diff --git a/benchmarks/bench_reference.cpp b/benchmarks/bench_reference.cpp
new file mode 100644
index 0000000..d3005fb
--- /dev/null
+++ b/benchmarks/bench_reference.cpp
@@ -0,0 +1,406 @@
+#include <memory>
+#include <thread>
+#include <vector>
+
+#include <nanobench.h>
+
+#include "reference.hpp"
+
+namespace {
+
+struct TestObject {
+  int64_t data = 42;
+
+  TestObject() = default;
+  explicit TestObject(int64_t value) : data(value) {}
+};
+
+// Trait helpers for templated benchmarks
+template <typename T> struct PointerTraits;
+
+template <typename T> struct PointerTraits<std::shared_ptr<T>> {
+  using pointer_type = std::shared_ptr<T>;
+  using weak_type = std::weak_ptr<T>;
+
+  template <typename... Args> static pointer_type make(Args &&...args) {
+    return std::make_shared<T>(std::forward<Args>(args)...);
+  }
+
+  static const char *name() { return "std::shared_ptr"; }
+  static const char *weak_name() { return "std::weak_ptr"; }
+};
+
+template <typename T> struct PointerTraits<Ref<T>> {
+  using pointer_type = Ref<T>;
+  using weak_type = WeakRef<T>;
+
+  template <typename... Args> static pointer_type make(Args &&...args) {
+    return make_ref<T>(std::forward<Args>(args)...);
+  }
+
+  static const char *name() { return "Ref"; }
+  static const char *weak_name() { return "WeakRef"; }
+};
+
+template <typename PtrType>
+void benchmark_creation(ankerl::nanobench::Bench &bench) {
+  using Traits = PointerTraits<PtrType>;
+
+  bench.run(std::string(Traits::name()) + " creation", [&] {
+    auto ptr = Traits::make(TestObject{123});
+    ankerl::nanobench::doNotOptimizeAway(ptr);
+  });
+}
+
+template <typename PtrType>
+void benchmark_copy(ankerl::nanobench::Bench &bench) {
+  using Traits = PointerTraits<PtrType>;
+
+  auto original = Traits::make(TestObject{123});
+  bench.run(std::string(Traits::name()) + " copy", [&] {
+    auto copy = original;
+    ankerl::nanobench::doNotOptimizeAway(copy);
+  });
+}
+
+template <typename PtrType>
+void benchmark_move(ankerl::nanobench::Bench &bench) {
+  using Traits = PointerTraits<PtrType>;
+
+  auto original = Traits::make(TestObject{123});
+  bench.run(std::string(Traits::name()) + " move", [&] {
+    auto moved = std::move(original);
+    ankerl::nanobench::doNotOptimizeAway(moved);
+    original = std::move(moved);
+    ankerl::nanobench::doNotOptimizeAway(original);
+  });
+}
+
+template <typename PtrType>
+void benchmark_weak_copy(ankerl::nanobench::Bench &bench) {
+  using Traits = PointerTraits<PtrType>;
+
+  auto strong_ptr = Traits::make(TestObject{123});
+  typename Traits::weak_type weak_original = strong_ptr;
+  bench.run(std::string(Traits::weak_name()) + " copy", [&] {
+    auto weak_copy = weak_original;
+    ankerl::nanobench::doNotOptimizeAway(weak_copy);
+  });
+}
+
+template <typename PtrType>
+void benchmark_weak_move(ankerl::nanobench::Bench &bench) {
+  using Traits = PointerTraits<PtrType>;
+
+  auto strong_ptr = Traits::make(TestObject{123});
+  typename Traits::weak_type weak_original = strong_ptr;
+  bench.run(std::string(Traits::weak_name()) + " move", [&] {
+    auto weak_moved = std::move(weak_original);
+    ankerl::nanobench::doNotOptimizeAway(weak_moved);
+    weak_original = std::move(weak_moved);
+    ankerl::nanobench::doNotOptimizeAway(weak_original);
+  });
+}
+
+template <typename PtrType>
+void benchmark_dereference(ankerl::nanobench::Bench &bench) {
+  using Traits = PointerTraits<PtrType>;
+
+  auto ptr = Traits::make(TestObject{456});
+  bench.run(std::string(Traits::name()) + " dereference",
