diff --git a/agent/eval_on_track.py b/agent/eval_on_track.py
new file mode 100644
index 0000000..c377090
--- /dev/null
+++ b/agent/eval_on_track.py
@@ -0,0 +1,95 @@
+"""
+eval_on_track.py — Evaluate a saved model on any track, zero-shot.
+
+Usage:
+    python3 eval_on_track.py --model models/wave4-champion/model.zip \
+                              --track donkey-generated-roads-v0 \
+                              --episodes 3 --max-steps 3000
+
+This is the proper zero-shot evaluation: load a trained model, connect
+to a track it has never seen, run N episodes, report reward and steps.
+"""
+import argparse, os, sys, time
+import numpy as np
+
+sys.path.insert(0, os.path.dirname(__file__))
+from stable_baselines3 import PPO
+from stable_baselines3.common.vec_env import DummyVecEnv, VecTransposeImage
+import gymnasium as gym
+
+from reward_wrapper import SpeedRewardWrapper
+from donkeycar_sb3_runner import ThrottleClampWrapper
+from multitrack_runner import StuckTerminationWrapper
+
+THROTTLE_MIN = 0.2
+SPEED_SCALE  = 0.1
+
+def make_env(env_id):
+    raw = gym.make(env_id)
+    env = ThrottleClampWrapper(raw, throttle_min=THROTTLE_MIN)
+    env = StuckTerminationWrapper(env, stuck_steps=80, min_displacement=0.5)
+    env = SpeedRewardWrapper(env, speed_scale=SPEED_SCALE)
+    return env
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--model',     required=True, help='Path to model.zip')
+    parser.add_argument('--track',     default='donkey-generated-roads-v0')
+    parser.add_argument('--episodes',  type=int, default=3)
+    parser.add_argument('--max-steps', type=int, default=3000)
+    args = parser.parse_args()
+
+    print(f'\n=== Zero-Shot Eval ===')
+    print(f'Model : {args.model}')
+    print(f'Track : {args.track}')
+    print(f'Episodes: {args.episodes} x max {args.max_steps} steps\n')
+
+    raw_env = make_env(args.track)
+    env = VecTransposeImage(DummyVecEnv([lambda: raw_env]))
+
+    model = PPO.load(args.model, env=env, device='auto')
+
+    all_rewards, all_steps = [], []
+    for ep in range(args.episodes):
+        obs, _ = env.reset()
+        total_reward, steps, done = 0.0, 0, False
+        pos_samples = []
+
+        while not done and steps < args.max_steps:
+            action, _ = model.predict(obs, deterministic=True)
+            obs, reward, terminated, truncated, info = env.step(action)
+            total_reward += float(reward[0])
+            steps += 1
+            done = bool(terminated[0] or truncated[0])
+            if steps % 100 == 0:
+                raw_info = info[0] if isinstance(info, (list,tuple)) else info
+                pos = raw_info.get('pos') if isinstance(raw_info, dict) else None
+                if pos is not None:
+                    pos_samples.append(np.array(list(pos)[:3]))
+
+        # Shuttle detection
+        note = ''
+        if len(pos_samples) >= 3:
+            net  = np.linalg.norm(pos_samples[-1] - pos_samples[0])
+            tot  = sum(np.linalg.norm(pos_samples[i+1]-pos_samples[i])
+                      for i in range(len(pos_samples)-1))
