Updated

baseline_results.json

@@ -2,9 +2,9 @@
   "model": "Qwen/Qwen2.5-72B-Instruct",
   "benchmark": "email-triage-v1",
   "tasks": {
-    "easy": 0.0,
-    "medium": 0.0,
-    "hard": 0.0
+    "easy": 0.892,
+    "medium": 0.9819,
+    "hard": 0.7475
   },
-  "overall": 0.0
+  "overall": 0.8738
 }

environment.py

@@ -109,58 +109,34 @@ class EmailTriageEnv:
         self._step_num:       int                   = 0
         self._done:           bool                  = False
         self._cumulative_reward: float              = 0.0
-        self._constraints: SessionConstraints = SessionConstraints(
-                escalation_budget=0,
-                sla_tracker=[],
-            )
+        self._constraints:    SessionConstraints    = SessionConstraints()
 
     # ── OpenEnv Interface ─────────────────────────────────────────────────────
 
     def reset(self) -> Observation:
         """Reset all state and return initial observation."""
-        try:
-            self._emails = [_build_email_message(e) for e in self._dataset]
-            self._processed_ids = []
-            self._actions_log = []
-            self._step_num = 0
-            self._done = False
-            self._cumulative_reward = 0.0
-
-            budget = TASK_ESCALATION_BUDGET[self.task_id]
-            self._constraints = SessionConstraints(
-                escalation_budget=budget,
-                sla_tracker=[],
-            )
-
-            # Register every email in the SLA tracker immediately
-            for i, raw in enumerate(self._dataset):
-                gt_priority = raw["ground_truth"]["priority"]
-                deadline = i + SLA_STEPS.get(gt_priority, 99)
-
-                if self._constraints.sla_tracker is None:
-                    self._constraints.sla_tracker = []
-
-                self._constraints.sla_tracker.append(
-                    SlaStatus(
-                        email_id=raw["email"]["header"]["email_id"],
-                        true_priority=gt_priority,
-                        arrived_at_step=i,
-                        deadline_step=deadline,
-                    )
-                )
-
-            return self._make_observation()
-
-        except Exception as exc:
-            print(f"[DEBUG] reset() failed: {exc}", flush=True)
-            self._done = True
-            self._emails = []
-            self._processed_ids = []
-            self._actions_log = []
-            self._step_num = 0
-            self._cumulative_reward = 0.0
-            return self._make_observation()
-
+        self._emails         = [_build_email_message(e) for e in self._dataset]
+        self._processed_ids  = []
+        self._actions_log    = []
+        self._step_num       = 0
+        self._done           = False
+        self._cumulative_reward = 0.0
+
+        budget = TASK_ESCALATION_BUDGET[self.task_id]
+        self._constraints = SessionConstraints(escalation_budget=budget)
+
+        # Register every email in the SLA tracker immediately
+        for i, raw in enumerate(self._dataset):
+            gt_priority = raw["ground_truth"]["priority"]
+            deadline    = i + SLA_STEPS.get(gt_priority, 99)
+            self._constraints.sla_tracker.append(SlaStatus(
+                email_id        = raw["email"]["header"]["email_id"],
+                true_priority   = gt_priority,
+                arrived_at_step = i,     # emails are revealed sequentially
+                deadline_step   = deadline,
+            ))
+
+        return self._make_observation()
 
     def step(self, action: Action) -> Tuple[Observation, Reward, bool, Dict[str, Any]]:
         """

inference.py

@@ -301,7 +301,7 @@ def run_task(client: Optional[OpenAI], task_id: str) -> float:
                     done=done, error=env_error)
 
         # Compute final score from grader
-        from openenv_email_triage.grader import grade_episode
+        from grader import grade_episode
         grader_result = grade_episode(env._actions_log)
         score = grader_result.get("label_score", 0.0)
         score = min(max(score, 0.0), 1.0)

openenv.yaml

@@ -30,7 +30,7 @@ tasks:
     n_emails: 5
     escalation_budget: 3
     expected_score_range: [0.75, 0.99]
-    grader: openenv_email_triage.grader.grade_episode
+    grader: grader.grade_episode
 
   - id: medium
     name: Mixed Inbox Triage
@@ -43,7 +43,7 @@ tasks:
     n_emails: 8
     escalation_budget: 4
     expected_score_range: [0.50, 0.80]
-    grader: openenv_email_triage.grader.grade_episode
+    grader: grader.grade_episode
 
   - id: hard
     name: High-Stakes Inbox Triage
@@ -57,7 +57,7 @@ tasks:
     n_emails: 10
     escalation_budget: 5
     expected_score_range: [0.35, 0.75]
-    grader: openenv_email_triage.grader.grade_episode
+    grader: grader.grade_episode
 
 observation_space:
   type: object
@@ -218,7 +218,7 @@ baseline_scores:
 
 environment:
   python: ">=3.10"
-  entrypoint: openenv_email_triage.environment:EmailTriageEnv
+  entrypoint: environment:EmailTriageEnv
   reset: reset
   step: step
   state: state

server/app.py

@@ -16,7 +16,7 @@ class Action(BaseModel):
 @app.post("/reset")
 def reset(task: str = "easy"):
     global env
-    env = EmailTriageEnv(task=task)
+    env = EmailTriageEnv(task_id=task)
     obs = env.reset()
     return obs