diff --git a/docs/plans/2026-03-01-io-pipeline-wiring-design.md b/docs/plans/2026-03-01-io-pipeline-wiring-design.md
new file mode 100644
index 000000000..878b15fb3
--- /dev/null
+++ b/docs/plans/2026-03-01-io-pipeline-wiring-design.md
@@ -0,0 +1,71 @@
+# I/O Pipeline Wiring — Design Document
+
+## Problem Statement
+
+GPU optimization infrastructure is fully built but 3 I/O pipeline components are unused:
+
+1. **EpochPrefetcher** (`cuda_pipeline/prefetch.rs`) — background thread loads next fold's data. DQNTrainer has `set_prefetcher()`/`take_prefetched_data()` methods but `train_baseline_rl.rs` never calls them.
+
+2. **GpuBufferPool** (`cuda_pipeline/mod.rs:278-367`) — pre-allocated CPU staging buffers for zero-alloc GPU uploads. DQNTrainer initializes it in `new()` but never references `self.buffer_pool`.
+
+3. **DoubleBufferedLoader** (`cuda_pipeline/double_buffer.rs`) — ping-pong GPU tensor slots for zero-downtime fold transitions. Not used by any trainer.
+
+All core GPU optimizations (mixed precision, dynamic batching, gradient accumulation, tensor core alignment, Rainbow DQN) are already wired and active. This work is the final I/O throughput piece.
+
+## Architecture
+
+### Fold-Level Pipeline (train_baseline_rl.rs)
+
+```
+for fold_idx in 0..num_folds:
+  // 1. Get data: from prefetcher (fold 1+) or fresh load (fold 0)
+  let fold_data = if let Some(prefetched) = trainer.take_prefetched_data() {
+      prefetched?
+  } else {
+      load_fold_data(fold_idx)  // first fold, no prefetch available
+  };
+
+  // 2. Kick off prefetch for NEXT fold (if not last)
+  if fold_idx + 1 < num_folds {
+      let next_path = fold_paths[fold_idx + 1].clone();
+      trainer.set_prefetcher(EpochPrefetcher::spawn(move || {
+          load_fold_data_from_path(&next_path)
+      }));
+  }
+
+  // 3. Train this fold (buffer_pool + double_buffer used internally)
+  trainer.train_with_preloaded_data(fold_data, val_data, callback).await?;
+```
+
+### Trainer-Internal Pipeline (DQNTrainer)
+
+Inside `train_with_preloaded_data()`:
+
+1. **GpuBufferPool**: Use `self.buffer_pool.upload_dqn()` instead of `DqnGpuData::upload()` to avoid per-fold heap allocation. The pool pre-allocates staging buffers for 100k bars once.
+
+2. **DoubleBufferedLoader**: Add `double_buffer: Option<DoubleBufferedLoader>` field. In the fold transition path, upload next fold's data to staging slot while current fold trains. `swap()` at fold boundary is O(1).
+
+### PPO Path
+
+PPO uses `PpoTrainer` which has a different data flow (trajectory collection, not preloaded features). The prefetcher and buffer pool don't directly apply. PPO already has gradient accumulation and mixed precision wired. No changes needed.
+
+## Files to Modify
+
+| File | Change |
+|---|---|
+| `crates/ml/examples/train_baseline_rl.rs` | Wire `set_prefetcher()`/`take_prefetched_data()` in DQN fold loop |
+| `crates/ml/src/trainers/dqn/trainer.rs` | Use `self.buffer_pool.upload_dqn()` in `train_with_preloaded_data()`, add `DoubleBufferedLoader` field, integrate in fold transitions |
+
+## Out of Scope
+
+- PPO I/O pipeline (different data flow — trajectory-based, not preloaded)
+- Multi-node distributed training (single-node multi-GPU kept as-is)
+- Pinned memory (Candle doesn't expose `cudaMallocHost`)
+
+## Success Criteria
+
+- `train_baseline_rl.rs` DQN path uses prefetcher for fold N+1 while fold N trains
+- DQNTrainer uses `buffer_pool.upload_dqn()` instead of direct `DqnGpuData::upload()`
+- DoubleBufferedLoader manages active/staging GPU slots during fold transitions
+- All 2444+ ml tests pass
+- Training binary compiles clean