NSTD.7 - Automated Narrativization and Story Planning

Type: DPF pattern body

Primary EntityOfConcern: planning, generating and inspecting a narrative produced with a tool, so that its content and form serve the intended use.

NSTD.7:1 - Problem frame

Use this pattern when LLM, NLG, graph-to-text, data-to-text, story-planning, schema-governed generation, or search is used to produce or repair narrative renderings.

First useful move: give the generator the material, the relations it should preserve and the intended reader’s task. Obtain a small draft and compare it with the material for omissions, inventions and misleading connections. Evaluate those defects before deciding how to use the draft.

What goes wrong if missed: fluent wording or a valid output schema hides an unsupported cause, missing condition or continuity error.

What this buys: useful automated drafting with a way to detect and repair the losses that matter to the reader.

NSTD.7:2 - Problem

Automated systems can produce fluent narratives that lose a source relation or violate a story constraint. A plan helps control generation, while inspection tests the actual result. Either can reveal a need to revise the other.

NSTD.7:3 - Forces

NSTD.7:4 - Solution

Work from a content plan to an inspected draft. For a short account, the plan can be a few instructions or marked source relations. Inspect a candidate before accepting, repairing or rejecting it; prior acceptance is not required to evaluate it. Use the optional note below when another worker, an experiment or a maintained workflow needs to recover those choices. Keep only the fields that serve that use.

GeneratedNarrativeUseNote@Context:
  sourceMaterialRef:
  generatedDraftRef:
  intendedReaderAndUse:
  requiredRelationsOrConstraints:
  generationMethodRef?:
  contentPlanRef?:
  discourseOrEventPlanRef?:
  schemaConstraintRefs?:
  sourceComparison:
  evaluationRef?:
  useDecisionOrRepair:
  consequentialUseDecisionMaker?:
  architectureSynthesisAssessmentRef?:
  additionalClaimGrounds?:
  revisitCondition?:

Choose additional checks from the claim being made:

The following operations describe a useful generation cycle. They can be interleaved: an inspected draft may reveal a better content selection or order. A short task can combine planning in one prompt and comparison in one reading. Separate stages or records are useful only when the task, tool or receiving use needs them.

For schema-governed generation, distinguish a format check from a content check. A schema can require scene fields, character roles, locations, source references, branches or game-engine inputs. Check whether those fields contain the right relations and satisfy the intended constraints. Record the constraints when later maintenance needs them; use NSTD.6 for narrative quality and C.34 when architecture-specific structural preservation is at issue. Neither a valid schema nor a complete note establishes factual accuracy.

For LLM-assisted analysis or theme generation, compare proposed themes with the source excerpts and serious alternative readings. The researcher decides which interpretation the report supports and explains its limitations. A generated plan can become input to later narrative work once its relevant claims or premises are suitable for that use; a named admission record is not a universal prerequisite.

Select a probe when its possible result can change the generation method, repair or intended use:

1. Source perturbation: change one relevant constraint and compare the next output. A failure to reflect the change exposes a grounding problem. A successful probe supports that tested relation; it does not prove fidelity for every output.
2. Structure recovery: ask a reader to reconstruct the relations needed for the task without seeing the prompt. Retain the first answer and actual assistance. A missing relation may require content or order repair.
3. Responsibility: for a consequential organizational use, establish who decides and can correct or withdraw it. A tool-generated assertion of permission supplies no such assignment.

NSTD.7:5 - Archetypal Grounding

NSTD.7:5.1 - Generated-narrative pipeline: graph-to-text case

An AI agent receives a concept graph and produces a polished explanation. The author compares the needed dependency and example relations with the draft, evaluates consequential losses and repairs the account. The optional record below supports repetition of that comparison.

