diff --git a/README.md b/README.md
index 0cbd5a2..8d85510 100644
--- a/README.md
+++ b/README.md
@@ -902,6 +902,16 @@ pnpm benchmark:request-local
 
 The command builds `dist` before reporting ten scenarios: sequential request-local hits, sequential process-local hits, enabled bounded fallbacks, request-local coalescing fan-out, process coalescing fan-out, remote-read-deadline coalescing fan-out, tracked Redis hits with shadow omitted, tracked Redis hits deterministically outside a partial shadow ramp, a ramped-down warm-hit confirmation, and a ramped-down clean-miss fill. Both shadow scenarios prove that the caller completes before detached Redis work. The benchmark is a maintainer tool and is not included in the published package. It asserts fallback counts, Redis behavior, coalescing state, timer cleanup, returned values, exactly-once SoT reuse, and conditional confirmation/fill without applying a timing threshold. Override its work sizes with `DIALCACHE_BENCH_ITERATIONS` and `DIALCACHE_BENCH_FANOUT`.
 
+### Redis write benchmark
+
+With a Redis reachable at `REDIS_URL` (default `redis://127.0.0.1:6379`, e.g. `docker run --rm -p 6379:6379 redis:6.2`), measure the local build's write path:
+
+```bash
+pnpm benchmark:redis-write
+```
+
+The command builds `dist`, then runs sequential tracked and untracked writes at 100 B, 10 KiB, 100 KiB, and 1 MiB payloads, reporting server-side command cost per write from `INFO commandstats` (the `EVALSHA` entry envelopes the stamp script's internal calls) and client-side p50/p95 latency. Like the cache-path benchmark it is a maintainer tool, is not part of the published package, and asserts no timing thresholds — absolute numbers depend on the machine, engine, and load, so compare runs only within one environment. Scale iteration counts with `DIALCACHE_BENCH_WRITE_SCALE`.
+
 ### Releasing
 
