DeepSeek paraphraser and patchwriting

The DeepSeek paraphraser is one call to a rewriting model, carrying the passage you paste plus up to 300 characters of instruction, returning a single completion. The system prompt requires it to preserve the original meaning and factual claims and to return only the rewritten text, with no preamble, commentary, quotation marks or markdown fence.

It is the same engine that serves the paragraph rewriter on this site, and the difference between the two routes is the material people bring to them. A rewriter is usually pointed at your own prose, where the only risk is meaning drift. A paraphraser is usually pointed at somebody else's, where there is a second risk that no amount of engine quality addresses.

The tool performs step two. Steps one and three are the ones that determine whether the output is defensible.

Step 1

Reconstruct the claim without the text in front of you. Working from the visible paragraph is what produces structural copying, independent of which tool runs next.

Step 2

Paste and constrain, within 1 to 12,000 characters and 300 characters of instruction. Name the terms that must survive and the claims that must not be strengthened.

Step 3

Verify against the source on facts, modality and evidential strength, then attribute. Neither verification nor attribution is performed by the engine.

The definition in circulation is precise enough to work with: patchwriting is taking a source passage, swapping out 30 to 40 percent of the words, restructuring two or three sentences, and submitting the result as your own paraphrase. The page that states it also says the practice violates the academic honesty policies of essentially every US, UK, Canadian and Australian university, with the citation present, and gives the reason: the underlying sentence structure and argumentative skeleton is still the source author's.

That reason is what makes word overlap the wrong metric. A paraphrasing tool can drive lexical similarity arbitrarily low while leaving the skeleton untouched, because reordering the argument is not what a rewriter is instructed to do. So a passage can come back with almost no shared vocabulary and still be a patchwrite, and the test is not how different it looks. The test is whether the sequence of moves is yours.

The two lists pull in opposite directions, which is why paraphrasing is harder than it looks and why a single instruction rarely covers it.

Must change: sentence architecture

Clause order, subordination, where the emphasis falls. This is the part that makes the construction yours.

Must change: the order of the argument

The part a tool will not do unless you ask, and the part the patchwriting definition actually turns on.

Must preserve: technical vocabulary

A defined term replaced by a synonym is an error, not a paraphrase. Name these in the instruction field.

Must preserve: modality and evidence

May, suggests, is associated with. Strengthening any of these misreports the source, which is a worse fault than copying it.

Must preserve: quantities and scope

Figures, dates, sample sizes and quantifiers. Verify these individually rather than reading for a general impression.

A competitor on this term advertises detection of four citation marker systems and reattachment of each one to the nearest semantic clause. Nothing of the kind exists here. This codebase contains no citation parser, no reference grammar and no post-processing step that inspects the output, so the number of citation systems understood is zero.

The practical alternatives are both manual. Ask in the instruction field that bracketed references be left exactly where they are, then verify each one by eye. Or strip the references, paraphrase the clean prose, and reinsert them yourself, which is slower to describe and faster to check, because you are auditing your own placement rather than a model's.

Ours: 5 runs per rolling 24 hours, each returning individually. 1 to 12,000 characters per run after trimming. 300 characters of instruction. One completion, with the response length capped, so an oversized input can return truncated output. No account, no card, no API.

From the competitors' own pages, for calibration. One tool's free tier is four runs in total, not per day, at 250 words per run, with a plan at 18 dollars a month offering 80 runs a month at 1,000 words per run. Another is a Word add-in with a 15-day trial and a full version at roughly 30 pounds, whose free version limits word count without stating the number. On input size this route is the most generous of the three, and on named features it is the least equipped, which is the trade being made.

Whether paraphrased text is detectable is the most searched question attached to this term, and one of the pages read for this route lists it as a heading outright. Our answer does not vary: we do not know, we are affiliated with no detection or similarity service, and we will not predict what any of them reports. There is no similarity index in this codebase, so we have no basis for an opinion even if we wanted one.

The more useful reframing is that detection is not the constraint that matters. An institution's policy is violated by presenting someone else's argument as your own, and that is true whether or not any software notices. A paraphrase that would survive a checker and not survive an examiner reading it next to the source has not solved the problem it was used for.

When the wording is the evidence, in which case quote it. When only the conclusion matters, in which case summarise and save the space. When you cannot restate the claim without the source in front of you, in which case the problem is comprehension and a tool will disguise it rather than solve it. And when the material is confidential, because this route uploads what you paste.

It is the right operation when you have understood the source, you need its detail rather than its voice, and the obstacle is producing a construction of your own at the length the argument requires.