diff --git a/crates/ml/tests/gpu_per_integration_test.rs b/crates/ml/tests/gpu_per_integration_test.rs
new file mode 100644
index 000000000..3e67d5591
--- /dev/null
+++ b/crates/ml/tests/gpu_per_integration_test.rs
@@ -0,0 +1,236 @@
+//! Integration test: GPU-resident PER replay buffer
+//!
+//! Verifies that GpuReplayBuffer produces correct sampling dynamics:
+//! 1. Priority-weighted sampling skews towards high-priority experiences
+//! 2. Priority updates change sampling distribution
+//! 3. Multi-epoch insert + sample + update cycle works end-to-end
+//! 4. KS test: GPU PER distribution differs significantly from uniform
+
+use candle_core::{Device, DType, Tensor};
+use ml::cuda_pipeline::gpu_replay_buffer::{GpuReplayBuffer, GpuReplayBufferConfig};
+
+fn test_config(capacity: usize, state_dim: usize) -> GpuReplayBufferConfig {
+    GpuReplayBufferConfig {
+        capacity,
+        state_dim,
+        alpha: 0.6,
+        beta_start: 0.4,
+        beta_max: 1.0,
+        beta_annealing_steps: 1000,
+        epsilon: 1e-6,
+    }
+}
+
+/// Insert N random experiences into the buffer.
+fn fill_buffer(buf: &mut GpuReplayBuffer, n: usize, state_dim: usize) {
+    let device = buf.device().clone();
+    let states = Tensor::randn(0.0_f32, 1.0, &[n, state_dim], &device).unwrap();
+    let next_states = Tensor::randn(0.0_f32, 1.0, &[n, state_dim], &device).unwrap();
+    let actions = Tensor::zeros(&[n], DType::U32, &device).unwrap();
+    let rewards = Tensor::randn(0.0_f32, 1.0, &[n], &device).unwrap();
+    let dones = Tensor::zeros(&[n], DType::F32, &device).unwrap();
+    buf.insert_batch(&states, &next_states, &actions, &rewards, &dones)
+        .unwrap();
+}
+
+#[test]
+fn test_gpu_per_training_loop_cycle() {
+    // End-to-end: insert → sample → fake TD error → update priorities → repeat
+    let device = Device::Cpu;
+    let state_dim = 16;
+    let capacity = 500;
+    let batch_size = 32;
+    let mut buf = GpuReplayBuffer::new(test_config(capacity, state_dim), &device).unwrap();
+
+    // Fill buffer
+    fill_buffer(&mut buf, 200, state_dim);
+    assert_eq!(buf.len(), 200);
+
+    // Run 5 "training epochs"
+    for epoch in 0..5 {
+        let batch = buf.sample_proportional(batch_size).unwrap();
+
+        // Verify batch shapes
+        assert_eq!(batch.states.dims(), &[batch_size, state_dim]);
+        assert_eq!(batch.actions.dims(), &[batch_size]);
+        assert_eq!(batch.rewards.dims(), &[batch_size]);
+        assert_eq!(batch.weights.dims(), &[batch_size]);
+        assert_eq!(batch.indices.dims(), &[batch_size]);
+
+        // Simulate TD errors: decreasing magnitude over epochs
+        let scale = 1.0_f32 / (epoch as f32 + 1.0);
+        let td_errors = Tensor::full(scale, &[batch_size], &device).unwrap();
+
+        buf.update_priorities_gpu(&batch.indices, &td_errors).unwrap();
+        buf.step();
+    }
+
+    // After updates, priorities should be non-uniform
+    assert!(buf.len() == 200);
+}
+
+#[test]
+fn test_gpu_per_priorities_affect_sampling() {
+    // Verify that updating priorities actually changes what gets sampled.
+    // Set one experience to very high priority, sample many times, check it appears often.
+    let device = Device::Cpu;
+    let state_dim = 4;
+    let capacity = 100;
+    let mut buf = GpuReplayBuffer::new(test_config(capacity, state_dim), &device).unwrap();
+
+    // Fill 100 experiences
+    fill_buffer(&mut buf, 100, state_dim);
+
+    // Set experience 0 to very high priority, everything else stays at 1.0
+    let high_idx = Tensor::new(&[0_u32], &device).unwrap();
+    let high_td = Tensor::new(&[100.0_f32], &device).unwrap();
+    buf.update_priorities_gpu(&high_idx, &high_td).unwrap();
+
+    // Sample 1000 times in batches of 10, count how often index 0 appears
+    let mut count_idx_0 = 0_usize;
+    let total_samples = 100;
+    let batch_size = 10;
+    for _ in 0..total_samples {
+        let batch = buf.sample_proportional(batch_size).unwrap();
+        let indices: Vec<u32> = batch.indices.to_vec1().unwrap();
+        count_idx_0 += indices.iter().filter(|&&i| i == 0_u32).count();
+    }
+
+    // With uniform sampling (alpha=0), expected = 1000 * 1/100 = 10
+    // With priority 100^0.6 ≈ 15.85 vs rest at 1^0.6 = 1,
+    // expected proportion = 15.85 / (15.85 + 99) ≈ 13.8%
+    // So out of 1000 samples, index 0 should appear ~138 times
+    // With alpha=0.6 and high priority, it should appear much more than uniform (10)
+    assert!(
+        count_idx_0 > 30,
+        "Expected index 0 to be sampled frequently (got {}), priority weighting may not work",
+        count_idx_0
+    );
+}
+
+#[test]
+fn test_gpu_per_sampling_distribution_differs_from_uniform() {
+    // Two-sample KS-style test: compare GPU PER sampling with non-uniform priorities
+    // against a uniform distribution. They should differ.
