diff --git a/scripts/bench-utils.mjs b/scripts/bench-utils.mjs
new file mode 100644
index 0000000..a124f3e
--- /dev/null
+++ b/scripts/bench-utils.mjs
@@ -0,0 +1,66 @@
+/**
+ * Shared utilities for benchmark formatting scripts.
+ */
+
+import { readFileSync } from "node:fs";
+
+export function formatTime(ns) {
+  if (ns >= 1e6) return `${(ns / 1e6).toFixed(2)} ms`;
+  if (ns >= 1e3) return `${(ns / 1e3).toFixed(2)} µs`;
+  return `${ns.toFixed(2)} ns`;
+}
+
+/**
+ * Classify a delta percentage (negative = experiment faster):
+ *   ⚪ within ±2% — indistinguishable from run-to-run noise
+ *   🟢 faster by more than 2%
+ *   🟠 slower by 2–5% — worth a look, not necessarily real
+ *   🔴 slower by 5% or more
+ */
+export function deltaEmoji(pct) {
+  const abs = Math.abs(pct);
+  if (abs < 2) return "⚪";
+  if (pct <= -5) return "🟢";
+  if (pct >= 5) return "🔴";
+  if (pct < 0) return "🟢";
+  return "🟠";
+}
+
+export const DELTA_LEGEND = "🟢 faster · 🔴 slower · 🟠 slightly slower · ⚪ within 2%";
+
+/**
+ * Parse benchmark JSON results into control/experiment pairs with deltas.
+ * Uses p50 (median) which is more robust to outliers than avg.
+ */
+export function parsePairs(json) {
+  const pairs = new Map();
+
+  for (const trial of json.benchmarks || []) {
+    for (const r of trial.runs || []) {
+      if (!r.stats) continue;
+      const m = r.name.match(/^(.+)\s+\((control|experiment)\)$/);
+      if (!m) continue;
+      const [, key, role] = m;
+      if (!pairs.has(key)) pairs.set(key, {});
+      pairs.get(key)[role] = r.stats;
+    }
+  }
+
+  const rows = [];
+  for (const [name, { control, experiment }] of pairs) {
+    if (!control || !experiment) continue;
+    const ctrlVal = control.p50 ?? control.avg;
+    const expVal = experiment.p50 ?? experiment.avg;
+    const delta = ((expVal - ctrlVal) / ctrlVal) * 100;
+    rows.push({ name, control: ctrlVal, experiment: expVal, delta });
+  }
+
+  return rows;
+}
+
+/**
+ * Read and parse the benchmark JSON results file.
+ */
+export function readBenchJSON(path) {
+  return JSON.parse(readFileSync(path, "utf8"));
+}
diff --git a/scripts/format-bench-cli.mjs b/scripts/format-bench-cli.mjs
index 5dc50ac..b169605 100644
--- a/scripts/format-bench-cli.mjs
+++ b/scripts/format-bench-cli.mjs
@@ -6,7 +6,7 @@
  *   BENCH_JSON_OUTPUT - Path to the JSON bench results
  */
 
-import { readFileSync } from "node:fs";
+import { DELTA_LEGEND, deltaEmoji, formatTime, parsePairs, readBenchJSON } from "./bench-utils.mjs";
 
 const jsonPath = process.env.BENCH_JSON_OUTPUT;
 
@@ -18,66 +18,13 @@ if (!jsonPath) {
 let json;
 
 try {
-  json = JSON.parse(readFileSync(jsonPath, "utf8"));
+  json = readBenchJSON(jsonPath);
 } catch (e) {
   console.error(`Could not read ${jsonPath}: ${e.message}`);
   process.exit(1);
 }
 
-function formatTime(ns) {
-  if (ns >= 1e6) return `${(ns / 1e6).toFixed(2)} ms`;
-  if (ns >= 1e3) return `${(ns / 1e3).toFixed(2)} µs`;
-
-  return `${ns.toFixed(2)} ns`;
-}
-
-function deltaEmoji(pct) {
-  const abs = Math.abs(pct);
-
-  if (abs < 1) return "⚪";
-  if (pct <= -5) return "🟢";
-  if (pct >= 5) return "🔴";
-
-  return "🟡";
-}
-
-// Group control/experiment pairs
-const pairs = new Map();
-
-for (const trial of json.benchmarks || []) {
-  for (const r of trial.runs || []) {
-    if (!r.stats) continue;
-
-    const m = r.name.match(/^(.+)\s+\((control|experiment)\)$/);
-
-    if (!m) continue;
-
-    const [, key, role] = m;
-
-    if (!pairs.has(key)) pairs.set(key, {});
-
-    pairs.get(key)[role] = r.stats;
-  }
-}
-
-if (pairs.size === 0) {
-  console.log("No comparison data found.");
-  process.exit(0);
-}
-
-// Build rows — use median (p50) which is far more robust to outliers from
-// CPU frequency scaling, GC pauses, and other system noise than the mean.
-const rows = [];
-
-for (const [name, { control, experiment }] of pairs) {
-  if (!control || !experiment) continue;
-
-  const ctrlVal = control.p50 ?? control.avg;
-  const expVal = experiment.p50 ?? experiment.avg;
-  const delta = ((expVal - ctrlVal) / ctrlVal) * 100;
-
-  rows.push({ name, control: ctrlVal, experiment: expVal, delta });
-}
+const rows = parsePairs(json);
 
