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

1 Introduction

• Direct prompt-to-text interfaces leave writers with limited means to inspect or negotiate assumptions about conflict, character, theme, pacing, and point of view across scenes.These decisions are often compressed into a prompt or left implicit in model parameters.Source: §1 · p. 1
• Prior systems explore continuations, co-writing, emotional-trajectory controls, and hierarchical script generation, leaving open what knowledge should populate intermediate representations and how writers can inspect its influence.The passage cites Creative Help, Wordcraft, TaleBrush, and Dramatron.Source: §1 · p. 1
• Incipit treats an axiom as a provisional, reusable proposition about human experience or craft, not a self-evident truth or mathematically guaranteed claim.Axioms may be culturally situated, contestable, or in tension with one another.Source: §1 · p. 2
• The workflow elicits intent, retrieves candidate axioms, lets writers establish a working set, generates a hierarchical blueprint, and uses it for scene writing and diagnosis.Its procedural commitment is that structure precedes prose, with writer intervention at each structural level.Source: §1 · p. 2
• The system contributes a shared inspectable vocabulary, a five-round dialogue and matching pipeline, and an editable three-level blueprint that carries selected propositions into planning.The blueprint supports creative premises and subsequent planning context, with partial tracing across levels.Source: §1 · p. 2
• The paper proposes an evaluation protocol to separate axiom grounding from dialogue and hierarchical planning, while presenting its worked example as illustrative rather than evaluative.The contribution is a system design and implementation account, not an outcome evaluation.Source: §1 · p. 2

2.1 Creativity support and mixed initiative

• Creativity-support and co-creative frameworks motivate a blueprint as a coordination artifact where models contribute candidates and structure while writers retain the ability to accept, edit, replace, or reject them.The cited accounts include Shneiderman (2007), Deterding et al. (2017), Yannakakis et al. (2014), and Rezwana and Maher (2023).Source: §2.1 · p. 2

2.2 AI-assisted writing and narrative planning

• Unlike tools offering sentence suggestions, free-form co-writing, emotional-trajectory controls, or hierarchical script generation, Incipit gives intermediate structure a curated semantic vocabulary.Writers can identify propositions supporting a core conflict or narrative technique rather than relying only on generic beats or latent model associations.Source: §2.2 · p. 3

2.3 Narrative formalisms and explicit knowledge

• Narrative formalisms make patterns operational but risk presenting contingent interpretations as universal rules; Incipit therefore treats axioms as deliberation prompts, not literary constraints.The representation includes contextual types and typed relations and does not claim exhaustive coverage of literary meaning.Source: §2.3 · p. 3

3.1 The knowledge resource

• The checked-in resource contains 1,455 axiom records, 1,464 work–axiom mappings across 149 works, and 472 inter-axiom relationships organized across 11 domains.Records have content or form categories, context types, abstraction levels, and relations including complements, tensions, contradicts, specializes, and evolves_from.Source: §3.1 · p. 3
• Axiom statements support retrieval and discussion, metadata supports filtering and diversity checks, and work mappings and relations provide selection context without establishing universal validity.The system permits selecting axioms in tension when that tension is creatively deliberate.Source: §3.1 · p. 3

3.2 Four design commitments

• The system displays selected axioms as named creative cores and carries their identifiers into the blueprint, exposing part of the generation context as an editable object.Figure 1’s dashed feedback denotes writer-led revision rather than automatic optimization.Source: §3.2 · p. 3, Figure 1 · p. 4
• Planning proceeds from direction to premise, beats, chapters, and scenes, with later stages receiving earlier artifacts as context; its effect on perceived coherence remains an empirical question.Writers can supplement answers, choose options, patch blueprint layers, request alternative cores, regenerate chapter outlines, and revise scene text; general branching and automatic versioning are outside the implementation.Source: §3.2 · pp. 3–4
• Automated diagnostic scores and suggestions rely on model judgments and simple heuristics, so the system presents them as revision support rather than measurements of literary merit.The paper calls for making this status explicit, particularly for culturally dependent concepts such as universality and originality.Source: §3.2 · p. 4

4 Workflow and System Design

• The workflow connects the knowledge, workflow, and interaction layers to turn user input into structured project state and expose dialogue, blueprint, and scene operations.The intended sequence makes planning artifacts available before prose production.Source: §4 · p. 4

4.1 Five-round direction dialogue

• The five-round dialogue elicits a story core, conflict, differentiation, expression style, and confirmation, storing responses and extracted fields to inform subsequent rounds.Rounds two through four allow either selectable options or custom responses.Source: §4.1 · p. 4

