Human Hypothesis Persistence and AI-Assisted Peripheral Exploration in Scientific Discovery

Advances in AI for Science have produced systems that partially or extensively automate literature search, hypothesis generation, hypothesis evaluation, experimental planning, analysis, and even the entire research cycle. In some domains, including biomolecular design, recent systems have achieved striking results. However, treating scientific discovery as a single search problem and asking, in one undifferentiated form, whether AI can autonomously conduct scientific research risks conflating distinct kinds of scientific exploration. This paper conceptually divides scientific discovery into two regimes. The first is derivable or statistically continuous discovery, in which candidates can be derived and evaluated in a relatively continuous manner from existing data, theories, representations, and evaluation criteria. The second is transformative or distribution-distant discovery, in which evidence is initially weak, the hypothesis lies far from the distribution of existing knowledge or current evaluation criteria, and progress may require changing the problem formulation or the evaluative framework itself. In the former regime, AI is especially well suited to rapid large-scale candidate generation, evaluation, rejection, and iterative optimization. In the latter, however, the problem is not limited to whether AI can ever generate a transformative candidate. A generated hypothesis may be discarded because of a low initial evaluation before it has been sufficiently developed. Accordingly, this paper transfers the established epistemic problems of transient diversity and premature convergence in scientific communities into the internal architecture of AI for Science, and operationalizes hypothesis persistence as an explicit and measurable design variable. It then distinguishes an early persistence boundary from a later integration boundary within the same pipeline. A particular hypothesis core explicitly preserved by a human as an anchor for further search is termed a Human-Anchored Hypothesis; the mechanism that maintains such an anchor across time is termed Human-Anchored Hypothesis Persistence. This paper further distinguishes two selection bottlenecks in scientific reasoning as parallel aspects derived from the same underlying problem setting as the author's earlier Integration-Boundary Hypothesis. Whereas the previous paper addressed the problem of integrating multiple strong surviving candidates into a finite output, the present paper addresses the problem of potentially important candidates disappearing before they have developed into strong candidates. On this basis, the paper proposes a research configuration in which a Human-Anchored Hypothesis provides an explicit origin for exploration, Human-Anchored Hypothesis Persistence preserves that hypothesis across time, and AI Peripheral Exploration conducts large-scale literature search, falsification, formalization, experimental design, and cross-disciplinary connection around the anchored core.