fix: emit strict score in END output and add scorer variance guard test

README.md

@@ -61,11 +61,11 @@ LLM mode is enabled by default in `inference.py` and requires:
 
 ```bash
 export API_BASE_URL="https://api.openai.com/v1"  # or validator-provided proxy URL
-export OPENAI_API_KEY="your-validator-provided-token"
+export HF_TOKEN="your-validator-provided-token"
 ```
 
 `MODEL_NAME` is optional and defaults to `meta-llama/Llama-3.1-8B-Instruct`.
-For compatibility with different validator versions, `API_KEY` and `HF_TOKEN` are also accepted.
+`HF_TOKEN` is required by this script.
 
 To run greedy mode locally without LLM calls:
 

inference.py

@@ -85,10 +85,10 @@ def log_step(step: int, action: str, reward: float, done: bool, error: Optional[
     )
 
 
-def log_end(success: bool, steps: int, rewards: List[float]) -> None:
+def log_end(success: bool, steps: int, score: float, rewards: List[float]) -> None:
     rewards_str = ','.join(f'{r:.2f}' for r in rewards)
     print(
-        f'[END] success={str(success).lower()} steps={steps} rewards={rewards_str}',
+        f'[END] success={str(success).lower()} steps={steps} score={score:.3f} rewards={rewards_str}',
         flush=True,
     )
 
@@ -250,7 +250,8 @@ async def run_episode(url: str, difficulty: str = 'medium', use_llm: bool = Fals
         print(f'[DEBUG] Episode error: {exc}', file=sys.stderr, flush=True)
 
     finally:
-        log_end(success=success, steps=steps_taken, rewards=rewards)
+        score = _strict_unit_interval(score, fallback=0.5)
+        log_end(success=success, steps=steps_taken, score=score, rewards=rewards)
 
     return score
 

tests/test_openenv_env.py

@@ -58,6 +58,25 @@ def test_score_in_range():
     assert 0.0 < env.score() < 1.0
 
 
+def test_score_varies_across_seeds():
+    scores = []
+    for seed in [1, 7, 13, 21, 42]:
+        env = ContainerYardEnvironment()
+        env.reset(difficulty="medium", seed=seed)
+        done = False
+        while not done:
+            stacks = as_dict(env._observe())["stack_states"]
+            chosen = next(
+                (i for i, stack in enumerate(stacks) if len(stack) < env.max_height), 0
+            )
+            obs = as_dict(env.step(ContainerAction(stack_index=chosen)))
+            done = obs["done"]
+        scores.append(env.score())
+
+    # Avoid disqualification: grader must not return a constant score.
+    assert len(set(scores)) > 1, f"Scores are constant across seeds: {scores}"
+
+
 @pytest.mark.parametrize("difficulty", ["easy", "medium", "hard"])
 def test_full_episode_completes(difficulty):
     env = ContainerYardEnvironment()