diff --git a/levsync/levsync.odin b/levsync/levsync.odin
index 30e441b..318b10c 100644
--- a/levsync/levsync.odin
+++ b/levsync/levsync.odin
@@ -2,6 +2,10 @@ package levsync
 
 import "base:intrinsics"
 
+// ---------------------------------------------------------------------------------------------------------------------
+// ----- Atomic Float Ops ------------------------
+// ---------------------------------------------------------------------------------------------------------------------
+
 @(private)
 Flop :: enum {
 	Add,
@@ -96,6 +100,30 @@ atomic_div_float :: #force_inline proc "contextless" (
 	return atomic_float_op_cas(dst, val, .Divide, ORDER)
 }
 
+// ---------------------------------------------------------------------------------------------------------------------
+// ----- Spinlock ------------------------
+// ---------------------------------------------------------------------------------------------------------------------
+
+// Spinlock for when you need manual control over trying to acquire. If you always loop try_lock then
+// Odin `core:sync -> Ticket_Mutex' is probably a better fit.
+Spinlock :: distinct bool
+
+// Returns true if lock successfuly acquired, false otherwise
+spinlock_try_lock :: #force_inline proc "contextless" (lock: ^Spinlock) -> bool {
+	was_locked, lock_acquired := intrinsics.atomic_compare_exchange_weak_explicit(
+		lock,
+		false,
+		true,
+		.Acq_Rel,
+		.Relaxed,
+	)
+	return lock_acquired
+}
+
+spinlock_unlock :: #force_inline proc "contextless" (lock: ^Spinlock) {
+	intrinsics.atomic_store_explicit(lock, false, .Release)
+}
+
 // ---------------------------------------------------------------------------------------------------------------------
 // ----- Tests ------------------------
 // ---------------------------------------------------------------------------------------------------------------------
@@ -389,3 +417,109 @@ test_atomic_release_acquire_publish_visibility :: proc(t: ^testing.T) {
 		testing.expectf(t, rd.data_valid, "Reader %d saw flag but data was not visible (memory ordering bug)", i)
 	}
 }
+
+@(test)
+test_spinlock_try_lock_mutual_exclusion :: proc(t: ^testing.T) {
+	// Stress test for spinlock_try_lock: N threads spin-acquire the lock and
+	// perform a deliberate non-atomic read-modify-write on shared data.
+	//
+	// If mutual exclusion holds:
+	//   - `counter` ends at exactly NUM_THREADS * ITERATIONS_PER_THREAD
+	//   - `concurrent_holders` never exceeds 1
+	//
+	// A multi-step RMW (read → relax → write) widens the critical section so
+	// any failure to exclude is virtually guaranteed to corrupt the counter.
+	NUM_THREADS :: 8
+	ITERATIONS_PER_THREAD :: 50_000
+
+	Shared :: struct {
+		lock:               Spinlock,
+		// Padding to avoid false sharing between the lock and the data it protects.
+		_pad0:              [64]u8,
+		counter:            int,
+		// Tracks how many threads believe they hold the lock simultaneously.
+		// Must never exceed 1.
+		concurrent_holders: int,
+		max_holders:        int,
+		_pad1:              [64]u8,
+	}
+
+	Thread_Data :: struct {
+		shared:  ^Shared,
+		barrier: ^sync.Barrier,
+	}
+
+	shared: Shared
+	barrier: sync.Barrier
+	sync.barrier_init(&barrier, NUM_THREADS)
+
+	thread_proc :: proc(th: ^thread.Thread) {
+		ctx := cast(^Thread_Data)th.data
+		s := ctx.shared
+
+		// All threads rendezvous here for maximum contention.
+		sync.barrier_wait(ctx.barrier)
+
+		for _ in 0 ..< ITERATIONS_PER_THREAD {
+			// Spin on try_lock until we acquire it.
+			for !spinlock_try_lock(&s.lock) {
+				intrinsics.cpu_relax()
+			}
+
+			// --- critical section start ---
+
+			// Atomically bump the holder count so we can detect overlapping holders.
+			holders := intrinsics.atomic_add_explicit(&s.concurrent_holders, 1, .Relaxed)
+
+			// Track the maximum we ever observed (relaxed is fine, this is
+			// purely diagnostic and protected by the spinlock for writes).
+			if holders + 1 > s.max_holders {
+				s.max_holders = holders + 1
+			}
+
+			// Non-atomic RMW: read, spin a tiny bit, then write.
+			// This deliberately creates a wide window where a second holder
+			// would cause a lost update.
+			val := s.counter
+			intrinsics.cpu_relax()
+			intrinsics.cpu_relax()
+			s.counter = val + 1
+
+			intrinsics.atomic_sub_explicit(&s.concurrent_holders, 1, .Relaxed)
+
+			// --- critical section end ---
+
+			spinlock_unlock(&s.lock)
+		}
+	}
+
+	td := Thread_Data{&shared, &barrier}
+
+	threads: [NUM_THREADS]^thread.Thread
+	for &th in threads {
+		th = thread.create(thread_proc)
+		th.data = &td
+	}
+	for th in threads {
+		thread.start(th)
+	}
+	for th in threads {
+		thread.join(th)
+		thread.destroy(th)
+	}
+
+	expected := NUM_THREADS * ITERATIONS_PER_THREAD
+	testing.expectf(
+		t,
+		shared.counter == expected,
+		"Counter mismatch: got %d, expected %d (mutual exclusion violated — lost updates)",
+		shared.counter,
+		expected,
+	)
+	testing.expectf(
+		t,
+		shared.max_holders == 1,
+		"Max concurrent lock holders was %d (expected 1 — lock was held by multiple threads simultaneously)",
+		shared.max_holders,
+	)
+}