diff --git a/paper/figures/fig_crossover_theory.pdf b/paper/tex/figures/fig_crossover_theory.pdf
similarity index 100%
rename from paper/figures/fig_crossover_theory.pdf
rename to paper/tex/figures/fig_crossover_theory.pdf
diff --git a/paper/figures/fig_regime_map.pdf b/paper/tex/figures/fig_regime_map.pdf
similarity index 99%
rename from paper/figures/fig_regime_map.pdf
rename to paper/tex/figures/fig_regime_map.pdf
index 20bb76a..8abdf45 100644
Binary files a/paper/figures/fig_regime_map.pdf and b/paper/tex/figures/fig_regime_map.pdf differ
diff --git a/paper/figures/fig_repair.pdf b/paper/tex/figures/fig_repair.pdf
similarity index 100%
rename from paper/figures/fig_repair.pdf
rename to paper/tex/figures/fig_repair.pdf
diff --git a/paper/figures/fig_repair_accuracy.pdf b/paper/tex/figures/fig_repair_accuracy.pdf
similarity index 100%
rename from paper/figures/fig_repair_accuracy.pdf
rename to paper/tex/figures/fig_repair_accuracy.pdf
diff --git a/paper/figures/fig_repair_kscaling.pdf b/paper/tex/figures/fig_repair_kscaling.pdf
similarity index 100%
rename from paper/figures/fig_repair_kscaling.pdf
rename to paper/tex/figures/fig_repair_kscaling.pdf
diff --git a/paper/tex/marc_aaai.tex b/paper/tex/marc_aaai.tex
index 77841d2..29b4868 100644
--- a/paper/tex/marc_aaai.tex
+++ b/paper/tex/marc_aaai.tex
@@ -352,10 +352,10 @@ \subsection{A factorization law predicts both results}
 instances per $n$ (Wilson intervals), with the same generators, polish, and checker as the
 solve-rate experiments; Appendix Table~\ref{tab:law} tabulates the dichotomy.
 
-% provenance: paper/figures/fig_crossover_theory.pdf, RESULTS.md R9
+% provenance: paper/tex/figures/fig_crossover_theory.pdf, RESULTS.md R9
 \begin{figure*}[t]
 \centering
-\includegraphics[width=0.62\textwidth]{../figures/fig_crossover_theory.pdf}
+\includegraphics[width=0.62\textwidth]{figures/fig_crossover_theory.pdf}
 \caption{The factorization law, measured. Left: $\log q(n)$ against $n$. The separable family
 is a line (slope $-1.03$, $R^2 = 0.98$), the coupled bilinear family is nearly flat
 ($-0.13$), and the geometry family, though syntactically coupled, collapses ($-0.77$,
@@ -414,13 +414,13 @@ \subsection{A factorization law predicts both results}
 first and ties; geometry has the first, fails the second, and ties. Figure~\ref{fig:regime}
 places every measured family on the two axes of the law.
 
-% provenance: paper/figures/fig_regime_map.pdf (scripts/plot_regime_map.py);
+% provenance: paper/tex/figures/fig_regime_map.pdf (scripts/plot_regime_map.py);
 % slopes from crossover_theory.json (R9), R27 slopes law-inverted from the LM arm of
 % crossover_families.json; outcomes from scaling.json (R15), crossover_families.json (R27),
 % coupled.json (R7), pointchain_learned.json (R25), real_systems.json (R26)
 \begin{figure}[ht]
 \centering
-\includegraphics[width=0.88\columnwidth]{../figures/fig_regime_map.pdf}
+\includegraphics[width=0.88\columnwidth]{figures/fig_regime_map.pdf}
 \caption{The regime map. Each measured family sits at its measured $\log q(n)$ slope
 (abscissa; the separable and coupled slopes are the fits of Table~\ref{tab:law}, the geometry slope that of Figure~\ref{fig:law}, the R27 slopes are
 inverted from the best-of-8 LM arm through Eq.~\eqref{eq:bestofk}) in its solution-structure
@@ -531,10 +531,10 @@ \subsection{Relocating the learned component: structural repair beats its contro
 \end{tabular}
 \end{table}
 
-% provenance: paper/figures/fig_repair_accuracy.pdf (scripts/plot_repair.py --panel left)
+% provenance: paper/tex/figures/fig_repair_accuracy.pdf (scripts/plot_repair.py --panel left)
 \begin{figure}[t]
 \centering
-\includegraphics[width=0.8\columnwidth]{../figures/fig_repair_accuracy.pdf}
+\includegraphics[width=0.8\columnwidth]{figures/fig_repair_accuracy.pdf}
 \caption{Table~\ref{tab:repair}, drawn: the ranker against its candidate-only and
 random controls; the dotted line is $K{=}4$ chance. Menu-size scaling is in Appendix
 Figure~\ref{fig:repair}.}
diff --git a/paper/tex/marc_aaai_appendix.tex b/paper/tex/marc_aaai_appendix.tex
index 96e7ed6..474ff03 100644
--- a/paper/tex/marc_aaai_appendix.tex
+++ b/paper/tex/marc_aaai_appendix.tex
@@ -138,10 +138,10 @@ \section{Repair: Menu-Size Scaling and Cost Accounting}
 (Figure~\ref{fig:repair}). Directly training at $K{=}16$ performs at chance, so the
 transferred checkpoint is selected and both negatives are kept in the record.
 
