11. K2 train-state residual retrieval, typed safe families + advantage margin 0.20
+12. K4 train-state residual retrieval, mean-by-type tangent consensus
+13. K4 mean-by-type residual retrieval + no-op prior plateau, canonical 0.03
+14. K2/K4 task-relative retrieval metric diagnostics
 15. K4 kernel-weighted residual consensus + no-op prior diagnostics
 16. K4 field-softmax residual barycenter + margin diagnostics
 17. K4 mean-by-type residual retrieval + wrong-gripper typed-prior diagnostics
@@ -97,7 +99,7 @@ Suggested claim:
 > DoVLA-CIL is not a better behavior-cloning policy; it is a local counterfactual action
 > selection rule. Deployment-clean K4 consensus residual transport with advantage
 > abstention and a small typed no-op prior plateau gives the strongest clean gain so far, while ungated KNN residual
-> retrieval, field-gradient ascent, broader non-expert BC targets, field-teacher/tangent distillation, z-score retrieval,
+> retrieval, field-gradient ascent, broader non-expert BC targets, field-teacher/tangent distillation, z-score/task-relative retrieval,
 > train-family reliability priors, policy-relative anchoring, residual+Gaussian hybrids,
 > tangent consensus, kernel-weighted tangent interpolation, field-softmax tangent barycenters, tangent ray-search, wrong-gripper typed priors, and same-state policy-baseline fallback fail to improve the main rows.
 > The large effect appears only when the field is queried on

results/paper_story_memo.md

@@ -36,6 +36,7 @@ when queried on proposal geometry that matches those local counterfactuals.
 | The proposal gap is now quantified | `paper_analysis.md` reports best clean +5.51 pp over canonical h16, same-state no-expert +27.25 pp, leaving a +21.74 pp clean-to-same-state gap | Core paper tension |
 | Policy fallback is not the same-state mechanism | adding a policy baseline candidate to the no-expert same-state lattice drops 56.99% to 40.70% even with margin 0.00 | Negative diagnostic |
 | Z-score retrieval metric does not help | z-score rows reach 32.23-32.81%, below raw retrieval | Negative diagnostic |
+| Task-relative actor-pose retrieval metric does not improve tangent transfer | K2 task-relative residual retrieval reaches 34.26% vs raw K2 35.01%; K4 task-relative mean-by-type + no-op reaches 34.43% vs raw K4 35.25% | Negative diagnostic |
 | Train-split residual family reliability does not recover the typed mask | after fixing threshold pass-through, scale-0.35 thresholds 0.10/0.25 reach 33.33%/33.28%, below typed safe residuals | Negative diagnostic |
 | Residual-tangent distillation does not solve clean proposal generation | aligned allmap tangent student reaches 28.87% despite low pseudo-target BC loss | Negative diagnostic |
 | Policy-relative residual anchoring does not improve the bridge | policy-anchor safe residual transport ties 33.74% rather than improving expert-anchor residuals | Negative diagnostic |
@@ -69,13 +70,14 @@ clean proposal result, the intended main rows are:
 21. K4 tight tangent ray-search: 34.55%
 22. Residual-tangent distillation policy: 28.87%
 23. Z-score residual retrieval: 32.23-32.81%
-24. Train-family reliability prior: 33.28-33.33%
-25. Residual+Gaussian hybrid K32/K64: 31.30% / 30.90%
-26. Lattice, near-miss only: 55.94%
-27. Lattice, no expert: 56.99%
-28. Lattice, no expert + policy baseline candidate: 40.70%
-29. Lattice, full: 69.33%
-30. Oracle ceiling: 86.78%
+24. Task-relative residual retrieval metric: 34.26-34.43%
+25. Train-family reliability prior: 33.28-33.33%
+26. Residual+Gaussian hybrid K32/K64: 31.30% / 30.90%
+27. Lattice, near-miss only: 55.94%
+28. Lattice, no expert: 56.99%
+29. Lattice, no expert + policy baseline candidate: 40.70%
+30. Lattice, full: 69.33%
+31. Oracle ceiling: 86.78%
 
 ## Novelty Framing
 
@@ -103,9 +105,9 @@ test-time search. The cleaner novelty is:
 
 ## Job Status
 
-Last checked: `2026-06-29 02:00 UTC`. The field-conditioned tangent-barycenter
-batch completed after passing CPU smokes, and the paper table rebuild now includes
-the field-softmax rows.
+Last checked: `2026-06-29 02:45 UTC`. The task-relative retrieval-metric batch
+completed after passing CPU/unit smokes, and the paper table rebuild now includes
+the task-relative rows.
 
