Get the rejection before the venue sends it.
Four critics, one human. The first and most expensive critic cannot be delegated: an offline human read, 30–60 minutes, social media off. The cheapest is an LLM told to be hostile. The loop assembles four machine critics: Reviewer 2, the venue's award winners, the empirical standards, and the ten nearest prior papers.
The Reviewer-2 prompt. The model plays Reviewer 2 at the target venue, and the prompt embeds the venue's own review criteria (for ICSE'27: novelty, rigor, relevance, verifiability and transparency, presentation). It must quote the text that fails each criterion and end with one of the track's real outcomes. Structuring the critique under the venue's headings matters: those headings are the exact form the real rejection would arrive in. Run on this paper's own generated draft, it returned reject — correctly flagging placeholder numbers, an unrunnable ground truth, and six uncited papers.
Then triage. A follow-up prompt divides the review's findings:
Close the loop. After fixes, one automatic gate guards the front door: recode the title and abstract with the same topic flags as the reading set; they must match the body's coding with zero flips. Fail → back to drafting. Pass → ship, with a replication package.
Combines with: 4 — write to a grammar produces what the critics attack; 7 — pause and reflect keeps part-2 decisions human; 8 — keep the receipts answers the verifiability criterion in advance.