Writing prose that reads as written

The lint in G6 catches habits, not authorship. It will flag a person who writes "it is
important to note" and miss a model that avoids every construction on its list. So passing
it is the floor, not the goal. What follows is the goal.

Say the thing

Most machine-written scientific prose fails in one way: it describes the significance of
a finding instead of stating the finding. Every listed tell is a variant of that.

The reporting odds ratio was elevated, underscoring the importance of continued
pharmacovigilance for this agent.

The clause after the comma asserts importance and adds nothing checkable. Cut it and put a
number in its place:

The reporting odds ratio was 3.84 (95% CI 2.89 to 5.12), based on 77 cases.

If a sentence would survive being deleted with nothing lost, delete it.

The specific habits

Do not write about writing. "It is important to note that", "it is worth mentioning",
"notably". If it were not worth mentioning it would not be in the paper.

Do not use "not just X but Y". It reads as emphasis and usually says less than X and Y
stated plainly. The same goes for "it's not A, it's B".

Do not attach an "-ing" tail asserting significance. "…, highlighting the need for…",
"…, reflecting the broader trend of…". These are almost never true claims; they are the
shape of a claim.

Use "is". "Serves as", "stands as", "functions as", "represents a" — occasionally one
of these is the right verb. Usually "is" was.

Attribute or cut. "Studies have shown", "experts argue", "it is widely accepted". Which
studies? Cite them. If you cannot cite them, you do not know it.

Watch the rate, not the word. "Robust", "crucial", "key", "comprehensive", "highlight"
are ordinary words with legitimate uses. Six of them in a paragraph is the tell. The lint
measures a rate for exactly this reason, and so should you.

Em dashes are fine in moderation. So is bold. The lint's thresholds are rates, not
prohibitions.

The formulaic conclusion

The source essay names a shape that is worth avoiding wholesale:

Despite these limitations, this study provides valuable insight into … Future research
should explore … Ultimately, these findings contribute to a growing body of evidence …

Three sentences, no content. A discussion that ends well ends with what the reader should
now believe, and what would change their mind:

Disproportionality cannot establish incidence, and these data contain no denominator.
A cohort study with prescription counts would settle whether the excess reflects risk or
reporting.

Checking your work

manuscript-guard check          # G6 among the rest

Findings name the rule and the reason. Disagree freely — several of these constructions
have defensible uses, and the warnings do not block a build. What you should not do is
change a sentence merely to satisfy the lint: if the rewrite is worse, keep the original
and say so.

Two things do block: model output artefacts (oaicite, [cite: 1], "as of my last
training data") and unfilled placeholders. Those have no defensible use in a submission.

If you are a model reading this

Draft freely, then re-read the draft against this page before showing it to anyone. The
constructions above will be in your first draft; that is what the page is for. The specific
thing to check for is the one at the top: sentences that assert importance instead of
stating a result. They are the hardest to notice and the most damaging to a paper, because
a reviewer reads them as padding and a reader learns nothing from them.