The Epistemic Chain Reaction: Human-AI Multiplication from Questions to Readout-Distinguishable Hypotheses — A Controlled Amplification Architecture for Scientific Discovery

Series. Paper IV of the Bounded Knower sequence. Paper I: From Problem to Hypothesis (DOI 10.5281/zenodo.22307148). Paper II: Knowledge Topology and the First Passage to Usable Hypotheses (DOI 10.5281/zenodo.22307561). Paper III: Before Evidence Can Decide (DOI 10.5281/zenodo.22307564). Status. GLOSA K0 conceptual and agenda paper — timestamped and citable, not independently reviewed, no priority claim; central augmentation laws remain [Open]. Produced with the glosa methodology (DOI 10.5281/zenodo.22301060); the author's raw lines are in the Blackbox Log (concept DOI 10.5281/zenodo.22302518). Abstract. Readout Universe; Readout Genesis Large language models and multi-agent systems can already generate, critique, rank, refine, and experimentally update scientific hypotheses. Yet the strongest form of human-AI augmentation may not be one-shot idea generation or autonomous replacement of scientists. Building on three companion papers, this manuscript proposes a controlled epistemic chain reaction: a human retains the problem and question, while heterogeneous AI processes recursively decompose the question, traverse distant knowledge structures, generate candidate relations, critique them, collapse readout-equivalent variants, and propagate only new candidate classes that clear declared grounding and discriminability gates. Each retained frontier item can then seed new subquestions or rival hypotheses. A finite epistemic multiplication factor, kepi , is introduced as the number of new live, readoutdistinguishable, minimally grounded child classes produced per retained frontier item after pruning. The proposed architecture distinguishes productive multiplication from verbosity: ten linguistically different hypotheses that predict the same accessible records count as one class, and a large kepi is not a truth score. The strongest proposed operating regime is not permanent expansion but controlled supercritical exploration (kepi > 1) followed by evidence- and human-governed contraction (kepi < 1) for appraisal and convergence. This paper synthesizes current evidence from AI Co-Scientist, autonomous discovery agents, LLM hypothesis-generation studies, human-AI cocreativity, question-generation systems, and documented diversity/hallucination failures. It argues that AI's distinctive leverage lies in parallel semantic transport, crossdomain bridge search, iterative mutation, adversarial critique, and low-cost branching, while human leverage remains strongest in problem ownership, question significance, value-sensitive constraints, interpretation of warrant, and revision commitments. The manuscript closes with falsifiable comparisons among human-only, one-shot AI, autonomous AI, and recursive human-AI workflows. The central hypothesis is that the largest augmentation will come from a role-separated, diversity-preserving, recursively gated system that increases the rate at which genuinely discriminable hypotheses become live, not merely the number of generated sentences. Version 3 (2026-09-05). Restores this record to Paper IV alone; the companion State of Evidence registry (PDF + Claim_Registry_E01_E25.csv) that version 2 briefly carried now lives in its own record, DOI 10.5281/zenodo.22308066 (founder ruling: keep them separate).