CategorysafetyEvidencereportedAttributed toEric Siu
ProblemEither a human writes everything, or an agent publishes unreviewed work. Both scale badly.
TechniqueSplit generation from selection from publication. The agent produces a fixed number of candidates. A human picks. Only the approved subset is published or scheduled, and the results are measured afterwards.
Why it worksIt puts the human at the one point that is irreversible, and it caps volume, so a bad run produces bad candidates rather than bad public output.
When to useAny lane whose output reaches an audience: content, outreach, listings, applications.
When not toRead-only research, where there is nothing to publish.