Auto-sync: 2026-06-28 01:28:06 (part 3)

results/paper_story_memo.md

@@ -20,6 +20,7 @@ when queried on proposal geometry that matches those local counterfactuals.
 | Gradient-based field optimization does not solve the clean proposal gap | `field_optim` best observed result is 25.39% | Negative diagnostic |
 | A broader non-expert proposal target does not reduce the proposal gap | direct broad non-expert policy is 27.88%; with field scoring it is 26.49% | Negative diagnostic |
 | Counterfactual residuals transfer better than absolute retrieved actions | nearest residual retrieval is 32.12% vs absolute retrieval 28.93%; KNN4 residual drops to 29.91% | Supported as a clean bridge |
+| Residual magnitude may be the next clean bottleneck | scale sweep for nearest residual transport is pending | Pending |
 | Seed-0 train-split field-teacher distillation does not solve the proposal gap | direct student is 26.84%; with field scoring it is 27.65% | Negative diagnostic |
 | All-split field-teacher distillation may still help checkpointing/coverage | allmap training has 100% train/val target coverage; rollout eval is pending | Pending |
 
@@ -68,7 +69,7 @@ test-time search. The cleaner novelty is:
 
 ## Active Jobs
 
-Last checked: `2026-06-28 05:20 UTC`.
+Last checked: `2026-06-28 05:26 UTC`.
 
 - `14858328`-`14858333`: completed train-split `field_selected_noexpert_bc5`;
   direct rollout is 26.84%, field-guided best is 27.65%.
@@ -79,6 +80,14 @@ Last checked: `2026-06-28 05:20 UTC`.
 - `14858451`/`14858452`: pending direct rollout evaluation and summary for allmap.
 - `14858453`/`14858454`: pending field-guided rollout sweep and summary for allmap.
 - `14858455`: rebuild `paper_table_status.*` after allmap summaries.
+- `14858978`: completed CPU Apptainer unit smoke for residual-scale selection.
+  Earlier smoke jobs `14858889`/`14858894` caught and fixed two scale wiring bugs
+  before rollout jobs started.
+- `14858875`/`14858876`: pending nearest residual scale `0.25` eval/summary.
+- `14858877`/`14858878`: pending nearest residual scale `0.50` eval/summary.
+- `14858879`/`14858880`: pending nearest residual scale `0.75` eval/summary.
+- `14858881`/`14858882`: pending nearest residual scale `1.25` eval/summary.
+- `14858883`: rebuild `paper_table_status.*` after residual-scale summaries.
 
 ## Decision Rule For Field-Teacher Jobs
 
@@ -90,3 +99,13 @@ Last checked: `2026-06-28 05:20 UTC`.
 - If it fails, keep the central paper story focused on the same-state mechanism
   and the clean-proposal bottleneck, with residual retrieval as the strongest
   deployment-clean bridge.
+
+## Decision Rule For Residual-Scale Jobs
+
+- If any residual scale beats 32.93%, promote tangent-transport residuals as the
+  best deployment-clean bridge.
+- If a smaller scale beats 32.12% but not 32.93%, present it as evidence that
+  counterfactual residuals transfer as local tangent directions with a calibrated
+  step length.
+- If all scales fail, keep scale `1.0` nearest residual retrieval as the clean
+  positive bridge and treat magnitude calibration as a negative ablation.

