diff --git a/packages/daemon/package.json b/packages/daemon/package.json
index ac145f40..709aac97 100644
--- a/packages/daemon/package.json
+++ b/packages/daemon/package.json
@@ -20,6 +20,7 @@
     "dev:fetch-ids": "FULLNODE_WEBSOCKET_BASEURL=${FULLNODE_HOST} node ../../scripts/fetch-fullnode-ids.js",
     "replay-balance": "yarn dlx ts-node src/scripts/replay-balance.ts",
     "bench:sync": "yarn dlx ts-node src/scripts/bench-sync.ts",
+    "bench:compare": "yarn dlx ts-node src/scripts/bench-compare.ts",
     "watch": "tsc -w",
     "test_images_up": "docker compose -f ./__tests__/integration/scripts/docker-compose.yml up -d",
     "test_images_down": "docker compose -f ./__tests__/integration/scripts/docker-compose.yml down",
diff --git a/packages/daemon/src/scripts/bench-compare.ts b/packages/daemon/src/scripts/bench-compare.ts
new file mode 100644
index 00000000..6c1df6b5
--- /dev/null
+++ b/packages/daemon/src/scripts/bench-compare.ts
@@ -0,0 +1,261 @@
+/**
+ * Copyright (c) Hathor Labs and its affiliates.
+ *
+ * This source code is licensed under the MIT license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+/**
+ * Bench comparator.
+ *
+ * Reads two bench-results JSON files (produced by bench-sync.ts) and emits a
+ * markdown report comparing per-metric medians. A 95% confidence interval on
+ * the median delta is computed via bootstrap resampling — a single point
+ * estimate would be misleading at the run counts this harness produces.
+ *
+ * This tool is informational only. It does not exit non-zero on regression;
+ * CI runner variance is too high for a hard gate to be useful. Emit to
+ * stdout, pipe to `gh pr comment`.
+ *
+ * Usage:
+ *   yarn bench:compare --baseline bench-results-master.json --candidate bench-results-branch.json
+ *   yarn bench:compare --baseline a.json --candidate b.json --bootstrap 10000
+ */
+
+import { readFileSync } from 'node:fs';
+
+interface Summary {
+  n: number;
+  min: number;
+  p50: number;
+  p95: number;
+  max: number;
+  mean: number;
+  samples: number[];
+}
+
+interface BenchOutput {
+  opts: {
+    scenario: string;
+    port: number;
+    lastEvent: number;
+    runs: number;
+    warmup: number;
+    label: string;
+  };
+  totalMs: Summary | null;
+  perSpan: Record<string, Summary | null>;
+}
+
+interface Opts {
+  baseline: string;
+  candidate: string;
+  bootstrap: number;
+  seed: number;
+}
+
+function parseArgs(): Opts {
+  const args = process.argv.slice(2);
+  const opts: Opts = {
+    baseline: '',
+    candidate: '',
+    bootstrap: 10000,
+    seed: 42,
+  };
+  for (let i = 0; i < args.length; i++) {
+    const a = args[i];
+    const v = args[i + 1];
+    switch (a) {
+      case '--baseline': opts.baseline = v; i++; break;
+      case '--candidate': opts.candidate = v; i++; break;
+      case '--bootstrap': opts.bootstrap = parseInt(v, 10); i++; break;
+      case '--seed': opts.seed = parseInt(v, 10); i++; break;
+      case '--help':
+      case '-h':
+        console.log(`Usage: bench-compare --baseline <a.json> --candidate <b.json> [--bootstrap 10000] [--seed 42]`);
+        process.exit(0);
+        break;
+      default:
+        console.error(`Unknown arg: ${a}`);
+        process.exit(1);
+    }
+  }
+  if (!opts.baseline || !opts.candidate) {
+    console.error('Both --baseline and --candidate are required');
+    process.exit(1);
