auto-sync 2026-07-02T18:12:23Z workspace

workspace/docs/cil_format.md

@@ -45,5 +45,15 @@ Benchmark tracks:
 - Deployment: execute deployment-clean proposals without same-state validation rewards.
 - Recovery: evaluate near-miss states and safety/abstention behavior.
 
+Generator baselines:
+
+- V0 transported residual retrieval copies train-state tangents into the current
+  state and lets the learned utility field select among them.
+- V1 utility-weighted residual retrieval changes proposal support before
+  selection by weighting train tangents with retrieval affinity and measured
+  source advantage, `exp(-distance / tau + rho * delta_utility)`.
+- The full CIL-Atlas generator should learn positive causal tangent support
+  directly, with V0/V1 kept as diagnostic baselines rather than the main novelty.
+
 JSONL shards should preserve complete groups. A group may exceed the target shard size, but it
 should never be split across shards unless an explicit future streaming mode opts into that tradeoff.

workspace/dovla_cil/eval/maniskill_policy_rollout.py

@@ -120,6 +120,7 @@ def evaluate_maniskill_policy_rollout(
     retrieval_residual_source_score_bonus_scale: float = 0.0,
     retrieval_residual_source_score_bonus_by_task: dict[str, float] | None = None,
     retrieval_residual_source_advantage_bonus_scale: float = 0.0,
+    retrieval_residual_source_advantage_weight_scale: float = 0.0,
     retrieval_residual_composite_l2_penalty_scale: float = 0.0,
     retrieval_residual_action_l2_penalty: float = 0.0,
     retrieval_residual_scale: float = 1.0,
@@ -284,6 +285,8 @@ def evaluate_maniskill_policy_rollout(
         )
     if retrieval_residual_source_advantage_bonus_scale < 0:
         raise ValueError("retrieval_residual_source_advantage_bonus_scale must be non-negative")
+    if retrieval_residual_source_advantage_weight_scale < 0:
+        raise ValueError("retrieval_residual_source_advantage_weight_scale must be non-negative")
     if retrieval_residual_composite_l2_penalty_scale < 0:
         raise ValueError("retrieval_residual_composite_l2_penalty_scale must be non-negative")
     if retrieval_residual_action_l2_penalty < 0:
@@ -447,6 +450,9 @@ def evaluate_maniskill_policy_rollout(
                 retrieval_residual_source_advantage_bonus_scale=(
                     retrieval_residual_source_advantage_bonus_scale
                 ),
+                retrieval_residual_source_advantage_weight_scale=(
+                    retrieval_residual_source_advantage_weight_scale
+                ),
                 retrieval_residual_composite_l2_penalty_scale=(
                     retrieval_residual_composite_l2_penalty_scale
                 ),
@@ -625,6 +631,11 @@ def evaluate_maniskill_policy_rollout(
             if selection_mode == "retrieval_residual"
             else 0.0
         ),
+        "retrieval_residual_source_advantage_weight_scale": (
+            retrieval_residual_source_advantage_weight_scale
+            if selection_mode == "retrieval_residual"
+            else 0.0
+        ),
         "retrieval_residual_composite_l2_penalty_scale": (
             retrieval_residual_composite_l2_penalty_scale
             if selection_mode == "retrieval_residual"
@@ -876,6 +887,7 @@ def _attach_retrieved_residual_candidates(
     retrieval_residual_source_score_bonus_scale: float = 0.0,
     retrieval_residual_source_score_bonus_by_task: dict[str, float] | None = None,
     retrieval_residual_source_advantage_bonus_scale: float = 0.0,
+    retrieval_residual_source_advantage_weight_scale: float = 0.0,
     retrieval_residual_composite_l2_penalty_scale: float = 0.0,
     retrieval_residual_anchor: str = "expert",
     retrieval_residual_direction: str = "candidate_minus_anchor",
@@ -909,6 +921,7 @@ def _attach_retrieved_residual_candidates(
                 list[list[list[float]]],
                 list[str],
                 list[float],
+                list[float],
             ]
         ],
     ] = defaultdict(list)
@@ -934,6 +947,7 @@ def _attach_retrieved_residual_candidates(
         residuals: list[list[list[float]]] = [np.zeros_like(anchor_action).tolist()]
         candidate_types = ["policy_residual"]
         residual_bonuses = [0.0]
+        residual_source_weights = [1.0]
         for record in records:
             if record.record_id == anchor.record_id:
                 continue
@@ -972,12 +986,25 @@ def _attach_retrieved_residual_candidates(
                 + float(source_score_bonus_scale) * source_score
                 + float(retrieval_residual_source_advantage_bonus_scale) * source_advantage
             )
+            residual_source_weights.append(
+                _source_advantage_weight(
+                    source_advantage,
+                    scale=retrieval_residual_source_advantage_weight_scale,
+                )
+            )
         feature = np.asarray(
             vectorize_toy_observation(records[0].observation_inline or {}, obs_dim=obs_dim),
             dtype=np.float32,
         )
         bank[next(iter(task_ids))].append(
-            (group_id, feature, residuals, candidate_types, residual_bonuses)
+            (
+                group_id,
+                feature,
+                residuals,
+                candidate_types,
+                residual_bonuses,
+                residual_source_weights,
+            )
         )
 