 - `14858328`-`14858333`: completed train-split `field_selected_noexpert_bc5`;
   direct rollout is 26.84%, field-guided best is 27.65%.
@@ -204,6 +206,20 @@ the field-softmax rows.
   negative/near-tie diagnostic below the 35.25% mean-consensus no-op plateau.
   Summary jobs `14893002`/`14893016`/`14893028` and rebuild job `14893069`
   completed.
+- `14893449`: completed the CPU Apptainer unit smoke for
+  `retrieval_metric=task_relative`, confirming the target/reference actor-pose
+  distance path in the container.
+- `14893458`: completed a 4-group CPU rollout smoke for K4 task-relative
+  residual retrieval with mean-by-type reduction, safe residual masks, and no-op
+  bonus 0.03. Earlier GPU arrays `14893473`/`14893475` were canceled/replaced
+  after an env mistake set `ALL_GROUPS=1`, which correctly triggered the
+  held-out split guard.
+- `14893787`/`14893789`: completed corrected task-relative retrieval GPU arrays.
+  K4 mean-by-type + no-op 0.03 reaches 34.43%, and K2 safe residual retrieval
+  reaches 34.26%, both below their raw-metric counterparts (35.25% and 35.01%).
+  Summary jobs `14893788`/`14893790` and rebuild job `14893791` completed. This
+  suggests that raw full-state similarity still carries useful robot/phase
+  information for residual transfer; object-only actor pose is too lossy here.
 - `14869627`: completed CPU Apptainer smoke for the new residual scale-grid
   selector. It selected index `3` on a two-residual/two-scale toy case and
   returned the expected action `0.20`, validating the candidate expansion and
@@ -230,7 +246,7 @@ the field-softmax rows.
   story.
 - Use `results/paper_analysis.md` for paired seed deltas, per-task gaps, and
   selection histograms when writing reviewer-facing tables.
-- Treat z-score retrieval, repaired train-family reliability priors, Gaussian hybrids,
+- Treat z-score and task-relative retrieval metrics, repaired train-family reliability priors, Gaussian hybrids,
   field optimization, field-teacher/tangent distillation, policy-relative anchoring, tangent consensus,
   kernel-weighted tangent interpolation, field-softmax tangent barycenters,
   wrong-gripper typed priors, and same-state policy-baseline fallback as negative

scripts/build_paper_table_status.py

@@ -455,6 +455,26 @@ SPECS = [
         story_role="current best clean typed sparse-intervention prior",
         pending_job="14883919/14883920",
     ),
+    ResultSpec(
+        key="retrieval_residual_k4_mean_noopbonus003_srcprog050",
+        label="K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op bonus 0.03, source progress >= 0.50",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_noopbonus0p03_srcprog0p50_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="train-source viability gate for sparse residual transport",
+        pending_job="14894093/14894094",
+    ),
+    ResultSpec(
+        key="retrieval_residual_k4_mean_noopbonus003_srcprog075",
+        label="K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op bonus 0.03, source progress >= 0.75",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_noopbonus0p03_srcprog0p75_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="strict train-source viability gate for sparse residual transport",
+        pending_job="14894095/14894096",
+    ),
     ResultSpec(
         key="retrieval_residual_taskrelative_k4_mean_noopbonus003",
         label="K4 task-relative mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.03",

scripts/eval_maniskill_policy_rollout.py

@@ -129,6 +129,13 @@ def main(argv: list[str] | None = None) -> int:
         help="Minimum train-split terminal success rate for a residual candidate family to be "
         "eligible in retrieval_residual mode. The policy_residual fallback is always kept.",
     )
+    parser.add_argument(
+        "--retrieval-residual-min-source-progress",
+        type=float,
+        default=0.0,
+        help="Minimum measured train-source progress for an individual residual candidate to "
+        "be eligible in retrieval_residual mode. The policy_residual fallback is always kept.",
+    )
     parser.add_argument(
         "--retrieval-residual-scale",
         type=float,
@@ -220,6 +227,7 @@ def main(argv: list[str] | None = None) -> int:
         retrieval_neighbors=args.retrieval_neighbors,
         retrieval_metric=args.retrieval_metric,
         retrieval_type_min_success=args.retrieval_type_min_success,
+        retrieval_residual_min_source_progress=args.retrieval_residual_min_source_progress,
         retrieval_residual_scale=args.retrieval_residual_scale,
         retrieval_residual_scales=retrieval_residual_scales,
         retrieval_residual_anchor=args.retrieval_residual_anchor,