4.2 Five-stage axiom matching

• A five-stage pipeline interprets the seed, generates thematic hypotheses, retrieves candidates, assembles a writer-selectable axiom set, and reviews it with heuristics and revision suggestions.Retrieval combines embedding search with optional candidates from domains implied by thematic hints.Source: §4.2 · pp. 4–5
• The retrieval contract retains at most 50 candidates, gives the first 30 to selection, and uses explicit bounds to define context size rather than an optimal policy.Selected records can include unreviewed proposals, and downstream provenance is not uniformly visible.Source: §4.2 · p. 5

4.3 Three-level blueprint

• The three-level blueprint carries selected propositions into editable premises and later planning, with structured levels validated before persistence.Its contracts establish objects and commitments at L1, order change at L2, and allocate change to bounded writing units at L3; valid references do not ensure artistic quality.Source: §4.3 · pp. 5–6

4.4 State transitions and intervention points

• Writers can intervene after completed stages, but edits update current records rather than creating a general branching graph, and blueprint edits can diverge from the axiom snapshot used downstream.The snapshot and blueprint do not form a fully synchronized provenance graph.Source: §4.4 · p. 6

4.5 Scene writing and diagnosis

• Scene generation uses the project axiom snapshot and other scene context, while diagnosis evaluates at most 6,000 characters and its scores and suggestions lack expert validation.The five diagnosis dimensions are self-consistency, universality, originality, consistent deduction, and connotation.Source: §4.5 · p. 7

4.6 Real-event abstraction

• The optional real-event route proposes story cores after abstraction, setting migration across a six-cell era–region grid, and residual-identifiability estimation; that score is not a privacy guarantee.Publication still requires human review for consent, defamation, confidentiality, and contextual harms.Source: §4.6 · p. 7

5 Implementation

• Incipit is a web application with a FastAPI backend, React frontend, PostgreSQL storage, JSONB artifacts, pgvector embeddings, and configurable model-provider integrations without automatic failover.The vector schema uses 1,024-dimensional columns, requiring a matching provider configuration or coordinated migration and re-embedding.Source: §5 · pp. 7–8

5.1 Prompt assembly and validation boundaries

• Generation helpers use a one-repair pattern for invalid JSON, while blueprint schemas constrain selected fields and references without guaranteeing persistent cross-level consistency or compelling scenes.The checks also do not enforce unique L2 beat identifiers or agreement between list position and order values.Source: §5.1 · p. 8

5.2 Planning and axiom grounding as separate variables

• The proposed protocol separates planning from axiom grounding so their contributions can be compared, but the prototype has not run these ablations.Software measures are partial, unvalidated proxies rather than one-to-one implementations of the theory’s five dimensions; the proposed study compares human ratings and writer experience against them.Source: §5.2 · pp. 9–10
• Table 4 illustrates one axiom’s propagation through the blueprint; only the axiom record comes from the checked-in resource, while the story artifacts are design examples.Source: Table 4 · p. 10

6 Worked Design Example

• The example is an authored illustration, not a user-study record or frozen model run, and makes no claim that its story is more original, coherent, or valuable.The trace was written for the paper and does not establish that a configured provider would generate the same words.Source: §6 · p. 10, §6.1 · p. 11
• A writer may select, reject, or replace a retrieved proposition as the working core for a premise about memory and a returning woman.The example uses a proposition about desire and the temporal illusion carried by its apparent object.Source: §6 · p. 10

6.1 From axiom to scene: a constructed trace

• In the constructed trace, the proposition is treated as a hypothesis for testing the protagonist’s interpretation, not as a lesson to announce.The writer’s note frames the axiom as a prompt for testing rather than a conclusion.Source: §6.1 · p. 11
• The trace carries a memory conflict from beats into a chapter and scene, where a ledger challenges Mara’s remembered account of her sister.The chapter links back to the axiom title rather than copying its statement into dialogue.Source: §6.1 · p. 11
• The scene is intended to realize the proposition through imagery, but diagnosis cannot establish artistic success or verify consistency beyond its 6,000-character excerpt.A writer could instead reject the archive, change point of view, or reconsider how the axiom’s tension develops.Source: §6.1 · p. 11

7 Evaluation Protocol

• No controlled user study, validated survey dataset, independently rated corpus, or preregistered comparison is available, so the paper makes no inferential claims from informal use.Primary outcomes, contrast, sample size, and analysis details must be fixed before preregistration and recruitment.Source: §7 · p. 11

7.1 Research questions and conditions

• The three conditions compare direct LLM writing, dialogue and L1–L3 planning without axiom knowledge, and the complete Incipit workflow.The questions address coherence, thematic integration and idea exploration, writer experience, and use of visible artifacts.Source: §7.1 · p. 12

