What problem does it solve? Standard score-based paper reviews produce balanced weakness lists that never commit to the single most damaging rejection argument a hostile reviewer would write. This Skill stress-tests a LaTeX theory paper by forcing one fresh reviewer to write the strongest ~200-word rejection memo, then having a second independent reviewer adjudicate each atomic rejection point against the current source. ## Core Features & Use Cases - Adversarial attack memo: A fresh cross-model reviewer (gpt-5.5 via Codex MCP, zero prior context) constructs the single strongest rejection paragraph targeting theorem validity, scope overclaim, missing proof obligations, and claim-vs-evidence gaps. - Point-by-point adjudication: A second fresh reviewer decomposes the attack into 3-7 atomic points and classifies each as answered_by_current_text, partially_answered, or still_unresolved, with file:line evidence and severity. - Computed verdict ledger: The skill maps per-point counts to a PASS/WARN/FAIL/NOT_APPLICABLE/BLOCKED/ERROR verdict and writes KILL_ARGUMENT.md, KILL_ARGUMENT.json, and optional HTML, with sha256 input hashes for staleness detection. - Use Case: Before submitting a NeurIPS theory paper that has plateaued at 7-8/10 in standard review rounds, run the kill-argument exercise to surface the headline-vs-body scope gap a senior area chair would seize on, then prepare rebuttal responses in advance. ## Quick Start Run the kill-argument adversarial review on my paper directory to find the strongest rejection argument and check whether the current text answers it.
