sync paper source after proxy gate refresh

latex/main.tex

@@ -9,12 +9,13 @@
 \usepackage{xcolor}
 \usepackage{hyperref}
 \hypersetup{
-  pdftitle={Counterfactual Action Atlas: Learning Local Causal Geometry for Vision-Language-Action Control},
+  pdftitle={Causal Tangent Transport: Learning Positive Do-Action Geometry from Same-State Counterfactual Charts},
   pdfauthor={DoVLA-CIL Working Draft},
-  pdfsubject={CIL-Atlas diagnostic draft with measured CTT rollout and utility-energy selector diagnostics},
+  pdfsubject={Causal Tangent Transport with same-state counterfactual chart evidence},
 }
 
-\title{Counterfactual Action Atlas: Learning Local Causal Geometry for Vision-Language-Action Control}
+\title{Causal Tangent Transport:\\
+Learning Positive Do-Action Geometry from Same-State Counterfactual Charts}
 \author{DoVLA-CIL Working Draft}
 \date{\today}
 
@@ -25,6 +26,7 @@
 \newcommand{\dovla}{\textsc{DoVLA}}
 \newcommand{\pptc}{\ensuremath{\mathrm{PPTC}}}
 \newcommand{\outcomeptr}{\ensuremath{\mathrm{OutcomePTR}}}
+\newcommand{\ncar}{\ensuremath{\mathrm{NCAR}}}
 
 \newtheorem{definition}{Definition}
 \newtheorem{theorem}{Theorem}
@@ -36,13 +38,14 @@
 \begin{abstract}
 Vision-language-action policies are usually trained from demonstrations that
 show what the robot did, but not what would have happened under nearby actions.
-We introduce the Counterfactual Action Atlas, a same-state interventional
-framework for measuring local do-action geometry and for auditing whether a
-deployment-clean generator actually reaches positive tangent support. A
-\bench{} chart restores the identical state and instruction, executes multiple
-action chunks, and measures which tangents cause recovery, progress, failure,
-collision, or success. The current six-task diagnostic shows why this object
-matters: direct h=16 behavior cloning reaches 29.74\% success,
+We study Causal Tangent Transport (\ctt{}), a deployment-clean generator that
+starts from measured train positive do-action tangents and transports them into
+a target same-state chart. The supporting \atlas{} framework measures the
+local causal geometry: a \bench{} chart restores the identical state and
+instruction, executes multiple action chunks, and records which tangents cause
+recovery, progress, failure, collision, or success. The current six-task
+diagnostic shows why this object matters: direct h=16 behavior cloning reaches
+29.74\% success,
 deployment-clean residual transport reaches 38.90\%, its top-8 proposal oracle
 reaches 44.35\%, and the hidden same-state no-expert chart reaches 56.99\%.
 This decomposes failure into proposal-support and selector gaps. The current
@@ -53,9 +56,22 @@ do-action tangent from a nearby source chart into the current target chart.
 Measured validation and test rollouts confirm the support story but not yet a
 final deployment method: with the visual-stat chart token, the held-out test
 proposal oracle reaches 51.39\% success and \outcomeptr{}@8 reaches 53.47\%.
-However, the score-only selected action remains below the measured base action,
-and the best train-calibrated learned dominance selector reaches only 31.25\%
-test success at 25.69\% coverage. We keep the metric boundary explicit:
+However, the score-only selected action remains below the measured base action.
+Under the clipped convention, the best train-calibrated learned dominance
+selector reaches 31.25\% test success at 25.69\% coverage. A bounded
+\texttt{tanh} execution diagnostic removes action-bound violations, and a
+train-calibrated context selector reaches 38.19\% selected test success, matching
+the tanh proposal oracle in micro success. The row-level selector-gap metric is
+still 12.50 points, and the proposal support itself is lower than the clipped
+visual-stat row. An explicit \texttt{env\_clip} diagnostic is action-bound-clean
+and preserves more proposal support. Increasing transported candidates from
+K=8 to K=16 raises the held-out test proposal oracle to 56.94\% and
+\outcomeptr{}@16 to 54.86\%, with zero action-bound unsafe rate, crossing the
+current support gate. The score-only selected action still reaches only
+27.78\%, below the 29.17\% base action. The best train-calibrated K=16
+selector reaches 35.42\% selected test success using train-only target--source
+RGB-stat chart compatibility, while still leaving a 24.31-point selector gap.
+We keep the metric boundary explicit:
 \outcomeptr{} is reported only after generated candidates are rolled out, while
 distance-only support diagnostics are \pptc{}. These runs are therefore
 diagnostic evidence for the Atlas thesis and the next selector/generator gate,
@@ -171,7 +187,10 @@ outcome vector
 \]
 for terminal success, dense progress, contact quality, safety violation,
 task-stage quality, smoothness, and recovery. A scalar utility $U(y_i)$ is used
-for ranking and CAR, while reports should keep the components visible.
+for ranking and CAR; reports keep the components visible.
+Metric scripts also report safety-label coverage. Missing or null
+\texttt{safety\_violation} fields are treated as unknown rather than safe, so an
+unsafe-rate number is reported only when safety labels are actually present.
 
 \subsection{Splits and Leakage Contract}
 
@@ -201,6 +220,27 @@ The current leakage audit over \texttt{data/cil\_charts/\{train,val,test\}}
 passes with zero violations. The train index contains 2,044 charts and 32,704
 rows; the complete split export contains 2,873 charts and 45,968 rows.
 
