How the full analysis works

This is how UnTXT checks your writing for AI patterns. Two frontier models argue over it passage by passage, and a third settles anything they cannot agree on between them.

It is not a separate tool you go and find. In the web app it is what happens when you press Analyse.

Why an argument rather than a score

A single reader, machine or otherwise, gives you one opinion carrying that reader's own blind spots, and you have no way of knowing which of the two you are looking at.

So your passage goes to Claude Opus 5.5 and GPT-6 Sol separately, each of them reading it cold, and that part matters rather more than it sounds: if the two reviewers started from a shared position then their agreeing later would tell you nothing at all. Each one makes the case that the passage reads as a person's, then takes apart what the other said, and then gives a verdict without being shown where the argument had got to. What they are arguing over is the evidence on the page: rhythm, word choice, how specific the detail is. Neither of them knows who wrote it, and neither do we.

When they land in the same place, you have two differently-built systems trained on different material reaching the same conclusion about how the same paragraph reads, which is a much harder thing to arrive at by accident than one system's opinion.

When they split, which happens more than you would expect, Claude Sonnet 5 reads the disputed passage on its own without being told which way either of the first two leaned, and breaks the tie.

What comes back

Passages, marked one of two ways.

Gold is a passage the reviewers agreed reads like one particular person: natural variation, a specific detail, a turn that a model would have smoothed out. Most AI detectors only ever look for risk, which means the best available outcome is that nothing was found. Marking what sounds like you gives you something to keep and something to write towards, and it is the part people tell us they did not expect.

Grey is a passage they agreed follows the patterns machine writing leans on. That describes the passage and not where it came from. Generic phrasing, structure that repeats, polish where polish does not belong, transitions you can see coming. Click it and you get what each reviewer said about that specific passage, including where they disagreed before settling.

Grey does not mean cheating, does not mean an accusation, and does not mean anybody has decided anything about you. It means two careful readers would question that paragraph, and it tells you which paragraph.

What it costs

A run on a piece of around 1,100 words comes to roughly 55 CVT when it runs straight away. There is no cheaper heuristic pass to choose instead, because a second, weaker answer sitting next to the real one turned out to be worth less than nothing, although the same analysis can run overnight for less, which is covered just below.

That gap is real work rather than a markup — three frontier models reading the same document, twice over, and arguing about it. The exact estimate is shown before you confirm, and it comes out of whatever balance you have, whether that is a monthly allowance or free credit.

Every new account gets one complimentary run that costs nothing against the balance, which exists so you can see what the thing actually does before deciding whether it is worth paying for.

Running it overnight

Between five in the evening and midnight your time, a second button, a crescent moon, appears beside Analyse. It queues the same analysis to run while you are asleep, with the same two reviewers and the same tiebreaker doing exactly the same job, and it costs about 40% fewer credits because it waits in the providers' overnight queues instead of jumping them. Nothing comes back in less than three hours, and everything is back within eight, so a document queued at ten is waiting for you by breakfast and we email you when it lands. Three can wait at a time on your own account, and forty on a school or company one, since that is roughly what a class set on deadline night looks like. So queue the thing, shut the laptop, and go and have your dinner.

When it is worth it

For anything with stakes on it. A dissertation, a piece of coursework that carries a module, an application, a document going to a client. Those are the moments where knowing which two paragraphs to reread is worth the credit.

For an early draft you are still restructuring, it is usually not. The passages will change before you submit anyway, and it is worth saving the run for the version you actually intend to hand in.

Very short pieces are turned away rather than guessed at. Below roughly 250 words there is not enough on the page for two reviewers to genuinely disagree about, and we would rather say no than hand you a confident reading of four sentences.

What it is not

It is not proof. When the reviewers agree, they agree about how a passage reads, never about who wrote it. The reading is probabilistic and stays probabilistic no matter how many models you point at it, and human writing genuinely does read as machine writing sometimes. Sataklela OÜ accepts no liability for decisions made on the basis of these results.

It does not publish an accuracy figure, and neither will we. Every accuracy number in this market comes from a benchmark chosen by whoever is quoting it, ours included, and we would rather hand you the reasoning behind each mark and let you judge it than hand you a number to trust.

It will not rewrite your work. There is no button that takes your draft and gives you back a cleaner one. Agency is yours to keep. Autonomy is ours to offer.

Talking about it afterwards

If a tutor asks how you checked your work, describe the process rather than quoting a result.

"I ran it through an AI pattern checker, where two independent models read each passage and argue it out, with a third one deciding when they disagree. Three passages came back marked. I rewrote two in my own words and left the third, because it says what I meant."

That describes something you did. It also happens to be the sort of engagement most tutors are pleased to hear about.

Questions we get asked

Does using it make my essay look suspicious?

No. It runs before you submit, on your own account, and nobody sees it but you.

It said my writing is fine and my tutor flagged it anyway. Now what?

Your tutor has context no model has — your previous work, what you said in a seminar, how the argument developed. That is not a contradiction so much as two different readings, and theirs includes things ours cannot see.

Why not four models, or ten?

Diminishing returns against real cost. Two frontier models built differently and trained differently already covers the ground, and the third earns its place only when the first two genuinely split.

Why show me the disagreements?

Because a mark that came from a real argument is worth more to you than a mark that came from a consensus you have to take on trust, and because seeing where two capable readers split is often the most informative thing on the page.

Try it — the first 1,000 words are free, and the first full analysis run is on us

Results are informational. Sataklela OÜ makes no representations about legal admissibility and accepts no liability for decisions made on the basis of them.