GeneratedNarrativeUseNote@GraphToTextTeaching:
  sourceMaterialRef: concept graph with dependency, example, counterexample and evidence links
  requiredRelationsOrConstraints: prerequisite chain, contrast pairs, evidence-return points
  generationMethodRef: LLM-assisted graph-to-text workflow
  contentPlanRef: selected nodes and relations
  discourseOrEventPlanRef: didactic dependency order with a contrast reveal
  generatedDraftRef: prose candidate
  sourceComparison: required relations compared with their sentences
  evaluationRef: relevant NSTD.6 questions and results
  useDecisionOrRepair: repair any missing prerequisite before the learner uses that inference
  revisitCondition: source graph, generator behavior or reader result changes

Pipeline steps:

1. Select the nodes, relations and returns the learner needs.
2. Choose an order through NSTD.2, then generate a short draft.
3. Compare the draft with the graph. If it says “B occurs because A” but the graph supports only “A precedes B”, replace the causal claim or obtain the missing causal account.
4. Use NSTD.6 to assess the repaired version for the intended reading. The draft need not have been accepted before this assessment.
5. Keep the adequate result. For repeated improvement, use E.23 with the version, changed part, protected characteristics and comparison basis.

NSTD.7:5.2 - Probe suite for generated narrative

NSTD.7:5.3 - Before and after repair: generated seminar outline

Before:

The generated outline sounds coherent and covers all important ideas, so it can be used as a DPF learning route.

Failure: neither the needed source relations nor the learner’s task is identified. “Covers all important ideas” has no comparison basis.

After:

The outline selects EntityOfConcern, forces, solution, neighboring-pattern exits and the improvement loop, ordered by their teaching prerequisites. Compare those relations with the source and try the intended reconstruction task through NSTD.8 and NSTD.6. A worked design estimate can support a provisional lesson; a claim about what learners actually recover requires a reading or teaching trial. C.35 becomes relevant only if the generated result is being used to inform architecture work.

NSTD.7:5.4 - Generated-storycraft boundary

For a franchise continuation, compare the generated scene with the selected continuity, premise and character-motivation constraints. A surprising event is acceptable when those relations support it; if it relies on a contradiction, repair the event or change the stated premise. Source perturbation can test a recurring generator problem. A private critique and a publication have different permissions questions. NSTD.6 can evaluate the candidate before either use is accepted.

NSTD.7:5.5 - Calibration for generated narrative

NSTD.7:5.6 - FPF owner teaching

NSTD.7 connects content planning, realization and evaluation. A generated account has claims expressed through a publication form; its fluency alone does not establish those claims. Use C.2.8 for reader recovery, A.10 for evidence claims and B.3 for assurance claims when they arise. C.35 is restricted to generated or discovered results intended to inform architecture work. Source changes and changed generator behavior can reopen an earlier comparison.

An LLM drafts an explanation of FPF pattern use from notes. Compare its selected relations, order and returns with those notes and the applicable pattern content. NSTD.6 can assess the candidate immediately; if the necessary subject basis is missing, it leaves that amount unassigned while retaining independently supported defect findings. An absent generation journal does not erase visible product qualities or defects.

A graph-to-text system turns an event graph into a match recap. The event graph, source timestamp, uncertainty markers, and official-result refresh route are admitted source basis for this rendering. The generated recap is a carrier. If the system adds causal explanations not in the graph, those claims are not admitted by graph-to-text success. Repair by lowering causal language, adding source return, or opening the evidence owner.

A game story-planning pipeline generates a branching scene. The schema may require objective, location, actors, traits, constraints, and available actions. NSTD.7 treats those fields as method and source-plan support, not as proof of playable, coherent, or ethically acceptable narrative. Structural, semantic, executable, and human probes remain separate from fluency.

An LLM proposes themes from interview notes for qualitative narrative analysis. The generated theme list is not the researcher’s interpretation by default. Human interpretive agency remains live: the researcher checks source excerpts, reflexive stance, alternative readings, and admissible use before any narrative rendering or report uses the generated material.

Use these examples to distinguish evaluation from acceptance:

When automated repair is used, preserve version identity. “Regenerate until better” destroys improvement evidence. Record the previous carrier, changed prompt or method, selected changed slice, expected value movement, protected trade-offs, and re-evaluation route. A generated variant can be more fluent and still worse on epiplexity, source return, or agency discipline.

Pipeline variants by source type:

If a pipeline variant requires a source type not covered by the current source pack, mark the case as a source-refresh trigger rather than silently generalizing. A graph-to-text claim, for example, may require a more specific graph-to-text source than a general NLG survey. A game narrative pipeline may need executable or playability probes that a plain text-generation source does not supply.