+\subsection{Benchmark Tracks and Branch Families}
+
+\bench{} is intended as a benchmark object, not only a private training set.
+It exposes four evaluator modes that all share the same chart primitive:
+Ranking Track gives participants an observation, instruction, and candidate
+actions and evaluates their ordering against hidden outcomes; Generation Track
+accepts generated action chunks and executes them from restored states;
+Deployment Track evaluates a closed-loop policy from initial task states; and
+Recovery Track starts from near-miss states and evaluates whether one proposed
+chunk causally recovers progress. Only the evaluator may read validation/test
+outcomes in all four modes.
+
+The current six-task export already records branch families such as base,
+expert, residual transport, random negative, wrong-direction, wrong-gripper,
+near-miss, and no-op branches. The intended \bench{}-Core schema generalizes
+these into base/stochastic-anchor branches, train-only expert branches,
+residual tangent branches, object-centric geometric branches, recovery
+tangents, negative anti-goal branches, and learned-generator branches. The
+method does not depend on these family labels at deployment; they define
+interventional coverage and ablation axes for learning local causal geometry.
+
 \section{Metrics}
 
 For one chart, let $a_b$ be the base action and let $U(\cdot)$ be measured
@@ -210,6 +250,16 @@ utility when outcomes are available. Branch Causal Action Regret is
   \qquad
   a^*_{\mathcal A}=\arg\max_{a\in\mathcal A} U(a).
 \]
+The normalized form reports the fraction of the base-to-oracle gap that remains
+unclosed:
+\[
+  \ncar =
+  \frac{U(a^*)-U(a_m)}
+       {|U(a^*)-U(a_b)|+\epsilon}.
+\]
+Evaluator artifacts report \ncar{} only on rows where the base-to-oracle gap is
+numerically meaningful; otherwise a tiny denominator can turn an uninformative
+chart into a misleadingly large normalized regret.
 
 \paragraph{Measured proposal metrics.}
 Generated candidates may be called successful only after they are rolled out or
@@ -231,6 +281,16 @@ Support Gap is
 \]
 These metrics are invalid for distance-only candidates.
 
+\paragraph{Safety coverage.}
+Measured rollout rows may include \texttt{safety\_violation} inside
+\texttt{base\_outcome} and \texttt{candidate\_outcomes}. The evaluator reports
+generated safety-label coverage, selected/base label-known rates, and unsafe
+rates over known labels only. The current implemented safety source is
+action-space validity: if a decoded chunk lies outside the ManiSkill action
+space before optional clipping, the row records
+\texttt{safety\_violation\_source=action\_bounds}. This is a deployment
+validity audit, not a collision/contact audit.
+
 \paragraph{Proxy geometry metrics.}
 When generated candidates have not been rolled out, this draft reports support
 geometry as \pptc{}:
@@ -285,20 +345,20 @@ a negative boundary penalty, and a cycle consistency term:
   + \lambda_-\max(0,m-d(\hat\xi_t,\Xi_t^-))
   + \lambda_c\|T_{\phi}(z_t,z_s,\hat\xi_t)-\xi_s^+\|_2^2.
 \]
-The first implementation uses exported base-action summaries as chart features.
-This is an engineering limitation, not the intended Atlas representation. The
-next chart export should include visual-language tokens, target/distractor
-object tokens, robot/contact-region tokens, and object-centric tangent frames.
-
-\paragraph{Deployment plan.}
-At test time, \atlas{} should sample or retrieve positive tangents, decode them
-into candidate action chunks, score their causal utility, and execute only if a
-calibrated lower confidence bound says the best tangent dominates the base
-action. The current \ctt{} implementation retrieves nearby train positive source
-tangents, transports them into the current chart, and decodes the public 21D
+The first implementation uses exported base-action summaries plus deterministic
+RGB/object statistics as chart features. This is an engineering limitation, not
+the full Atlas representation: the missing representation components are
+visual-language tokens, target/distractor object tokens, robot/contact-region
+tokens, and object-centric tangent frames.
+
+\paragraph{Deployment algorithm.}
+At test time, the implemented \ctt{} pipeline retrieves nearby train positive
+source tangents, transports them into the current chart, decodes the public 21D
 tangent summary as three residual keyframes with linear interpolation into an
-action chunk. This is an auditable engineering decoder, not a lossless
-reconstruction of the hidden branch action.
+action chunk, scores the candidates with the current utility field, and applies
+the calibrated dominance fallback when that rule is enabled. The decoder is an
+auditable engineering decoder, not a lossless reconstruction of the hidden
+branch action.
 
 \input{../paper/sections/theory}
 
@@ -326,8 +386,8 @@ reranking, barycentric chart synthesis, CVAE generators, and flow generators are
 diagnostic baselines. V1 does not beat V0. Negative-margin reranking does not
 replace local positive support. Raw-action CVAE and spline flow variants can be
 safe under NegativeNear, but collapse strict positive support. These failures
-motivate \ctt{}: positive support should be transported from measured positive
-chart tangents, with negative tangents defining boundaries, rather than sampled
+motivate \ctt{}: positive support is transported from measured positive chart
+tangents, with negative tangents defining boundaries, rather than sampled
 from ambient noise.
 
 Previous distance-only memory and local-atlas diagnostics must be read as
@@ -345,7 +405,7 @@ success.
 \caption{First residual \ctt{} proxy smoke on train self-target charts. This
 table is an artifact check, not validation/test performance and not
 \outcomeptr{}. The 0.20 NegativeNear value exceeds the current safety gate, so
-the run should not be claimed as method success.}
+the run is not claimed as method success.}
 \label{tab:ctt-smoke}
 \scriptsize
 \input{../runs/ctt_residual_smoke_proxy/table}
@@ -420,6 +480,209 @@ candidates underperform the measured base action on that chart. This negative
 smoke is useful because it protects the paper from reclassifying \pptc{} proxy
 evidence as outcome success.
 