     output: list[_RolloutCase] = []
@@ -1023,11 +1050,17 @@ def _attach_retrieved_residual_candidates(
                 source_residuals,
                 source_candidate_types,
                 source_residual_bonuses,
+                source_residual_weights,
             ) = entry
             source_group_ids.append(source_group_id)
             residuals.extend(source_residuals)
             candidate_types.extend(source_candidate_types)
-            residual_weights.extend([float(source_weight)] * len(source_residuals))
+            residual_weights.extend(
+                [
+                    float(source_weight) * float(residual_source_weight)
+                    for residual_source_weight in source_residual_weights
+                ]
+            )
             residual_bonuses.extend(source_residual_bonuses)
         if retrieval_residual_reduce in _NUMPY_RESIDUAL_REDUCERS:
             residuals, candidate_types, residual_bonuses = _reduce_residual_candidates_by_type(
@@ -1036,6 +1069,7 @@ def _attach_retrieved_residual_candidates(
                 mode=retrieval_residual_reduce,
                 weights=residual_weights,
                 bonuses=residual_bonuses,
+                mean_weights=retrieval_residual_source_advantage_weight_scale > 0.0,
                 consensus_penalty_scale=retrieval_residual_consensus_penalty_scale,
                 composite_l2_penalty_scale=retrieval_residual_composite_l2_penalty_scale,
             )
@@ -1091,6 +1125,7 @@ def _reduce_residual_candidates_by_type(
     mode: str,
     weights: list[float] | None = None,
     bonuses: list[float] | None = None,
+    mean_weights: bool = False,
     consensus_penalty_scale: float = 0.0,
     composite_l2_penalty_scale: float = 0.0,
 ) -> tuple[list[list[list[float]]], list[str]] | tuple[
@@ -1112,6 +1147,8 @@ def _reduce_residual_candidates_by_type(
         raise ValueError("weights and residuals must have the same length")
     if bonuses is not None and len(bonuses) != len(residuals):
         raise ValueError("bonuses and residuals must have the same length")
+    if mean_weights and weights is None:
+        raise ValueError("mean_weights requires weights")
     if consensus_penalty_scale < 0:
         raise ValueError("consensus_penalty_scale must be non-negative")
     if composite_l2_penalty_scale < 0:
@@ -1136,11 +1173,8 @@ def _reduce_residual_candidates_by_type(
         if not values:
             continue
         stack = np.stack(values, axis=0)
-        if mode in {"mean_by_type", "compose_mean_by_type"}:
-            reduced = np.mean(stack, axis=0)
-            reduced_bonus = float(np.mean(value_bonuses)) if value_bonuses else 0.0
-        elif mode == "kernel_mean_by_type":
-            if value_weights:
+        if mode in {"mean_by_type", "compose_mean_by_type", "kernel_mean_by_type"}:
+            if value_weights and (mode == "kernel_mean_by_type" or mean_weights):
                 np_weights = np.asarray(value_weights, dtype=np.float32)
                 weight_sum = float(np.sum(np_weights))
                 if weight_sum > 1e-12:
@@ -1239,13 +1273,13 @@ def _source_reward_score(reward: Any, *, progress: float) -> float:
 
