Llama assignment checker, examined

One request carrying route, family, a text input bounded at 1 to 12,000 characters, and an optional instruction. On the server a language model is told to analyse only the supplied text for the qualities the tool name implies, to give concise findings, and, in the prompt's own words, never to claim certainty about authorship, plagiarism or which model wrote text, describing observable signals and limitations instead.

There is no index, no corpus and no lookup of any kind. Nothing is compared against published work, prior submissions or a web crawl, because no such component exists in this codebase. A similarity score is not something this route produces badly; it is something it cannot produce at all.

Remote by resolution rather than by name: the route's registry family reads as an analyzer, and the surface resolver sends every non-watermark analyzer route to the server classifier. So text is transmitted, the allowance is 5 free runs per rolling 24 hours, and no local claim is available.

The search term resolves to two populations that do not overlap, and a page that does not say which it belongs to is misleading by omission.

One population is student-facing: plagiarism and grammar checkers, advertising database coverage and unlimited length. One page in that group claims plagiarism detection, grammar, content evaluation and citation checking together, scanning against a database it does not describe. The other is instructor-facing guidance about deciding whether a student used AI, and the best of it is notably cautious.

This route is neither. It reads text and describes what is observable in it. Stating that first is not modesty, it is the only way a reader can tell whether this page is useful to them before spending a run.

Ordered so the manual work happens where it is cheapest.

Step 1

Verify sources by hand first. Whether a citation exists and supports its claim is outside what any text-only analysis can determine, so it never belongs in the automated step.

Step 2

Paste one structurally complete section, 1 to 12,000 characters, with a narrow question in the 300-character instruction field. Unspecified scope produces generic findings.

Step 3

Treat each finding as a hypothesis and test it against the draft. A finding you cannot confirm by looking is an observation about text and carries no further weight.

The distinction that governs this whole route. Four things are readable from a document alone, and the thing everybody wants to know is not among them.

Observable: uniformity

Paragraph lengths, sentence openings, section shapes. Structural regularity is the most reliably visible pattern and the one least affected by editing.

Observable: modality

Whether claims are hedged or asserted flat. Academic prose qualifies constantly, and a document that never does reads distinctively.

Observable: analytical depth

Whether paragraphs advance a claim or restate a premise more elegantly. Tidy structure over surface insight is visible in the text itself.

Not observable: origin

Who or what produced the document. No arrangement of words establishes this, which is why the prompt forbids asserting it and why no number is returned.

Competing tools in the adjacent detector category return a figure, and the figure is the product. It travels: into a screenshot, into an email, into a meeting where nobody present can interrogate how it was computed.

This route returns prose. That is a worse artefact to forward and a better one to use, because every statement in it names something in the text that can be checked. A finding that the third paragraph asserts without support points at the third paragraph. A score of 82 percent points at nothing.

The absence is enforced rather than stylistic. The system prompt requires observable signals and limitations, and forbids certainty about authorship or plagiarism, so there is no configuration in which this route emits a confidence figure.

The most rigorous instructor-facing page in this result set states the position plainly: no detector result should be treated as proof on its own, and a detector should support an investigation rather than serve as the sole basis for an accusation or penalty. It also places running a tool fifth in a seven-step process that begins with confirming what was actually permitted and ends with deciding on multiple pieces of evidence.

That ordering is the correct one and it deliberately reduces the weight on outputs like ours. Verifying facts and sources comes before the tool. Talking to the writer comes after it.

Process evidence dominates textual evidence, and it is not close. Drafts, notes, a reading list and document version history describe how a piece was produced. Text analysis describes only how the finished document reads, and reading evenly is not an offence.

Any system reporting signals will report them for text a person wrote, and the burden is not evenly distributed. Both instructor-facing pages fetched for this brief raise it, and one names the group directly: non-native English speakers, and writers with an unusual or highly consistent style.

The mechanism is not mysterious. A smaller working vocabulary and safer syntactic choices produce regularity. So does being taught a paragraph template and applying it well. Both generate exactly the uniformity that reads as machine-produced, and neither is evidence of anything at all.

This is the strongest practical argument for descriptions over scores. A description of regularity invites the reader to ask whether regularity means anything here. A percentage invites them to stop asking.

No plagiarism database, no source matching, no citation verification, no web lookup. No similarity score, AI probability or confidence figure of any kind. No grammar correction and no rewriting: this route classifies, and transformation happens elsewhere.

No file upload, so documents are pasted. No batch queue, no API, no LMS integration, no history and no account. Five runs per rolling 24 hours against a hashed identifier, and 12,000 characters per run enforced before transmission.

And no model claim. The configured default is a general model set by an environment variable. The route does not run Llama, holds no affiliation with Meta, and predicts no result from Turnitin, GPTZero, Copyleaks or anyone else.