+\begin{table}[t]
+\centering
+\caption{Action-bound audit for the leakage-audited RGB-reference chart DB.
+The bound is the ManiSkill \texttt{pd\_ee\_delta\_pose} action space,
+$[-1,1]^7$. This is an action-representation validity audit, not a
+collision/contact safety audit.}
+\label{tab:action-bound-audit}
+\scriptsize
+\input{../runs/action_bound_audit_rgb_refs/table}
+\end{table}
+
+Table~\ref{tab:action-bound-audit} exposes a previously hidden rollout
+confound: every base branch in the exported RGB-reference chart DB exceeds the
+nominal ManiSkill action bounds, and about 30.8\% of branch actions exceed the
+same bounds. The per-dimension audit shows the largest violations in the
+rotation/gripper-like dimensions, with a global max-fit scale of about 0.215
+for train/validation/test. The measured rollout evaluator now logs these
+pre-clipping violations and can run with \texttt{--disable-action-clipping} to
+compare faithful raw-action replay against clipped deployment execution. It also
+supports an explicit \texttt{--execution-action-scale} diagnostic so scaled
+raw-action replay is a separate, logged convention rather than an implicit fix.
+A one-chart \texttt{base\_context\_obs} smoke at scale 0.215 confirms 100\%
+action-bound label coverage and zero action-bound violations for base and
+generated candidates. Full scaled validation/test refresh jobs completed and
+are aggregated in
+Tables~\ref{tab:ctt-base-context-obs-val-scaled-rollout}
+and~\ref{tab:ctt-base-context-obs-test-scaled-rollout}. The global scale
+diagnostic rejects a naive ``scale everything'' fix: it removes base action-bound
+violations, but support and selected success drop sharply and about 20\% of
+generated candidates still violate action bounds. The rollout evaluator now also
+supports a bounded \texttt{--execution-action-transform tanh} decoder
+diagnostic, which maps decoded controls smoothly into finite action bounds
+before validity checks. Full tanh validation/test refresh jobs completed and
+are aggregated in
+Tables~\ref{tab:ctt-base-context-obs-val-tanh-rollout}
+and~\ref{tab:ctt-base-context-obs-test-tanh-rollout}; they have zero
+base/generated/selected action-bound violations under known labels, but lower
+proposal support than the clipped/no-clipping visual-stat rows. The full
+no-clipping
+validation/test refresh jobs completed and are aggregated in
+Tables~\ref{tab:ctt-base-context-obs-val-noclip-rollout}
+and~\ref{tab:ctt-base-context-obs-test-noclip-rollout}; the clipped rollout
+tables below therefore remain support/selector diagnostics under the current
+deployment clipping policy, not final safety evidence.
+
+\begin{table}[t]
+\centering
+\caption{No-clipping validation refresh for \texttt{base\_context\_obs}, K=8.
+This replays decoded raw actions from restored states and records action-bound
+validity labels before any clipping. The labels are action-space violations, not
+collision/contact outcomes.}
+\label{tab:ctt-base-context-obs-val-noclip-rollout}
+\scriptsize
+\resizebox{\linewidth}{!}{\input{../runs/ctt_base_context_obs_val_noclip_rollout_comparison/table}}
+\end{table}
+
+\begin{table}[t]
+\centering
+\caption{No-clipping held-out test refresh for \texttt{base\_context\_obs},
+K=8. The proposal oracle remains nonzero, but selected success does not improve
+over the raw-replay base and almost all known labels indicate action-space
+bound violations.}
+\label{tab:ctt-base-context-obs-test-noclip-rollout}
+\scriptsize
+\resizebox{\linewidth}{!}{\input{../runs/ctt_base_context_obs_test_noclip_rollout_comparison/table}}
+\end{table}
+
+The no-clipping refresh changes the measured action execution distribution but
+does not rescue the deployment claim. On validation, \outcomeptr{}@8 is 45.41\%,
+proposal-oracle success is 41.06\%, selected success is 22.71\%, and base
+success is 28.99\%. On held-out test, \outcomeptr{}@8 is 54.17\% and
+proposal-oracle success is 50.00\%, but selected success equals the raw-replay
+base at 25.00\%. Known action-bound violation rates are 97.58\%--100.00\% for
+selected/base rows and 98.13\%--98.44\% for generated rows. Thus raw-action
+replay confirms proposal support remains measurable while action scaling and
+selection remain blockers.
+
+\begin{table}[t]
+\centering
+\caption{Scaled raw-action validation refresh for
+\texttt{base\_context\_obs}, K=8, with \texttt{--execution-action-scale 0.215}
+and clipping disabled. This is a diagnostic convention suggested by the
+action-bound audit, not a final action representation.}
+\label{tab:ctt-base-context-obs-val-scaled-rollout}
+\scriptsize
+\resizebox{\linewidth}{!}{\input{../runs/ctt_base_context_obs_val_scaled0215_rollout_comparison/table}}
+\end{table}
+
+\begin{table}[t]
+\centering
+\caption{Scaled raw-action held-out test refresh for
+\texttt{base\_context\_obs}, K=8, with \texttt{--execution-action-scale 0.215}
+and clipping disabled. The global scale improves base action-bound validity but
+does not preserve proposal support.}
+\label{tab:ctt-base-context-obs-test-scaled-rollout}
+\scriptsize
+\resizebox{\linewidth}{!}{\input{../runs/ctt_base_context_obs_test_scaled0215_rollout_comparison/table}}
+\end{table}
+
+The scaled refresh is a negative diagnostic. On validation, base action-bound
+violations fall to zero, but \outcomeptr{}@8 drops to 27.54\%,
+proposal-oracle success to 12.08\%, selected success to 6.76\%, and generated
+action-bound violations remain 19.20\% over known labels. On held-out test,
+\outcomeptr{}@8 is 37.50\%, proposal-oracle success is 16.67\%, selected
+success is 9.03\%, base success is 22.92\%, and generated action-bound
+violations are 19.97\%. Therefore the next fix must be an action-representation
+or decoder correction, not a global post-hoc scale.
+
+\begin{table}[t]
+\centering
+\caption{Per-dimension train-max scaled validation refresh for
+\texttt{base\_context\_obs}, K=8, with clipping disabled. The scale vector is
+fit only from the train split action-bound audit.}
+\label{tab:ctt-base-context-obs-val-perdim-rollout}
+\scriptsize
+\resizebox{\linewidth}{!}{\input{../runs/ctt_base_context_obs_val_perdim_trainmax_rollout_comparison/table}}
+\end{table}
+
+\begin{table}[t]
+\centering
+\caption{Per-dimension train-max scaled held-out test refresh for
+\texttt{base\_context\_obs}, K=8, with clipping disabled. This diagnostic tests
+whether per-dimension max-fit scaling preserves more support than the global
+scale.}
+\label{tab:ctt-base-context-obs-test-perdim-rollout}
+\scriptsize
+\resizebox{\linewidth}{!}{\input{../runs/ctt_base_context_obs_test_perdim_trainmax_rollout_comparison/table}}
+\end{table}
+
+The per-dimension train-max scale is also negative. It is fit from train
+base-branch action bounds only and keeps the measured base action bound-valid,
+but generated candidates still violate action bounds on 27.42\% of validation
+rows and 30.30\% of held-out test rows. Support collapses further than under
+\texttt{tanh}: validation proposal-oracle success is 16.91\%, selected success
+is 9.18\%, and \outcomeptr{}@8 is 30.43\%; held-out test proposal-oracle
+success is 22.22\%, selected success is 11.81\%, and \outcomeptr{}@8 is
+36.11\%. This rejects another simple post-processing convention and leaves the
+decoder/action representation itself as the blocker.
+
+\begin{table}[t]
+\centering
+\caption{Explicit \texttt{env\_clip} validation refresh for
+\texttt{base\_context\_obs}, K=8, with clipping disabled. The declared decoder
+convention clips decoded controls to action-space bounds before validity checks,
+so action-bound labels measure the declared convention rather than silent
+simulator clipping.}
+\label{tab:ctt-base-context-obs-val-envclip-rollout}
+\scriptsize
+\resizebox{\linewidth}{!}{\input{../runs/ctt_base_context_obs_val_envclip_rollout_comparison/table}}
+\end{table}
+
+\begin{table}[t]
+\centering
+\caption{Explicit \texttt{env\_clip} held-out test refresh for
+\texttt{base\_context\_obs}, K=8, with clipping disabled. This convention is
+action-bound-clean and preserves more proposal support than bounded
+\texttt{tanh}, but score-only selection remains below base.}
+\label{tab:ctt-base-context-obs-test-envclip-rollout}
+\scriptsize
+\resizebox{\linewidth}{!}{\input{../runs/ctt_base_context_obs_test_envclip_rollout_comparison/table}}
+\end{table}
+
+The explicit \texttt{env\_clip} refresh is the strongest bounded-action support
+diagnostic so far. Generated, selected, and base action-bound violation rates
+are all 0.00\% with 100\% known labels. On validation, \outcomeptr{}@8 is
+50.24\%, proposal-oracle success is 41.55\%, selected success is 22.22\%, and
+base success is 27.54\%. On held-out test, \outcomeptr{}@8 is 50.69\%,
+proposal-oracle success is 47.92\%, selected success is 23.61\%, and base
+success is 29.17\%. Thus \texttt{env\_clip} largely removes the action-bound
+validity confound without the severe support collapse of per-dimension scaling
+or bounded \texttt{tanh}, but it exposes the remaining selector/dominance
+failure.
+
+\begin{table}[t]
+\centering
+\caption{Bounded \texttt{tanh} validation refresh for
+\texttt{base\_context\_obs}, K=8, with clipping disabled. This action convention
+maps decoded controls into finite action bounds before validity checks.}
+\label{tab:ctt-base-context-obs-val-tanh-rollout}
+\scriptsize
+\resizebox{\linewidth}{!}{\input{../runs/ctt_base_context_obs_val_tanh_rollout_comparison/table}}
+\end{table}
+
+\begin{table}[t]
+\centering
+\caption{Bounded \texttt{tanh} held-out test refresh for
+\texttt{base\_context\_obs}, K=8, with clipping disabled. The convention is
+action-bound-clean but selected success remains below the tanh base action.}
+\label{tab:ctt-base-context-obs-test-tanh-rollout}
+\scriptsize
+\resizebox{\linewidth}{!}{\input{../runs/ctt_base_context_obs_test_tanh_rollout_comparison/table}}
+\end{table}
+
+The bounded tanh refresh removes the action-bound validity confound in this
+diagnostic convention: generated, selected, and base action-bound violation
+rates are all 0.00\% with 100\% known labels on validation and held-out test.
+It does not by itself solve \ctt{} deployment. On validation,
+\outcomeptr{}@8 is 42.51\%, proposal-oracle success is 36.71\%, selected
+success is 21.26\%, and base success is 27.54\%. On held-out test,
+\outcomeptr{}@8 is 41.67\%, proposal-oracle success is 38.19\%, selected
+success is 25.00\%, and base success is 33.33\%. This makes tanh useful as an
+action-bound-clean diagnostic, but not as a support-preserving decoder fix.
+
 \begin{table}[t]
 \centering
 \caption{Measured residual \ctt{} rollout on 69 validation positive-support
@@ -461,6 +724,10 @@ points. The selector failure remains: selected success is 24.15\%, below the
 27.54\% base success, and the success selector gap grows to 16.43 points. Thus
 deterministic visual statistics help proposal support, but they do not solve
 deployment-clean action selection.
+The regenerated measured-rollout tables expose safety-label fields only for
+artifacts generated after the action-bound audit landed. Older aggregate tables
+therefore remain valid for success/support comparisons, but not for
+unsafe-contact measurement.
 