+            eff  = net/tot if tot > 0.1 else 1.0
+            if eff < 0.3 and steps >= 500:
+                note = f'  ⚠️  SHUTTLE EXPLOIT? macro_eff={eff:.2f}'
+
+        status = '✅ RAN FULL EVAL' if steps >= args.max_steps else '❌ CRASHED'
+        print(f'  ep{ep+1}: {total_reward:.1f} reward / {steps} steps '
+              f'({total_reward/max(steps,1):.2f}/step)  {status}{note}')
+        all_rewards.append(total_reward)
+        all_steps.append(steps)
+        time.sleep(0.5)
+
+    print(f'\n  Mean reward : {np.mean(all_rewards):.1f}')
+    print(f'  Mean steps  : {np.mean(all_steps):.0f}')
+    print(f'  {"✅ DRIVES" if np.mean(all_steps) > 500 else "❌ CRASHES"}')
+
+    env.close()
+    time.sleep(2)
+
+if __name__ == '__main__':
+    main()
diff --git a/agent/outerloop-results/autoresearch_phase4_log.txt b/agent/outerloop-results/autoresearch_phase4_log.txt
index 8850b45..2c94c20 100644
--- a/agent/outerloop-results/autoresearch_phase4_log.txt
+++ b/agent/outerloop-results/autoresearch_phase4_log.txt
@@ -766,3 +766,47 @@
 [2026-04-16 17:13:24] [Wave4] Proposed params: {'learning_rate': 0.0003194840316570061, 'steps_per_switch': 4507, 'total_timesteps': 66965}
 [2026-04-16 17:13:26] [Wave4] Launching trial 23: {'learning_rate': 0.0003194840316570061, 'steps_per_switch': 4507, 'total_timesteps': 66965}
 [2026-04-16 17:13:26] [Wave4] Command: python3 /home/paulh/projects/donkeycar-rl-autoresearch/agent/multitrack_runner.py --total-timesteps 66965 --steps-per-switch 4507 --learning-rate 0.0003194840316570061 --eval-episodes 3 --save-dir /home/paulh/projects/donkeycar-rl-autoresearch/agent/models/wave4-trial-0023
+[2026-04-16 18:26:11] [Wave4] Trial 23 finished in 4365.0s, rc=0
+[2026-04-16 18:26:11] [Wave4] Parsed: combined=150.773 mini_monaco=150.773
+[2026-04-16 18:26:11] [Wave4] ===== Trial 23 Summary =====
+[2026-04-16 18:26:11]   GP data points : 23
+[2026-04-16 18:26:11]   Wave4 Champion: trial=3 score=1943.10 params={'learning_rate': 0.0006852550685205609, 'steps_per_switch': 17499, 'total_timesteps': 157743}
+[2026-04-16 18:26:11]   Top 5:
+[2026-04-16 18:26:11]     score=1943.10  params={'learning_rate': 0.0006852550685205609, 'steps_per_switch': 17499, 'total_timesteps': 90000}
+[2026-04-16 18:26:11]     score=1573.07  params={'learning_rate': 0.00033896311195395496, 'steps_per_switch': 5448, 'total_timesteps': 69359}
+[2026-04-16 18:26:11]     score=1435.04  params={'learning_rate': 0.0007252855740444645, 'steps_per_switch': 6851, 'total_timesteps': 89893}
+[2026-04-16 18:26:11]     score=230.98  params={'learning_rate': 0.0006672844816013197, 'steps_per_switch': 4747, 'total_timesteps': 64179}
+[2026-04-16 18:26:11]     score=193.23  params={'learning_rate': 0.0010680884757571496, 'steps_per_switch': 3269, 'total_timesteps': 81912}
+[2026-04-16 18:26:13] 
+[Wave4] ========== Trial 24/25 ==========
+[2026-04-16 18:26:13] [Wave4] GP UCB top-5 proposals:
+[2026-04-16 18:26:13]   UCB=7.5247 mu=6.2951 σ=0.6148 params={'learning_rate': 0.0003596012644521025, 'steps_per_switch': 8013, 'total_timesteps': 60054}