results/paper_table_status.json

@@ -247,6 +247,120 @@
       "best_config": null,
       "gain_vs_h16_policy": 0.02376811594202899
     },
+    {
+      "key": "retrieval_residual_scale025",
+      "label": "Train-state residual retrieval, scale 0.25",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p25_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "tangent transport scale ablation",
+      "fallback_success": null,
+      "pending_job": "14858875/14858876",
+      "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_scale050",
+      "label": "Train-state residual retrieval, scale 0.50",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "tangent transport scale ablation",
+      "fallback_success": null,
+      "pending_job": "14858877/14858878",
+      "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_scale075",
+      "label": "Train-state residual retrieval, scale 0.75",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p75_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "tangent transport scale ablation",
+      "fallback_success": null,
+      "pending_job": "14858879/14858880",
+      "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_scale125",
+      "label": "Train-state residual retrieval, scale 1.25",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale1p25_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "tangent transport scale ablation",
+      "fallback_success": null,
+      "pending_job": "14858881/14858882",
+      "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_hybrid_k32",
+      "label": "Train-state residual + Gaussian proposals, K32 sigma0.35",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_hybrid_k32_sigma0p35_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "hybrid tangent/local proposal bridge",
+      "fallback_success": null,
+      "pending_job": "14859042/14859043",
+      "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_hybrid_k64",
+      "label": "Train-state residual + Gaussian proposals, K64 sigma0.50",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_hybrid_k64_sigma0p50_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "hybrid tangent/local proposal bridge",
+      "fallback_success": null,
+      "pending_job": "14859044/14859045",
+      "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_knn4",
       "label": "KNN counterfactual residual retrieval",

results/paper_table_status.md

@@ -16,6 +16,12 @@ Baseline h=16 policy: 29.74%
 | field_selected_noexpert_policy_allmap | Field-selected no-expert distillation policy, aligned validation | pending 14858449/14858450/14858451/14858452 | pending | pending | yes | no | no | field-teacher student with aligned checkpoint selection |
 | field_selected_noexpert_policy_allmap_field | Field-selected no-expert distillation + field, aligned validation | pending 14858449/14858450/14858453/14858454 | pending | pending | yes | no | no | aligned field-teacher student with field scoring |
 | retrieval_residual | Train-state counterfactual residual retrieval | complete | 32.12% | +2.38 pp | yes | no | no | transferable local tangent proposal |
+| retrieval_residual_scale025 | Train-state residual retrieval, scale 0.25 | pending 14858875/14858876 | pending | pending | yes | no | no | tangent transport scale ablation |
+| retrieval_residual_scale050 | Train-state residual retrieval, scale 0.50 | pending 14858877/14858878 | pending | pending | yes | no | no | tangent transport scale ablation |
+| retrieval_residual_scale075 | Train-state residual retrieval, scale 0.75 | pending 14858879/14858880 | pending | pending | yes | no | no | tangent transport scale ablation |
+| retrieval_residual_scale125 | Train-state residual retrieval, scale 1.25 | pending 14858881/14858882 | pending | pending | yes | no | no | tangent transport scale ablation |
+| retrieval_residual_hybrid_k32 | Train-state residual + Gaussian proposals, K32 sigma0.35 | pending 14859042/14859043 | pending | pending | yes | no | no | hybrid tangent/local proposal bridge |
+| retrieval_residual_hybrid_k64 | Train-state residual + Gaussian proposals, K64 sigma0.50 | pending 14859044/14859045 | pending | pending | yes | no | no | hybrid tangent/local proposal bridge |
 | retrieval_residual_knn4 | KNN counterfactual residual retrieval | complete | 29.91% | +0.17 pp | yes | no | no | KNN tangent proposal ablation |
 | near_miss_only_lattice | Same-state lattice, near-miss only | complete | 55.94% | +26.20 pp | no | yes | no | minimal mechanism result |
 | no_expert_lattice | Same-state lattice, no expert | complete | 56.99% | +27.25 pp | no | yes | no | main conservative mechanism result |

scripts/build_paper_table_status.py

@@ -145,6 +145,66 @@ SPECS = [
         story_role="transferable local tangent proposal",
         pending_job="14857111/14857112/14857113",
     ),
+    ResultSpec(
+        key="retrieval_residual_scale025",
+        label="Train-state residual retrieval, scale 0.25",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p25_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="tangent transport scale ablation",
+        pending_job="14858875/14858876",
+    ),
+    ResultSpec(
+        key="retrieval_residual_scale050",
+        label="Train-state residual retrieval, scale 0.50",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="tangent transport scale ablation",
+        pending_job="14858877/14858878",
+    ),
+    ResultSpec(
+        key="retrieval_residual_scale075",
+        label="Train-state residual retrieval, scale 0.75",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p75_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="tangent transport scale ablation",
+        pending_job="14858879/14858880",
+    ),
+    ResultSpec(
+        key="retrieval_residual_scale125",
+        label="Train-state residual retrieval, scale 1.25",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale1p25_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="tangent transport scale ablation",
+        pending_job="14858881/14858882",
+    ),
+    ResultSpec(
+        key="retrieval_residual_hybrid_k32",
+        label="Train-state residual + Gaussian proposals, K32 sigma0.35",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_hybrid_k32_sigma0p35_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="hybrid tangent/local proposal bridge",
+        pending_job="14859042/14859043",
+    ),
+    ResultSpec(
+        key="retrieval_residual_hybrid_k64",
+        label="Train-state residual + Gaussian proposals, K64 sigma0.50",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_hybrid_k64_sigma0p50_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="hybrid tangent/local proposal bridge",
+        pending_job="14859044/14859045",
+    ),
     ResultSpec(
         key="retrieval_residual_knn4",
         label="KNN counterfactual residual retrieval",