-% provenance: paper/figures/fig_repair_kscaling.pdf (scripts/plot_repair.py --panel right), RESULTS.md R10
+% provenance: paper/tex/figures/fig_repair_kscaling.pdf (scripts/plot_repair.py --panel right), RESULTS.md R10
 \begin{figure}[ht]
 \centering
-\includegraphics[width=0.6\textwidth]{../figures/fig_repair_kscaling.pdf}
+\includegraphics[width=0.6\textwidth]{figures/fig_repair_kscaling.pdf}
 \caption{Menu-size scaling behind main-text Figure~\ref{fig:repair-acc}: the $K{=}4$
 checkpoint evaluated zero-shot at larger menus retains an accuracy edge at $K{=}8$ that
 closes by $K{=}16$, while the blind enumeration it displaces grows from $2.5$ to $9.1$
diff --git a/scripts/plot_hard_eval.py b/scripts/plot_hard_eval.py
index 0cf3e2f..73b9880 100644
--- a/scripts/plot_hard_eval.py
+++ b/scripts/plot_hard_eval.py
@@ -3,7 +3,7 @@
 row; refine variants are labelled classical baselines (fixing-plan A2 house rule).
 
 Run:  python scripts/plot_hard_eval.py
-Writes paper/figures/fig_hard_suite.pdf and paper/figures/hard_suite_table.md.
+Writes paper/tex/figures/fig_hard_suite.pdf and paper/tex/figures/hard_suite_table.md.
 """
 import json
 from pathlib import Path
@@ -11,7 +11,7 @@
 from marc.eval.metrics import two_proportion_z
 
 SRC = Path("results/p_hard/hard_eval.json")
-FIGDIR = Path("paper/figures")
+FIGDIR = Path("paper/tex/figures")
 
 
 def _cell(m):
diff --git a/scripts/plot_regime_map.py b/scripts/plot_regime_map.py
index 17a2bc7..c1c1d55 100644
--- a/scripts/plot_regime_map.py
+++ b/scripts/plot_regime_map.py
@@ -151,7 +151,7 @@ def main():
               bbox_to_anchor=(0.01, 0.12))
 
     fig.tight_layout()
-    out = REPO / "paper" / "figures" / "fig_regime_map.pdf"
+    out = REPO / "paper" / "tex" / "figures" / "fig_regime_map.pdf"
     fig.savefig(out)
     print(f"wrote {out}")
     for name, slope, y, oc, *_ in pts:
diff --git a/scripts/plot_repair.py b/scripts/plot_repair.py
index 8f78064..ef13f6f 100644
--- a/scripts/plot_repair.py
+++ b/scripts/plot_repair.py
@@ -97,7 +97,7 @@ def main():
     args = ap.parse_args()
     labels, gen = load_generalization()
     ks, full, rand, calls = load_scaling()
-    figs = REPO / "paper" / "figures"
+    figs = REPO / "paper" / "tex" / "figures"
 
     if args.panel == "left":
         fig, ax = plt.subplots(figsize=(4.6, 3.4)); _accuracy(ax, labels, gen)
diff --git a/scripts/plot_results.py b/scripts/plot_results.py
index 88ca09d..c3fbe3d 100644
--- a/scripts/plot_results.py
+++ b/scripts/plot_results.py
@@ -2,7 +2,7 @@
 """Render P2 paper figures + summary table from ``results/p2_main/`` JSON.
 
 Reads the suite/ablation JSON written by ``run_main_eval.py`` and produces, under
-``paper/figures/``:
+``paper/tex/figures/``:
 
   * fig_generalization.pdf        — in-distribution vs. held-out solve rate + gap.
   * fig_perturbation.pdf          — solve rate vs. perturbation magnitude per split.
@@ -14,7 +14,7 @@
 
 Usage:
     python scripts/plot_results.py
-    python scripts/plot_results.py --results results/p2_main --figures paper/figures
+    python scripts/plot_results.py --results results/p2_main --figures paper/tex/figures
 """
 
 from __future__ import annotations
@@ -209,7 +209,7 @@ def build_summary_table(
 def main() -> None:
     parser = argparse.ArgumentParser(description="Render P2 figures + summary table")
     parser.add_argument("--results", default="results/p2_main")
-    parser.add_argument("--figures", default="paper/figures")
+    parser.add_argument("--figures", default="paper/tex/figures")
     args = parser.parse_args()
 
