Index of everything on this site: https://ai.hiina.space/index.md
The idea behind this report is a story autoencoder, modelled on game decompilation. Decompile a film or a story into a structured "source code" (scenes, characters, beats, tropes, perhaps an AI-generated TV Tropes page), regenerate the story from that source, and check that the round trip holds. Then remix at the source level, the way decompiled Super Mario 64 gets modded: "Hamlet, but Lovecraftian and gender-swapped." The questions are which parts of that loop already exist, which representations survive the trip, and how to tell a faithful remix from a fluent rewrite.
Both halves of a story autoencoder have prior art, but in our searches we found nobody who joins them. On the decompile side, film-to-screenplay systems exist: ScreenWriter (2024) works on whole films and OmniScript (2026) on 5 to 30 minute clips. There is also a long line of text-to-structure extractors for Propp functions, ATU tale types, Thompson motifs, turning points and TV Tropes. On the recompile side, plan-conditioned generators run from Plan-and-Write (2019) through Dramatron and DOC to Agents' Room (2025). What we found none of is a study that does the whole loop at story or screenplay scale: extract a structure from an existing story, regenerate from it, score the reconstruction against the source with controls, then edit the structure and score the edit.
The near-misses are worth knowing exactly. Shen & Ying compress Chinese web novels to outlines and expand them back, measuring distortion, but they make no edits and use no decoys. Chen & Si use story reconstruction as a feedback signal to refine annotation prompts (abstract only). Rewriting Video closes a video-to-text-to-video loop, but only on 8-second clips.
Across the separate pieces, the evidence points one way. Coarse, boundary-tolerant representations come back reliably: event and scene lists, and coarse type labels queried one at a time as a checklist. Named functions, exact beats, tropes, and causal or character-state links do not. GPT-4 reaches at best 27.1 F1 on 95 film tropes, against about 65 for humans. A 2019 model matches turning points 41% of the time on synopses, against 64% for humans. Decoders realise only about 72 to 78% of plan items, and the best book-scale claim checker reaches an F1 of 58.2. That makes decoy controls mandatory, for the same reason any LLM claim checker needs shifted controls. The edit side is thinner still: the only established benchmark edits the ending of a five-sentence story, and LLM "contemporary adaptations" of stories double or nearly quadruple their plot-hole rate. The project is open, cheap to start on text through the API, and its main methodological risk is that a decoder already knows Hamlet, so canonical works will appear to reconstruct from almost nothing.
Next section: https://ai.hiina.space/reports/story-decompiling-and-remixing/02-film-to-script-decompilers-recover-dialogue-well.md
Text licensed CC BY 4.0 (attribution: Hiina, ai.hiina.space). Source: https://github.com/hiinaspace/ai-hiina-space