 
 def _nearest_retrieval_entries(
-    candidates: list[tuple[Any, np.ndarray, Any, Any]],
+    candidates: list[tuple[Any, ...]],
     query: np.ndarray,
     *,
     retrieval_neighbors: int,
     retrieval_metric: str,
     task_id: str | None = None,
-) -> list[tuple[Any, np.ndarray, Any, Any]]:
+) -> list[tuple[Any, ...]]:
     return [
         entry
         for entry, _distance in _nearest_retrieval_entries_with_distances(
@@ -1259,13 +1293,13 @@ def _nearest_retrieval_entries(
 
 
 def _nearest_retrieval_entries_with_distances(
-    candidates: list[tuple[Any, np.ndarray, Any, Any]],
+    candidates: list[tuple[Any, ...]],
     query: np.ndarray,
     *,
     retrieval_neighbors: int,
     retrieval_metric: str,
     task_id: str | None = None,
-) -> list[tuple[tuple[Any, np.ndarray, Any, Any], float]]:
+) -> list[tuple[tuple[Any, ...], float]]:
     if retrieval_metric == "raw":
         scored = [
             (item, float(np.mean((item[1] - query) ** 2)))
@@ -1378,6 +1412,17 @@ def _kernel_weights_from_distances(distances: list[float]) -> list[float]:
     return weights.astype(np.float32).tolist()
 
 
+def _source_advantage_weight(source_advantage: float, *, scale: float) -> float:
+    if scale <= 0.0:
+        return 1.0
+    exponent = float(scale) * float(source_advantage)
+    exponent = min(max(exponent, -50.0), 50.0)
+    weight = math.exp(exponent)
+    if not math.isfinite(weight) or weight <= 0.0:
+        return 1.0
+    return weight
+
+
 def _evaluate_task_cases(
     task_id: str,
     cases: list[_RolloutCase],