 \begin{table}[t]
 \centering
@@ -473,7 +740,7 @@ validation row.}
 \resizebox{\linewidth}{!}{\input{../runs/ctt_base_context_obj_val_rollout_comparison/table}}
 \end{table}
 
-Table~\ref{tab:ctt-base-context-obj-val-rollout} closes the loop on the
+Table~\ref{tab:ctt-base-context-obj-val-rollout} completes the
 object-layout proxy diagnostic. Despite passing the proxy gate, the measured
 proposal oracle is only 38.16\%, below the RGB-stat row's 40.58\%, while
 \outcomeptr{}@8 remains 50.24\% and selected success drops to 20.29\%. This is
@@ -483,9 +750,9 @@ missing visual-language or object-centric causal chart token.
 \begin{table}[t]
 \centering
 \caption{Measured residual \ctt{} rollout on 48 test positive-support charts
-across three train seeds, K=8. The generated proposal oracle crosses the
-internal 50\% support target, but the selected action fails because the current
-score/dominance rule chooses poor candidates.}
+across three train seeds, K=8. The generated proposal oracle exceeds 50\%, but
+the selected action fails because the current score/dominance rule chooses poor
+candidates.}
 \label{tab:ctt-test-rollout}
 \scriptsize
 \resizebox{\linewidth}{!}{\input{../runs/ctt_test_rollout_comparison/table}}
@@ -547,8 +814,8 @@ does not transfer as a reliable dominance certificate.
 \caption{Learned dominance fallback trained on validation measured rows and
 evaluated on held-out test rows. Features are deployment-visible candidate
 features: utility-energy scores, score margins to base, rank, and tangent
-norms. This improves over the base test success but remains far below the
-paper gate.}
+norms. This improves over the base test success but remains a partial selector
+diagnostic.}
 \label{tab:ctt-learned-dominance}
 \scriptsize
 \resizebox{\linewidth}{!}{\input{../runs/ctt_learned_dominance_val_to_test/table}}
@@ -559,7 +826,7 @@ can repair the selector without new rollouts. A ridge dominance model trained
 only on validation measured rows reaches 30.56\% held-out test selected success
 at 24.31\% coverage, improving over the 28.47\% base policy and the 22.22\%
 score-only selector. This is a real selector improvement, but it is still far
-from the 47--50\% target and leaves a 25.69-point success selector gap. The
+from a method-success claim and leaves a 25.69-point success selector gap. The
 evidence therefore narrows the bottleneck: transported proposals contain useful
 actions, and lightweight dominance helps, but the final method needs a stronger
 train-only utility-energy model and richer visual/object-centric chart tokens.
@@ -576,7 +843,8 @@ checkpoints are negative across three seeds: when
 
 \begin{table}[t]
 \centering
-\caption{Best validation-calibrated dominance diagnostic so far: learned
+\caption{Best clipped-convention validation-calibrated dominance diagnostic:
+learned
 context dominance over the measured \texttt{base\_context\_obs} visual-stat
 rollout rows. The calibrator is fit on validation measured rows only and
 evaluated on held-out test rows.}
@@ -594,6 +862,198 @@ dominance row. The gain is real but not sufficient: proposal-oracle success is
 remains 24.31 points. The remaining problem is still reliable deployment-clean
 dominance, not merely generating more nearby tangents.
 
