How this library was written

The xResearch library was produced through human-directed, AI-assisted research and writing. The process used a personal model of the human author's writing and research heuristics, a retrieval system that made relevant heuristics available at decision points, multiple generative-AI agents, external scholarly sources, structured evidence records, mechanical verification tools, and repeated human direction.

This note states who did what. That transparency matters because a polished artifact does not reveal its production history. Draxler and colleagues found an “AI ghostwriter effect” in studies with 30 and 96 participants: people often claimed authorship of AI-generated text while omitting disclosure, even when they reported limited ownership; greater influence over the text increased ownership (2024). Process influence matters, but no interaction log can settle authorship by itself.

The note therefore avoids two shortcuts. It does not call AI systems authors: current ICMJE guidance excludes AI from authorship and leaves humans responsible for accuracy, attribution, integrity, and disclosure. It also does not claim that a visible human intervention proves understanding, originality, learning, or quality. A trace can support attribution and audit. The strength of the underlying contribution requires separate judgment and evidence.

The heuristic model was derived from the human author's own writing and scholarship, including dissertations, theses, peer-reviewed publications, and related research work. Those source materials capture recurring choices about how to frame a problem, find a gap through a field's own debates, order evidence, distinguish constructs, build an argumentative turn, use tables and figures, adapt the same analytic core for different audiences, and end by sizing a contribution against an unfinished problem.

The model does not replace those works or reproduce their private text in this library. It represents recurring guidance abstracted from them. During production, the Human-Augmented Guidance (HAG) system retrieved named heuristics relevant to a particular decision. The writer or research agent examined the guidance, checked any recorded reasoning pathway, and then applied, adapted, or rejected it. Retrieval made a heuristic available; it did not make the decision.

This distinction is central to the public claim. The model is not a generic style prompt placed at the beginning of a session. It can bring a human-derived rule to a particular question—such as whether a table is an evidence store or an argumentative step, whether a conclusion should return to a stakeholder, or whether an essay should change load order without changing the evidence. Yet the source of a heuristic does not guarantee that it fits the current problem. The decision record therefore names both the guidance and what was done with it.

The following table is a disclosure of functional roles, not a legal determination of authorship. It uses the logic of ANSI/NISO CRediT—make contribution types visible—while distinguishing human responsibility from tool assistance.

The program used a repeated six-stage cycle. The human author first defined the question, intended audience, claim boundary, and excluded material. HAG guidance was then retrieved for consequential research or writing choices. AI agents completed bounded research, drafting, or editorial work. Sources and artifacts were checked. A separate challenge pass attempted to find prior art, rival explanations, claim inflation, disclosure failures, and leakage. The human author then reviewed, revised, rejected, redirected, or approved the work. A review could reopen any earlier stage.

Parallel agents were used only for separable tasks such as source checking, family-level revision, ledger reconciliation, and independent challenge. Central thesis, construct definitions, cross-paper terminology, audience rules, and final integration remained shared decisions. This allocation was a control against parallel work producing multiple incompatible editorial programs; its effectiveness has not been independently evaluated.

Four records make the process inspectable.

The HAG decision log identifies the publication decision, model, heuristic name and identifier, whether it was applied or adapted, and its concrete influence. Each applied heuristic's recorded reasoning pathway was checked; an absent relationship is recorded as absent rather than silently invented. The number of entries is a process count. It is not a measure of human contribution, text quality, novelty, or learning. This note does not use an interim corpus-wide count. A release-specific count, when reported, is produced deterministically from unique HAG decision rows in the linked log and includes the total and identifier span; the log remains the audit source.

For this prepared corpus, the linked log contains 111 unique decision rows, HAG-001 through HAG-111: 109 are recorded as applied and two as adapted. None is recorded as rejected at the final decision-row level, although several applied heuristics led to rejecting a display, inference, or wording choice. These statuses describe how retrieved guidance entered the editorial record. They do not estimate its causal influence or the human share of authorship.

