Oracle Gap and Signal Fidelity: A Fixed-Pool Diagnostic for Test-Time LLM Collaboration

Before paying for multi-agent AI verifiers, check the math first: on some benchmarks there's barely any correct answer left to recover.