Llama style analyzer

The analyzer reads a passage and reports on its surface mechanics: which register the vocabulary sits in, how tightly sentence lengths cluster, which syntactic frames recur, how punctuation is distributed, and whether any of that changes across the passage. The report is written prose, generated at a low temperature with a ceiling of 1,400 completion tokens.

The word describes is doing deliberate work in that sentence. Nothing here is computed in the sense that a readability formula is computed. A formula takes syllables and sentence counts and returns a number by arithmetic. This route returns a reading, and the difference matters when you decide how much weight to put on it.

Quantitative stylometry works on four families of features: lexical, grammatical, syntactic and punctuation patterns. They are chosen because they are largely independent of subject matter, so a writer's habits show through whether the topic is a quarterly report or a short story.

Those families are the right way to think about what a style reading can and cannot see. Habits of construction are visible in them. Intent, accuracy and originality are not, which is why a style analysis is never evidence about where a text came from, only a description of how it is built.

Lexical

Word choice, repetition, the size of the vocabulary actually in use rather than available.

Grammatical

Voice, tense, modality, how often constructions are standardised into the same safe form.

Syntactic

Clause structure, sentence openings, the distribution of sentence lengths across a passage.

Punctuation

Rhythm markers. Which separators do the work, and whether one of them dominates.

This route is served remotely. The passage is posted to our AI server, read, and the report returned in the same response.

Step 1

Paste a passage between 1 and 12,000 characters. The bound is enforced before the model is called, so an oversized paste costs nothing.

Step 2

Add up to 300 characters of instruction to narrow the reading, for example to one feature family rather than all four.

Step 3

Read the findings against the passage itself. Anything the report names should be locatable in the text, and if it is not, discard it.

A study on short samples tested human text against GPT-3.5, GPT-4, Llama 2, Llama 3, Orca and Falcon, using a human designed feature set alongside an n-gram pipeline over lexical, grammatical, syntactic and punctuation patterns. On ten sentence samples it reported up to .87 Matthews correlation in a seven class setting and binary accuracy between .79 and 1.0, with one Wikipedia against GPT-4 pairing reaching .98.

Two things follow, and only two. First, the sample size in that work is a useful floor for ours: ten sentences, not two. Second, those numbers belong to that pipeline and that corpus. They are not this tool's accuracy, this tool has no measured accuracy, and no figure of that kind is produced anywhere in this interface.

Five runs per rolling 24 hours, counted against a hashed identifier. Twelve thousand characters per run, roughly two thousand words, so a long document has to be sampled rather than submitted. One passage at a time: no batch, no file upload, no export and no API.

The text is transmitted. Several tools in this category advertise browser only processing and no uploads, and our watermark and hidden character routes genuinely do work that way. This one does not, because the reading requires a model, and a page that blurred those two cases would be lying about one of them.

A grade level, a lexical density figure or a tone percentage would all be trivial to display and none of them would be real. This surface does not calculate them. Competing pages print a readability score and a fixed tone scale running from formal through casual to aggressive, backed by a rubric they do not publish.

There is a second reason beyond honesty. A score compresses a passage into one number and then invites you to move the number, which is how prose gets optimised into something worse. A written finding points at a sentence. Pointing at a sentence is actionable in a way that a decimal is not.

No, and the constraint is written into the system prompt this route runs under: never claim certainty about authorship, plagiarism, or which AI model wrote text. Describe observable signals and limitations instead.

The research position is consistent with that. Discriminating features are learned on a specific corpus with a specific pipeline, and detection trained on one model's output degrades when it is pointed at another's. What a style reading supports is a statement about the text: this passage is uniform, its clauses cluster, its openers repeat. What it does not support is a statement about the author.

Editors running a consistency pass across contributions of mixed origin. Researchers and teachers who want the observable features of a submission named without a verdict attached to them. Writers diagnosing why a draft that is technically clean still reads as inert.

The last case is the most common and the least served elsewhere. Flatness is not an error, so no correctness tool reports it. It shows up as clustered sentence lengths, repeated openers and grammar standardised into the same safe construction, which is precisely what a style reading is able to name.