Methodology.

The research corpus is a library of professionally produced, publicly released screenplays: films and television episodes whose scripts studios and publishers have made available. Each new release requires recorded source provenance and a metadata review covering year, format, writers and primary genre. Reviews are tied to the held file bytes and the exact measurement result; changing either requires a new review. Writing credits describe screenplay and revision writers on the held draft, excluding literary and story-only authors. When a cover omits an author, the review explicitly identifies the external credit source used. Earlier releases predate that review ledger; missing historical provenance and incomplete credits are disclosed in their packages. Each script has one primary genre from a fixed list of ten.

We publish statistics derived from these scripts. We do not republish the scripts themselves, and no dataset we release contains screenplay text.

Screenplay structural statistics are counted deterministically from the screenplay text by our measurement engine: page counts, scene headings, scene lengths in eighths of a page, locations, speaking characters, dialogue and action volume. Run the engine twice on the same script and it returns the same numbers. PDF page counts include title pages, blank pages and other front matter in the held file; they are not counts of numbered story pages. Files with unseparated prose or artwork appendices are excluded from the reviewed benchmark. Repeatability does not establish that the extraction is correct. No LLM judgement scores are included. Runtime is a fitted estimate, explained separately below, not an observed film running time.

The engine is versioned (current studies use v1.5), and each measurement records how confidently the script’s layout was read. The cohort excludes low-confidence parses. Family-specific guards remove flagged measurements from affected statistics; accepted measurements can still contain extraction errors.

Each study defines its cohort before any number is computed. The current flagship cohort is: Produced feature screenplays released 2000-2026, measured from the file we hold, excluding parses the engine rated low confidence.

Features and television scripts are never pooled into one figure. No genre cut is published below n=20: eligible samples under the floor are counted and shown, but their statistics are held back. This is a publication rule, not proof that a sample above the floor represents the wider industry.

Of 557 scripts in the corpus, 428 are in the current flagship cohort and 129 are excluded, each with a stated reason. The most common:

Screenplay studies use frozen data; the cost-of-entry study explains its live directory calculation separately. Each screenplay study renders from a frozen snapshot that records its cohort rules, every inclusion and exclusion, the engine version and a checksum of the data. When the corpus grows or the engine changes, a study updates by moving to a new dated snapshot, and its update history says so. Publication requires an explicit approved version and checksums are verified on load and build.

Browse the dated release history and public manifests. The current flagship snapshot is anatomy-of-a-produced-screenplay-2026-09-v5, generated 20 September 2026.

Screenplay PDFs are not uniform, and extraction can fail for one part of a script while succeeding for the rest. A script whose scene headings are set in a way our parser does not recognise may still yield usable dialogue counts; a scene flag alone does not establish whether another measurement is accurate.

We apply plausibility thresholds to each family of measurements. A feature averaging more than five pages per scene, or containing fewer than twenty speech blocks, may indicate missed headings or dialogue. These are heuristics for likely extraction problems, not mathematical impossibilities. A failed family is excluded from measures that depend on it; other eligible measures remain. Each chart reports its own resulting sample size.

The 20 September source review withdrew the earlier dialogue exemption for No One Will Save You: its apparent speakers were action directions. Shared cue checks now flag prose-like leading cues, unresolved name variants and substantial unattributed speech. Source-reviewed failures, including lyrics mistaken for speakers, receive additional family exclusions. These conservative checks can also withhold valid material; they do not estimate an error rate. The package includes the thresholds, flags and review reasons. Of the 428 scripts in the current cohort, 200 carry at least one flag, and every one of them is named with its reason in the snapshot manifest, alongside the dataset.