+[2026-04-16 18:26:13]   UCB=6.0685 mu=4.6817 σ=0.6934 params={'learning_rate': 0.0001643952288780874, 'steps_per_switch': 9023, 'total_timesteps': 61369}
+[2026-04-16 18:26:13]   UCB=5.8942 mu=5.3342 σ=0.2800 params={'learning_rate': 0.00036103199241236257, 'steps_per_switch': 6908, 'total_timesteps': 64716}
+[2026-04-16 18:26:13]   UCB=5.6180 mu=4.5299 σ=0.5441 params={'learning_rate': 0.0006214434613010804, 'steps_per_switch': 7433, 'total_timesteps': 61101}
+[2026-04-16 18:26:13]   UCB=5.2906 mu=4.8505 σ=0.2200 params={'learning_rate': 0.000363303441736895, 'steps_per_switch': 6643, 'total_timesteps': 65509}
+[2026-04-16 18:26:13] [Wave4] Proposed params: {'learning_rate': 0.0003596012644521025, 'steps_per_switch': 8013, 'total_timesteps': 60054}
+[2026-04-16 18:26:15] [Wave4] Launching trial 24: {'learning_rate': 0.0003596012644521025, 'steps_per_switch': 8013, 'total_timesteps': 60054}
+[2026-04-16 18:26:15] [Wave4] Command: python3 /home/paulh/projects/donkeycar-rl-autoresearch/agent/multitrack_runner.py --total-timesteps 60054 --steps-per-switch 8013 --learning-rate 0.0003596012644521025 --eval-episodes 3 --save-dir /home/paulh/projects/donkeycar-rl-autoresearch/agent/models/wave4-trial-0024
+[2026-04-16 19:09:39] [Wave4] Trial 24 finished in 2604.0s, rc=0
+[2026-04-16 19:09:39] [Wave4] Parsed: combined=37.824 mini_monaco=37.824
+[2026-04-16 19:09:39] [Wave4] ===== Trial 24 Summary =====
+[2026-04-16 19:09:39]   GP data points : 24
+[2026-04-16 19:09:39]   Wave4 Champion: trial=3 score=1943.10 params={'learning_rate': 0.0006852550685205609, 'steps_per_switch': 17499, 'total_timesteps': 157743}
+[2026-04-16 19:09:39]   Top 5:
+[2026-04-16 19:09:39]     score=1943.10  params={'learning_rate': 0.0006852550685205609, 'steps_per_switch': 17499, 'total_timesteps': 90000}
+[2026-04-16 19:09:39]     score=1573.07  params={'learning_rate': 0.00033896311195395496, 'steps_per_switch': 5448, 'total_timesteps': 69359}
+[2026-04-16 19:09:39]     score=1435.04  params={'learning_rate': 0.0007252855740444645, 'steps_per_switch': 6851, 'total_timesteps': 89893}
+[2026-04-16 19:09:39]     score=230.98  params={'learning_rate': 0.0006672844816013197, 'steps_per_switch': 4747, 'total_timesteps': 64179}
+[2026-04-16 19:09:39]     score=193.23  params={'learning_rate': 0.0010680884757571496, 'steps_per_switch': 3269, 'total_timesteps': 81912}
+[2026-04-16 19:09:41] 
+[Wave4] ========== Trial 25/25 ==========
+[2026-04-16 19:09:41] [Wave4] GP UCB top-5 proposals:
+[2026-04-16 19:09:41]   UCB=2.8514 mu=2.5523 σ=0.1495 params={'learning_rate': 0.0003128257557719074, 'steps_per_switch': 6836, 'total_timesteps': 62683}
+[2026-04-16 19:09:41]   UCB=2.6831 mu=2.3606 σ=0.1612 params={'learning_rate': 0.0004069124594262667, 'steps_per_switch': 6035, 'total_timesteps': 62228}
+[2026-04-16 19:09:41]   UCB=2.5561 mu=2.1065 σ=0.2248 params={'learning_rate': 0.00011678403675177319, 'steps_per_switch': 6295, 'total_timesteps': 60168}
+[2026-04-16 19:09:41]   UCB=2.4849 mu=2.1243 σ=0.1803 params={'learning_rate': 0.0002227611562953955, 'steps_per_switch': 6484, 'total_timesteps': 60292}