workspace/latex/main.aux

@@ -22,17 +22,17 @@
 \@writefile{toc}{\contentsline {subsection}{\numberline {2.1}Same-State Action Charts}{2}{subsection.2.1}\protected@file@percent }
 \@writefile{toc}{\contentsline {subsection}{\numberline {2.2}Tracks}{2}{subsection.2.2}\protected@file@percent }
 \@writefile{toc}{\contentsline {section}{\numberline {3}Metrics}{2}{section.3}\protected@file@percent }
-\@writefile{lot}{\contentsline {table}{\numberline {1}{\ignorespaces Core results. Clean rows do not use same-state validation proposals or expert proposals at deployment.}}{3}{table.1}\protected@file@percent }
-\newlabel{tab:main-results}{{1}{3}{Core results. Clean rows do not use same-state validation proposals or expert proposals at deployment}{table.1}{}}
-\@writefile{lot}{\contentsline {table}{\numberline {2}{\ignorespaces Causal Action Regret decomposition on the current six-task diagnostic. Support is the proposal-generation gap to the hidden same-state no-expert oracle; selector is the clean proposal-oracle headroom left by the deployed selector.}}{3}{table.2}\protected@file@percent }
-\newlabel{tab:car-decomposition}{{2}{3}{Causal Action Regret decomposition on the current six-task diagnostic. Support is the proposal-generation gap to the hidden same-state no-expert oracle; selector is the clean proposal-oracle headroom left by the deployed selector}{table.2}{}}
 \@writefile{toc}{\contentsline {section}{\numberline {4}CIL-Atlas}{3}{section.4}\protected@file@percent }
 \@writefile{toc}{\contentsline {section}{\numberline {5}Current Six-Task Diagnostic}{3}{section.5}\protected@file@percent }
-\@writefile{lot}{\contentsline {table}{\numberline {3}{\ignorespaces Train-source score prior sensitivity for the current K6 transported residual field. A very small prior calibrates the selector; stronger priors tie or degrade success.}}{4}{table.3}\protected@file@percent }
-\newlabel{tab:source-score-sweep}{{3}{4}{Train-source score prior sensitivity for the current K6 transported residual field. A very small prior calibrates the selector; stronger priors tie or degrade success}{table.3}{}}
-\@writefile{lot}{\contentsline {table}{\numberline {4}{\ignorespaces Selector calibration ablations on the current K6 transported residual field. All rows are deployment-clean and use only train-split calibration signals.}}{4}{table.4}\protected@file@percent }
-\newlabel{tab:selector-calibration}{{4}{4}{Selector calibration ablations on the current K6 transported residual field. All rows are deployment-clean and use only train-split calibration signals}{table.4}{}}
-\@writefile{toc}{\contentsline {subsection}{\numberline {5.1}Selector Diagnostics}{4}{subsection.5.1}\protected@file@percent }
-\@writefile{toc}{\contentsline {section}{\numberline {6}Roadmap to the Full Paper}{4}{section.6}\protected@file@percent }
+\@writefile{toc}{\contentsline {subsection}{\numberline {5.1}Selector Diagnostics}{3}{subsection.5.1}\protected@file@percent }
+\@writefile{toc}{\contentsline {section}{\numberline {6}Roadmap to the Full Paper}{3}{section.6}\protected@file@percent }
+\@writefile{lot}{\contentsline {table}{\numberline {1}{\ignorespaces Core results. Clean rows do not use same-state validation proposals or expert proposals at deployment.}}{4}{table.1}\protected@file@percent }
+\newlabel{tab:main-results}{{1}{4}{Core results. Clean rows do not use same-state validation proposals or expert proposals at deployment}{table.1}{}}
+\@writefile{lot}{\contentsline {table}{\numberline {2}{\ignorespaces Causal Action Regret decomposition on the current six-task diagnostic. Support is the proposal-generation gap to the hidden same-state no-expert oracle; selector is the clean proposal-oracle headroom left by the deployed selector.}}{4}{table.2}\protected@file@percent }
+\newlabel{tab:car-decomposition}{{2}{4}{Causal Action Regret decomposition on the current six-task diagnostic. Support is the proposal-generation gap to the hidden same-state no-expert oracle; selector is the clean proposal-oracle headroom left by the deployed selector}{table.2}{}}
 \@writefile{toc}{\contentsline {section}{\numberline {7}Limitations}{4}{section.7}\protected@file@percent }
+\@writefile{lot}{\contentsline {table}{\numberline {3}{\ignorespaces Train-source score prior sensitivity for the current K6 transported residual field. A very small prior calibrates the selector; stronger priors tie or degrade success.}}{5}{table.3}\protected@file@percent }
+\newlabel{tab:source-score-sweep}{{3}{5}{Train-source score prior sensitivity for the current K6 transported residual field. A very small prior calibrates the selector; stronger priors tie or degrade success}{table.3}{}}
+\@writefile{lot}{\contentsline {table}{\numberline {4}{\ignorespaces Selector calibration ablations on the current K6 transported residual field. All rows are deployment-clean and use only train-split calibration signals.}}{5}{table.4}\protected@file@percent }
+\newlabel{tab:selector-calibration}{{4}{5}{Selector calibration ablations on the current K6 transported residual field. All rows are deployment-clean and use only train-split calibration signals}{table.4}{}}
 \gdef \@abspage@last{5}

