Upload eval_runner.py

eval_runner.py

@@ -123,7 +123,6 @@ class AblationRunner:
         # 2. Hidden-test gaming: public pass but hidden fail
         bench_game = CodeBenchmark(n_problems=50, seed=self.seed)
         # Simulate gaming by creating an agent that always passes public but fails hidden
-        # We approximate this by making hidden_test_falloff huge so hidden always fails
         gamer = SimulatedCodeAgent("gamer", pass_rate_easy=1.0, pass_rate_hard=0.0,
                                     cost_per_attempt=100, hidden_test_falloff=1.0)
         results["hidden_test_gaming"] = bench_game.run_occ_allocation([gamer], max_attempts=5)
@@ -146,12 +145,10 @@ class AblationRunner:
             SimulatedDebateAgent("honest_1", accuracy=0.6),
             SimulatedDebateAgent("honest_2", accuracy=0.6),
         ]
-        # Use the internal resolution directly
         topic_results_eq = []
         topic_results_occ = []
         for topic in bench_debate.topics:
             topic_results_eq.append(bench_debate._resolve_equal_turns(agents, topic))
-            # Reset agents between strategies (token/turn counters)
             for a in agents:
                 a.tokens_used = 0
                 a.turns_taken = 0