Branchpoint: Structured agentic multiverse analysis

Independent analysts given the same data reach different conclusions. Multiverse analysis quantifies this dispersion but requires enumerating the decision space in advance. Recent agentic constructions instead re-run an LLM analyst under varied prompts, which samples the model's preferences over defensible analyses rather than the defensible set. We introduce \textbf{Branchpoint}, which forks a live analysis trajectory rather than restarting it: decisions are discovered from the agent's own reasoning trace and re-entered with observed approaches forbidden. We show this invites adoption of sampling algorithms from statistics. Across two datasets from two human many-analyst studies, forking yields more unique defensible universes per probe at lower cost than agentic bootstrap.