7.2 Study design

• The proposed counterbalanced within-participant study uses comparable briefs, randomized condition order, controlled model settings and budgets, and a prespecified power analysis.Different briefs per writer are intended to limit carryover; recruitment would include writers with varied experience.Source: §7.2 · p. 12
• Artifacts would include an outline and prose excerpt, with writers allowed to edit outputs so the study evaluates collaboration rather than generation alone.Blinded literary judges would rate five dimensions; writer-facing measures include the Creativity Support Index, agency, ownership, workload, and interviews.Source: §7.2 · p. 12, §7.3 · p. 12

7.3 Measures and analysis

• The proposed instruments are unvalidated; the artifact rubric uses five-point anchors, and interaction logs would track edits, acceptance, regeneration, and time.The protocol proposes mixed-effects analysis, uncertainty intervals, reliability reporting, and judges blind to system heuristic scores.Source: §7.3 · p. 12, §7.4 · p. 13, Table 5 · p. 13

7.4 Proposed rubric and ablation record

• Plan versus Direct estimates dialogue, interface, and planning together, while Axiom versus Plan estimates incremental grounding; a failed manipulation check would hinder interpretation.The protocol also calls for matched briefs, common prompt components, and records of model settings, budgets, condition-specific prompts, and writer edits.Source: §7.4 · p. 13

8.1 A shared vocabulary, not a quality oracle

• Writers can point to, dispute, combine, or replace axioms, supporting reflection even when they reject a system recommendation.Explicit propositions provide a coordination option that latent model representations cannot offer directly.Source: §8.1 · p. 14

8.2 Scope of the current prototype

• The prototype demonstrates an executable workflow, not causal improvement, and its behavior depends on third-party models while several capabilities remain unimplemented.The knowledge base’s sources, extraction, translations, and taxonomy limit coverage; 149 mapped works cannot represent global literary diversity.Source: §8.2 · p. 14

8.3 Ethics, authorship, and data governance

• Real-event abstraction requires safeguards beyond a generated identifiability score, while generated prose and canonical mappings retain risks of stereotypes and narrow representation.The paper calls for retention controls, export and deletion mechanisms, human review, and appropriate ethics review and informed consent for future studies.Source: §8.3 · p. 14

9 Conclusion

• Incipit’s implemented architecture connects writer intent to planning and prose through an explicit, editable knowledge layer.The paper characterizes its contribution as an inspectable path from selected propositions to creative artifacts.Source: §9 · p. 14
• The prototype combines dialogue, axiom matching, a three-level blueprint, scene writing, real-event abstraction, and advisory diagnosis.The conclusion describes an architecture and workflow, not a measured outcome.Source: §9 · p. 14
• Whether the approach improves coherence, exploration, or writer agency requires the paper’s proposed controlled comparison.The conclusion does not claim those improvements have been established.Source: §9 · p. 15

• The hosted demonstration lets readers inspect and try the reported workflow, subject to service availability, account requirements, provider configuration, and deployment changes.Source: Artifact Availability · p. 15
• The live service is not a versioned reproducibility artifact; source code, curated data, prompt templates, and deployment configuration remain private, with no archival identifier or license claimed.Source: Artifact Availability · p. 15
• A future release should include versioned software and data, prompt templates, schema migrations, evaluation materials, provenance documentation, and licenses after source and rights review.Source: Artifact Availability · p. 15

• The appendix records implementation facts for the paper's described version as a compact manifest, not a permanently reproducible release.Source: A Implementation Manifest · p. 15
• The manifest is explicitly scoped to implementation facts that could otherwise be lost as the prototype evolves.Source: A Implementation Manifest · p. 15
• Table 6 presents configuration and boundary facts derived from the checked-in backend, not performance measurements.Source: Table 6 · p. 16

B Selected Prompt and Schema Excerpts

• The complete templates remain private; excerpts show constraint types and checked machine-readable keys without implying prompts alone control model behavior.Source: B Selected Prompt and Schema Excerpts · p. 17
• Table 7 pairs prompt constraints with corresponding implementation checks.Source: Table 7 · p. 17
• JSON keys, enum values, IDs, and other machine-readable fields stay in English while generated content may use the selected language.Source: B Selected Prompt and Schema Excerpts · p. 17
• Translating an identifier or enum value would make a literary choice appear to be a schema error.Source: B Selected Prompt and Schema Excerpts · p. 17

C Proposed Study Data Dictionary

• Future evaluation should archive enough information to distinguish writers' choices from system defaults; the proposed fields are not currently available as a study dataset.Source: C Proposed Study Data Dictionary · p. 18
• Table 8 lists proposed archival fields for a future preregistered comparison, with data collection requiring ethics review, consent, and a retention policy.Source: Table 8 · p. 18