+  }
+  return opts;
+}
+
+function median(values: number[]): number {
+  if (values.length === 0) return NaN;
+  const sorted = [...values].sort((a, b) => a - b);
+  const mid = sorted.length >>> 1;
+  return sorted.length % 2 === 0 ? (sorted[mid - 1] + sorted[mid]) / 2 : sorted[mid];
+}
+
+/** Mulberry32 — small, deterministic PRNG. Good enough for bootstrap resampling. */
+function mulberry32(seed: number) {
+  let s = seed >>> 0;
+  return () => {
+    s = (s + 0x6d2b79f5) >>> 0;
+    let t = s;
+    t = Math.imul(t ^ (t >>> 15), t | 1);
+    t ^= t + Math.imul(t ^ (t >>> 7), t | 61);
+    return ((t ^ (t >>> 14)) >>> 0) / 4294967296;
+  };
+}
+
+function resample(values: number[], rng: () => number): number[] {
+  const out = new Array(values.length);
+  for (let i = 0; i < values.length; i++) {
+    out[i] = values[Math.floor(rng() * values.length)];
+  }
+  return out;
+}
+
+interface CompareResult {
+  metric: string;
+  baselineP50: number;
+  candidateP50: number;
+  deltaPct: number;
+  ciLowPct: number;
+  ciHighPct: number;
+  flag: '🟢' | '🔴' | '⚪' | '⚠️';
+  note?: string;
+}
+
+function compareSamples(
+  metric: string,
+  baseline: number[],
+  candidate: number[],
+  bootstrap: number,
+  rng: () => number,
+): CompareResult {
+  const baselineP50 = median(baseline);
+  const candidateP50 = median(candidate);
+
+  if (baselineP50 === 0 || Number.isNaN(baselineP50)) {
+    return {
+      metric,
+      baselineP50,
+      candidateP50,
+      deltaPct: NaN,
+      ciLowPct: NaN,
+      ciHighPct: NaN,
+      flag: '⚠️',
+      note: 'baseline median is 0/NaN',
+    };
+  }
+
+  const deltaPct = ((candidateP50 - baselineP50) / baselineP50) * 100;
+
+  // Bootstrap 95% CI on the median delta %.
+  const deltas = new Array(bootstrap);
+  for (let i = 0; i < bootstrap; i++) {
+    const b = median(resample(baseline, rng));
+    const c = median(resample(candidate, rng));
+    deltas[i] = b === 0 ? NaN : ((c - b) / b) * 100;
+  }
+  deltas.sort((a, b) => a - b);
+  const ciLowPct = deltas[Math.floor(0.025 * bootstrap)];
+  const ciHighPct = deltas[Math.floor(0.975 * bootstrap)];
+
+  let flag: CompareResult['flag'] = '⚪';
+  if (ciHighPct < 0) flag = '🟢';
+  else if (ciLowPct > 0) flag = '🔴';
+
+  return { metric, baselineP50, candidateP50, deltaPct, ciLowPct, ciHighPct, flag };
+}
+
+function fmtMs(v: number): string {
+  if (!Number.isFinite(v)) return 'n/a';
+  if (v >= 100) return v.toFixed(1);
+  if (v >= 10) return v.toFixed(2);
+  return v.toFixed(3);
+}
+
+function fmtPct(v: number): string {
+  if (!Number.isFinite(v)) return 'n/a';
+  const sign = v >= 0 ? '+' : '';
+  return `${sign}${v.toFixed(1)}%`;
+}
+
+function renderTable(results: CompareResult[]): string {
+  const lines = [
+    '| metric | baseline p50 (ms) | candidate p50 (ms) | Δ | 95% CI | |',
+    '|---|---:|---:|---:|---|:---:|',
+  ];
+  for (const r of results) {
+    const ci = Number.isFinite(r.ciLowPct) && Number.isFinite(r.ciHighPct)
+      ? `[${fmtPct(r.ciLowPct)}, ${fmtPct(r.ciHighPct)}]`
+      : 'n/a';
+    const note = r.note ? ` *(${r.note})*` : '';
+    lines.push(
+      `| ${r.metric} | ${fmtMs(r.baselineP50)} | ${fmtMs(r.candidateP50)} | ${fmtPct(r.deltaPct)} | ${ci} | ${r.flag}${note} |`,
+    );
+  }
+  return lines.join('\n');
+}
+
+function main() {
+  const opts = parseArgs();