+\begin{table}[t]
+\centering
+\caption{Bounded-tanh selector diagnostic over already measured candidates.
+The learned context+tangent selector is fit on validation tanh rows and
+evaluated once on held-out test tanh rows. It is action-bound-clean, but not a
+train-clean deployment selector.}
+\label{tab:ctt-base-context-obs-tanh-learned-dominance}
+\scriptsize
+\resizebox{\linewidth}{!}{\input{../runs/ctt_base_context_obs_learned_dominance_context_tangent_success_tanh_val_to_test/table}}
+\end{table}
+
+Table~\ref{tab:ctt-base-context-obs-tanh-learned-dominance} asks whether the
+bounded tanh action convention can be paired with a better selector. The
+context+tangent success calibrator reaches 35.42\% selected held-out test
+success at 10.42\% coverage, above the tanh base action at 33.33\% and the
+tanh score-only selector at 25.00\%. The result is useful but narrow: proposal
+oracle success is only 38.19\%, hidden chart oracle success is 72.92\%, and the
+success support gap is 36.81 points. It is therefore a bounded-action selector
+diagnostic, not the final Atlas method.
+
+\begin{table}[t]
+\centering
+\caption{Train-calibrated bounded-tanh selector diagnostic. Calibration uses
+train-split tanh measured rollout rows with same-chart and same-state source
+retrieval excluded, then evaluates once on held-out test tanh rows.}
+\label{tab:ctt-base-context-obs-tanh-learned-train-dominance}
+\scriptsize
+\resizebox{\linewidth}{!}{\input{../runs/ctt_base_context_obs_learned_dominance_context_success_tanh_train_to_test/table}}
+\end{table}
+
+Table~\ref{tab:ctt-base-context-obs-tanh-learned-train-dominance} removes the
+validation-calibration crutch. The train-calibrated context-success selector
+uses 432 train measured tanh rows for calibration and reaches 38.19\% selected
+success on held-out test at 18.06\% coverage. This matches the bounded-tanh
+proposal oracle and exceeds the tanh base action at 33.33\%. The interpretation
+is precise: under this bounded action convention, the current lightweight
+selector can match the proposal set in micro success, but the row-level
+success selector gap remains 12.50 points and CTT no longer generates enough
+successful proposals. Support, with residual selection error, remains the main
+blocker for this convention.
+
+\begin{table}[t]
+\centering
+\caption{Train-calibrated \texttt{env\_clip} selector diagnostic. Calibration
+uses train-split \texttt{env\_clip} measured rollout rows with same-chart and
+same-state source retrieval excluded, then evaluates once on held-out test
+\texttt{env\_clip} rows.}
+\label{tab:ctt-base-context-obs-envclip-learned-train-dominance}
+\scriptsize
+\resizebox{\linewidth}{!}{\input{../runs/ctt_base_context_obs_learned_dominance_basic_envclip_train_to_test/table}}
+\end{table}
+
+Table~\ref{tab:ctt-base-context-obs-envclip-learned-train-dominance} tests
+whether the better bounded proposal support translates through train-clean
+selection. It does not yet. The best current env-clip selector is the basic
+utility-margin ridge row: selected held-out test success reaches 31.94\% at
+22.22\% coverage, above the env-clip base action at 29.17\% but far below the
+47.92\% proposal oracle. The success selector gap remains 21.53 points. The
+method evidence therefore separates the action-convention fix from the
+remaining dominance problem: \texttt{env\_clip} preserves bounded proposal
+support, but current selector features do not exploit it.
+A train-clean nonlinear dominance sweep under the same \texttt{env\_clip}
+rows is also negative: the best row reaches 30.56\% selected held-out test
+success, below the ridge row. This suggests the current deployment-visible
+features, not only linear model capacity, are the selector bottleneck.
+
+\begin{table}[t]
+\centering
+\caption{Train-calibrated \texttt{env\_clip} source-evidence selector
+diagnostic. This selector may read train-only source-chart positive/negative
+tangent statistics because \ctt{} proposals are transported from measured train
+positive source tangents; it does not read validation/test outcomes.}
+\label{tab:ctt-base-context-obs-envclip-source-evidence-dominance}
+\scriptsize
+\resizebox{\linewidth}{!}{\input{../runs/ctt_base_context_obs_learned_dominance_source_envclip_train_to_test/table}}
+\end{table}
+
+Table~\ref{tab:ctt-base-context-obs-envclip-source-evidence-dominance} adds a
+CTT-specific source-evidence feature family: source positive count, source
+positive utility statistics, and the transported tangent's distance back to the
+train source chart's positive/negative tangent sets. This is deployment-clean
+because train chart outcomes are available to the CTT source memory, while
+target validation/test outcomes remain evaluator-only. The row improves
+selected held-out test success to 33.33\% at 23.61\% coverage, above the
+env-clip base action at 29.17\% and the prior env-clip ridge selector at
+31.94\%. The selector gap remains 20.14 points, so source evidence is useful
+but not enough; the remaining failure still calls for richer chart
+representation and dominance features.
+
+\paragraph{K=16 support extension.}
+The advisor's decomposition says the main scientific gap is proposal support,
+not only selector calibration. We therefore reran the same declared
+\texttt{env\_clip} convention with K=16 transported candidates. The train
+calibration, validation, and held-out test arrays all completed with three
+seeds and action-bound labels known for every base, selected, and generated
+candidate. The K=16 validation split reaches 46.38\% proposal-oracle success
+and \outcomeptr{}@16 of 53.62\%. The train-calibration split reaches 53.70\%
+proposal-oracle success and \outcomeptr{}@16 of 54.63\%. The held-out test row
+is the important support result: proposal-oracle success rises to 56.94\%,
+\outcomeptr{}@16 is 54.86\%, the success support gap falls to 20.14 points, and
+the generated action-bound unsafe rate remains 0.00\%.
+
+\begin{table}[t]
+\centering
+\caption{Held-out test \texttt{env\_clip} measured rollout at K=16.
+The table is generated by \texttt{scripts/eval\_metrics.py}; candidates are
