Fan-Out Harvest Protocol (FOH): A Modular Method for Parallel, Standardized AI-Assisted Research Synthesis

The Fan-Out Harvest Protocol (FOH) is a reusable workflow for parallel, AI-assisted research. It is designed for research questions that can be broken into several independent sub-questions, explored in separate AI chat sessions, and then harvested into one comparable body of findings. FOH addresses four common problems in AI-assisted research: Context overload: large research tasks can exceed what a single AI chat window can reliably hold, track, and synthesize. Context anchoring: when several questions are asked in one long chat, earlier answers can bias later ones. Inconsistent outputs: when each AI session answers in a different format, results are difficult to compare, stack, or synthesize. Lossy synthesis: when large outputs are copied into a new synthesis window, material can be missed, distorted, or accidentally invented if part of the source material is omitted. FOH addresses these problems by separating the research task into clean, independent windows while preserving coherence through standardized Run Cards, a fixed return structure, shared confidence tags, and a mechanical harvest step. Each sub-question is run in its own session to reduce cross-contamination. Each session returns findings in the same nine-field structure, anchored by a worked example rather than a bare template. Synthesis then happens in the same harvest window that already contains all completed returns, rather than in a new chat where context must be reconstructed. FOH specifies a complete operating pattern for this process. It includes a Phase 0 framing lock to settle the research goal, gating decisions, assumptions, and kill criteria before writing any Run Cards. It provides a Run Card template for creating one self-contained prompt per sub-question. It defines a nine-field return skeleton covering the run header, TL;DR, key findings, evidence quality, open questions, dead ends and quarantine, seeds for the next pack, an optional operator note, and sources. The protocol also includes a two-axis confidence tag, [source: H/M/L | claim: H/M/L], which separates the reliability of a source from the strength of a specific claim. It defines ID governance rules so that the operator, not the model, controls numbering. Model-suggested follow-ups remain unnumbered until the operator chooses to promote them. FOH is intended as a practical research workflow, not as a formally validated research methodology. It does not include reproducibility data. Its defaults, including the three-sub-question minimum, H/M/L confidence tiers, and numbering conventions, should be adjusted for different models, domains, and research contexts. Development referenced a four-run internal pilot, but that pilot is not included in this deposit and is not required to use the protocol. FOH is a sub-protocol of Iterative AI Production (IAP) (https://zenodo.org/records/21969467), but it can be used on its own. The two concepts it borrows from IAP, kill criteria and provenance tagging, are defined inside the FOH materials, so no companion document is needed to run it. This Zenodo deposit contains the FOH protocol (in both PDF and Markdown formats) and a plain-language PDF companion. The Markdown files are recommended when copying Run Cards, since PDF copy-paste can alter quotation marks or formatting in ways that may break prompt syntax, and Markdown is easier for the AI to read.