fix: clamp scores to (0.001, 0.999) for Phase 2 strict range validation

environment.py

@@ -73,7 +73,7 @@ class Action(BaseModel):
 
 class Reward(BaseModel):
     """Deterministic grading result returned after each step."""
-    total_score: float = Field(..., ge=0.0, le=1.0, description="Final score 0.0–1.0")
+    total_score: float = Field(..., gt=0.0, lt=1.0, description="Final score in (0, 1) exclusive")
     breakdown: Dict[str, float] = Field(..., description="Score breakdown by category")
 
 
@@ -228,7 +228,7 @@ class CloudSREEnv:
         self._health = max(0.0, self._health - 5.0)
 
         reward = Reward(
-            total_score=0.0,
+            total_score=0.001,
             breakdown={
                 "status": 0.0,
                 "message_investigating": 0.0,
@@ -335,7 +335,7 @@ class CloudSREEnv:
             breakdown["penalty_system_crash"] = -SYSTEM_CRASH_PENALTY
 
         raw = sum(v for k, v in breakdown.items() if not k.startswith("penalty_"))
-        total = max(0.0, min(1.0, round(raw - penalty, 4)))
+        total = max(0.001, min(0.999, round(raw - penalty, 4)))
 
         breakdown["budget_spent"] = self._budget
         breakdown["final_health"] = self._health

inference.py

@@ -269,7 +269,7 @@ def run_evaluation():
 
                 if done:
                     reward = step_data["reward"]
-                    score = reward["total_score"]
+                    score = max(0.001, min(0.999, float(reward["total_score"])))
                     results[task] = score
                     print(f"[STEP] step={obs['turn_number']} reward={score}", flush=True)
                     print(f"[END] task={task} score={score} steps={obs['turn_number']}", flush=True)
@@ -282,10 +282,10 @@ def run_evaluation():
                     obs = step_data["observation"]
 
             except Exception as exc:
-                print(f"[STEP] step={obs['turn_number']} reward=0.0", flush=True)
-                print(f"[END] task={task} score=0.0 steps={obs['turn_number']}", flush=True)
+                print(f"[STEP] step={obs['turn_number']} reward=0.001", flush=True)
+                print(f"[END] task={task} score=0.001 steps={obs['turn_number']}", flush=True)
                 print(f"  ERROR: Environment step failed: {exc}")
-                results[task] = 0.0
+                results[task] = 0.001
                 done = True
 
         time.sleep(0.5)

server/environment.py

@@ -73,7 +73,7 @@ class Action(BaseModel):
 
 class Reward(BaseModel):
     """Deterministic grading result returned after each step."""
-    total_score: float = Field(..., ge=0.0, le=1.0, description="Final score 0.0–1.0")
+    total_score: float = Field(..., gt=0.0, lt=1.0, description="Final score in (0, 1) exclusive")
     breakdown: Dict[str, float] = Field(..., description="Score breakdown by category")
 
 
@@ -228,7 +228,7 @@ class CloudSREEnv:
         self._health = max(0.0, self._health - 5.0)
 
         reward = Reward(
-            total_score=0.0,
+            total_score=0.001,
             breakdown={
                 "status": 0.0,
                 "message_investigating": 0.0,
@@ -335,7 +335,7 @@ class CloudSREEnv:
             breakdown["penalty_system_crash"] = -SYSTEM_CRASH_PENALTY
 
         raw = sum(v for k, v in breakdown.items() if not k.startswith("penalty_"))
-        total = max(0.0, min(1.0, round(raw - penalty, 4)))
+        total = max(0.001, min(0.999, round(raw - penalty, 4)))
 
         breakdown["budget_spent"] = self._budget
         breakdown["final_health"] = self._health