+actually rolled out, so \outcomeptr{}, SupportGap, and SelectorRegret are
+measured rather than proxy quantities.}
+\label{tab:ctt-base-context-obs-envclip-k16-test-rollout}
+\scriptsize
+\resizebox{\linewidth}{!}{\input{../runs/ctt_base_context_obs_test_envclip_k16_rollout_comparison/table}}
+\end{table}
+
+Table~\ref{tab:ctt-base-context-obs-envclip-k16-test-rollout} is a stronger
+support result than the K=8 \texttt{env\_clip} row: the proposal oracle moves
+from 47.92\% to 56.94\% on the same held-out test charts while preserving
+action-bound validity. The score-only selector moves only from 23.61\% to
+27.78\%, still below the measured base action. This is exactly the failure mode
+the Causal Action Regret decomposition is meant to expose: CTT now finds many
+positive measured tangents, but the present utility selector often chooses the
+wrong one.
+
+\begin{table}[t]
+\centering
+\caption{Train-calibrated lower-confidence dominance fallback on the same
+held-out K=16 \texttt{env\_clip} measured rows. The conformal residual
+quantile and threshold are fit on train-calibration rows only. The artifact now
+reports action-bound unsafe execution and within-chart pairwise causal
+calibration error.}
+\label{tab:ctt-base-context-obs-envclip-k16-lcb-dominance}
+\scriptsize
+\resizebox{\linewidth}{!}{\input{../runs/ctt_base_context_obs_dominance_envclip_k16_train_to_test/table}}
+\end{table}
+
+Table~\ref{tab:ctt-base-context-obs-envclip-k16-lcb-dominance} is the clean
+Part-G decision-rule diagnostic for the K=16 support result. It executes a
+generated candidate only when a calibrated lower confidence bound on
+$F(a)-F(a_b)$ exceeds the learned threshold. The rule is action-bound clean
+under the available labels: safety label coverage is complete and unsafe
+execution rate is 0.00\%. It does not solve selection. The auto threshold
+covers 13.19\% of rows and reaches 27.78\% selected success; the fixed
+$\tau=0$ rerun falls back 88.89\% of the time and reaches the 29.17\% base
+rate. Both retain a 29.86-point success selector gap to the 56.94\% proposal
+oracle. The within-chart pairwise causal calibration error is 0.1633 over
+12{,}434 measured pairs, so this score-source LCB is a safety fallback
+diagnostic, not a reliable dominance certificate.
+
+\begin{table}[t]
+\centering
+\caption{Best current train-calibrated \texttt{env\_clip} K=16 selector
+diagnostic. Calibration uses only train-split K=16 measured rollout rows with
+same-chart and same-state source retrieval excluded, then evaluates once on
+held-out test K=16 rows.}
+\label{tab:ctt-base-context-obs-envclip-k16-learned-train-dominance}
+\scriptsize
+\resizebox{\linewidth}{!}{\input{../runs/ctt_base_context_obs_learned_dominance_chartcompat_obs_utility_task_envclip_k16_train_to_test/table}}
+\end{table}
+
+Table~\ref{tab:ctt-base-context-obs-envclip-k16-learned-train-dominance}
+tests whether a train-clean selector can exploit the improved K=16 support. It
+cannot yet, but the representation ablation now identifies a useful selector
+signal. The best row is a ridge selector over deployment-visible target--source
+RGB-stat chart compatibility, trained on utility-margin targets with Mondrian
+task-specific thresholds: 35.42\% selected success at 65.97\% coverage,
+compared with a 29.17\% base action and 56.94\% proposal oracle. The same
+artifact now reports pairwise causal calibration error over within-chart
+candidate contrasts; the held-out test error is 0.0150 over 12{,}434 measured
+pairs. Thus the remaining 24.31-point selector gap is not explained away by
+omitting the requested calibration metric. The previous basic
+success-weighted task-threshold row reached 32.64\%. RGB-stat
+compatibility with a success-weighted target reaches 33.33\%; object-layout
+compatibility reaches 31.25\%; and the combined RGB-stat/object-layout
+compatibility row ties 32.64\%. Source-evidence and nonlinear K=16 variants do
+not improve this: source utility-margin reaches 30.56\%, source success reaches
+30.56\%, context-source nonlinear ties 31.94\%, and task-scoped context/source
+variants are lower. We also tested a
+train-clean within-chart pairwise causal-ranking objective for the same ridge
+selector. Pairwise-only reaches 29.17\%, while hybrid pointwise-plus-pairwise
+basic, tangent, and source-evidence variants reach 31.25\%, 30.56\%, and
+29.17\%, respectively. Thus replacing the pointwise loss by a pairwise chart
+ranking loss is insufficient with the current features. We also fixed the
+nonlinear dominance diagnostic so chart-compatibility feature sets load the
+same deployment-visible target/source chart maps instead of zero-valued
+compatibility vectors. This does not improve the result: HGB classifier
+variants over RGB-stat compatibility reach 28.47\%, source+RGB-stat
+compatibility reaches 26.39\%, and the best regressor-only RGB-stat
+compatibility row reaches 32.64\% with pairwise calibration errors between
+0.0086 and 0.0222 for the rerun regressor diagnostics. The
+conclusion is sharper than the K=8 result: support is now strong enough to make
+the paper story credible, train-only visual chart compatibility is useful, but
+selector/utility energy is still the dominant remaining bottleneck.
+
 The feature-source audits turn this negative result into a concrete engineering
 target. In the original \texttt{data/cil\_charts} indexes, scene ids and
 instructions are present, but observation embeddings and raw observation
@@ -626,9 +1086,8 @@ train-calibrated learned selector reaches 31.25\% selected success at 25.69\%
 coverage, compared with 29.17\% measured base success and 51.39\% proposal
 oracle success. This is a small clean improvement over base, but it is lower
 than the validation-calibrated context diagnostic and leaves a 25.69-point
-success selector gap. The paper should therefore treat train-only dominance as
-evidence that the current selector is underpowered, not as a solved deployment
-method.
+success selector gap. We therefore treat train-only dominance as evidence that
+the current selector is underpowered, not as a solved deployment method.
 
