What's this about?
DCP tests whether an AI research result is both useful and unrecoverable by matched agents given the same starting information but not the target research history.
What exactly is a “recovery witness”?
It is an independently reproduced route to the claimed numerical result. If comparison agents recover it, the result may be derivable from existing information rather than evidence of a novel discovery process.
Why is this worth noting?
Two controlled audits each saw zero recoveries in 96 episodes and used an LLM-free verifier over frozen evidence. The protocol offers stronger evidence language than reporting a higher benchmark score alone.