workspace/latex/main.fdb_latexmk

@@ -1,5 +1,5 @@
 # Fdb version 4
-["pdflatex"] 1783013516 "main.tex" "main.pdf" "main" 1783013516 0
+["pdflatex"] 1783017127 "main.tex" "main.pdf" "main" 1783017127 0
   "/cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/enc/dvips/cm-super/cm-super-ts1.enc" 1136849721 2900 1537cc8184ad1792082cd229ecc269f4 ""
   "/cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/map/fontname/texfonts.map" 1577235249 3524 cb3e574dea2d1052e39280babc910dc8 ""
   "/cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/tfm/jknappen/ec/tcrm1000.tfm" 1136768653 1536 e07581a4bb3136ece9eeb4c3ffab8233 ""
@@ -34,6 +34,7 @@
   "/cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/cm/cmbx12.pfb" 1248133631 32080 340ef9bf63678554ee606688e7b5339d ""
   "/cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/cm/cmbx9.pfb" 1248133631 32298 c6d25bb16d1eac01ebdc6d7084126a1e ""
   "/cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/cm/cmcsc10.pfb" 1248133631 32001 6aeea3afe875097b1eb0da29acd61e28 ""
+  "/cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/cm/cmex10.pfb" 1248133631 30251 6afa5cb1d0204815a708a080681d4674 ""
   "/cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/cm/cmmi10.pfb" 1248133631 36299 5f9df58c2139e7edcf37c8fca4bd384d ""
   "/cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/cm/cmmi5.pfb" 1248133631 37912 77d683123f92148345f3fc36a38d9ab1 ""
   "/cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/cm/cmmi7.pfb" 1248133631 36281 c355509802a035cadc5f15869451dcee ""
@@ -106,9 +107,9 @@
   "/cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/web2c/texmf.cnf" 1692823885 39639 818dd8e5a0dce3a72597a7e6bad45d55 ""
   "/cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/var/lib/texmf/fonts/map/pdftex/updmap/pdftex.map" 1692821535 5031293 07973d5e761f645996ac32ebf7b8d1f3 ""
   "/cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/var/lib/texmf/web2c/pdftex/pdflatex.fmt" 1692821530 1386016 abd9ba6eb53c47bc5de8d672b6ac87d3 ""
-  "main.aux" 1783013516 3883 ff508ab370eaa8e9feeafda994006f38 "pdflatex"
-  "main.out" 1783013516 1245 b3d461b4fec67164f7152e09acaee494 "pdflatex"
-  "main.tex" 1783013512 9958 486b6a340f6a633503c9c60ae76b4c1d ""
+  "main.aux" 1783017127 3883 188704cafa467156903b53ef9791ac88 "pdflatex"
+  "main.out" 1783017127 1245 b3d461b4fec67164f7152e09acaee494 "pdflatex"
+  "main.tex" 1783017077 10821 d88c97be7ba8ff9760de11b80af38e29 ""
   "tables/car_decomposition.tex" 1783013498 662 bac24c0014c2ba573860ce199b776934 ""
   "tables/main_results.tex" 1783004655 1061 19bfdcc844614aadefc921dae15c700d ""
   "tables/selector_calibration.tex" 1783011198 788 1b54a8eb88519c50862f4acf718798ab ""

workspace/latex/main.fls

@@ -663,6 +663,7 @@ INPUT /cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fo
 INPUT /cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/cm/cmbx12.pfb
 INPUT /cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/cm/cmbx9.pfb
 INPUT /cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/cm/cmcsc10.pfb
+INPUT /cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/cm/cmex10.pfb
 INPUT /cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/cm/cmmi10.pfb
 INPUT /cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/cm/cmmi5.pfb
 INPUT /cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/cm/cmmi7.pfb

workspace/latex/main.log

@@ -1,4 +1,4 @@
-This is pdfTeX, Version 3.141592653-2.6-1.40.22 (TeX Live 2021 Gentoo Linux) (preloaded format=pdflatex 2023.8.23)  2 JUL 2026 13:31
+This is pdfTeX, Version 3.141592653-2.6-1.40.22 (TeX Live 2021 Gentoo Linux) (preloaded format=pdflatex 2023.8.23)  2 JUL 2026 14:32
 entering extended mode
  restricted \write18 enabled.
  %&-line parsing enabled.
@@ -483,19 +483,19 @@ File: umsb.fd 2013/01/14 v3.01 AMS symbols B
 
