Published October 19, 2026 | Version 2.3

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

Authors/Creators

Description

Version 2.3 of Adversarial Multi-AI Falsification is a bounded case study of AI-assisted scholarly workflow governance. It distinguishes selected public verbatim excerpts and message/line maps from author-reported recovery findings and hashes identifying internal artifacts. The public deposit does not contain the full captured transcripts or executed rubric.

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. E2 remains partial. The principal execution/adjudication sequence remains author-reconstructed, and panel unanimity cannot be independently recomputed from the public deposit. The operational proposition remains untested; this revision does not establish it as true.

The paper proposes a three-field outcome record and six workflow controls, including a two-turn assignment-echo checkpoint. These are unvalidated design propositions, not demonstrated interventions. This is a case report and protocol demonstration, not a benchmark, product comparison, prevalence estimate or complete reproducibility study.

AMAF_v2_3.pdf is the current controlling manuscript, with AMAF_v2_3.docx as its editable counterpart. All seven v2.2 files are retained unchanged as historical records, including the old preparation-candidate manuscript, README, manifest and checksum list. Read 00_README_AMAF_v2_3_RELEASE.txt for the current controlling-version guide; AMAF_v2_3_SHA256SUMS.txt lists this release's file identities. The historical v1.8 paper is retained only for its Section C dual-track summary. No additional private records, new experiment or new empirical validation are released.

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 retains responsibility for the sources, interpretations, arguments and conclusions.

Files

AMAF_v2_3.pdf

Files (1.4 MB)

Name Size
md5:ad584d7ae5e8f66047f7775daea2760c
2.3 kB Preview Download
md5:c9e28359a4ab4de223ca4708fdd26e43
170.7 kB Preview Download
md5:aa88362bb2112e43fcb7775296db6ff3
316.0 kB Preview Download
md5:3d2baef965a61a8b3a70d47d9638c816
46.3 kB Preview Download
md5:b0cf10ab7d29b6be6f2baec74a02fc2d
38.3 kB Download
md5:513981daebd6d7674703c464f0e2d1f6
174.2 kB Preview Download
md5:9878ff05f828f3f4c26b82a3fe675767
1.0 kB Preview Download
md5:e320088abf2b938f3319359dcffb6e6c
806 Bytes Preview Download
md5:07d48dd7fc987ce6171a31a6376d8107
1.6 kB Preview Download
md5:2e61d3d14887b584ec19143e8698b181
216.2 kB Preview Download
md5:5a4a13508af076c97a42257f66cb9e04
442.9 kB Preview Download

Additional details

Related works

Is supplement to
Working paper: 10.5281/zenodo.21567334 (DOI)