diff --git a/docs/index.rst b/docs/index.rst
index c59b425..554ce24 100644
--- a/docs/index.rst
+++ b/docs/index.rst
@@ -176,11 +176,11 @@ LitXBench
        <td><a href="https://github.com/hasan-sayeed/KnowMat2">KnowMat2 (GPT-5.2 High)</a></td>
        <td>0.52</td>
        <td>0.43</td>
-       <td>0.43 &plusmn; 0.29</td>
-       <td>0.28</td>
-       <td>0.66</td>
-       <td>0.66</td>
-       <td>0.19</td>
+       <td>0.43 &plusmn; 0.06</td>
+       <td>0.27</td>
+       <td>0.67</td>
+       <td>0.65</td>
+       <td>0.17</td>
        <td>&mdash;</td>
        <td>19.40</td>
        <td class="links-cell"><a href="https://github.com/curtischong/KnowMat2">code</a> <a href="https://github.com/curtischong/KnowMat2/tree/master/std_1_high/knowmat2">run</a> <a href="https://chemrxiv.org/engage/chemrxiv/article-details/6902772ea482cba122c41c14">paper</a> <a href="https://github.com/Radical-AI/litxbench/pull/4">pr</a></td>
diff --git a/scripts/paper/combined_results/process_f1_summary.csv b/scripts/paper/combined_results/process_f1_summary.csv
index cf42329..c06aeb8 100644
--- a/scripts/paper/combined_results/process_f1_summary.csv
+++ b/scripts/paper/combined_results/process_f1_summary.csv
@@ -1,5 +1,5 @@
 method	process_f1
-KnowMat2	0.6592 +/- 0.6302
+KnowMat2	0.6734 +/- 0.1376
 Claude Haiku 4.5	0.8426 +/- 0.0586
 GPT 5 Mini Medium	0.8422 +/- 0.0698
 Gemini 3 Flash	0.8560 +/- 0.0197
diff --git a/scripts/paper/combined_results/zero_shot_summary_with_ci.csv b/scripts/paper/combined_results/zero_shot_summary_with_ci.csv
index e36c866..fcfda77 100644
--- a/scripts/paper/combined_results/zero_shot_summary_with_ci.csv
+++ b/scripts/paper/combined_results/zero_shot_summary_with_ci.csv
@@ -1,5 +1,5 @@
 method	overall_precision	overall_recall	overall_f1	meas_f1	process_f1	material_f1	config_f1	avg_attempts	cost_usd
-KnowMat2	0.5231	0.4286	0.4320 +/- 0.2916	0.2758	0.6592	0.6589	0.1876	1.00	0
+KnowMat2	0.5245	0.4258	0.4288 +/- 0.0586	0.2666	0.6734	0.6511	0.1745	1.00	0
 Claude Haiku 4.5	0.6417	0.6862	0.6510 +/- 0.0130	0.5089	0.8426	0.9380	0.3840	2.21	1.7177
 GPT 5 Mini Medium	0.6706	0.7023	0.6764 +/- 0.0434	0.5222	0.8422	0.9422	0.4051	2.49	3.46597
 Gemini 3 Flash	0.7404	0.7560	0.7376 +/- 0.0489	0.6092	0.8560	0.9692	0.5166	2.58	1.73078
diff --git a/scripts/paper/pareto_front.py b/scripts/paper/pareto_front.py
index c2c9bf8..baf14ea 100644
--- a/scripts/paper/pareto_front.py
+++ b/scripts/paper/pareto_front.py
@@ -1,4 +1,4 @@
-# %%
+import math
 import re
 from dataclasses import dataclass
 
