Incipit: Axiom-Grounded Scaffolding for Human-AI Literary Creation

Synthesis: The paper asks how a work's conceptual commitments connect creation, interpretation and evaluation, and answers with an intermediate unit: a proposition-like axiom realized through characters, events, language and form, then reconstructed differently by situated readers. It turns that account into an inspectable artifact — 1,455 axiom records, 1,464 mappings to 149 works, 472 typed links — and audits it. The audit confirms structural integrity and names gaps: the context taxonomy cannot separate Types C and D, cross-work reuse is sparse, and provenance is absent. Five dimensions and five falsifiable predictions are offered as a research program, not results.

Key Findings

The unit of analysis is a premise, not a theme. Each of the 1,455 axiom identifiers states a provisional organizing proposition and carries at least one literary example and one work mapping.
Relations, not membership, carry the organization. The 472 typed links comprise 350 complements, 55 tensions, 39 specializations, 21 evolutions and 7 contradictions; curator-supplied strengths average 0.839 (range 0.70–0.93).
The ontology is work-local, not cross-work. 1,464 mappings cover 149 works — 432 primary, 507 secondary, 525 technique roles — yet 1,448 of the 1,455 axioms map to exactly one work.
Content records outnumber formal ones. The snapshot holds 933 content records (64.1%) against 522 form records (35.9%), labeled 482 fundamental, 871 intermediate and 102 specific across 11 primary domains.
The context taxonomy does not work as coded. All 1,189 Type A records carry empty era and region lists, and all 47 Type C records populate both, so the stored fields cannot distinguish Type C from Type D.
The empirical claims are predictions, not findings. No expert or reader study was run; the protocol proposes five raters on a 35-record stratified sample, and H1–H5 each state what would count against them.

Four claims and the level between theme and text

The framework rests on four claims. Configurational organization treats a work as a selected set of premises and relations. Realization mediates meaning: choices of character, event, affect, language and form can qualify, test or transform a premise rather than illustrate it. Reading reconstructs and differentiates, so a reconstruction can recover an organizing commitment while altering its scope or relations. Evaluation is relational and multidimensional. Together they define an intermediate level between an abstract theme and a particular text — the level at which a writer revising literary creation can change an organizing commitment, a reader can justify a reconstruction, and a critic can compare two works. The care-and-autonomy miniature is invented, not observed; it shows that attribution and relation do explanatory work a premise inventory cannot, and that the configuration scaffolds revision rather than dictating it.

What the audit establishes, and what it cannot

Recomputed from the JSON files, the audit is structurally sound: 1,455 unique identifiers, every axiom carrying at least one example and one mapping, all 1,464 mapping endpoints and all 944 relationship endpoints resolving, and no duplicate (work, axiom) pair or self-link. Works carry 8–13 mappings (median 10; mean 9.83). The Great Gatsby stores 12 mappings whose roles separate central claims from technique, and the Hamlet and To Live slices contrast epistemic contamination with endurance. The audit also shows what the Knowledge Graph is not: only seven axioms map to more than one work, so reuse is sparse. The checks establish structural consistency, not interpretive quality: they cannot say whether a claim is insightful. Reading is situated, and the artifact records one curation, so treat the snapshot as a Benchmark candidate rather than a benchmark.

Five dimensions and a validation program that has not run

Coherence, contextual reach, distinctiveness, realization fidelity and interpretive richness stay separate rather than combining into a scalar measure; the dataset lacks the comparison corpus a novelty estimate would need, and a broad context label is explicitly not a proxy for quality. The mathematical vocabulary is bounded: the open premise space is not assumed enumerable, the records are interpretations rather than premises in a deductive system, and the strength field is a confidence-like annotation, not a truth value. The paper rejects the combinatorial analogy as an explanation of creative search and, citing Brooks's heresy of paraphrase, treats records as indexes into readings rather than texts. Validation is staged as a research program: expert rating of records, reader and comparison studies for H1–H5, then a creative-process study comparing configuration-guided with theme-based planning. Language models proposed candidate formulations while human curators selected, rewrote and grounded them, but no study has tested whether the structures change decisions.

