About

Narrative Structures is an exploration of one intersection of the humanities with AI — using language models as collaborators in the structural analysis of film. It is part of a broader inquiry into the role of AI in humanities education: where the technology helps the humanist do the work, where it doesn't, and what the work looks like when both partners are honest about what they bring. Each essay is the product of that collaboration, with the AI's contributions documented and the human's framing decisions named.

What "AI-assisted" means here

The pipeline behind these essays is built around a small set of specialized sub-agents, each with a narrow job and a verification loop attached. One agent annotates a film's subtitle file scene by scene, identifying speakers, visual action, and natural beat boundaries. A second agent reads that annotated transcript and the film's screenplay (when available) and produces a forty-beat structural breakdown — a chronological partition of the film, with timestamps, organized around a small fixed set of structural pivots called rivets.

That breakdown is then audited by a separate fact-checking agent. The auditor doesn't trust its own earlier reasoning; it goes back to the subtitle file as ground truth and verifies every factual claim line by line. When the screenplay disagrees with the finished film, the finished film wins, and the discrepancy is noted. Claims that can't be verified from the available evidence get flagged in place rather than smoothed over.

The essays draw from the verified beat breakdowns and from a parallel wiki of supporting research — actor pages, production histories, signature-scene write-ups, era and lineage essays — that the AI helps populate but that a human curates. Quoted material is verified against its alleged source. Dialogue is cited to specific subtitle timestamps. Where the AI hallucinates, the verification loops are designed to catch it; where they don't, the human is supposed to.

What the human still does

Framing. Judgment. The decision about which film, and why now, and what the essay is actually for. The decision about which structural reading is interesting and which is merely correct. The taste-level work of selecting which beat is a midpoint and which is just a strong scene that happens to fall near the middle. The willingness to throw out an analysis when it isn't catching anything the unaided reader couldn't see.

Most importantly: the test of whether the framework is doing real work. A structural reading is only worth publishing if it changes how you see the film — if it surfaces a pattern the AI's pattern-matching alone wouldn't have surfaced, or if applying the framework consistently to a film exposes a genuine choice the filmmakers made. When the analysis comes back generic, it gets cut.

Why film structure

Film is a domain where the AI's strengths and the human's strengths are mostly non-overlapping. The AI is patient with hours of subtitle data; it can hold all forty beats in attention at once; it will count things and cross-reference them and not get bored. The human has taste, history, the felt sense of what a particular scene is doing, the willingness to be wrong out loud about what a film means. Neither does what the other does.

Structural analysis sits in the seam. It needs the AI's diligence with the subtitle file and the human's judgment about which structural pattern is worth naming. When the seam is honest, the result is better than either could produce alone. When it isn't, the result is plausible-sounding nonsense — confidently structured, factually adrift, telling you nothing you didn't already know. The verification loops exist because the difference between the two is easy to miss from the inside.

The framework

The structural framework these essays use is described in Methodology. Briefly: every film is read as a sequence of ten pivots — the rivets — that partition the film into clear functional regions, with a quadrant assignment that says something about whether the protagonist's initial approach was sufficient and whether better tools were available. The rivets are not the three-act structure. They're a more granular partition that survives contact with films the three-act model fumbles.

Colophon

Essays are written in Markdown with embedded interactive figures, then rendered to HTML by a small Python pipeline. The supporting wiki — beat breakdowns, actor and crew pages, themed essays, scene-by-scene research — is published separately. Links between the two are deliberate; the wiki is where the receipts live, and this site is where the readings get written down.