Pick a paper, then run the agent live against its registered repository or replay a completed run. Every verdict is clamped to what the sandbox actually observed.