+            [&] { ankerl::nanobench::doNotOptimizeAway(ptr->data); });
+}
+
+template <typename PtrType>
+void benchmark_weak_lock_success(ankerl::nanobench::Bench &bench) {
+  using Traits = PointerTraits<PtrType>;
+
+  auto strong_ptr = Traits::make(TestObject{789});
+  typename Traits::weak_type weak_ptr = strong_ptr;
+  bench.run(std::string(Traits::weak_name()) + " lock success", [&] {
+    auto locked = weak_ptr.lock();
+    ankerl::nanobench::doNotOptimizeAway(locked);
+  });
+}
+
+template <typename PtrType>
+void benchmark_weak_lock_failure(ankerl::nanobench::Bench &bench) {
+  using Traits = PointerTraits<PtrType>;
+
+  typename Traits::weak_type weak_ptr;
+  {
+    auto strong_ptr = Traits::make(TestObject{999});
+    weak_ptr = strong_ptr;
+  }
+  bench.run(std::string(Traits::weak_name()) + " lock failure", [&] {
+    auto locked = weak_ptr.lock();
+    ankerl::nanobench::doNotOptimizeAway(locked);
+  });
+}
+
+template <typename PtrType>
+void benchmark_multithreaded_copy(ankerl::nanobench::Bench &bench,
+                                  int num_threads) {
+  using Traits = PointerTraits<PtrType>;
+
+  // Create the shared object outside the benchmark
+  auto ptr = Traits::make(TestObject{456});
+
+  // Create background threads that will create contention
+  std::atomic<bool> keep_running{true};
+  std::vector<std::thread> background_threads;
+
+  for (int i = 0; i < num_threads - 1; ++i) {
+    background_threads.emplace_back([&]() {
+      while (keep_running.load(std::memory_order_relaxed)) {
+        auto copy = ptr;
+        ankerl::nanobench::doNotOptimizeAway(copy);
+      }
+    });
+  }
+
+  // Benchmark the foreground thread under contention
+  bench.run(std::string(Traits::name()) + " copy under contention", [&] {
+    auto copy = ptr;
+    ankerl::nanobench::doNotOptimizeAway(copy);
+  });
+
+  // Clean up background threads
+  keep_running.store(false, std::memory_order_relaxed);
+  for (auto &t : background_threads) {
+    t.join();
+  }
+}
+
+template <typename PtrType>
+void benchmark_multithreaded_weak_lock(ankerl::nanobench::Bench &bench,
+                                       int num_threads) {
+  using Traits = PointerTraits<PtrType>;
+
+  // Create the shared object and weak reference outside the benchmark
+  auto strong_ptr = Traits::make(TestObject{789});
+  typename Traits::weak_type weak_ptr = strong_ptr;
+
+  // Create background threads that will create contention
+  std::atomic<bool> keep_running{true};
+  std::vector<std::thread> background_threads;
+
+  for (int i = 0; i < num_threads - 1; ++i) {
+    background_threads.emplace_back([&]() {
+      while (keep_running.load(std::memory_order_relaxed)) {
+        auto locked = weak_ptr.lock();
+        ankerl::nanobench::doNotOptimizeAway(locked);
+      }
+    });
+  }
+
+  // Benchmark the foreground thread under contention
+  bench.run(std::string(Traits::weak_name()) + " lock under contention", [&] {
+    auto locked = weak_ptr.lock();
+    ankerl::nanobench::doNotOptimizeAway(locked);
+  });
+
+  // Clean up background threads
+  keep_running.store(false, std::memory_order_relaxed);
+  for (auto &t : background_threads) {
+    t.join();
+  }
+}
+
+template <typename PtrType>
+void benchmark_weak_copy_with_strong_contention(ankerl::nanobench::Bench &bench,
+                                                int num_threads) {
+  using Traits = PointerTraits<PtrType>;
+
+  // Create the shared object and weak reference outside the benchmark
+  auto strong_ptr = Traits::make(TestObject{456});
+  typename Traits::weak_type weak_ptr = strong_ptr;
