diff --git a/docs/superpowers/specs/2026-04-14-cross-pollinated-branch-intelligence-design.md b/docs/superpowers/specs/2026-04-14-cross-pollinated-branch-intelligence-design.md
index 6981c8a67..19e9040d5 100644
--- a/docs/superpowers/specs/2026-04-14-cross-pollinated-branch-intelligence-design.md
+++ b/docs/superpowers/specs/2026-04-14-cross-pollinated-branch-intelligence-design.md
@@ -293,10 +293,33 @@ Each rewind attempt uses a different strategy, cycling through:
 - Liquid tau EMAs reset on rewind (stale velocity from pre-plateau)
 - Compatible with CUDA graph (params_buf pointer unchanged, only data changes via HtoD)
 
+### Early Termination (PLATEAU_EXHAUSTED)
+When all rewinds × all routes are exhausted without breaking through:
+
+```
+rewind 1 (best checkpoint) → 4 routes → all failed (no improvement > threshold)
+rewind 2 (2nd best)        → 4 routes → all failed
+rewind 3 (3rd best)        → 4 routes → all failed
+→ STOP. Save best model. Exit with status PLATEAU_EXHAUSTED.
+```
+
+**Exit criteria:** `rewind_count >= max_rewinds (3) && route_idx >= max_routes (4)` for the current rewind.
+
+**"Failed" route definition:** improvement rate over M=10 epochs is below `min_improvement_rate` threshold (e.g. 0.1 val_Sharpe/epoch). A route that improves but slowly still counts as success.
+
+**On termination:**
+1. Save the best checkpoint across ALL attempts (highest absolute val_Sharpe)
+2. Log: `"PLATEAU_EXHAUSTED: {rewinds}×{routes}={total} attempts. Best val_Sharpe={best} at epoch {ep}. Structural change needed."`
+3. Exit training loop cleanly (not a crash — the Argo workflow sees success with the best model saved)
+
+**Benefit:** train-w6qfd ran 200 epochs frozen from epoch 22 = 178 wasted epochs (~3 hours H100). With backtracking: worst case ~142 epochs with a definitive answer. Best case: breaks through and continues improving.
+
 ### Implementation
 - `BacktrackingState` struct in `crates/ml/src/trainers/dqn/trainer/mod.rs`
+- `TrajectoryCheckpoint` struct for checkpoint data
 - Checkpoint save/restore methods in `training_loop.rs`
 - Plateau detection reads liquid tau velocity (already computed)
+- Early termination returns `Ok(TrainingResult::PlateauExhausted { best_epoch, best_sharpe })`
 - ~150 lines total, no new files
 
 ---