+[2026-04-16 19:09:41]   UCB=2.3964 mu=2.0203 σ=0.1880 params={'learning_rate': 0.0004890619502018291, 'steps_per_switch': 6956, 'total_timesteps': 63892}
+[2026-04-16 19:09:41] [Wave4] Proposed params: {'learning_rate': 0.0003128257557719074, 'steps_per_switch': 6836, 'total_timesteps': 62683}
+[2026-04-16 19:09:43] [Wave4] Launching trial 25: {'learning_rate': 0.0003128257557719074, 'steps_per_switch': 6836, 'total_timesteps': 62683}
+[2026-04-16 19:09:43] [Wave4] Command: python3 /home/paulh/projects/donkeycar-rl-autoresearch/agent/multitrack_runner.py --total-timesteps 62683 --steps-per-switch 6836 --learning-rate 0.0003128257557719074 --eval-episodes 3 --save-dir /home/paulh/projects/donkeycar-rl-autoresearch/agent/models/wave4-trial-0025
diff --git a/agent/outerloop-results/autoresearch_results_phase4.jsonl b/agent/outerloop-results/autoresearch_results_phase4.jsonl
index 94359c7..6f94b2c 100644
--- a/agent/outerloop-results/autoresearch_results_phase4.jsonl
+++ b/agent/outerloop-results/autoresearch_results_phase4.jsonl
@@ -20,3 +20,5 @@
 {"trial": 20, "timestamp": "2026-04-16T14:10:06.493399", "params": {"learning_rate": 0.0007104769870597334, "steps_per_switch": 5822, "total_timesteps": 73160}, "combined_test_score": 20.9587, "mini_monaco_reward": 20.9587, "model_path": "/home/paulh/projects/donkeycar-rl-autoresearch/agent/models/wave4-trial-0020/model.zip", "champion": false, "run_status": "ok", "elapsed_sec": 3578.8967142105103}
 {"trial": 21, "timestamp": "2026-04-16T15:41:17.902221", "params": {"learning_rate": 0.0006240263317913531, "steps_per_switch": 6248, "total_timesteps": 81430}, "combined_test_score": 86.1012, "mini_monaco_reward": 86.1012, "model_path": "/home/paulh/projects/donkeycar-rl-autoresearch/agent/models/wave4-trial-0021/model.zip", "champion": false, "run_status": "ok", "elapsed_sec": 5465.586006164551}
 {"trial": 22, "timestamp": "2026-04-16T17:13:22.634461", "params": {"learning_rate": 0.0010680884757571496, "steps_per_switch": 3269, "total_timesteps": 81912}, "combined_test_score": 193.2341, "mini_monaco_reward": 193.2341, "model_path": "/home/paulh/projects/donkeycar-rl-autoresearch/agent/models/wave4-trial-0022/model.zip", "champion": false, "run_status": "ok", "elapsed_sec": 5520.664301395416}
+{"trial": 23, "timestamp": "2026-04-16T18:26:11.713980", "params": {"learning_rate": 0.0003194840316570061, "steps_per_switch": 4507, "total_timesteps": 66965}, "combined_test_score": 150.773, "mini_monaco_reward": 150.773, "model_path": "/home/paulh/projects/donkeycar-rl-autoresearch/agent/models/wave4-trial-0023/model.zip", "champion": false, "run_status": "ok", "elapsed_sec": 4365.0226583480835}
+{"trial": 24, "timestamp": "2026-04-16T19:09:39.779549", "params": {"learning_rate": 0.0003596012644521025, "steps_per_switch": 8013, "total_timesteps": 60054}, "combined_test_score": 37.824, "mini_monaco_reward": 37.824, "model_path": "/home/paulh/projects/donkeycar-rl-autoresearch/agent/models/wave4-trial-0024/model.zip", "champion": false, "run_status": "ok", "elapsed_sec": 2604.015403985977}