     res = Path(args.results)
diff --git a/scripts/run_crossover_theory.py b/scripts/run_crossover_theory.py
index 184cb97..d57dc4f 100644
--- a/scripts/run_crossover_theory.py
+++ b/scripts/run_crossover_theory.py
@@ -49,11 +49,11 @@
 3. Predicts P_random(n;K) = 1-(1-q(n))^K and compares to the observed random-restart
    rates in results/p_scaling/scaling.json and results/p_coupled/coupled.json.
 4. Predicts the crossover n* from v and the learned ceiling p_L, compares to observed.
-5. Writes results/p_crossover/crossover_theory.json and paper/figures/fig_crossover_theory.pdf.
+5. Writes results/p_crossover/crossover_theory.json and paper/tex/figures/fig_crossover_theory.pdf.
 6. (--k-sweep) Optionally sweeps the restart budget K to trace the *restart curve* as
    solve-rate vs wall-clock: at each K the mean wall-clock and restarts a best-of-K run
    spends (it stops at first success) against the rate it reaches. Writes the curve to the
-   JSON payload and paper/figures/fig_restart_curve.pdf. The single-K measure also now
+   JSON payload and paper/tex/figures/fig_restart_curve.pdf. The single-K measure also now
    records wall-clock per instance (``random_mean_ms``) so cost is reported alongside rate.
 
 Run:  python scripts/run_crossover_theory.py [--trials 300] [--K 8]
@@ -287,7 +287,7 @@ def main() -> None:
     ap.add_argument("--k-sweep", type=str, default="",
                     help="comma list of restart budgets K to sweep, e.g. '1,2,4,8,16,32' — "
                          "traces the restart curve (solve-rate vs wall-clock) into the JSON "
-                         "payload and paper/figures/fig_restart_curve.pdf")
+                         "payload and paper/tex/figures/fig_restart_curve.pdf")
     args = ap.parse_args()
 
     indep = measure_family("indep", NS_INDEP, INDEP_START, args.trials, args.K, args.seed)
@@ -449,7 +449,7 @@ def _plot(indep, coupled, v, K, n_star, learned_indep, geometry=None) -> None:
     ax[1].set_title("(b) parameter-free prediction of the crossover")
     ax[1].legend(fontsize=7, loc="center right")
     fig.tight_layout()
-    d = Path("paper/figures"); d.mkdir(parents=True, exist_ok=True)
+    d = Path("paper/tex/figures"); d.mkdir(parents=True, exist_ok=True)
     fig.savefig(d / "fig_crossover_theory.pdf")
     print(f"wrote {d/'fig_crossover_theory.pdf'}")
 
@@ -479,7 +479,7 @@ def _plot_restart_curve(sweeps, Ks) -> None:
         a.set_title(f"restart curve — {fam}  (K∈[{min(Ks)},{max(Ks)}])")
         a.legend(fontsize=7)
     fig.tight_layout()
-    d = Path("paper/figures"); d.mkdir(parents=True, exist_ok=True)
+    d = Path("paper/tex/figures"); d.mkdir(parents=True, exist_ok=True)
     fig.savefig(d / "fig_restart_curve.pdf")
     print(f"wrote {d/'fig_restart_curve.pdf'}")
 
diff --git a/scripts/run_dimension_scaling.py b/scripts/run_dimension_scaling.py
index bd2cf1a..00d48b2 100644
--- a/scripts/run_dimension_scaling.py
+++ b/scripts/run_dimension_scaling.py
@@ -30,7 +30,7 @@
 rate carries a 95% Wilson CI; learned-vs-baseline comparisons carry two-proportion
 z-test p-values (house rules, paper/RESULTS.md).
 
-Outputs: results/p_scaling/scaling.json and paper/figures/fig_dimension_scaling.pdf.
+Outputs: results/p_scaling/scaling.json and paper/tex/figures/fig_dimension_scaling.pdf.
 Run:  python scripts/run_dimension_scaling.py [--quick] [--seeds N]
 """
 from __future__ import annotations
@@ -249,7 +249,7 @@ def _plot(rows) -> None:
     ax.set_title("Learned inference vs. classical refinement & prior (95% CI)")
     ax.legend(fontsize=8, loc="upper right")
     fig.tight_layout()
-    fig_dir = Path("paper/figures")
+    fig_dir = Path("paper/tex/figures")
     fig_dir.mkdir(parents=True, exist_ok=True)
     fig.savefig(fig_dir / "fig_dimension_scaling.pdf")
     print(f"wrote {fig_dir/'fig_dimension_scaling.pdf'}")
diff --git a/scripts/run_overnight.py b/scripts/run_overnight.py
index 006c57c..1c249b0 100644
--- a/scripts/run_overnight.py
+++ b/scripts/run_overnight.py
@@ -61,7 +61,7 @@
 ]
 
 # Dirs scanned after each phase to attribute freshly written outputs.
-OUTPUT_DIRS = ["results", "checkpoints", "paper/figures"]
+OUTPUT_DIRS = ["results", "checkpoints", "paper/tex/figures"]
 
 
 # ------------------------------------------------------------------- utilities