graders result clampped between 0 and 1

inference.py

@@ -155,6 +155,9 @@ def run_task(
     success    = False
     last_error = None
 
+    def clamp_score(value: float) -> float:
+        return max(0.0001, min(0.9999, round(value, 4)))
+
     while not done and step < MAX_STEPS:
         step += 1
         last_error = None
@@ -193,7 +196,7 @@ def run_task(
             result  = env.step(action)
             obs     = result.observation.model_dump()
             done    = result.done
-            reward  = result.reward.total
+            reward  = clamp_score(result.reward.total)
             rewards.append(reward)
 
             # Check if operation failed
@@ -227,8 +230,8 @@ def run_task(
     print(f"[END] success={success_str} steps={step} rewards={rewards_str}")
 
     raw_score = rewards[-1] if rewards else 0.0001
-    final_score = max(0.0001, min(0.9999, round(raw_score, 4)))
-    safe_rewards = [max(0.0001, min(0.9999, r)) for r in rewards]
+    final_score = clamp_score(raw_score)
+    safe_rewards = [clamp_score(r) for r in rewards]
     return {
         "task_id":      task_id,
         "final_score":  final_score,