scripts/eval_maniskill_policy_rollout.py

@@ -103,6 +103,12 @@ def main(argv: list[str] | None = None) -> int:
         default=1,
         help="Nearest train states to use for retrieval_lattice/retrieval_residual proposals.",
     )
+    parser.add_argument(
+        "--retrieval-residual-scale",
+        type=float,
+        default=1.0,
+        help="Scale applied to train-state residuals before adding them to the policy mean.",
+    )
     parser.add_argument(
         "--lattice-exclude-types",
         default="",
@@ -132,6 +138,7 @@ def main(argv: list[str] | None = None) -> int:
         field_optim_trust_radius=args.field_optim_trust_radius,
         field_optim_l2_penalty=args.field_optim_l2_penalty,
         retrieval_neighbors=args.retrieval_neighbors,
+        retrieval_residual_scale=args.retrieval_residual_scale,
         lattice_exclude_types=lattice_exclude_types,
     )
     print(json.dumps({key: value for key, value in result.items() if key != "rows"}, indent=2))

scripts/slurm/eval_maniskill_policy_rollout.sbatch

@@ -49,6 +49,7 @@ FIELD_OPTIM_STEP_SIZE="${FIELD_OPTIM_STEP_SIZE:-0.05}"
 FIELD_OPTIM_TRUST_RADIUS="${FIELD_OPTIM_TRUST_RADIUS:-0.5}"
 FIELD_OPTIM_L2_PENALTY="${FIELD_OPTIM_L2_PENALTY:-0.0}"
 RETRIEVAL_NEIGHBORS="${RETRIEVAL_NEIGHBORS:-1}"
+RETRIEVAL_RESIDUAL_SCALE="${RETRIEVAL_RESIDUAL_SCALE:-1.0}"
 LATTICE_EXCLUDE_TYPES="${LATTICE_EXCLUDE_TYPES:-}"
 if [[ -n "${LATTICE_EXCLUDE_TYPES_COLON:-}" ]]; then
   LATTICE_EXCLUDE_TYPES="${LATTICE_EXCLUDE_TYPES_COLON//:/,}"
@@ -97,5 +98,6 @@ apptainer exec --nv \
     --field-optim-trust-radius "$FIELD_OPTIM_TRUST_RADIUS" \
     --field-optim-l2-penalty "$FIELD_OPTIM_L2_PENALTY" \
     --retrieval-neighbors "$RETRIEVAL_NEIGHBORS" \
+    --retrieval-residual-scale "$RETRIEVAL_RESIDUAL_SCALE" \
     --lattice-exclude-types "$LATTICE_EXCLUDE_TYPES" \
     "${EXTRA_ARGS[@]}"