 if (rows.length === 0) {
   console.log("No comparison data found.");
@@ -122,5 +69,5 @@ for (const row of rows) {
 }
 
 console.log();
-console.log("🟢 faster · 🔴 slower · 🟡 within 5% · ⚪ within 1%");
+console.log(DELTA_LEGEND);
 console.log();
diff --git a/scripts/format-bench-comment.mjs b/scripts/format-bench-comment.mjs
index 80b0eae..b5bfec0 100644
--- a/scripts/format-bench-comment.mjs
+++ b/scripts/format-bench-comment.mjs
@@ -13,6 +13,7 @@
  */
 
 import { readFileSync } from "node:fs";
+import { DELTA_LEGEND, deltaEmoji, formatTime, parsePairs, readBenchJSON } from "./bench-utils.mjs";
 
 const marker = "<!-- bench-compare -->";
 
@@ -43,77 +44,30 @@ const jsonPath = process.env.BENCH_JSON_OUTPUT;
 
 if (jsonPath) {
   try {
-    const json = JSON.parse(readFileSync(jsonPath, "utf8"));
-    summarySection = buildSummary(json);
+    const rows = parsePairs(readBenchJSON(jsonPath));
+
+    if (rows.length > 0) {
+      const tableRows = rows.map(({ name, control, experiment, delta }) => {
+        const emoji = deltaEmoji(delta);
+        const sign = delta > 0 ? "+" : "";
+        return `| ${emoji} | ${name} | ${formatTime(control)} | ${formatTime(experiment)} | ${sign}${delta.toFixed(1)}% |`;
+      });
+
+      summarySection = [
+        "",
+        "| | Benchmark | Control (p50) | Experiment (p50) | Δ |",
+        "|---|---|---:|---:|---:|",
+        ...tableRows,
+        "",
+        `> ${DELTA_LEGEND}`,
+        "",
+      ].join("\n");
+    }
   } catch {
     // JSON not available or malformed — skip summary
   }
 }
 
-function formatTime(ns) {
-  if (ns >= 1e6) return `${(ns / 1e6).toFixed(2)} ms`;
-  if (ns >= 1e3) return `${(ns / 1e3).toFixed(2)} µs`;
-  return `${ns.toFixed(2)} ns`;
-}
-
-function deltaEmoji(pct) {
-  const abs = Math.abs(pct);
-  // negative pct means experiment is faster (lower time = better)
-  if (abs < 1) return "⚪";
-  if (pct <= -5) return "🟢";
-  if (pct >= 5) return "🔴";
-  return "🟡";
-}
-
-function buildSummary(json) {
-  const benchmarks = json.benchmarks || [];
-
-  // In comparison mode, benchmarks come in pairs inside summary groups.
-  // Each benchmark alias is like "gts small (control)" / "gts small (experiment)".
-  // Group them by stripping the suffix.
-  const pairs = new Map();
-
-  for (const trial of benchmarks) {
-    for (const r of trial.runs || []) {
-      if (!r.stats) continue;
-      const m = r.name.match(/^(.+)\s+\((control|experiment)\)$/);
-      if (!m) continue;
-      const [, key, role] = m;
-      if (!pairs.has(key)) pairs.set(key, {});
-      pairs.get(key)[role] = r.stats;
-    }
-  }
-
-  if (pairs.size === 0) return "";
-
-  const rows = [];
-  for (const [name, { control, experiment }] of pairs) {
-    if (!control || !experiment) continue;
-    // Use p50 (median) — far more robust to GC pauses and noisy-neighbor
-    // spikes on shared CI runners than the mean.
-    const ctrlVal = control.p50 ?? control.avg;
-    const expVal = experiment.p50 ?? experiment.avg;
-    const delta = ((expVal - ctrlVal) / ctrlVal) * 100;
-    const emoji = deltaEmoji(delta);
-    const sign = delta > 0 ? "+" : "";
-    rows.push(
-      `| ${emoji} | ${name} | ${formatTime(ctrlVal)} | ${formatTime(expVal)} | ${sign}${delta.toFixed(1)}% |`,
-    );
-  }
-
-  if (rows.length === 0) return "";
-
-  return [
-    "",
-    "| | Benchmark | Control (p50) | Experiment (p50) | Δ |",
-    "|---|---|---:|---:|---:|",
-    ...rows,
-    "",
-    "> 🟢 faster · 🔴 slower · 🟡 within 5% · ⚪ within 1%",
-    "",
-  ].join("\n");
-}
-
 // ---------------------------------------------------------------------------
 // Assemble comment
 // ---------------------------------------------------------------------------