The Process Is the Product: A Human-Directed, Multi-Model Adversarial Workflow for Trustworthy AI-Assisted Research

This paper reports a single-operator case study in directing several artificial-intelligence systems through a disciplined research and software-development process, and argues a narrow thesis: that the process—the documented, reproducible, accountable workflow—is a product in its own right, distinct from the papers and code it produced and valuable on its own terms. Over roughly three months an author with no formal programming or thermal-engineering background moved from a first, recoverable misstep, through a five-gate non-converging review streak, to a ratified methodology with an explicit stopping rule that reached three consecutive gates withzero unaccepted product blockers and was declared converged. The claim is comparative and defeasible rather than a proof: one operator’s history, retrospectively assembled, with model identity confounded with task assignment, and cross-model agreement is not independent validation. What the history supports is that an adversarial, multi-model, human-directed loop can fail to converge for a structural reason—fixing the probe rather than the defect class—that the failure is first visible in the finding-count trajectory, and that a few cheap, machine-checkable invariants plus a taxonomized stopping rule are what make an otherwise unbounded adversarialsurface tractable. I situate the workflow against Ethan Mollick’s co-intelligence framing and the “jagged frontier” evidence, against established project-management and engineering practice, and against the failure modes—AI “slop” and unreviewed “vibe coding”—it is designed to avoid. Its originating idea is borrowed from adversarial machine learning. The contribution is not a new component but a composition: different models in different institutional roles, with production separated from judgment by a literal seam.