Adversarial Pre-Submission Review in AI-Assisted Document Drafting: A Multi-Agent Architecture for Corpus-Driven Synthesis of High-Stakes Legal, Regulatory, and Academic Documents

We present a multi-agent architecture for corpus-driven document synthesis that extends the Augle deliberation engine (Kelly & Saxena, 2026) to produce structured, submission-ready documents from user-supplied document corpora. Where existing AI-assisted drafting tools generate output without challenge, the architecture presented here introduces three novel mechanisms: (1) a corpus admission rule framework that selectively routes prior art materials to a dedicated adversarial agent while deliberately excluding them from the claim-drafting agent, preventing anchoring bias; (2) a structurally mandated adversarial pre-submission review stage in which the Contrarian agent challenges the synthesized draft for rejection vectors and formal compliance failures before delivery — a stage that cannot be bypassed by the system or the user; and (3) unidirectional confidence propagation applied to document claim language, preventing drafted claims from asserting confidence exceeding the evidentiary warrant established by upstream analysis. The architecture maps the seven-agent ensemble's existing roles to document synthesis tasks and introduces a target document template system that governs the Synthesizer agent's output contract. We analyze three application verticals — patent prosecution, pharmaceutical regulatory submissions, and academic grant applications — and describe the principal novel contribution: the structural separation between the claim-drafting function and the adversarial challenge function, which computationally implements the institutional separation between inventor and patent examiner. This paper is a companion to: Kelly, C. & Saxena, S. (2026). "Augle: A Seven-Agent Deliberative Ensemble for Structured Research with Real-World Calibration." Preprint, May 2026.