NSTD.7:6 - Bias-Annotation

Fluency can conceal a missing condition or invented cause. Inspect the relation that matters to the reader, rather than inferring source fidelity from style. The opposite error is to require a complete production dossier before evaluating an observable defect. Select additional checks from the claim and consequence, and keep candidate evaluation available throughout drafting.

NSTD.7:7 - Conformance Checklist

NSTD.7:8 - Common Anti-Patterns and How to Avoid Them

NSTD.7:9 - Consequences

The benefit is automated drafting with a tractable source comparison and repair cycle. Its cost depends on the relations and consequences being checked; a small adequate account can be retained without a separate dossier.

NSTD.7:10 - Rationale

Planning makes selected content and order available for control; inspecting the resulting narrative checks whether that control succeeded. Evaluation therefore belongs inside the generation cycle and can support a later acceptance decision.

NSTD.7:11 - SoTA-Echoing

NSTD.7:11.1 - Operational comparison against domain vocabulary

Narratology distinguishes story, discourse and presentation; generation methods distinguish selected content, its order and its realization. Use these distinctions to locate a defect: omitted material, misleading sequence, unsupported relation or wording. A tool may combine these operations internally, so evaluate the output without inventing a production history.

Keep domain terms when they help select or repair the operation. Resolve an ambiguity where it changes the claim or intended use; another FPF pattern is useful only if its particular question is at issue.

Gatt and Krahmer’s NLG survey (2018) distinguishes planning and realization. Alabdulkarim, Li and Peng’s story-generation survey (2021) reviews controllability, commonsense, characters and evaluation limits. Cardona Rivera, Jhala, Porteous and Young’s narrative-planning review (2024) examines plan-based narrative methods. Kumaran, Rowe, Mott and Lester’s SceneCraft (2023) generates interactive scenes from author instructions; Rahman, Yu and Cho’s G-KMS (2026) tests structured generation, repair and execution in a compact Unity RPG setting. Ma et al.’s survey (2026) identifies continuing coherence, consistency, diversity, control and evaluation problems. Nguyen-Trung and Nguyen’s NITA (2026) assigns interpretive judgement to the researcher in its qualitative-analysis method. These contributions support different operations and conditions, not one compulsory workflow for every narrative.

Operational payload:

• From NLG, keep content planning, discourse planning, and realization separate. If a tool returns only final prose, inspect its source relation and ordering; reconstruct a plan when it helps repair or reproduce the result. A present-product judgement does not establish the tool’s hidden production history.
• Story-generation surveys motivate testing the characteristics that matter to the output’s use, including coherence, control or diversity. Their findings do not create a universal admission form.
• A plot or event plan guides generation. Compare the realized story with its intended goals and constraints rather than infer success from the plan.
• For an output that must execute in an engine, inspect parsing, state reachability and relevant semantic constraints. A successful schema check alone leaves other narrative qualities untested.
• From interactive game generation, executable or playability probes may be needed when the narrative must function inside an engine or workflow; text fluency alone is the wrong evidence.
• In the NITA method, the qualitative researcher interprets proposed themes in relation to excerpts, questions and alternatives. Preserve that responsibility when using this method; it is not a claim that all automated narratives require a human interpreter.

The practical consequence is to preserve the speed of automated drafting while examining the particular relations, constraints and outcomes on which the receiving use depends.

NSTD.7:12 - Relations

Uses G.2, C.35, A.6.3.NAR, C.2.8 for structural amount, C.33 for architecture-description adequacy, C.34, NSTD.6, A.19.ECS, C.16, E.22, E.23, A.10, B.3, D.1 through D.5, and G.11. E.22/E.23 use the evaluated version, changed part and relevant NSTD.6 results; a retry is an improvement only when comparison supports that conclusion. C.35 applies when the result informs architecture work; C.34 to architecture-specific preservation; G.2 to SoTA harvesting and synthesis across traditions. Reopen when the selected material, generator, method, schema, intended use, evaluation result or relevant research changes. Support-map entry: open Source Use And Refresh Map when generation, NLG, story-planning, schema, or source-pack claims are relied on; open DPF Precision Restoration And Owner Map when generated source plan, plot plan, schema constraint, admission, correspondence, or responsibility words blur object kinds; open Semiotic And Language-Precision Bridge when prompt output changes language state, coarsening, cue, or quality wording.

NSTD.7:End