 \begin{table}[t]
 \centering
@@ -679,20 +1138,34 @@ The current draft is backed by explicit artifacts:
   \texttt{scripts/slurm/eval\_ctt\_generated\_rollout.sbatch} implement the
   measured generated-candidate rollout path, including a self-source exclusion
   flag for train-split calibration and metadata loading for deployment-visible
-  chart features such as \texttt{base\_context\_obs}.
+  chart features such as \texttt{base\_context\_obs}. They also record
+  action-bound validity labels and support raw-action replay through
+  \texttt{--disable-action-clipping} plus explicit scaled replay through
+  \texttt{--execution-action-scale}; the same path logs bounded execution
+  diagnostics such as \texttt{--execution-action-transform tanh}.
+  \item \texttt{scripts/eval\_nonlinear\_dominance\_selector.py} shares the
+  chart-compatibility feature-loading path with the ridge selector, including
+  selector chart-feature mode and source split hashes, so nonlinear selector
+  diagnostics are leakage-audited rather than zero-feature fallbacks.
+  \item \texttt{scripts/audit\_action\_bounds.py} produces
+  \texttt{runs/action\_bound\_audit\_rgb\_refs}, which audits whether exported
+  chart actions lie inside the deployment action space before measured rollout.
   \item \texttt{scripts/build\_ctt\_rollout\_comparison.py} aggregates
   measured validation/test rollouts and reports selected success, proposal
-  oracle success, hidden chart oracle success, success support gap, and success
-  selector gap.
+  oracle success, hidden chart oracle success, \ncar{}, success support gap,
+  and success selector gap.
   \item \texttt{scripts/eval\_dominance\_selector.py} calibrates a dominance
-  fallback rule on validation measured rows and evaluates it on held-out test
+  fallback rule on calibration measured rows and evaluates it on held-out
   measured rows; it can use rollout row scores or recompute scores from a
-  train-only utility-energy checkpoint.
+  train-only utility-energy checkpoint, and its current artifacts report
+  action-bound unsafe execution, pairwise causal calibration ECE, and
+  coverage/fallback rates for the LCB decision rule.
   \item \texttt{scripts/eval\_learned\_dominance\_selector.py} trains a small
   validation- or train-calibrated dominance model over deployment-visible
   candidate features and evaluates it on held-out measured test rows; it also
-  logs feature/target ablations, including context metadata and tangent-code
-  variants, for selector diagnostics.
+  logs feature/target ablations, context metadata, tangent-code variants,
+  Mondrian task thresholds, and within-chart pairwise causal-ranking variants
+  for selector diagnostics.
   \item \texttt{scripts/eval\_nonlinear\_dominance\_selector.py} runs the
   train-calibrated nonlinear selector diagnostics in
   \texttt{runs/ctt\_base\_context\_obs\_nonlinear\_dominance\_*}.
@@ -700,22 +1173,47 @@ The current draft is backed by explicit artifacts:
   \texttt{scripts/build\_ctt\_proxy\_comparison.py} generate the local-atlas
   baseline and validation proxy gate table.
   \item \texttt{scripts/check\_tangent\_reconstruction.py} verifies that the
-  exported 21D tangent codes are deterministic summaries of \texttt{delta\_action}.
+  exported 21D tangent codes are deterministic summaries of
+  \texttt{delta\_action}; \texttt{runs/tangent\_reconstruction} and
+  \texttt{runs/tangent\_reconstruction\_rgb\_refs} each check 45{,}968 rows
+  with zero failures and log chart data/split hashes.
+  \item \texttt{scripts/audit\_cil\_charts.py} writes the leakage reports
+  \texttt{runs/leakage\_audit} and
+  \texttt{runs/leakage\_audit\_rgb\_refs}, both currently passing with no
+  violations or hash-artifact warnings.
   \item \texttt{scripts/train\_utility\_energy.py} and
   \texttt{scripts/calibrate\_dominance.py} implement the utility/scoring and
   dominance-calibration path.
   \item \texttt{scripts/summarize\_ctt\_runs.py} generates
-  \texttt{runs/summary\_ctt.csv} and \texttt{runs/summary\_ctt.md}.
-  \item \texttt{paper/notes/theory\_ctt.md} and
-  \texttt{paper/sections/theory.tex} state the theory obligations.
+  \texttt{runs/summary\_ctt.csv}; the persistent prose overview is consolidated
+  in \texttt{README.md}.
+  \item \texttt{scripts/audit\_ctt\_paper\_artifacts.py} audits forbidden
+  wording, paper table inputs, required implementation paths, and run artifact
+  contracts, writing \texttt{runs/paper\_ctt\_audit/audit.json} and a TeX table.
+  \item \texttt{scripts/backfill\_paper\_run\_artifacts.py} transparently
+  backfills non-Markdown run metadata such as missing grouped-metric placeholders,
+  config stubs, and log stubs for paper-referenced run dirs while preserving
+  existing files and not recreating deleted Markdown reports.
+  \item \texttt{paper/sections/theory.tex} states the theory obligations.
 \end{itemize}
 
