When you build software with a large language model, the same system that writes the code also writes the prose describing it — and that prose arrives in confident, claim-ready language by default. The result is that an LLM-collaborative portfolio accumulates "novel contributions" faster than anyone audits them. This short paper argues that the responsible counter-practice is adversarial self-audit before publication: take your own novelty claims, run each through a textbook-prior-art check and a literature search, reclassify honestly, and publish the result even when it deflates you. The author offers a worked case: a physics-heavy project carrying forty-nine numbered intellectual-property claims, of which, after audit, zero survived as confident novel contributions. The paper argues this outcome is a feature — the deflation is what makes the one surviving system-level claim credible — and that auditing-to-zero is a concrete, transferable discipline that does not require trusting the author's good intentions, only their willingness to publish the audit.
Brad M Lindsey (Sat,) studied this question.