Clichés, dead citations and unsourced figures as badges you can open

A findings list is a second document

The usual shape of a review is a list: twenty findings, each with a category and a suggestion, and a draft somewhere above it. To act on any one of them you read the finding, hold the wording in your head, scroll back, find the sentence, work out which of six rules applied, and then rewrite.

Replacing the flagged phrase with an interactive component removes the hunt. The reason, the suggested replacement and the original wording sit on the same line as the sentence that needs changing, so reading and acting happen in one place.

The density becomes legible too, which is something a list cannot show. Six badges in one paragraph is a paragraph that needs rebuilding. Six badges spread across two pages is a vocabulary problem. Identical counts, completely different work, and only the inline view distinguishes them.

Mechanical fixes and judgement, kept apart

Some flags have a safe replacement. A cliché usually wants plain wording, and deleting it costs nothing: the sentence was carrying filler rather than meaning. Those are the ones an apply button can handle, because nothing about the piece's subject changes when the phrase goes.

Citations and figures are deliberately left untouched. A fabricated-sounding citation and an unsourced percentage need a source, and no rule can supply one. Rewording around an unverified number makes the problem harder to catch next time, because the sentence now reads smoothly and the missing evidence is buried an edit deeper.

Keeping that boundary is the difference between a helpful assistant and a machine that launders a problem. The tool says the fix is a citation; the fixes pass skips it and says so.

The same principle applies to vague attribution. According to experts is a flag whose correct repair is a name, and a rewrite that turns it into research suggests is a worse sentence with the same empty space behind it.

• Safe to fix mechanically: stock phrases, filler, redundant qualifiers.
• Not safe to fix mechanically: unsourced figures, vague attribution, placeholder citations.
• Never automatic: anything that would change what the sentence claims.

The fixes view is a preview, not an edit

Toggling the fixes shows the draft with the mechanical changes applied, which is useful for one specific judgement: does this read better, or merely different? Restyling has a way of feeling like improvement when you watch it happen, and the only honest test is to read the result.

It is also a way to check how much work is left. If applying every mechanical fix leaves the piece still full of badges, the remaining problems are the ones that need a person, which is the correct ordering of the afternoon.

Nothing is written back to your text. Switch back to annotated and the draft is exactly as pasted, which means the fixes are a suggestion you can accept sentence by sentence in your own editor rather than a bulk operation you have to undo.

Citation shapes that read as sourced

There are four or five shapes that make a sentence look sourced without sourcing it: a bracketed number, an author-and-year pair, a phrase like a 2023 study found, a bare link, a footnote marker. Every one of them is a promise that something exists at the other end.

A model produces all five fluently, and it produces them for the same reason it produces confident prose: the shape is what it learned, not the underlying fact. The most convincing is the author-and-year, because it is specific enough to look checked and vague enough that nobody checks it in a hurry.

The badge cannot distinguish a real citation from a fabricated one; it only knows a source was named in a recognised shape. Treat every badge of this kind as a to-do rather than a finding, and resolve it by typing two words into a search engine: the author and the year. Either a paper exists or it does not, and that takes less time than arguing about it.

The same test catches the citation that exists but says something else. Finding the paper proves it exists; reading the abstract proves whether it supports the sentence. The second step is the one that separates a review from a lookup.

Why the fixes pass skips numbers

It is tempting to let a rewrite smooth over an unsourced figure. The sentence is awkward with the caveat in it, the number is probably right, and a tidy version reads better. That temptation is the reason this tool refuses.

Rewording an unverified number removes the thing that would have prompted someone to check it. In a month, the sentence reads like every other sentence, and the only remaining evidence that it was never sourced is a mark that has been overwritten.

The correct repair for an unsourced figure is a source, a deletion, or a decision to label it as an estimate. Three options, all of them actions, none of them a rewording. If none of those is possible today, the honest state is that the claim is unresolved, and leaving a badge on it is the accurate representation of that.

This is also why the fixes view is a preview rather than an edit. Watching the mechanical fixes apply shows you how much of the draft is a wording problem, and how much is an evidence problem, and the second list is the one that matters.

Badge density as a decision tool

Before reading a single badge, count them per paragraph. A paragraph with four or more is usually a paragraph that was generated first and edited second, and the right response is to rewrite it rather than to open four explanations.

Density also tells you where to stop. A draft with two badges in the first section and nineteen in the second is not uniformly bad, it is a piece where one section was produced differently, and that is a much easier thing to fix.

The badges themselves are most useful on prose that is otherwise fine. Once a paragraph needs rebuilding, the specific reason each phrase was flagged stops mattering, because the phrase is going anyway.

Which suggests a simple rule of thumb: badges for editing, density for deciding whether to edit at all.

What a badge cannot see about a sentence

A badge knows a phrase. It does not know whether the sentence around it is doing any work. In a piece that has lost its argument, the badges are a distraction: rewriting six clichés produces a cleaner version of a paragraph that should not exist.

Read for the argument first and mark second. Two minutes of reading tells you whether the paragraph makes a claim; the badges then tell you how the claim is dressed. Doing it in the other order produces a tidy draft with nothing in it.

The same applies to a flagged number. The badge tells you a figure has no source, and the sentence around it tells you whether the figure matters. A number in an aside can be cut; a number in a headline has to be found.

This is why the badges are worth opening rather than merely counting. The explanation is short and it is usually about the sentence rather than about the phrase, which is the part you would have missed.

What a good replacement looks like

The replacement for a cliché is never another phrase of the same length. In most cases the cliché is occupying the space where a fact should be: the sentence that says our solution is robust could say what it survives.

So read the badge as a question rather than a swap. What would need to be true for this sentence to be worth writing? The answer is usually a specification, a number or an example, and none of those can be generated from the phrase itself.

What it cannot know

It cannot tell a real citation from a fake one. It only knows whether a source was named at all, in a shape that resembles attribution. A reference to a real study and a reference to a study the model invented look identical here, and that is not a limitation to design around: verifying a source requires opening it.

It cannot judge whether a claim is true, whether a cliché was used ironically, or whether a piece of industry jargon is the right register for its audience. All three are contextual judgements and every rule is context-free by construction.

What it can do is make the shape of the draft visible quickly enough that you can decide where to spend real attention. For the truth of a claim, compare the passage against its source or take the section into the claim tools. For anything with consequences, a named reviewer is the last step and not the first one to skip.