 {/cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/var/lib/texmf/fonts/map/pdf
 tex/updmap/pdftex.map}] [2] (./tables/main_results.tex) (./tables/car_decomposi
-tion.tex) [3] (./tables/source_score_sweep.tex)
-(./tables/selector_calibration.tex) [4] [5] (./main.aux)
+tion.tex) (./tables/source_score_sweep.tex)
+(./tables/selector_calibration.tex) [3] [4] [5] (./main.aux)
 Package rerunfilecheck Info: File `main.out' has not changed.
 (rerunfilecheck)             Checksum: B3D461B4FEC67164F7152E09ACAEE494;1245.
  ) 
 Here is how much of TeX's memory you used:
  9899 strings out of 480884
  146741 string characters out of 5900692
- 453545 words of memory out of 5000000
+ 458545 words of memory out of 5000000
  26980 multiletter control sequences out of 15000+600000
  414083 words of font info for 68 fonts, out of 8000000 for 9000
  36 hyphenation exceptions out of 8191
- 71i,8n,74p,371b,391s stack positions out of 5000i,500n,10000p,200000b,80000s
+ 71i,8n,74p,371b,393s stack positions out of 5000i,500n,10000p,200000b,80000s
 {/cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/font
 s/enc/dvips/cm-super/cm-super-ts1.enc}</cvmfs/soft.computecanada.ca/gentoo/2023
 /x86-64-v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/cm/cmbx10.pfb></cvm
@@ -504,29 +504,30 @@ fs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1
 -v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/cm/cmbx9.pfb></cvmfs/soft.
 computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/
 amsfonts/cm/cmcsc10.pfb></cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr
-/share/texmf-dist/fonts/type1/public/amsfonts/cm/cmmi10.pfb></cvmfs/soft.comput
+/share/texmf-dist/fonts/type1/public/amsfonts/cm/cmex10.pfb></cvmfs/soft.comput
 ecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/amsfon
-ts/cm/cmmi5.pfb></cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/t
-exmf-dist/fonts/type1/public/amsfonts/cm/cmmi7.pfb></cvmfs/soft.computecanada.c
-a/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/cm/cmr
-10.pfb></cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist
-/fonts/type1/public/amsfonts/cm/cmr12.pfb></cvmfs/soft.computecanada.ca/gentoo/
-2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/cm/cmr17.pfb></
-cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/ty
-pe1/public/amsfonts/cm/cmr7.pfb></cvmfs/soft.computecanada.ca/gentoo/2023/x86-6
-4-v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/cm/cmr8.pfb></cvmfs/soft.
-computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/
-amsfonts/cm/cmr9.pfb></cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/sh
-are/texmf-dist/fonts/type1/public/amsfonts/cm/cmsy10.pfb></cvmfs/soft.computeca
+ts/cm/cmmi10.pfb></cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/
+texmf-dist/fonts/type1/public/amsfonts/cm/cmmi5.pfb></cvmfs/soft.computecanada.
+ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/cm/cm
+mi7.pfb></cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dis
+t/fonts/type1/public/amsfonts/cm/cmr10.pfb></cvmfs/soft.computecanada.ca/gentoo
+/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/cm/cmr12.pfb><
+/cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/t
+ype1/public/amsfonts/cm/cmr17.pfb></cvmfs/soft.computecanada.ca/gentoo/2023/x86
+-64-v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/cm/cmr7.pfb></cvmfs/sof
+t.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/publi
+c/amsfonts/cm/cmr8.pfb></cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/
+share/texmf-dist/fonts/type1/public/amsfonts/cm/cmr9.pfb></cvmfs/soft.computeca
 nada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/
-cm/cmsy7.pfb></cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texm
-f-dist/fonts/type1/public/amsfonts/symbols/msbm10.pfb></cvmfs/soft.computecanad
-a.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/cm-super/sfr
-m1000.pfb>
-Output written on main.pdf (5 pages, 199333 bytes).
+cm/cmsy10.pfb></cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/tex
+mf-dist/fonts/type1/public/amsfonts/cm/cmsy7.pfb></cvmfs/soft.computecanada.ca/
+gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/symbols/
+msbm10.pfb></cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-
+dist/fonts/type1/public/cm-super/sfrm1000.pfb>
+Output written on main.pdf (5 pages, 209497 bytes).
 PDF statistics:
- 150 PDF objects out of 1000 (max. 8388607)
- 124 compressed objects within 2 object streams
+ 154 PDF objects out of 1000 (max. 8388607)
+ 127 compressed objects within 2 object streams
  21 named destinations out of 1000 (max. 500000)
  81 words of extra memory for PDF output out of 10000 (max. 10000000)
 

workspace/latex/main.pdf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:575ebbe026fa95190f2b69f456bb87115625d0774d9ec3250cd699f1484d02c6
+oid sha256:0191fdb33c2a848626b82b9c3ff8f04c07019d727321e0a7c1d73ca4c950e732
 size 199333

workspace/latex/main.tex

@@ -36,9 +36,10 @@ field, reaches 38.90\%. A diagnostic top-8 oracle over the same clean proposal
 prefix reaches 44.35\%, while the hidden same-state no-expert chart reaches
 56.99\%. This reveals the central bottleneck: current selector headroom is 5.45
 points, but the proposal support gap is 12.64 points. The next method step is
-therefore not another selector prior; it is a learned positive tangent
-generator that closes support gap without same-state validation rewards,
-expert proposals, or hidden simulator state at deployment.
+therefore not another selector prior; it is proposal support: first
+utility-weighted positive tangent retrieval, then a learned tangent generator
+that closes support gap without same-state validation rewards, expert
+proposals, or hidden simulator state at deployment.
 \end{abstract}
 