+
+  // Create background threads copying the strong pointer
+  std::atomic<bool> keep_running{true};
+  std::vector<std::thread> background_threads;
+
+  for (int i = 0; i < num_threads - 1; ++i) {
+    background_threads.emplace_back([&]() {
+      while (keep_running.load(std::memory_order_relaxed)) {
+        auto copy = strong_ptr;
+        ankerl::nanobench::doNotOptimizeAway(copy);
+      }
+    });
+  }
+
+  // Benchmark weak reference copying under strong reference contention
+  bench.run(std::string(Traits::weak_name()) + " copy with strong contention",
+            [&] {
+              auto weak_copy = weak_ptr;
+              ankerl::nanobench::doNotOptimizeAway(weak_copy);
+            });
+
+  // Clean up background threads
+  keep_running.store(false, std::memory_order_relaxed);
+  for (auto &t : background_threads) {
+    t.join();
+  }
+}
+
+template <typename PtrType>
+void benchmark_strong_copy_with_weak_contention(ankerl::nanobench::Bench &bench,
+                                                int num_threads) {
+  using Traits = PointerTraits<PtrType>;
+
+  // Create the shared object and weak reference outside the benchmark
+  auto strong_ptr = Traits::make(TestObject{789});
+  typename Traits::weak_type weak_ptr = strong_ptr;
+
+  // Create background threads copying the weak pointer
+  std::atomic<bool> keep_running{true};
+  std::vector<std::thread> background_threads;
+
+  for (int i = 0; i < num_threads - 1; ++i) {
+    background_threads.emplace_back([&]() {
+      while (keep_running.load(std::memory_order_relaxed)) {
+        auto weak_copy = weak_ptr;
+        ankerl::nanobench::doNotOptimizeAway(weak_copy);
+      }
+    });
+  }
+
+  // Benchmark strong reference copying under weak reference contention
+  bench.run(std::string(Traits::name()) + " copy with weak contention", [&] {
+    auto strong_copy = strong_ptr;
+    ankerl::nanobench::doNotOptimizeAway(strong_copy);
+  });
+
+  // Clean up background threads
+  keep_running.store(false, std::memory_order_relaxed);
+  for (auto &t : background_threads) {
+    t.join();
+  }
+}
+
+} // anonymous namespace
+
+void compare_creation() {
+  ankerl::nanobench::Bench bench;
+  bench.title("Creation performance comparison");
+  bench.relative(true);
+  benchmark_creation<std::shared_ptr<TestObject>>(bench);
+  benchmark_creation<Ref<TestObject>>(bench);
+}
+
+void compare_copy() {
+  ankerl::nanobench::Bench bench;
+  bench.title("Copy performance comparison");
+  bench.relative(true);
+  benchmark_copy<std::shared_ptr<TestObject>>(bench);
+  benchmark_copy<Ref<TestObject>>(bench);
+}
+
+void compare_move() {
+  ankerl::nanobench::Bench bench;
+  bench.title("Move performance comparison");
+  bench.relative(true);
+  benchmark_move<std::shared_ptr<TestObject>>(bench);
+  benchmark_move<Ref<TestObject>>(bench);
+}
+
+void compare_weak_copy() {
+  ankerl::nanobench::Bench bench;
+  bench.title("Weak copy performance comparison");
+  bench.relative(true);
+  benchmark_weak_copy<std::shared_ptr<TestObject>>(bench);
+  benchmark_weak_copy<Ref<TestObject>>(bench);
+}
+
+void compare_weak_move() {
+  ankerl::nanobench::Bench bench;
+  bench.title("Weak move performance comparison");
+  bench.relative(true);
+  benchmark_weak_move<std::shared_ptr<TestObject>>(bench);
+  benchmark_weak_move<Ref<TestObject>>(bench);
+}
+
+void compare_dereference() {
+  ankerl::nanobench::Bench bench;
+  bench.title("Dereference performance comparison");
+  bench.relative(true);
+  benchmark_dereference<std::shared_ptr<TestObject>>(bench);
