diff --git a/crates/ml-alpha/src/rl/gpu_hindsight.rs b/crates/ml-alpha/src/rl/gpu_hindsight.rs
new file mode 100644
index 000000000..85e90b5c5
--- /dev/null
+++ b/crates/ml-alpha/src/rl/gpu_hindsight.rs
@@ -0,0 +1,38 @@
+use std::sync::Arc;
+use anyhow::{Context, Result};
+use cudarc::driver::{CudaSlice, CudaStream};
+use crate::heads::HIDDEN_DIM;
+
+pub const HINDSIGHT_RING_LEN: usize = 512;
+pub const CLOSED_RING_SIZE: usize = 256;
+pub const CLOSED_RING_FIELDS: usize = 5;
+
+pub struct GpuHindsight {
+    pub mid_ring_d: CudaSlice<f32>,
+    pub ring_write_idx_d: CudaSlice<u32>,
+    pub peak_mid_d: CudaSlice<f32>,
+    pub entry_mid_d: CudaSlice<f32>,
+    pub position_dir_d: CudaSlice<i32>,
+    pub closed_ring_d: CudaSlice<f32>,
+    pub closed_h_t_d: CudaSlice<f32>,
+    pub closed_write_head_d: CudaSlice<u32>,
+    pub closed_count_d: CudaSlice<u32>,
+    pub closed_step_d: CudaSlice<u32>,
+}
+
+impl GpuHindsight {
+    pub fn new(stream: &Arc<CudaStream>, b_size: usize) -> Result<Self> {
+        Ok(Self {
+            mid_ring_d: stream.alloc_zeros(b_size * HINDSIGHT_RING_LEN).context("her mid_ring")?,
+            ring_write_idx_d: stream.alloc_zeros(b_size).context("her ring_write_idx")?,
+            peak_mid_d: stream.alloc_zeros(b_size).context("her peak_mid")?,
+            entry_mid_d: stream.alloc_zeros(b_size).context("her entry_mid")?,
+            position_dir_d: stream.alloc_zeros(b_size).context("her position_dir")?,
+            closed_ring_d: stream.alloc_zeros(CLOSED_RING_SIZE * CLOSED_RING_FIELDS).context("her closed_ring")?,
+            closed_h_t_d: stream.alloc_zeros(CLOSED_RING_SIZE * HIDDEN_DIM).context("her closed_h_t")?,
+            closed_write_head_d: stream.alloc_zeros(1).context("her closed_write_head")?,
+            closed_count_d: stream.alloc_zeros(1).context("her closed_count")?,
+            closed_step_d: stream.alloc_zeros(CLOSED_RING_SIZE).context("her closed_step")?,
+        })
+    }
+}
diff --git a/crates/ml-alpha/src/rl/isv_slots.rs b/crates/ml-alpha/src/rl/isv_slots.rs
index 2896384b9..089cacfa9 100644
--- a/crates/ml-alpha/src/rl/isv_slots.rs
+++ b/crates/ml-alpha/src/rl/isv_slots.rs
@@ -998,6 +998,36 @@ pub const RL_NOISY_SIGMA_INIT_INDEX: usize = 547;
 /// 0.0 at ISV allocation (alloc_zeros); first step sets it to 1.0.
 pub const RL_STEP_COUNTER_ISV_INDEX: usize = 548;
 
+// ─── Hindsight Experience Replay (HER) — bidirectional ──────────────
+//
+// Backward peak: the mid-ring tracks the recent price trajectory so a
+// closed trade can be replayed with a goal set at the peak mid seen
+// during the hold. Forward continuation: the closed_ring stores the
+// hidden state + trajectory at close, enabling synthetic transitions
+// that continue from the close point toward a later mid-price goal.
+
+/// HER activation threshold. When the raw PnL (before scaling) of a
+/// closed trade falls below this value, HER injection is suppressed
+/// for that episode (no benefit in replaying catastrophic losses as
+/// near-successes). Default: -50.0.
+pub const RL_HINDSIGHT_THRESHOLD_INDEX: usize = 549;
+
+/// Priority boost applied to HER-injected transitions in the PER
+/// priority tree. Multiplied onto the TD-error priority so hindsight
+/// transitions are sampled more aggressively than the same transition
+/// without goal relabelling. Default: 2.0.
+pub const RL_HINDSIGHT_PRIORITY_BOOST_INDEX: usize = 550;
+
+/// Diagnostic: number of HER transitions injected into the replay
+/// buffer this step. Written by the HER injection kernel; read by
+/// diag JSONL to monitor hindsight utilisation rate.
+pub const RL_HINDSIGHT_INJECT_COUNT_INDEX: usize = 551;
+
+/// Forward-continuation lookahead in steps. When replaying a closed
+/// trade with a forward goal, the kernel looks ahead this many steps
+/// in the closed_ring to find a trajectory segment. Default: 64.
+pub const RL_HINDSIGHT_FORWARD_LOOKAHEAD_INDEX: usize = 552;
+
 /// Last RL-allocated slot index (exclusive). The integrated trainer
 /// extends `ISV_TOTAL_DIM` to at least this value at trainer init time.
-pub const RL_SLOTS_END: usize = 549;
+pub const RL_SLOTS_END: usize = 553;
diff --git a/crates/ml-alpha/src/rl/mod.rs b/crates/ml-alpha/src/rl/mod.rs
index ac4908191..323962579 100644
--- a/crates/ml-alpha/src/rl/mod.rs
+++ b/crates/ml-alpha/src/rl/mod.rs
@@ -17,6 +17,7 @@
 
 pub mod common;
 pub mod dqn;
+pub mod gpu_hindsight;
 pub mod gpu_replay;
 pub mod frd;
 pub mod iqn;