 \section{Introduction}
@@ -72,9 +73,10 @@ deployment leakage.
   selector calibration.
   \item We propose \atlas{}, an object-centric representation of local causal
   action geometry learned from counterfactual intervention charts.
-  \item We frame transported residual retrieval as Generator V0 and identify
-  learned positive tangent generation as the main path beyond the current
-  ceiling.
+  \item We frame transported residual retrieval as Generator V0, implement a
+  utility-weighted Generator V1 baseline that changes proposal geometry rather
+  than selector scores, and identify learned positive tangent generation as the
+  main path beyond the current ceiling.
   \item We report a six-task diagnostic showing a 9.16 point clean gain over
   h=16 behavior cloning, while exposing substantially larger proposal-generation
   headroom than selector headroom.
@@ -161,8 +163,25 @@ transport them around the current policy mean,
   \tilde a_{j,\alpha}=a_{\mathrm{base}}+\alpha
   (a^{\mathrm{train}}_j-a^{\mathrm{train}}_{\mathrm{anchor}}).
 \]
-This is useful evidence, not the final novelty. The intended \atlas{} generator
-learns
+Generator V1 is the minimal support-side intervention. Instead of changing the
+selector potential after candidates are formed, it changes the retrieved
+tangent field itself:
+\[
+  w_j(o,\ell)\propto
+  \exp\left(-d(z(o,\ell),z_j)/\tau_k+
+  \rho\,[U(y_j)-U(y_{\mathrm{anchor},j})]\right),
+\]
+\[
+  \bar\delta_{t}(o,\ell)=
+  \frac{\sum_{j:b_j=t}w_j(o,\ell)
+  (a^{\mathrm{train}}_j-a^{\mathrm{train}}_{\mathrm{anchor},j})}
+  {\sum_{j:b_j=t}w_j(o,\ell)}.
+\]
+This keeps the method in the same measured causal chart language: positive
+train interventions make the local transported tangent stronger, negative
+ones are suppressed, and no validation-state reward is read at deployment.
+V1 is still a diagnostic bridge, not the final novelty. The intended
+\atlas{} generator learns
 \[
   q_{\phi}(\delta a\mid o,\ell)
 \]
@@ -207,9 +226,9 @@ source priors to the role of diagnostic baselines.
 The full A*/Q1 paper should scale \bench{} from this six-task diagnostic to a
 large ManiSkill3 core benchmark, add one external long-horizon or embodiment
 benchmark, and include real robot near-miss recovery. Method development should
-proceed in the order implied by the decomposition: weighted residual retrieval
-as Generator V1, then a positive-tangent CVAE or flow-matching generator in
-object-centric spline tangent space, trained with pairwise/listwise utility
+proceed in the order implied by the decomposition: evaluate Generator V1,
+then replace retrieval with a positive-tangent CVAE or flow-matching generator
+in object-centric spline tangent space, trained with pairwise/listwise utility
 field losses and calibrated dominance. The acceptance bar is selected clean
 success above 47--50\% on the current tasks, generator top-8 oracle above 50\%,
 selector gap below 3 points, and a real near-miss recovery gain of at least

workspace/latex/tables/car_decomposition.tex

@@ -0,0 +1,15 @@
+\begin{table}[t]
+\centering
+\caption{Causal Action Regret decomposition on the current six-task diagnostic. Support is the proposal-generation gap to the hidden same-state no-expert oracle; selector is the clean proposal-oracle headroom left by the deployed selector.}
+\label{tab:car-decomposition}
+\small
+\begin{tabular}{@{}lrrrrrr@{}}
+\toprule
+Method & Base & Prop. oracle & Selected & State oracle & Support & Selector \\
+\midrule
+Current CIL-Atlas V0 & 29.74 & 44.35 & 38.90 & 56.99 & 12.64 & 5.45 \\
+\bottomrule
+\end{tabular}
+\vspace{2pt}
+\footnotesize Clean gain is 9.16 points; gap closed is 33.6\%; 65--75\% closure targets are 47.45--50.17.
+\end{table}