scripts/slurm/smoke_retrieval_residual_scale_unit.sbatch

@@ -0,0 +1,74 @@
+#!/bin/bash
+#SBATCH --job-name=smoke_residual_scale
+#SBATCH --account=def-yalda
+#SBATCH --nodes=1
+#SBATCH --ntasks=1
+#SBATCH --cpus-per-task=1
+#SBATCH --mem=1G
+#SBATCH --time=00:05:00
+#SBATCH --output=outputs/hpc/logs/%x_%j.out
+#SBATCH --error=outputs/hpc/logs/%x_%j.err
+
+set -euo pipefail
+
+PROJECT_DIR="${PROJECT_DIR:-$SLURM_SUBMIT_DIR}"
+SCRATCH_ROOT="/scratch/$USER/dovla"
+SIF="${SIF:-$SCRATCH_ROOT/containers/pytorch_2.7.1_cuda12.8.sif}"
+PYTHON="${PYTHON:-$SCRATCH_ROOT/envs/maniskill/bin/python}"
+
+module load StdEnv/2023 apptainer/1.4.5
+cd "$PROJECT_DIR"
+mkdir -p outputs/hpc/logs
+
+export OMP_NUM_THREADS=1
+export OPENBLAS_NUM_THREADS=1
+export MKL_NUM_THREADS=1
+export DOVLA_TORCH_THREADS=1
+
+apptainer exec \
+  --env "OMP_NUM_THREADS=1,OPENBLAS_NUM_THREADS=1,MKL_NUM_THREADS=1,DOVLA_TORCH_THREADS=1,PYTHONDONTWRITEBYTECODE=1" \
+  -B "$PROJECT_DIR:$PROJECT_DIR" \
+  -B "/scratch/$USER:/scratch/$USER" \
+  "$SIF" "$PYTHON" - <<'PY'
+import torch
+
+from dovla_cil.eval.maniskill_policy_rollout import _select_action_chunk
+
+
+class StubModel:
+    def __init__(self, mean, target_offset):
+        self.mean = mean
+        self.target = mean + target_offset
+
+    def forward_policy(self, observations, instructions):
+        del observations, instructions
+        return self.mean
+
+    def forward_field(self, observations, instructions, action):
+        del observations, instructions
+        distance = ((action - self.target) ** 2).reshape(action.shape[0], -1).sum(dim=1)
+        return {"potential": -distance}
+
+
+mean = torch.full((1, 1, 3), 0.1)
+residual = torch.full_like(mean, 0.8)
+scale = 0.5
+model = StubModel(mean, target_offset=residual * scale)
+residuals = torch.stack([torch.zeros_like(mean), residual], dim=1)
+actions, index = _select_action_chunk(
+    model,
+    observations=torch.zeros(1, 3),
+    instructions=["pick"],
+    torch=torch,
+    selection_mode="retrieval_residual",
+    num_candidates=1,
+    candidate_sigma=0.0,
+    selection_seed=0,
+    action_candidates=residuals,
+    retrieval_residual_scale=scale,
+)
+expected = mean + residual * scale
+assert torch.allclose(actions, expected), (actions, expected)
+assert index.tolist() == [1], index
+print({"status": "ok", "scale": scale, "actions": actions.tolist(), "index": index.tolist()})
+PY

scripts/slurm/summarize_h16_field_sweep.sbatch

@@ -51,6 +51,7 @@ for result_path in sorted(run_root.glob("k*_sigma*/seed_*/online_rollout.json"))
             "field_optim_trust_radius": data.get("field_optim_trust_radius", 0.0),
             "field_optim_l2_penalty": data.get("field_optim_l2_penalty", 0.0),
             "retrieval_neighbors": data.get("retrieval_neighbors", 0),
+            "retrieval_residual_scale": data.get("retrieval_residual_scale", 0.0),
             "num_groups": data.get("num_groups"),
             "policy_rollout_success_rate": data.get("policy_rollout_success_rate", 0.0),
             "policy_rollout_progress": data.get("policy_rollout_progress", 0.0),

