diff --git a/README.md b/README.md
index 133ebc5..cb71a03 100644
--- a/README.md
+++ b/README.md
@@ -188,6 +188,17 @@ DynamoDbClient client = AlternatorDynamoDbClient.builder()
 `connectionAcquisitionTimeoutMs` or `connectionTimeoutMs` to 0 with CRT will fall back to
 SDK defaults (an info message is logged).
 
+**CRT sync 5xx connection note:** The AWS SDK CRT sync response handler treats HTTP
+5xx server errors as a condition where the native CRT connection may be closed instead
+of returned to the pool. There is no `AwsCrtHttpClient.Builder` option to disable this;
+`maxConcurrency`, idle timeout, TCP keep-alive, and CRT connection health settings do
+not change that status-code close policy. In this repository's probe using AWS SDK
+2.42.2, five sequential 500 responses reused the same TCP connection locally, while CI
+observed one replacement connection across the same five responses. Treat CRT sync 5xx
+reuse as best-effort: the current probe does not show one new connection per 5xx, but
+application code should still assume a connection can rotate after any 5xx response.
+Use Apache sync if strict single-connection reuse after 5xx responses is required.
+
 These settings are applied as defaults before any customizer callback runs, so the
 customizer can override them if needed.
 
diff --git a/src/test/java/com/scylladb/alternator/internal/HttpClientNon2xxConnectionReuseTest.java b/src/test/java/com/scylladb/alternator/internal/HttpClientNon2xxConnectionReuseTest.java
index e1b8af8..b33fb0d 100644
--- a/src/test/java/com/scylladb/alternator/internal/HttpClientNon2xxConnectionReuseTest.java
+++ b/src/test/java/com/scylladb/alternator/internal/HttpClientNon2xxConnectionReuseTest.java
@@ -1,6 +1,7 @@
 package com.scylladb.alternator.internal;
 
 import static org.junit.Assert.assertEquals;
+import static org.junit.Assert.assertTrue;
 
 import com.scylladb.alternator.AlternatorConfig;
 import java.io.IOException;
@@ -141,6 +142,19 @@ private void assertConnectionReuse(String clientName, int status, ReuseProbeServ
         clientName + " status " + status + " should reach the server",
         REQUESTS,
         server.requestCount());
+    if (allowsCrtSyncServerErrorReconnect(clientName, status)) {
+      // AWS CRT sync documents 5xx server errors as a condition where a connection may be
+      // closed instead of returned to the pool. The native path does not rotate on every 5xx
+      // in this short probe, but CI has observed one replacement connection across five
+      // sequential 500 responses.
+      assertTrue(
+          clientName + " status " + status + " should reuse TCP connections",
+          server.acceptedConnections() <= 2);
+      assertTrue(
+          clientName + " status " + status + " should reuse client TCP ports",
+          server.uniqueRemotePorts() <= 2);
+      return;
+    }
     assertEquals(
         clientName + " status " + status + " should reuse one TCP connection",
         1,
@@ -151,6 +165,10 @@ private void assertConnectionReuse(String clientName, int status, ReuseProbeServ
         server.uniqueRemotePorts());
   }
 
+  private boolean allowsCrtSyncServerErrorReconnect(String clientName, int status) {
+    return "crt-sync".equals(clientName) && status >= 500;
+  }
+
   private void drainAndClose(AbortableInputStream body) throws IOException {
     try (AbortableInputStream stream = body) {
       byte[] buffer = new byte[1024];