Know the instrument.
Know its limits.

What a Gadjit score is

Gadjit splits a manuscript into passages of about 1,000 words and asks, for each passage, how closely its writing habits match machine-generated fiction: sentence-length rhythm, vocabulary richness, punctuation and contraction habits, dialogue share, function-word profile, repetition of phrases, and a lexicon of words and tics that language models over-use. Each passage gets a machine-signal score from 0 to 100. The manuscript report summarises those scores: a mean with a 95 percent interval, and how many passages reach the flag threshold.

The score is calibrated on passages scored by a model that never saw their author, with human and machine passages weighted equally, so it says how machine-like a passage reads assuming even odds. It is not the chance that a manuscript was machine-written; that depends on how much machine writing is in the pile being screened.

The threshold is not a round number. It is set on human-written calibration passages, each scored by a model that was not fitted on it, so that at most 1 percent of the contemporary ones reach it. Public-domain prose scores lower, so across all the calibration passages pooled, it flags a smaller share. The model card below gives the threshold for each era, says which one is used, and gives the share of calibration passages it flags. Every rate on this page is then measured on a test fold of books used for neither fitting nor calibration. Each rate is shown with the number of passages behind it, and with the number of books wherever the model card records one.

Gadjit refuses four kinds of text rather than guess: texts under 1,500 words; text that does not read as English, since it is trained on English fiction only; text with almost no sentence boundaries, meaning end punctuation and capital letters, which it needs to measure sentence rhythm; and text from which no passage of at least 700 words can be formed, such as a text made only of very short fragments. Inside a manuscript it scores, any passage with almost no sentence boundaries, and any passage that does not read as English, is left unscored; the report warns about it and counts those words as not scored. It never reports a single headline percentage as if it were proof. Verse and scripts are scored with a warning when the layout shows them, because they are the cases where every stylometric detector, Gadjit included, is weakest.

In the report, each passage also shows what moved its score. The model adds up its measures with fixed weights, so every group of measures, such as sentence rhythm or punctuation, pushes the score a known amount toward human or toward machine. The page lists the strongest pushes and marks the machine-favoured words it counted. These explain the score; they do not change it.

Current model card

External test: contemporary fiction the model never saw

Licensed contemporary novels and collections (Cory Doctorow, Peter Watts, Charles Stross, Rudy Rucker, Brandon Sanderson, Kelly Link and others), used under non-commercial licences for measurement only and never for training. This is the number a living novelist should care about.

Source matching

Beside the machine-signal score, a report can carry a second, separate scale: text the manuscript shares with other works. The two are never combined. A passage can be human-written and copied, or machine-written and original.

Every book in the index is cut into overlapping runs of five words. A fingerprint of a fixed share of those runs is kept, chosen by winnowing, the method behind the MOSS plagiarism detector (Schleimer, Wilkerson and Aiken, 2003). Every five-word run of the manuscript is then looked up. Each hit is confirmed word by word against the book itself, and exact runs close together are joined across small edits, so a lightly edited copy is found as a near-verbatim match. Any stretch copied word for word for 24 words or more is always found; shorter copies down to 15 words, and edited copies, usually are. Phrases found in five or more different books are treated as stock phrases and ignored.

Each match gives the source's title, author, chapter where the book marks one, a link to the source, and the source's own wording. A match is marked as having its source named nearby when the author's full name or the title appears close to it, as in an epigraph or a credited quotation.

What is searched is stated in every report. It is public-domain English fiction from Project Gutenberg: the most-read works on its list, and the books in Gadjit's own training corpus. The open web, in-copyright books and translations are not searched yet. Source matching finds reused text, not reused ideas, plots or characters, which no text comparison can trace reliably. Reusing public-domain text is lawful; the scale reports reuse and leaves the judgment to the reader. Reports include this section only where the source index is available.

Published limits

• Era gap. The human-written seed is mostly public-domain fiction, which means 19th- and early-20th-century prose, plus a small set of openly licensed contemporary works. The machine side is current. Some of the separation the model learned is era, not authorship: contemporary human prose, with its contractions and present-day vocabulary, scores higher than Austen does, even though some modern habits, such as frequent one-sentence paragraphs, push the score down. The external test above measures the rate a contemporary writer would see; until licensed modern backlist from publisher partners is in the seed, treat the test-fold rates as a best case.
• Style, not provenance. A stylometric score cannot see drafts, keystrokes or a writer's process. Heavy human editing of machine text lowers the score; a human imitating machine habits raises it.
• Generators change. Recall is reported per generator model and per prompting mode, each with its sample size. The style-instructed mode only tells the generator to avoid common machine tells; no paraphrasing or humanizer tool has been tested, and human editing of machine text is not covered. New models need new mirrors.

Data policy

No text from pirated corpora or shadow libraries, ever. The current seed is public-domain novels and stories, contemporary works released under the CC0, CC BY and CC BY-SA licences, and synthetic machine mirrors generated from those human passages. The model card above gives the counts.

Two further sources are planned and are not in the seed yet: backlist licensed by partner publishers under contract, and manuscripts that authors choose to contribute.

Manuscripts submitted for scoring are processed in memory and not stored. The sample on the home page is the opening of The Turn of the Screw by Henry James, a public-domain book the model never trained on.

Due process

A Gadjit report is one input to a conversation with an author, never grounds on its own for an editorial or contractual decision. The highest-scoring human-written passages in the test fold, flagged or not, are published on this page with every model release so anyone can reproduce them.

A signal for a conversation, never the final word.

Begin a reading ↗