Fix all declared and runtime 0.0/1.0 score references

- openenv.yaml: reward_range [0.0,1.0] -> [0.01,0.99], target_score 1.0 -> 0.99
- environment.py: all episode_score init/reset from 0.0 -> 0.01, intermediate info 0.0 -> 0.01

The LLM evaluator reads openenv.yaml and sees declared range/targets
that include boundary values 0.0 and 1.0, failing the strict open
interval check before even calling the API.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

app/environment.py

@@ -48,7 +48,7 @@ class APClerkEnvironment:
         self._observation:   Optional[APObservation] = None
         self._step_count:    int                     = 0
         self._done:          bool                    = False
-        self._episode_score: float                   = 0.0
+        self._episode_score: float                   = 0.01
         self._max_steps:     int                     = 1
         self._context_store: List[str]               = []
 
@@ -62,7 +62,7 @@ class APClerkEnvironment:
         self._task_id       = task_id
         self._step_count    = 0
         self._done          = False
-        self._episode_score = 0.0
+        self._episode_score = 0.01
         self._max_steps     = spec.max_steps
 
         obs = spec.generator(seed=seed)
@@ -125,7 +125,7 @@ class APClerkEnvironment:
             info: Dict[str, Any] = {
                 "task_id":    self._task_id,
                 "step_count": self._step_count,
-                "episode_score": 0.0,
+                "episode_score": 0.01,
                 "intermediate": True,
             }
             return self._observation, reward, False, info

openenv.yaml

@@ -26,61 +26,61 @@ tasks:
     name: "Perfect Three-Way Match"
     difficulty: easy
     expected_decision: APPROVE_FULL
-    target_score: 1.0
+    target_score: 0.99
 
   - id: easy_no_po_found
     name: "No Purchase Order Found"
     difficulty: easy
     expected_decision: REJECT
-    target_score: 1.0
+    target_score: 0.99
 
   - id: medium_quantity_shortfall
     name: "Quantity Shortfall Reconciliation"
     difficulty: medium
     expected_decision: APPROVE_PARTIAL
-    target_score: 1.0
+    target_score: 0.99
 
   - id: medium_price_discrepancy
     name: "Unit Price Discrepancy"
     difficulty: medium
     expected_decision: REJECT
-    target_score: 1.0
+    target_score: 0.99
 
   - id: medium_split_delivery
     name: "Split Delivery Reconciliation"
     difficulty: medium
     expected_decision: APPROVE_FULL
-    target_score: 1.0
+    target_score: 0.99
 
   - id: medium_vendor_mismatch
     name: "Vendor Name Mismatch"
     difficulty: medium
     expected_decision: REJECT
-    target_score: 1.0
+    target_score: 0.99
 
   - id: hard_policy_violation
     name: "Policy Violation — Unauthorized Freight Charge"
     difficulty: hard
     expected_decision: REJECT
-    target_score: 1.0
+    target_score: 0.99
 
   - id: hard_duplicate_invoice
     name: "Duplicate Invoice Detection"
     difficulty: hard
     expected_decision: REJECT
-    target_score: 1.0
+    target_score: 0.99
 
   - id: hard_partial_po_match
     name: "Partial PO Coverage"
     difficulty: hard
     expected_decision: APPROVE_PARTIAL
-    target_score: 1.0
+    target_score: 0.99
 
   - id: hard_tax_discrepancy
     name: "Unauthorized Tax Charge"
     difficulty: hard
     expected_decision: REJECT
-    target_score: 1.0
+    target_score: 0.99
 
 action_space:
   type: APAction
@@ -117,7 +117,7 @@ observation_space:
     action_history:   List[dict]
     context_notes:    List[string]
 
-reward_range: [0.0, 1.0]
+reward_range: [0.01, 0.99]
 reward_type: APReward
 reward_partial_credit: true
 episode_termination: terminal_action_or_max_steps