A Working Framework for Disciplined Nonlinear Inquiry in Human–AI Collaboration: Steve's Law

People do not always discover useful ideas by moving through a predetermined sequence. An observation may prompt an association, which opens a question in another domain, which generates a hypothesis that only later becomes intelligible as part of a coherent chain. Conventional reasoning practices often respond to this pattern in one of two unhelpful ways: they suppress the nonlinear path as distraction, or they accept an appealing chain without requiring adequate evidence. Steve's Law proposes a third option. It separates permissive discovery from disciplined validation and expresses that separation as a repeatable cycle: observation, association, question, hypothesis, investigation, evidence, attempted disproof, rejection or refinement, iteration, conclusion, and proportionate action. The framework synthesizes established ideas from the philosophy of science, creativity research, metacognition, cognitive offloading, assistive technology, neurodiversity scholarship, and human–artificial intelligence interaction. Its proposed contribution is not a new rule of logic, nor a theory of how novel ideas are generated. It is a procedural architecture governing the transition from unconstrained association to epistemically accountable claims and proportionate action — with explicit status labeling, disconfirmation, evidence provenance, risk-sensitive action thresholds, and an AI-mediated translation layer. The paper's claims are ordered in three tiers: (1) the general inquiry architecture, the primary contribution; (2) human–AI cognitive scaffolding as one implementation of that architecture; and (3) cross-neurotype communication as a specific application hypothesis. Artificial intelligence may support the process by externalizing working memory, translating between expressive forms, organizing claims, locating sources, generating alternatives, and prompting disconfirmation. In this role, it may function as an AI-mediated cognitive scaffold or assistive cognitive technology: not a replacement for thought, but a user-directed aid that helps a person access, organize, test, and communicate thought. It must not be treated as an evidentiary authority. Recent evidence that unguarded AI scaffolding can degrade unaided performance, homogenize output, and atrophy metacognition motivates the framework's insistence on friction, verification, and human authority over action. This conceptual paper defines the framework, locates it in adjacent literature — including Peircean abduction, the Geneplore model, divergent/convergent thinking, design thinking, extended-mind and distributed-cognition accounts, and contemporary research on AI cognitive scaffolding — proposes applications in education and work, outlines cautious mental-health-adjacent uses, presents four retrospective case narratives as process illustrations rather than validation, and develops falsifiable research propositions. Particular attention is given to cross-neurotype communication as an application hypothesis: an AI-mediated validation layer may help a person preserve an idiosyncratic path of discovery while translating the resulting argument into a form others can inspect. A final section sketches a future direction — the framework's procedural specification as a control policy for AI agent architectures — supported by a first literature audit, which found no verified system combining the architecture's seven elements; it is offered as a provisional research hypothesis, not a novelty claim. The paper concludes that Steve's Law is best treated as a working theory and research program. Its value depends on whether controlled studies show improvements in idea diversity, evidentiary quality, calibration, communication, and action without unacceptable increases in error, dependence, masking pressure, or cognitive burden.