+  const baseline: BenchOutput = JSON.parse(readFileSync(opts.baseline, 'utf8'));
+  const candidate: BenchOutput = JSON.parse(readFileSync(opts.candidate, 'utf8'));
+
+  if (baseline.opts.scenario !== candidate.opts.scenario) {
+    console.error(`⚠️  Scenario mismatch: baseline=${baseline.opts.scenario} candidate=${candidate.opts.scenario}`);
+    process.exit(1);
+  }
+
+  const rng = mulberry32(opts.seed);
+  const results: CompareResult[] = [];
+
+  if (baseline.totalMs && candidate.totalMs) {
+    results.push(
+      compareSamples('totalMs', baseline.totalMs.samples, candidate.totalMs.samples, opts.bootstrap, rng),
+    );
+  }
+
+  const allSpanNames = new Set([
+    ...Object.keys(baseline.perSpan),
+    ...Object.keys(candidate.perSpan),
+  ]);
+  const spanNames = [...allSpanNames].sort();
+  for (const name of spanNames) {
+    const b = baseline.perSpan[name];
+    const c = candidate.perSpan[name];
+    if (!b || !c) {
+      results.push({
+        metric: name,
+        baselineP50: b?.p50 ?? NaN,
+        candidateP50: c?.p50 ?? NaN,
+        deltaPct: NaN,
+        ciLowPct: NaN,
+        ciHighPct: NaN,
+        flag: '⚠️',
+        note: b ? 'missing in candidate' : 'missing in baseline',
+      });
+      continue;
+    }
+    results.push(compareSamples(name, b.samples, c.samples, opts.bootstrap, rng));
+  }
+
+  const greens = results.filter((r) => r.flag === '🟢').length;
+  const reds = results.filter((r) => r.flag === '🔴').length;
+  const noise = results.filter((r) => r.flag === '⚪').length;
+  const warn = results.filter((r) => r.flag === '⚠️').length;
+
+  console.log('## Sync benchmark comparison');
+  console.log('');
+  console.log(`**Scenario:** \`${baseline.opts.scenario}\` (${baseline.opts.lastEvent} events)  `);
+  console.log(`**Runs:** baseline=${baseline.opts.runs} (label: \`${baseline.opts.label}\`), candidate=${candidate.opts.runs} (label: \`${candidate.opts.label}\`), warmup=${baseline.opts.warmup}/${candidate.opts.warmup}  `);
+  console.log(`**Bootstrap samples:** ${opts.bootstrap}, seed: ${opts.seed}  `);
+  console.log(`**Verdict:** 🟢 ${greens} improvement${greens === 1 ? '' : 's'} · 🔴 ${reds} regression${reds === 1 ? '' : 's'} · ⚪ ${noise} noise · ⚠️ ${warn} skipped`);
+  console.log('');
+  console.log(renderTable(results));
+  console.log('');
+  console.log('_🟢/🔴 mean the 95% CI is fully on one side of 0. ⚪ means the CI crosses 0 — the difference is indistinguishable from noise at this run count. This report is informational only; CI runner variance makes hard gates unreliable at the run counts we can afford in CI._');
+}
+
+main();
diff --git a/packages/daemon/src/scripts/bench-sync.ts b/packages/daemon/src/scripts/bench-sync.ts
index 63c5a69c..727ad541 100644
--- a/packages/daemon/src/scripts/bench-sync.ts
+++ b/packages/daemon/src/scripts/bench-sync.ts
@@ -236,6 +236,9 @@ interface Summary {
   p95: number;
   max: number;
   mean: number;
+  /** Raw per-run values, retained so downstream tools (e.g. bench-compare)
+   *  can bootstrap confidence intervals on the median delta. */
+  samples: number[];
 }
 
 function summarize(values: number[]): Summary | null {
@@ -249,6 +252,7 @@ function summarize(values: number[]): Summary | null {
     p95: pick(0.95),
     max: sorted[sorted.length - 1],
     mean: values.reduce((a, b) => a + b, 0) / values.length,
+    samples: values,
   };
 }