@@ -20,16 +20,16 @@ class ModelPoint:
 
 
 models: tuple[ModelPoint, ...] = (
-    ModelPoint("Gemini CLI (Gemini-3.1 Pro Preview)", 0.7952, 6.45507, GREEN, "s"),
-    ModelPoint("Gemini 3.1 Pro", 0.7701, 4.17163, BLUE, "s"),
-    ModelPoint("Gemini 3 Flash", 0.7371, 1.73078, RED, "s"),
-    ModelPoint("Claude Code (Opus 4.6)", 0.7754, 26.1145, GREEN, "o"),
-    ModelPoint("Claude Opus 4.6", 0.7203, 5.37406, BLUE, "o"),
-    ModelPoint("Claude Haiku 4.5", 0.6488, 1.7177, RED, "o"),
-    ModelPoint("Codex (GPT 5.2 Codex High)", 0.7246, 4.17361, GREEN, "^"),
-    ModelPoint("GPT 5.2 High", 0.7244, 4.99104, BLUE, "^"),
-    ModelPoint("GPT 5 Mini Med.", 0.6731, 3.46597, RED, "^"),
-    ModelPoint("KnowMat2", 0.4320, 19.40, GREEN, "p"),
+    ModelPoint("Gemini CLI (Gemini-3.1 Pro Preview)", 0.7958, 6.45507, GREEN, "s"),
+    ModelPoint("Gemini 3.1 Pro", 0.7708, 4.17163, BLUE, "s"),
+    ModelPoint("Gemini 3 Flash", 0.7376, 1.73078, RED, "s"),
+    ModelPoint("Claude Code (Opus 4.6)", 0.7766, 26.1145, GREEN, "o"),
+    ModelPoint("Claude Opus 4.6", 0.7231, 5.37406, BLUE, "o"),
+    ModelPoint("Claude Haiku 4.5", 0.6510, 1.7177, RED, "o"),
+    ModelPoint("Codex (GPT 5.2 Codex High)", 0.7268, 4.17361, GREEN, "^"),
+    ModelPoint("GPT 5.2 High", 0.7273, 4.99104, BLUE, "^"),
+    ModelPoint("GPT 5 Mini Med.", 0.6764, 3.46597, RED, "^"),
+    ModelPoint("KnowMat2", 0.4288, 19.40, GREEN, "p"),
 )
 
 
@@ -44,9 +44,24 @@ def pareto_front(points: tuple[ModelPoint, ...]) -> tuple[ModelPoint, ...]:
     return tuple(pareto_points)
 
 
-pareto = pareto_front(models)
-pareto_cost = [point.cost for point in pareto]
-pareto_f1 = [point.f1 for point in pareto]
+def upper_convex_front(points: tuple[ModelPoint, ...]) -> tuple[ModelPoint, ...]:
+    front: list[ModelPoint] = []
+
+    def slope(left: ModelPoint, right: ModelPoint) -> float:
+        x_left = math.log10(left.cost)
+        x_right = math.log10(right.cost)
+        return (right.f1 - left.f1) / (x_right - x_left)
+
+    for point in pareto_front(points):
+        while len(front) >= 2 and slope(front[-2], front[-1]) <= slope(front[-1], point):
+            front.pop()
+        front.append(point)
+    return tuple(front)
+
+
+convex_front = upper_convex_front(models)
+front_cost = [point.cost for point in convex_front]
+front_f1 = [point.f1 for point in convex_front]
 
 plt.rcParams.update(
     {
@@ -79,7 +94,7 @@ def clean_text(text: str) -> str:
 fig, ax = plt.subplots(figsize=(8, 8))
 ax.set_box_aspect(1)
 
-ax.plot(pareto_cost, pareto_f1, color=LINE_C, linewidth=3.2, linestyle="--", zorder=2, label="Pareto front")
+ax.plot(front_cost, front_f1, color=LINE_C, linewidth=3.2, linestyle="--", zorder=2, label="Convex Pareto front")
 
 for point in models:
     ax.scatter(
@@ -131,4 +146,3 @@ def clean_text(text: str) -> str:
 plt.savefig("pareto_front.pdf", bbox_inches="tight")  # vector output
 plt.savefig("pareto_front.png", bbox_inches="tight", dpi=1200)  # high-res raster
 plt.show()
-# %%