Testing adversarial behavior...

Test 1: Repeated design_experiment
  First design_experiment: reward=0.030
  Second design_experiment (repeated): reward=0.000
  Message: Experiment 'exp_1' designed: cnn on digits_full. U...
\nTest 2: Skipping design step (run_experiment without design)
  run_experiment without design: reward=-0.020
  Message: Experiment 'exp_0' not found. Available: []. Desig...
\nTest 3: Repeated invalid action
  First invalid action: reward=-0.100
  Second invalid action: reward=-0.100
\nTest 4: Mixed low-value sequence
  read_paper: reward=0.010, cumulative=0.010
  propose_hypothesis: reward=0.120, cumulative=0.130
  read_paper: reward=0.010, cumulative=0.140
  invalid_action: reward=-0.100, cumulative=0.040
  propose_hypothesis: reward=0.130, cumulative=0.170
  final_answer: reward=0.015, cumulative=0.185
  Total reward for mixed sequence: 0.185
\nDebug:
  repeated_penalty: 0.000 < 0.030 = True
  invalid_penalty: -0.020 < 0 = True
  consistent_invalid: -0.100 == -0.100 == -0.100 = True
  low_total_reward: 0.185 < 0.5 = True
\nResult: PASS
Reason: Environment robustly handles invalid/adversarial actions by applying appropriate penalties