diff --git a/src/tasks/batch_write_task.cpp b/src/tasks/batch_write_task.cpp
index d7093b51..6edf4309 100644
--- a/src/tasks/batch_write_task.cpp
+++ b/src/tasks/batch_write_task.cpp
@@ -1482,6 +1482,16 @@ KvError BatchWriteTask::WriteLargeValue(const WriteDataEntry &entry)
             std::get_if<IoStringBuffer>(&entry.large_val_);
         iosb != nullptr)
     {
+        // The on-disk large-value header encodes the length in 31 bits (bit 31
+        // is the has-metadata flag), so a value longer than
+        // kLargeValueLengthMask (2 GiB - 1) cannot be represented: casting to
+        // uint32 would truncate it and/or collide with the metadata flag,
+        // corrupting the blob so every later read fails. EloqStore does not
+        // support values this large; reject.
+        if (iosb->Size() > kLargeValueLengthMask)
+        {
+            return KvError::InvalidArgs;
+        }
         const auto &fragments = iosb->Fragments();
         const uint32_t num_segments = fragments.size();
         std::vector<const char *> ptrs(num_segments);
@@ -1504,6 +1514,12 @@ KvError BatchWriteTask::WriteLargeValue(const WriteDataEntry &entry)
         const size_t size = pinned->second;
         assert(base != nullptr);
         assert(size > 0);
+        // See the IoStringBuffer path above: values longer than
+        // kLargeValueLengthMask (2 GiB - 1) cannot be encoded on disk.
+        if (size > kLargeValueLengthMask)
+        {
+            return KvError::InvalidArgs;
+        }
         const uint32_t num_segments =
             static_cast<uint32_t>((size + seg_size - 1) / seg_size);
 
diff --git a/tests/large_value_concurrency.cpp b/tests/large_value_concurrency.cpp
index a97681e1..32b5c72d 100644
--- a/tests/large_value_concurrency.cpp
+++ b/tests/large_value_concurrency.cpp
@@ -978,3 +978,33 @@ TEST_CASE(
     store->Stop();
     CleanupStore(opts);
 }
+
+// A large value whose length exceeds
+// kLargeValueLengthMask (2 GiB - 1) cannot be represented in the 31-bit
+// on-disk length field (bit 31 is the has-metadata flag). WriteLargeValue must
+// reject it up front instead of casting the length to uint32 -- which would
+// truncate it and/or collide with the metadata flag, corrupting the blob so
+// every later read fails. The guard runs before any memory is dereferenced, so
+// a huge declared length over a small real pinned chunk exercises it without
+// allocating gigabytes.
+TEST_CASE("pinned large value over the 2 GiB length cap is rejected",
+          "[large-value-concurrency][pinned]")
+{
+    PinnedMultiShardHarness h(kSegmentSize, /*num_shards=*/1);
+    eloqstore::KvOptions opts =
+        MakePinnedOpts(h, /*pinned_tail_scratch_slots=*/2);
+    eloqstore::EloqStore *store = InitStore(opts);
+
+    eloqstore::TableIdent tbl{"lv-cap", 0};
+    // 2 GiB == kLargeValueHasMetadataBit: one byte past the 31-bit length cap.
+    const size_t too_large = size_t{1} << 31;
+    std::pair<char *, size_t> dst{h.Base(0), too_large};
+
+    eloqstore::BatchWriteRequest req;
+    AsyncPinnedWrite(store, req, tbl, "big", dst, /*metadata=*/"", /*ts=*/1);
+    req.Wait();
+    REQUIRE(req.Error() == eloqstore::KvError::InvalidArgs);
+
+    store->Stop();
+    CleanupStore(opts);
+}