+\begin{table}[t]
+\centering
+\caption{Claim-to-artifact audit for this draft. The audit is generated by
+\texttt{scripts/audit\_ctt\_paper\_artifacts.py}. Warnings track the advisor's
+full run-contract fields such as per-run Markdown reports and logs; the current
+workspace policy keeps persistent prose consolidated in \texttt{README.md}.}
+\label{tab:paper-ctt-artifact-audit}
+\scriptsize
+\input{../runs/paper_ctt_audit/table}
+\end{table}
+
 \section{Limitations and Next Steps}
 
 This draft is not yet a final deployment method. The current \ctt{} evidence
 includes validation and test measured generated-candidate rollouts, but the
-selected action fails the internal paper gate even when the test proposal oracle
-passes 50\%. The missing method component is calibrated dominance and a better
+selected action remains weak even when the clipped-convention test proposal
+oracle passes 50\%. The missing method component is calibrated dominance and a better
 utility selector, not another proxy-only support plot. The first standalone
 train-only utility-energy checkpoints and context-metadata ridge variants do
 not solve this transfer problem, so the chart token must move beyond base-action
@@ -723,17 +1221,28 @@ summaries and coarse task metadata toward exported visual-language and
 object-centric geometry. The RGB-reference export now provides a leakage-audited
 visual-stat token and improves measured support plus validation-calibrated
 selected success, but this does not qualify as the needed learned object-centric
-representation because the best held-out selected success is only 32.64\% and
-the selector gap remains large. Related work experiments, external
-benchmarks, real robot near-miss recovery, unsafe-contact measurement, and a
-dominance rule that approaches the internal success gate remain to be completed
-before submission.
+representation because the best K=16 \texttt{env\_clip} train-clean selected
+success is still only 35.42\% and the selector gap remains 24.31 points. The
+bounded tanh diagnostic now
+has a train-calibrated selector reaching 38.19\% selected success, but with lower
+proposal support than the clipped visual-stat row.
+The no-clipping refresh further shows that most
+decoded actions violate the nominal deployment action bounds before clipping, so
+action scaling/decoding must be resolved before safety or deployment claims are
+made. A global max-fit execution scale of 0.215 fixes base action-bound
+violations but collapses proposal support and leaves nonzero generated
+action-bound violations, so the next fix must be representational rather than a
+post-hoc scalar. The bounded tanh decoder is action-bound-clean on full
+validation/test, but it sacrifices too much support to be promoted as the
+decoder fix. Related work experiments, external benchmarks, real robot
+near-miss recovery, unsafe-contact measurement, and a stronger train-clean
+dominance rule remain to be completed before submission.
 
 The next experimental step is concrete: replace the current weak train-only
 utility-energy selector with visual-language/object-centric chart features,
-rerun held-out measured dominance selection, and add unsafe execution and
-fallback-rate metrics. Proxy evidence may open the rollout gate; it cannot
-replace rollout measurement.
+rerun held-out measured dominance selection, and collect actual unsafe-contact
+labels for the safety/fallback metrics now present in the evaluator. Proxy
+evidence may open the rollout gate; it cannot replace rollout measurement.
 
 \section{Conclusion}