The inline AI highlighter: editing against marks instead of reports

Why a findings list is the wrong shape for the fix

A verification report describes a document. An editor instructs you. Those are different products, and the difference shows up in the five steps a writer takes after reading a report: read the finding, remember the wording it quoted, scroll back through the draft, find the sentence, decide what to write instead. Only the last two are thinking. The first three are transport, and transport is where a review quietly turns into a guess.

Inline marks remove the transport. The flag sits on the words that caused it, in the sentence that has to change, which means the edit is one gesture rather than a hunt. It sounds like a minor piece of interface design. In practice it is the difference between fixing fourteen problems and fixing the four you could still remember at the time.

It matters most for the defects that only exist at sentence scale. A hedge attached to a qualifier, an announcement connective that carries no content, an absolute claim nobody can support: none of those are findable with a search and replace, because they are shapes rather than tokens. A shape is easiest to judge while you are rewriting it, and the rewrite is exactly when an inline mark is in front of you.

There is a second, quieter advantage. A report is a separate document, and the moment you begin editing, every excerpt and line reference in it goes stale. Marks cannot drift from the text, because they are rendered from the text. The evidence and the thing being judged are the same object.

The four families of marks, and what each one asks

Every mark in the editor comes from one of four pattern families, and the colour is the instruction. Turquoise marks stock vocabulary: the twenty or so words a model reaches for when it has nothing specific to say. That is an invitation to replace the word with the thing itself. Not every instance needs replacing, but every instance is a place where a specific would have been better.

Amber marks hedges and stacked qualifiers: may, arguably, potentially, typically, it is worth noting. Here the instruction is not replacement but decision. Either you know what happened or you do not, and a sentence that hedges twice is usually a sentence where the writer has not decided yet.

Purple marks announcements: furthermore, moreover, additionally, ultimately. These describe the structure of an argument instead of making one. The fix is almost always deletion. Cut the opener, start on the subject, and the paragraph usually reads as though the connective was never needed, because it was not.

Red marks absolutes: always, never, guaranteed, proven, everyone. These are the marks with consequences, because they make promises a publisher has to keep. The instruction is to source the claim or soften it, and softening is a legitimate decision rather than a failure of nerve. The alternative — a marketing sentence that becomes a contractual one — is how complaints start.

• Turquoise: stock vocabulary. Replace with the specific thing, or delete the sentence if there is nothing specific to say.
• Amber: hedges. Decide. Two qualifiers in one sentence usually cancel each other.
• Purple: announcements. Delete the connective and open on the subject.
• Red: absolutes. Attach a source, or make a smaller claim you can defend.

Density is the number to watch, not the count

Twelve flags in a three-hundred-word product description and twelve in a two-thousand-word report are not the same document. The count rewards length; density does not. Flags per thousand words is the figure that lets you compare a draft against its own previous version, and compare two drafts of different lengths on one scale.

The bands are worth knowing roughly. Under three flags per thousand reads as ordinary professional prose: plenty of careful writers use these words, and a technical document may legitimately repeat terms. Between three and eight, readers begin to feel a pattern even when they cannot name it, which is the moment a piece starts sounding like a template. Above eight, word-level repair stops working, because the vocabulary is a symptom of paragraphs that had nothing particular to assert.

Watch the figure across an edit rather than once. If the density does not fall while the prose improves, you are substituting synonyms rather than finding specifics, and the second reader will feel the same absence under new words. The number is a diagnostic for whether your edit was real.

None of this makes density a grade. A legal paragraph may need three hedges to be accurate, and a sentence about a guarantee may need the word. The figure tells you where to look; it does not tell you what to do when you get there.

A five-minute pass that holds up

Work on a section rather than the whole document. The marks are most useful at a scale you can hold in your head, and a section is also the unit you will actually finish in one sitting. Paste it, read it once without touching anything, then start from the risky end.

Take the red marks first. Every absolute is a promise, and promises are the part of a draft that other people can hold you to. Then the amber ones, because stacked hedges are where meaning quietly disappears. Then the purple connectives, which are deletions and therefore fast. Vocabulary comes last, and only where you can name the specific thing the sentence was trying to say.

That ordering is deliberate. If you start with vocabulary you will spend the whole pass making the prose sound different without making it more true, and you will run out of attention before you reach the claims.

After the edits, run the pass again. Density should be lower, and the marks that remain should be ones you consciously decided to keep. A mark you decided to keep is not a problem; it is a decision you can explain to an editor.

Working in registers the marks were not built for

Legal, medical and regulatory prose uses these words on purpose. A contract says may and shall because those words carry different obligations, and a clinical summary hedges because the evidence is genuinely limited. Run a marketing-tuned tell list over that material and it will flag accuracy as a defect.

The fix is not to ignore the marks but to choose which families you are reading. In a contract, turn the amber family off entirely and keep the red one: absolutes still matter, because a contract that promises something it cannot deliver is a real problem. Keep the purple family as optional, since connectives in legal prose usually mean the clause order matters.

Technical documentation sits in the middle. Procedures want some repetition and some uniformity, so vocabulary marks are mostly noise, while stacked hedges and runaway sentence length are genuine risks. A specification that says the system should generally attempt to may be an honest description or an unfinished one, and the mark is asking you which.

In a multi-author document the marks do something else useful: they show where a second writer's style differs from the first. A section that suddenly lights up in amber and purple is a section written by somebody else or generated separately, and that is worth knowing before you edit for consistency.

The marks as a teaching tool

After four or five passes, something changes that is worth more than any single edit: you start hearing the sentence before the mark appears. The tool has done its job at that point, and the honest response is to use it less.

Keep the ten patterns that recur for you, written down somewhere you look at, and stop re-learning them every quarter. Most writers have a set of about that size; the rest of a tell list is other people's habits rather than your own.

Do not turn the list into a rule for other writers. Two people editing against the same vocabulary list produce the same voice, which is the failure the list was meant to prevent. The marks are a reading aid for an individual, and a style guide for a team should be written in sentences rather than in banned words.

There is a commercial argument here too. A review that can point at specific marks, in specific sentences, with a specific suggested replacement is a review a client can act on. A review that says this reads like AI is a disagreement, and disagreements about taste take longer than the edits would have.

What the marks cannot see

Marks find shape, not truth. A sentence can carry three flags and be exactly right, and a sentence can be completely clean and completely false. The tool has no view on whether a number is accurate, whether a quotation exists or whether a named study was ever run, because none of those questions are answerable from the vocabulary.

They are also not evidence of authorship. Every word on every list is ordinary English, used constantly by people writing carefully. Treating a phrase list as a detector is how public accusations get made against writers whose only crime was a formal register, and it is why this tool is described as an editing instrument rather than a classifier.

What the marks genuinely miss is the fabricated specific: the plausible citation, the confident figure, the expert who does not exist. Those failures are invisible lexically and expensive commercially. They are what the claim review editor, the verification matrix and the source comparison exist for.

So the editor is one station on a line rather than the whole inspection. Marks make the prose readable and honest at sentence scale. The console checks structure, claims and formatting. A named reviewer takes whatever carries consequences — anything about health, money, law, safety or a named person. Doing the first of those three well is what makes the other two short.