Triangulating for Truthiness

Working effectively with large language models requires something that the speed and fluency of those models can easily obscure: rigorous human judgment. In this paper, Jonathan Snyder argues that LLMs are not deterministic systems—they operate on statistical probability, not hard logic—and that this makes them, without proper guardrails, a source of systemic risk in software development and operations. The antidote is not to use them less, but to govern them more deliberately, treating the human operator as the accountable epistemic anchor and the LLM as a well-informed contributor whose outputs are always hypotheses until proven otherwise.

Drawing on his own experiments refactoring a React application with LLM assistance, Snyder offers a set of concrete principles for effective AI collaboration: codifying architectural specifications as the source of truth, applying the five whys in a data-driven, staged approach, using automation as an Andon cord that halts work when AI-generated changes exceed what testing can safely evaluate, and resisting the "vending machine" prompt in favor of a Socratic dialogue that stress-tests assumptions and triangulates toward truth.

This paper is part of the Spring 2026 Enterprise Technology Leadership Journal, a collection of guidance papers from IT Revolution authors and past presenters at the Enterprise Technology Leadership Summit.