Decrease MAX_STEPS to 250 and change LLM frame skip to 10

inference.py

@@ -346,9 +346,9 @@ def run_episode(task_id: int) -> float:
         info         = {}
 
         while not done:
-            # Frame skipping: only invoke the LLM every 5 steps to prevent 30-min evaluation timeouts.
+            # Frame skipping: only invoke the LLM every 10 steps to prevent 30-min evaluation timeouts.
             # Step skips use the heuristics to keep episode run-time blazing fast.
-            if step_count % 5 == 0:
+            if step_count % 10 == 0:
                 action = llm_agent(obs)
             else:
                 action = heuristic_agent(obs)

openenv.yaml

@@ -46,20 +46,20 @@ tasks:
     numeric_id: 0
     grader: time_to_detection
     max_score: 1.0
-    episode_length: 500
+    episode_length: 250
     description: Detect sinusoidal FDI attack within 100 steps of attack start
   - id: multi_attack_classification
     difficulty: medium
     numeric_id: 1
     grader: classification_accuracy
     max_score: 1.0
-    episode_length: 500
+    episode_length: 250
     description: Classify attack type from observation window
   - id: stealthy_attack_detection
     difficulty: hard
     numeric_id: 2
     grader: pre_lock_loss_detection
     max_score: 1.0
-    episode_length: 500
+    episode_length: 250
     description: Detect stealthy low-amplitude attack before PLL loss-of-lock
-episode_length: 500
+episode_length: 250

src/env.py

@@ -27,7 +27,7 @@ from src.detector import AdaptiveDetector
 
 
 WINDOW_SIZE = 20
-MAX_STEPS = 500
+MAX_STEPS = 250
 LOCK_LOSS_THRESHOLD = 0.0873  # 5 degrees in radians
 
 DETECTION_THRESHOLD = 2.0