scripts/slurm/summarize_h16_policy_ckpt.sbatch

@@ -60,6 +60,7 @@ for result_path in sorted(base_dir.glob(f"seed_*/{out_name}")):
             "field_optim_trust_radius": data.get("field_optim_trust_radius", 0.0),
             "field_optim_l2_penalty": data.get("field_optim_l2_penalty", 0.0),
             "retrieval_neighbors": data.get("retrieval_neighbors", 0),
+            "retrieval_residual_scale": data.get("retrieval_residual_scale", 0.0),
             "policy_rollout_success_rate": data.get("policy_rollout_success_rate", 0.0),
             "policy_rollout_progress": data.get("policy_rollout_progress", 0.0),
             "oracle_success_rate": data.get("oracle_success_rate", 0.0),
@@ -120,17 +121,18 @@ lines = [
     f"Mean progress: {summary['mean_progress']:.2%}",
     f"Mean action MSE to best: {summary['mean_action_mse_to_best']:.3f}",
     "",
-    "| seed | mode | k | retrieval K | sigma | opt steps | trust | success | progress | oracle | action MSE |",
-    "|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|",
+    "| seed | mode | k | retrieval K | residual scale | sigma | opt steps | trust | success | progress | oracle | action MSE |",
+    "|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|",
 ]
 for row in rows:
     lines.append(
-        "| {seed} | {mode} | {k} | {retrieval} | {sigma:.2f} | {steps} | {trust:.2f} | "
+        "| {seed} | {mode} | {k} | {retrieval} | {scale:.2f} | {sigma:.2f} | {steps} | {trust:.2f} | "
         "{success:.2%} | {progress:.2%} | {oracle:.2%} | {mse:.3f} |".format(
             seed=row["seed"],
             mode=row.get("selection_mode") or "policy",
             k=row.get("num_candidates") or 1,
             retrieval=row.get("retrieval_neighbors") or 0,
+            scale=row.get("retrieval_residual_scale") or 0.0,
             sigma=row.get("candidate_sigma") or 0.0,
             steps=row.get("field_optim_steps") or 0,
             trust=row.get("field_optim_trust_radius") or 0.0,

tests/test_maniskill_policy_rollout.py

@@ -18,9 +18,11 @@ from dovla_cil.eval.maniskill_policy_rollout import (
     _RolloutCase,
     _adapt_action_dim,
     _attach_retrieved_residual_candidates,
+    _effective_lattice_candidate_count,
     _load_state_archive,
     _numeric_action_values,
     _select_action_chunk,
+    _selected_candidate_type,
     _summarize_rows,
 )
 
@@ -370,6 +372,63 @@ def test_retrieval_residual_mode_translates_residuals_around_policy_mean() -> No
     assert index.tolist() == [1]
 
 
+def test_retrieval_residual_mode_scales_residuals() -> None:
+    import torch
+
+    mean = torch.full((1, 1, 3), 0.1)
+    residual = torch.full_like(mean, 0.8)
+    target_offset = residual * 0.5
+    model = _StubModel(torch, mean, best_offset=target_offset)
+    residuals = torch.stack([torch.zeros_like(mean), residual], dim=1)
+    actions, index = _select_action_chunk(
+        model,
+        observations=torch.zeros(1, 3),
+        instructions=["a"],
+        torch=torch,
+        selection_mode="retrieval_residual",
+        num_candidates=1,
+        candidate_sigma=0.0,
+        selection_seed=0,
+        action_candidates=residuals,
+        retrieval_residual_scale=0.5,
+    )
+
+    assert torch.allclose(actions, mean + target_offset)
+    assert index.tolist() == [1]
+
+
+def test_retrieval_residual_gaussian_candidates_are_counted_and_labeled() -> None:
+    case = _RolloutCase(
+        group_id="g",
+        task_id="PickCube-v1",
+        source_dataset=Path("."),
+        state={},
+        observation={"features": [0.0]},
+        instruction="pick",
+        oracle_score=1.0,
+        oracle_success=True,
+        expert_score=1.0,
+        expert_success=True,
+        best_action_values=[[0.0]],
+        candidate_action_values=[[[0.0]], [[0.2]]],
+        candidate_types=["policy_residual", "residual_near_miss"],
+    )
+
+    assert (
+        _effective_lattice_candidate_count(
+            case,
+            selection_mode="retrieval_residual",
+            num_candidates=4,
+            candidate_sigma=0.2,
+        )
+        == 5
+    )
+    assert (
+        _selected_candidate_type(case, selected_index=4, selection_mode="retrieval_residual")
+        == "retrieval_residual_gaussian"
+    )
+
+
 def test_retrieval_residual_candidates_use_knn_train_residuals() -> None:
     def record(group_id: str, candidate_type: str, action_value: float, feature: float):
         return SimpleNamespace(