How the diagnostics work, and where they fall short
Every instrument here is built on a named, published theory of how stories work on the mind. The theory shapes what an instrument looks for; it doesn't prove what the instrument concludes about your script. This page sets out what is measured, what is inferred, what is experimental, and what we have not proven yet.
What actually happens to your pages
Your screenplay is split into scenes, then analysed by a large language model following each instrument's protocol: what it reads for, the evidence it must cite, the theory it applies. The model reads your script the way a trained reader briefed on the theory would. Findings go into the report with the scenes they rest on, and the lines it flags are quoted.
Technical detail: analysis runs on a large language model through OpenRouter, on a provider endpoint that does not train on your script or retain it. Instrument protocols are versioned prompts; the model and instrument versions used for your report are the ones live at generation time.
Measured, inferred, experimental
Some readings in a report deserve more weight than others. Every instrument has one of three classifications, shown on its page, here, and above its section of a report.
Structural: counted from the text
Counted directly from the text: scene counts, who appears with whom, dialogue distribution. If you counted by hand, you would get substantially the same numbers.
Interpretive: read by a model
Read by the analysis model the way a careful human reader would read: subtext, sympathy, emotional rhythm. Grounded in named theory, but the band is a judgement, not a measurement.
Experimental: a research-stage estimate
Research-stage estimates: how closely audiences would follow a script together, which audiences a story shape suits, and identity read from the text. The research behind them is real. Nobody has yet checked what these estimates say about a script against what real audiences do, so treat them as hypotheses.
Some structural facts (like character co-presence) also appear inside interpretive and experimental instruments. The classification describes the instrument's headline readings, the ones you might be tempted to treat as measurements.
How to read a band
Every model estimate in a report is given as a band: strong, moderate or weak. A fit, to a genre, an audience or a story shape, reads strong fit, partial fit or weak fit. The model works on a scale of 0 to 100 and rounds to fives, so a figure like 73 or 95 would claim a precision it does not have. Counts stay as numbers: how many scenes, how many characters share a scene. In practice:
- Results vary between runs. Re-analysing the same script can move a reading that sits near a boundary into the next band. Trust a pattern (“these scenes sit well below their neighbours”) over any single reading.
- Compare inside one script. The model reads each screenplay on its own terms, so a “strong” in your draft tells you nothing about a “strong” in someone else's.
- Check the evidence. Findings cite the scenes behind them, and the lines a report flags are quoted. If the evidence doesn't convince you, the band shouldn't either.
- Well-known screenplays are a special case. The model may recognise a famous script from its training data, which can contaminate analysis of published work. Your unpublished draft does not have this problem, because the model has never seen it.
Where the instruments are weakest
- Non-standard formatting can mis-segment scenes, which skews every reading made scene by scene.
- Deliberate technique can read as defect: intentional tonal collision, static characters, or diffuse focus may be flagged. The flag means “confirm this is intended”, not “this is wrong”.
- Identity attributes read from text are inferences and are marked stated-vs-inferred; they can be wrong, which is why Representation & Power is classified experimental.
- Non-English and heavily idiomatic scripts get less reliable dialogue and subtext analysis.
- Audience-segment and synchrony estimates have no ground-truth validation yet. They are hypotheses drawn from arc shape and text.
What we have not validated yet
We have not yet published test-retest consistency figures, agreement rates with experienced human readers, or benchmark results on a fixed script set. Until we do, treat every report as a well-read second opinion that shows its evidence, and weigh it the way you would weigh notes from a reader you trust.
Our commitment: as validation work lands, results will be published on this page, including the unflattering ones, and instruments will carry visible version numbers so results can be compared across revisions.