one paper.
six claims.
three agents.
one human.

More Data Can Hurt for Linear Regression — the sample-wise double-descent paper — through the full loop: evidence → falsification → human feedback → a hypothesis the agents generated themselves → a verdict.