 Publishing starts by manually running the `Release` workflow from current `main`. After the package checks pass, Semantic Release selects the next version from Conventional Commits since the highest stable `vX.Y.Z` tag. While the package is pre-1.0, breaking changes bump minor — their `BREAKING CHANGE:` footers still drive full release notes without forcing 1.0.0 — `feat` bumps minor, and every other normal PR-title type (`fix`, `perf`, `docs`, `style`, `refactor`, `test`, `build`, `chore`, `ci`, and `revert`) bumps patch. The highest required bump wins. Major bumps return when 1.0.0 is cut; `release.config.mjs` implements this table and must change together with this section.
diff --git a/package.json b/package.json
index 9df31aa..343db59 100644
--- a/package.json
+++ b/package.json
@@ -77,6 +77,7 @@
   ],
   "scripts": {
     "benchmark:request-local": "pnpm build && node scripts/benchmark-request-local.mjs",
+    "benchmark:redis-write": "pnpm build && node scripts/benchmark-redis-write.mjs",
     "build": "tsup src/index.ts src/datadog.ts src/node-redis.ts src/prometheus.ts src/redis-protocol.ts src/valkey-glide.ts --format esm,cjs --dts --clean",
     "check": "pnpm typecheck && pnpm test && pnpm build && pnpm test:package",
     "typecheck": "tsc --noEmit",
diff --git a/scripts/benchmark-redis-write.mjs b/scripts/benchmark-redis-write.mjs
new file mode 100644
index 0000000..7a05fe7
--- /dev/null
+++ b/scripts/benchmark-redis-write.mjs
@@ -0,0 +1,112 @@
+// Maintainer benchmark for the Redis write path. Measures the local build's
+// tracked and untracked writes against a live Redis and reports server-side
+// command cost per write (INFO commandstats; the EVALSHA entry envelopes
+// script-internal calls) alongside client-side latency percentiles. It
+// asserts nothing and applies no timing thresholds: absolute numbers are
+// machine-, engine-, and load-dependent, so compare runs only against the
+// same environment.
+//
+// Requires a reachable Redis, e.g.: docker run --rm -p 6379:6379 redis:6.2
+// Usage: pnpm benchmark:redis-write            (REDIS_URL to override)
+// DIALCACHE_BENCH_WRITE_SCALE scales iteration counts (default 1).
+import { createClient } from "redis";
+
+import { createNodeRedisDialCacheClient, dialcacheRedisScripts } from "../dist/node-redis.js";
+
+const REDIS_URL = process.env.REDIS_URL ?? "redis://127.0.0.1:6379";
+const SCALE = Number(process.env.DIALCACHE_BENCH_WRITE_SCALE ?? "1");
+const SIZES = [
+  { name: "100 B", bytes: 100, n: 4_000 },
+  { name: "10 KiB", bytes: 10 * 1024, n: 2_000 },
+  { name: "100 KiB", bytes: 100 * 1024, n: 600 },
+  { name: "1 MiB", bytes: 1024 * 1024, n: 150 },
+];
+const WARMUP = 30;
+
+function percentile(sortedAscending, p) {
+  const index = Math.min(sortedAscending.length - 1, Math.floor((p / 100) * sortedAscending.length));
+  return sortedAscending[index];
+}
+
+async function commandStats(client) {
+  const raw = await client.sendCommand(["INFO", "commandstats"]);
+  const stats = {};
+  for (const line of String(raw).split("\n")) {
+    const match = /^cmdstat_([a-z|]+):calls=(\d+),usec=(\d+)/.exec(line.trim());
+    if (match !== null) {
+      stats[match[1]] = { calls: Number(match[2]), usec: Number(match[3]) };
+    }
+  }
+  return stats;
+}
+
+const client = createClient({
+  url: REDIS_URL,
+  scripts: dialcacheRedisScripts,
+  disableOfflineQueue: true,
+  socket: { connectTimeout: 2_000 },
+});
+client.on("error", () => undefined);
+try {
+  await client.connect();
+} catch (error) {
+  console.error(`Could not reach Redis at ${REDIS_URL}; start one first, e.g. docker run --rm -p 6379:6379 redis:6.2`);
+  throw error;
+}
+const adapter = createNodeRedisDialCacheClient(client);
+
+const rows = [];
+for (const mode of ["tracked", "untracked"]) {
+  for (const size of SIZES) {
+    const iterations = Math.max(1, Math.round(size.n * SCALE));
+    const payload = "x".repeat(size.bytes);
+    const valueKey = `benchmark:write:${mode}:${size.bytes}:value`;
+    const watermarkKey = `benchmark:write:${mode}:${size.bytes}:watermark`;
+    const request = mode === "tracked"
+      ? { valueKey, watermarkKey, cacheTtlMs: 60_000, value: payload }
+      : { valueKey, cacheTtlMs: 60_000, value: payload };
+
+    for (let i = 0; i < WARMUP; i += 1) {
+      await adapter.write(request);
+    }
+    await client.sendCommand(["CONFIG", "RESETSTAT"]);
+
+    const latenciesUsec = [];
+    for (let i = 0; i < iterations; i += 1) {
+      const start = process.hrtime.bigint();
+      await adapter.write(request);
+      latenciesUsec.push(Number(process.hrtime.bigint() - start) / 1_000);
+    }
+
+    // Sum only the commands the client dispatches top-level (SET, EVALSHA,
+    // and the EVAL recovery). Script-internal calls surface in commandstats
+    // too, but the EVALSHA entry already envelopes their execution time.
+    const stats = await commandStats(client);
+    const serverUsec = (stats.set?.usec ?? 0)
+      + (stats.evalsha?.usec ?? 0)
+      + (stats.eval?.usec ?? 0);
+    latenciesUsec.sort((a, b) => a - b);
+    rows.push({
+      mode,
+      size: size.name,
+      writes: iterations,
+      serverUsecPerWrite: serverUsec / iterations,
+      clientP50Usec: percentile(latenciesUsec, 50),
+      clientP95Usec: percentile(latenciesUsec, 95),
+    });
+  }
+}
+await client.quit();
+
+console.log(`Redis write benchmark — ${REDIS_URL}`);
+console.log("mode       size      writes   server µs/write   client p50 µs   client p95 µs");
+for (const row of rows) {
+  console.log(
+    row.mode.padEnd(10)
+    + row.size.padEnd(10)
+    + String(row.writes).padEnd(9)
+    + row.serverUsecPerWrite.toFixed(1).padEnd(18)
+    + row.clientP50Usec.toFixed(0).padEnd(16)
+    + row.clientP95Usec.toFixed(0),
+  );
+}