Building a story atlas without confusing sources with AI retellings Şafak Akyol, developer of the Black Horse Stories app, built a public Fairy Tale Atlas that separates historical source texts from AI-generated retellings and original fiction using two distinct content stores and explicit provenance metadata. The system gives historical-text records precedence over generated narratives with matching identifiers, and renders language-specific notices labeling adaptations as AI-assisted retellings and original fiction as invented. A November 15 revision added an optional film mode, exposed only when matching language content exists. Disclosure: I’m Şafak Akyol, the developer of Black Horse Stories, a subscription app that uses AI to create personalised illustrated and audio stories. An AI assistant drafted this article from my project records, with implementation details checked against the current source code and content files. It describes implementation choices, not an independent review or a scholarly study. My daughter Alara’s requests for bedtime adventures about a black horse started the project. Alongside the app, I built a public Fairy Tale Atlas. That introduced a different problem from generating a readable story: how could the interface explain what a reader was actually looking at? A historical text, a translation, a condensed adaptation and a newly invented story can all occupy the same screen. Giving each one a source link does not make them interchangeable. The web implementation has two distinct content stores. Historical text records live in PHP files under content/atlas/texts/ . Finished editorial narratives live in JSON files under content/atlas/narratives/ . AtlasTaleNarrative::record checks whether an identifier already exists in AtlasTaleText::all . If it does, the narrative loader returns null. This gives the historical-text store precedence instead of silently substituting a generated narrative for that record. A historical record can contain the source language, author, edition year, source URL, original paragraphs, translations with credits, and modern retellings. The narrative records instead carry fields such as kind , multilingual titles and paragraphs, source basis , and editorial notes . The distinction is useful because “we have a reference” and “we reproduce the referenced text” are different claims. The Alice narrative is marked adaptation . Its source basis identifies the Project Gutenberg edition of Lewis Carroll’s complete book, and complete plot in source is true. Editorial notes describe the condensed retelling and its changes. The Manas-related record makes a deliberately different claim: { "kind": "original fiction", "source basis": { "complete plot in source": false } } Its source reference is a UNESCO record concerning oral transmission and transcription. The content file says that this reference does not provide a complete tale. The reading text, “The Unfinished Melody,” is new fiction inspired by the broad theme of recording an oral narrative. Its invented characters and events are not presented as an episode of Manas. That boundary should survive rendering. A contextual reference must not become apparent evidence for dialogue the system invented. AtlasTaleNarrative::copy provides separate notices in Turkish, English and German. Adaptations are identified as AI-assisted retellings, not exact historical texts or faithful translations. Original fiction gets a different notice explaining that its characters, dialogue and events are invented. The “Original text” tab does not manufacture an original when none is reproduced. Its notice explains the limitation. The sources notice also distinguishes contextual records from complete texts: a link may lead to a catalogue, research or a performance. For historical records, AtlasTaleText::paragraphs handles selection separately. The original mode uses the requested translation when available, otherwise the stored original; the source mode returns the source-language paragraphs. A fallback is not evidence that a translation exists. The November 15 revision adds another relevant distinction: selected entries have an optional film block. Alice, for example, has a separately named modern-version section associated with the 1951 animated film, described as a retelling in our own words rather than a screenplay. AtlasTaleText::modes exposes that extra mode only when the corresponding language content exists. It remains separate from both the book-based retelling and historical text. The same revision also reworked adaptations to retain major story elements. Alice’s current editorial notes document restoring the Queen, croquet and trial. Those earlier omissions should not be described as unresolved behavior in the current version. The separation extends beyond visible labels. Narrative structured data describes a ShortStory using the relevant notice. It adds isBasedOn for an adaptation. Historical-text structured data represents the original and retelling separately, with translation credits and translationOfWork where applicable. None of this proves historical provenance or guarantees editorial accuracy. It makes the application’s claims inspectable. The project notes describe AI-assisted checking, not universal human editorial review. For me, the reusable design principle is simple: store the relationship you can support, label the text you actually display, and leave uncertainty visible.