+  benchmark_dereference<Ref<TestObject>>(bench);
+}
+
+void compare_weak_lock_success() {
+  ankerl::nanobench::Bench bench;
+  bench.title("Weak lock success performance comparison");
+  bench.relative(true);
+  benchmark_weak_lock_success<std::shared_ptr<TestObject>>(bench);
+  benchmark_weak_lock_success<Ref<TestObject>>(bench);
+}
+
+void compare_weak_lock_failure() {
+  ankerl::nanobench::Bench bench;
+  bench.title("Weak lock failure performance comparison");
+  bench.relative(true);
+  benchmark_weak_lock_failure<std::shared_ptr<TestObject>>(bench);
+  benchmark_weak_lock_failure<Ref<TestObject>>(bench);
+}
+
+void compare_multithreaded_copy(int num_threads) {
+  ankerl::nanobench::Bench bench;
+  bench.title("Copy performance under contention");
+  bench.relative(true);
+  bench.minEpochIterations(500000);
+  benchmark_multithreaded_copy<std::shared_ptr<TestObject>>(bench, num_threads);
+  benchmark_multithreaded_copy<Ref<TestObject>>(bench, num_threads);
+}
+
+void compare_multithreaded_weak_lock(int num_threads) {
+  ankerl::nanobench::Bench bench;
+  bench.title("Weak lock performance under contention");
+  bench.relative(true);
+  bench.minEpochIterations(500000);
+  benchmark_multithreaded_weak_lock<std::shared_ptr<TestObject>>(bench,
+                                                                 num_threads);
+  benchmark_multithreaded_weak_lock<Ref<TestObject>>(bench, num_threads);
+}
+
+void compare_weak_copy_with_strong_contention(int num_threads) {
+  ankerl::nanobench::Bench bench;
+  bench.title("Weak copy performance under strong reference contention");
+  bench.relative(true);
+  bench.minEpochIterations(500000);
+  benchmark_weak_copy_with_strong_contention<std::shared_ptr<TestObject>>(
+      bench, num_threads);
+  benchmark_weak_copy_with_strong_contention<Ref<TestObject>>(bench,
+                                                              num_threads);
+}
+
+void compare_strong_copy_with_weak_contention(int num_threads) {
+  ankerl::nanobench::Bench bench;
+  bench.title("Strong copy performance under weak reference contention");
+  bench.relative(true);
+  bench.minEpochIterations(500000);
+  benchmark_strong_copy_with_weak_contention<std::shared_ptr<TestObject>>(
+      bench, num_threads);
+  benchmark_strong_copy_with_weak_contention<Ref<TestObject>>(bench,
+                                                              num_threads);
+}
+
+int main() {
+  const int num_threads = 3;
+
+  // Each comparison has std::shared_ptr as 100% baseline
+  compare_creation();
+  compare_copy();
+  compare_move();
+  compare_weak_copy();
+  compare_weak_move();
+  compare_dereference();
+  compare_weak_lock_success();
+  compare_weak_lock_failure();
+  compare_multithreaded_copy(num_threads);
+  compare_multithreaded_weak_lock(num_threads);
+  compare_weak_copy_with_strong_contention(num_threads);
+  compare_strong_copy_with_weak_contention(num_threads);
+
+  return 0;
+}
diff --git a/tools/check_snake_case.py b/tools/check_snake_case.py
index 9120ecc..996076e 100755
--- a/tools/check_snake_case.py
+++ b/tools/check_snake_case.py
@@ -66,6 +66,8 @@ def check_snake_case_violations(filepath, check_new_only=True):
     exclusions = [
         # C++ standard library and common libraries
         r"\b(std::|weaseljson|simdutf|doctest)",
+        # Nanobench library API (external camelCase API)
+        r"\b(nanobench::|doNotOptimizeAway|minEpochIterations)\b",
         # Template parameters and concepts
         r"\b[A-Z][a-zA-Z0-9_]*\b",
         # Class/struct names (PascalCase is correct)