diff --git a/src/sample.h b/src/sample.h
index 86553c05..0b8deb15 100644
--- a/src/sample.h
+++ b/src/sample.h
@@ -186,8 +186,10 @@ class sample {
 
 	/// Decrement ref count and reclaim if unreferenced.
 	friend void intrusive_ptr_release(sample *s) {
-		if (s->refcount_.fetch_sub(1, std::memory_order_release) == 1) {
-			std::atomic_thread_fence(std::memory_order_acquire);
+		// Acquire earlier owners' releases before publishing the sample for reuse.
+		// A release decrement followed by an acquire fence is also valid C++, but
+		// TSan does not model that fence and reports false races on recycled data.
+		if (s->refcount_.fetch_sub(1, std::memory_order_acq_rel) == 1) {
 			s->factory_->reclaim_sample(s);
 		}
 	}
diff --git a/testing/int/samples.cpp b/testing/int/samples.cpp
index 270b1f59..ac120f1d 100644
--- a/testing/int/samples.cpp
+++ b/testing/int/samples.cpp
@@ -72,3 +72,32 @@ TEST_CASE("sample conversion", "[basic]") {
 		values[1] = (double)(-buf[0]);
 	}
 }
+
+TEST_CASE("sample recycling acquires earlier readers", "[sample][threads]") {
+	lsl::factory fac(cft_int32, 1, 1);
+	for (int i = 0; i < 32; ++i) {
+		auto sample = fac.new_sample(42.0, false);
+		auto *address = sample.get();
+		std::atomic<bool> released{false};
+		double observed = 0.0;
+		std::thread reader([copy = sample, &released, &observed]() mutable {
+			observed = copy->timestamp();
+			copy.reset();
+			released.store(true, std::memory_order_relaxed);
+		});
+		struct join_thread {
+			std::thread &thread;
+			~join_thread() { if (thread.joinable()) thread.join(); }
+		} joiner{reader};
+
+		// Only control the release order here. An acquire/release handshake or
+		// joining before reuse would hide synchronization missing from refcount_.
+		while (!released.load(std::memory_order_relaxed)) std::this_thread::yield();
+		sample.reset(); // Last owner must acquire the earlier reader's release.
+		auto recycled = fac.new_sample(84.0, true);
+		reader.join();
+		CHECK(recycled.get() == address);
+		CHECK(observed == 42.0);
+		CHECK(recycled->timestamp() == 84.0);
+	}
+}
diff --git a/testing/int/sendbuffer.cpp b/testing/int/sendbuffer.cpp
index 6e517859..3bb221bb 100644
--- a/testing/int/sendbuffer.cpp
+++ b/testing/int/sendbuffer.cpp
@@ -29,6 +29,7 @@
 #include "send_buffer.h"
 #include <atomic>
 #include <catch2/catch_all.hpp>
+#include <mutex>
 #include <thread>
 #include <vector>
 
@@ -136,6 +137,7 @@ TEST_CASE("multi-threaded send_buffer stress", "[queue][regression][send_buffer]
 	const int iterations_per_thread = 200;
 
 	lsl::factory fac(lsl_channel_format_t::cft_float32, 4, buffer_size * 2);
+	std::mutex factory_mut;
 	auto sendbuf = std::make_shared<lsl::send_buffer>(buffer_size);
 
 	std::vector<std::shared_ptr<lsl::consumer_queue>> queues;
@@ -148,7 +150,13 @@ TEST_CASE("multi-threaded send_buffer stress", "[queue][regression][send_buffer]
 	// Producer threads push samples concurrently
 	auto producer = [&]() {
 		for (int i = 0; i < iterations_per_thread; ++i) {
-			auto sample = fac.new_sample(static_cast<double>(i), true);
+			lsl::sample_p sample;
+			{
+				// The factory permits only one allocator at a time. Keep the
+				// send_buffer pushes concurrent without racing its sample pool.
+				std::lock_guard<std::mutex> lock(factory_mut);
+				sample = fac.new_sample(static_cast<double>(i), true);
+			}
 			sendbuf->push_sample(sample);
 			push_count.fetch_add(1, std::memory_order_relaxed);
 		}