Index of everything on this site: https://ai.hiina.space/index.md
The question is not only "has anyone decompiled a story" but "what
would make story decompilation as checkable as a matching game decomp."
The answer the evidence gives is uncomfortable: nothing will, because no
story has a hash. What can be built is the analogue of a
NON_MATCHING build with a test suite. The suite is claim
recall, same-story judgements and contradiction checks. It earns trust
only through decoys, because LLM judges have been shown repeatedly to
reward length, overlap, names and their own prose. Other LLM-judged
annotation work, in music for one, has found the same pattern and the
same fix. Seen this way, the round trip is less a training objective
than a ground-truth generator for a field whose human gold sets are
measured in dozens of films and fifteen folktales.
The more novel insight is about where the latent should live. Folklore's type, motif and ecotype layering, industry's bible-plus-beats split, and the decomp world's code-data split all locate remixability in separating persistent identity from narrative function and from local surface, not in a dense global embedding. If experiments 1 to 3 show that a typed scene-card schema beats a prose synopsis at equal budget, and that schema-side edits produce fewer plot holes than prompt-side rewrites, that would be the first quantitative evidence that "story source code" exists in a useful sense. If not, the result is still publishable and useful: it would show that frontier models carry plot in prose about as well as in any schema, and that the remaining remix problem is consistency checking, not representation.
Text licensed CC BY 4.0 (attribution: Hiina, ai.hiina.space). Source: https://github.com/hiinaspace/ai-hiina-space