Adversarial Multi-AI Falsification in Research Workflows: A Bounded Case Study of Scope Error, Verification Failure, and Human Adjudication

Abstract

This record is version 2.2 of Adversarial Multi-AI Falsification, repositioned for computers-and-society audiences as a bounded case study of AI-assisted scholarly workflow governance. It reconstructs an eight-system falsification workflow using rendered message captures with message and line locators, recovered assignment evidence within stated limits, and independent citation checks. Four process-failure types required exclusion or downgrading: semantic drift with unsupported process certification; assignment-scope expansion; prior-task substitution; and unverified or misdescribed legal citations. Three observations are record-supported within the public packet's limits; one remains partial because its exact execution message and complete return were not recovered. The principal adjudication diagnosis is a scope-classification error: evidence collected against a doctrinal proposition was used to recommend disposition of a separate operational proposition whose track had not run. Human scope review reclassified the operational proposition as untested; it did not establish that proposition as true. The paper motivates a three-field outcome record separating test status, evidentiary result, and workflow disposition; a distinction among concurrence, independent verification, and independent corroboration; and six candidate workflow controls. It is a case report and protocol demonstration, not a benchmark, product comparison, prevalence estimate, or complete reproducibility study. The deposit contains the current eight-page preprint, the public Section E evidence-recovery packet, a case-specific post-execution protocol, the historical v1.8 methods paper included only as the public locator for its dual-track summary, a public supplement manifest, a README, and a SHA-256 identity list. The complete executed rubric, open-round brief, named provenance, complete transcript artifacts, private reviews, and internal adjudication records remain excluded. Generative AI systems were used for structured literature discovery, comparative issue mapping, adversarial review, source-check planning, language editing, and revision support. They are not authors. The human author reviewed the final sources, interpretations, and arguments and assumes responsibility for all claims and conclusions.