The source ledgers keep one record per study or authoritative source, including citation, review status, population, design, the authors' reported finding, effect size when available, conditions, criticism, workflow use, and a bounded claimable sentence. Duplicate reports are reconciled rather than averaged as though they were independent studies. Heterogeneous estimates remain study-level unless a cited review legitimately synthesized them.

The display and dependency registers state what argumentative job a table or figure performs, which evidence it uses, what readers must not infer, and where ideas or displays are reused. A no-chart decision is recorded when numbers are incommensurable or a visual would imply an effect that has not been observed.

The QA and revision records document challenges and corrections. These include source mismatches, unsupported numerical claims, internal-register language, accessibility problems, and material that crossed a public boundary. A correction does not erase the fact that the error occurred; it shows why layered checking is necessary.

The trace establishes that named human-derived heuristics were made available, that agents reported applying or adapting particular guidance, that research and drafting labor was distributed, and that source and editorial checks were performed. It can support a more specific disclosure than “AI was used.” It also permits readers to inspect whether the published structure is consistent with the declared reasoning rules.

It does not establish that the human author personally composed every sentence. It does not show that every logged heuristic caused the final choice, that unlogged influences were absent, or that a decision was intellectually good because it was human-derived. It does not prove understanding, originality, ownership, a unique voice, learning, complementarity, or superior scholarship. Counts of prompts, agents, decisions, citations, or revisions would add precision without necessarily adding validity.

Nor is the resulting library evidence that HHITL works. It is an operational case of the proposed pattern. Independent comparisons are required before attributing quality, efficiency, learning, or novelty to heuristic guidance.

This publication was developed through human-directed, AI-assisted research and writing. A personal HeuriSight heuristic model derived from the human author's prior scholarship supplied named guidance on framing, evidence, structure, displays, and conclusions. Generative-AI agents assisted with literature searching, source records, drafting, editing, and checks. Retrieved guidance could be applied, adapted, or rejected; consequential claims and publication decisions remained human responsibilities. Sources and process records are linked where possible. AI systems are not authors, and the human author accepts responsibility for the final publication.

In a released version, the final sentence records that the human author has reviewed and accepted the publication. It is an accountability statement, not evidence that the workflow caused quality, originality, understanding, or learning.

HeuriSight developed the mechanisms examined across this research program and operates the xResearch library. It therefore has a direct organizational interest in how the evidence, proposed constructs, and research agenda are interpreted. Each public item either carries an explicit disclosure or links to this note. No paper in the current library demonstrates a HeuriSight effect. Process transparency reduces ambiguity; it does not remove the need for skeptical review, independent validation, or correction.

Whether heuristic guidance improves scholarship is an empirical question. A credible comparison would assign multiple writers and multiple research tasks to at least four conditions: human work without generative AI; human work with generic AI; human work with the same heuristics as a static guide; and human work with contestable, point-of-decision HAG guidance. A model-only condition would establish the solo-system benchmark where the task permits it.

The primary outcome should be independently and blindly scored scholarly quality using a preregistered rubric. Citation accuracy, unsupported claims, prior-art coverage, originality judgments, time, workload, revision burden, and disclosure completeness should remain separate outcomes. Writers should later explain and defend the argument without the system; that measure would test understanding rather than artifact quality. New tasks and delayed assessments would be needed before making transfer or learning claims. Writers and tasks should both be treated as sources of variation.

The design must also include conditions in which retrieved guidance is weak, outdated, or inapplicable. A system that increases compliance with the author's past habits while suppressing better new reasoning would be a failure, even if the prose became more consistent. The strongest result would not be a distinctive house style. It would be evidence that explicit, contestable human heuristics improve work beyond generic AI and simpler guidance without weakening source fidelity, independent understanding, or the capacity to reject bad advice.

That evidence does not yet exist. This library documents a method that can now be evaluated, not a result that has already been earned.