What this means for practice

Instructors. Ask students to state one organizing premise for a text, point to the consequence that makes it available, and supply a plausible alternative reading, because the schema keeps statement, description, example and context separate.
Researchers. Separate record quality from evidence about reading: rate a stratified sample on single focus, clarity, grounding, context adequacy and analytic usefulness, and treat low agreement as a finding about underspecified labels or legitimate plural readings, not annotator failure.
Software developers. Preserve provenance before shipping this knowledge base into a writing interface — versioned candidates, prompt and model manifests, source-passage identifiers, adjudication logs — because the current snapshot cannot reconstruct its own pipeline.

Limitations

No expert ratings or human-subject results exist in the repository; the validation plan proposes five raters on a 35-record stratified sample, so every dimension claim is untested.
The snapshot preserves no versioned candidate pool, prompt or model manifests, source-passage identifiers, curator identities or adjudication logs, so candidate counts, curator numbers and rejection percentages cannot be reported.
All 47 Type C and all 123 Type D records populate both era and region lists, so the context labels cannot support comparative analysis until the taxonomy is repaired.
1,448 of 1,455 axioms map to exactly one work, so co-occurrence and graph analysis are sparse, and the artifact has no held-out annotation set, no calibrated strengths (curator-supplied mean 0.839) and no public archival identifier or dataset license.

Connected Concepts

Connected Articles

MindCopilot: Towards Formalizing and Evaluating Granular Human-LLM Co-Writing — MindCopilot: Towards Formalizing and Evaluating Granular Human-LLM Co-Writing
Enhancing creative writing with robot-LLM integration: The interplay of embodiment, AI creativity and user engagement — Enhancing creative writing with robot-LLM integration
Generativism: Toward a Learning Theory for the Age of Generative Artificial Intelligence — Generativism: Toward a Learning Theory for the Age of Generative Artificial Intelligence
What do you mean by human-AI collaboration: Prerequisite functions and the affordances needed to achieve it — What do you mean by human-AI collaboration: Prerequisite functions and the affordances needed to achieve it
Evidence-Grounded Multimodal Knowledge Graph Construction for Multi-Lecture Educational Reasoning — Evidence-Grounded Multimodal Knowledge Graph Construction for Multi-Lecture Educational Reasoning
Ontology-Based Layered Hybrid AI-Driven Knowledge Model for Personalized E-Learning — Ontology-Based Layered Hybrid AI-Driven Knowledge Model for Personalized E-Learning
Bypass, Offload, or Scaffold: A Conceptual Model of How Large Language Models Shape Learning — Bypass, Offload, or Scaffold: A Conceptual Model of How Large Language Models Shape Learning
Rethinking Scaffolding in LLM Tutors: The Interactional Mismatch Between Benchmarks and Real-World Deployments — Rethinking Scaffolding in LLM Tutors: The Interactional Mismatch Between Benchmarks and Real-World Deployments
TriKoNet: The Trivalence Model of Potential Co-Creativity in Socio-Technical Networks — TriKoNet: The Trivalence Model of Potential Co-Creativity in Socio-Technical Networks
Agreement Is Not Quality: Blind Expert Verification of Human and LLM Qualitative Coding When Human Consensus Is Not Ground Truth — Agreement Is Not Quality: Blind Expert Verification of Human and LLM Qualitative Coding
Academic Erasure: The Disappearance of Complexity Under AI-Supported Writing — Academic Erasure: The Disappearance of Complexity Under AI-Supported Writing
Generative Artificial Intelligence and Creativity in K–12 Education: A Systematic Scoping Review — Generative Artificial Intelligence and Creativity in K–12 Education: A Systematic Scoping Review

Citation

Liu, Q., & Zhao, C. (2026). Incipit: Axiom-Grounded Scaffolding for Human-AI Literary Creation. arXiv:2609.25504.

Metadata

Study design	Theoretical Analysis, System Development
Discipline	Humanities Education, Writing Education
Level	Higher Ed
Intended audience	Instructors, Researchers, Software Developers
Learning and instruction	Creativity, Scaffolding, Storytelling in Education, Situated Learning
Technologies and techniques	Knowledge Graph, Large Language Models (LLMs), Educational NLP
Assessment and measurement	Evaluative Judgment, Educational Measurement
Research methods and evaluation	Benchmark, Research Methods in AIED, Qualitative Research
Equity, ethics, and responsible use	Bias Mitigation, AI Use and Disclosure Statements
Foundations of AI in education	Theory Development in AI in Education, Theories and Frameworks, Human AI Collaboration
Page kind	Framework

Source History Edit Chat