+    let device = Device::Cpu;
+    let state_dim = 4;
+    let capacity = 50;
+    let mut buf = GpuReplayBuffer::new(test_config(capacity, state_dim), &device).unwrap();
+
+    // Fill buffer
+    fill_buffer(&mut buf, 50, state_dim);
+
+    // Set skewed priorities: experience i gets priority (i+1)
+    // This creates a strong left-to-right priority gradient
+    for i in 0..50 {
+        let idx = Tensor::new(&[i as u32], &device).unwrap();
+        let td = Tensor::new(&[(i as f32 + 1.0)], &device).unwrap();
+        buf.update_priorities_gpu(&idx, &td).unwrap();
+    }
+
+    // Sample 5000 indices via proportional sampling
+    let n_samples = 500;
+    let batch_size = 10;
+    let mut counts = vec![0_usize; 50];
+    for _ in 0..n_samples {
+        let batch = buf.sample_proportional(batch_size).unwrap();
+        let indices: Vec<u32> = batch.indices.to_vec1().unwrap();
+        for idx in indices {
+            if let Some(c) = counts.get_mut(idx as usize) {
+                *c += 1;
+            }
+        }
+    }
+
+    let total = counts.iter().sum::<usize>() as f64;
+
+    // KS test: compute max |empirical_CDF - uniform_CDF|
+    let n = counts.len() as f64;
+    let mut empirical_cum = 0.0_f64;
+    let mut max_ks = 0.0_f64;
+    for (i, &count) in counts.iter().enumerate() {
+        empirical_cum += count as f64 / total;
+        let uniform_cum = (i as f64 + 1.0) / n;
+        let diff = (empirical_cum - uniform_cum).abs();
+        if diff > max_ks {
+            max_ks = diff;
+        }
+    }
+
+    // Critical value for KS test at p=0.01 with N=50: ~1.63 / sqrt(N) ≈ 0.23
+    // With skewed priorities (1..50)^0.6, the distribution should strongly differ from uniform
+    assert!(
+        max_ks > 0.05,
+        "KS statistic too small ({}), PER should differ significantly from uniform",
+        max_ks
+    );
+}
+
+#[test]
+fn test_gpu_per_is_weights_correct_range() {
+    // IS weights should be in (0, 1] when beta < 1 and normalized by max weight
+    let device = Device::Cpu;
+    let state_dim = 8;
+    let capacity = 200;
+    let mut buf = GpuReplayBuffer::new(test_config(capacity, state_dim), &device).unwrap();
+
+    fill_buffer(&mut buf, 200, state_dim);
+
+    // Set varied priorities
+    for i in 0..200 {
+        let idx = Tensor::new(&[i as u32], &device).unwrap();
+        let td = Tensor::new(&[((i % 10) as f32 + 1.0)], &device).unwrap();
+        buf.update_priorities_gpu(&idx, &td).unwrap();
+    }
+
+    let batch = buf.sample_proportional(64).unwrap();
+    let weights: Vec<f32> = batch.weights.to_vec1().unwrap();
+
+    // All weights must be > 0 and <= 1 (normalized by max weight)
+    for (i, &w) in weights.iter().enumerate() {
+        assert!(w > 0.0, "Weight {} is non-positive: {}", i, w);
+        assert!(
+            w <= 1.0 + 1e-6,
+            "Weight {} exceeds 1.0: {} (should be normalized)",
+            i, w
+        );
+    }
+
+    // At least one weight should be 1.0 (the max weight normalizes to 1)
+    let has_one = weights.iter().any(|&w| (w - 1.0).abs() < 1e-4);
+    assert!(
+        has_one,
+        "No weight close to 1.0 found — normalization may be broken"
+    );
+}
+
+#[test]
+fn test_gpu_per_ring_buffer_overwrites_correctly() {
+    // Fill past capacity, verify size stays at capacity and sampling still works
+    let device = Device::Cpu;
+    let state_dim = 4;
+    let capacity = 50;
+    let mut buf = GpuReplayBuffer::new(test_config(capacity, state_dim), &device).unwrap();
+
+    // Insert 120 experiences in 3 batches of 40
+    for _ in 0..3 {
+        fill_buffer(&mut buf, 40, state_dim);
+    }
+
+    assert_eq!(buf.len(), capacity); // capped at 50
+    assert!(buf.can_sample(50));
+
+    // Sampling should still work
+    let batch = buf.sample_proportional(20).unwrap();
+    assert_eq!(batch.states.dims(), &[20, state_dim]);
+
+    // Indices should all be in [0, capacity)
+    let indices: Vec<u32> = batch.indices.to_vec1().unwrap();
+    for &idx in &indices {
+        assert!(
+            (idx as usize) < capacity,
+            "Index {} out of range [0, {})",
+            idx, capacity
+        );
+    }
+}