Llama thesis checker

One pasted section, read for two things at once. At the level of argument: whether a central claim is present and specific, whether the supporting paragraphs bear on it, whether the closing lands where the opening pointed. At the level of surface: sentence length distribution, recurring constructions, hedging density, the register the vocabulary sits in.

The result is written at a low temperature with a ceiling of 1,400 completion tokens, so it arrives as a short report rather than an annotated document. Every finding in it should be locatable in the passage you submitted. Anything you cannot find there should be discarded, and that test is the correct way to use a reading of this kind.

The term is overloaded, and the confusion costs people time. One kind grades a thesis statement against a rubric and returns criteria scores. One kind is a manuscript service: grammar, academic tone, style guide compliance down to the edition, similarity checking, and a proofread file returned with track changes. One kind is an authorship detector that takes a PDF, reports an overall probability and highlights individual sentences by confidence.

This route is a fourth thing, and the smallest of the four: a reading. It describes a section and produces nothing else. Knowing that before you paste saves you from expecting a corrected file or a verdict that is not coming.

Rubric graders

Return criteria scores and an overall quality rating. There is no rubric on this surface.

Manuscript services

Return an edited file, a revision table and style guide compliance. Nothing is edited here.

Authorship detectors

Return a probability and per sentence confidence. No probability is produced here.

This route

Returns written findings on argument and prose for one section of up to 12,000 characters.

Served remotely. The section is posted to our AI server, read there, and the report comes back in the same response.

Step 1

Paste a section of 1 to 12,000 characters. The bound is enforced before the model is called, so an oversized paste costs no allowance.

Step 2

Constrain the reading with up to 300 characters of instruction, for example to argument structure only rather than prose.

Step 3

Verify each finding against the passage before acting on it, then re-run after the edit if the change was substantial.

Detectors in this category publish an overall probability plus a sentence by sentence confidence breakdown. The better ones also publish their error rates, and those numbers are the argument against leaning on the score. One detector reports 96.2% F1 on research paper text with a 7.2% false-positive rate on research papers, and openly states that performance varies by model, citing 99.4% F1 in one case against 77.1% in another. Those are that vendor's figures for that vendor's system, not ours.

Read the 7.2% as what it is: on a hundred honestly written papers, several are flagged. There is no appeal from a number, which is why a system prompt that forbids authorship claims is the right design here. This route has no measured accuracy, publishes none, and describes the writing instead.

Academic sections are conventionalised by design, and the conventions overlap precisely with the features people associate with machine prose. A literature review cycles through sources in a fixed rhythm. A methods chapter enumerates materials and steps with deliberate uniformity. Both produce clustered sentence lengths, heavy hedging and repeated constructions.

So the same finding means different things in different chapters. Uniformity in a methods section is compliance with the genre. Uniformity in a discussion chapter is a writer who has stopped making choices. A report that names the feature is useful; a score that averages both into one figure for the document is not.

Twelve thousand characters per run, roughly two thousand words. A 60,000 word thesis cannot be submitted, and there is no PDF or DOCX upload to submit it through. Five runs per rolling 24 hours, keyed to a hashed identifier. One section at a time: no batch, no export, no API, no integration with a word processor or a learning management system.

No similarity search, no citation parsing, no style guide, no rubric, no grade. The text is transmitted to our AI server, unlike the watermark and hidden character routes on this site, which do their work locally. We claim no compliance certifications, because nothing in this codebase would support the claim.

A score is a single ordering over a passage, and once it exists the natural response is to move it. Editing to move a number produces prose optimised for the number, which in academic writing usually means shorter sentences and fewer qualifications in places where the qualifications were doing real work.

A finding points at something. This paragraph opens the same way as the two before it. The claim in the first sentence is not the claim the section defends. Both are checkable against the page in front of you, and both survive being wrong, because you can look and disagree. That is the whole reason this surface returns prose.

Candidates in the final editing weeks, checking that an introduction rewritten several times still matches the conclusion it is supposed to set up. Supervisors triaging a chapter before committing comments. Writers who assembled a section from model output and want the residue described rather than adjudicated.

The common requirement is a second reading without a judgement attached. Institutions decide what counts as acceptable assistance, detectors decide what they will flag, and neither decision is made here. What is available here is a description of the section, delivered in a few seconds, with the document left untouched.