Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Repair Windows errors before they cause bigger problems3Scan for outdated or missing drivers - takes under a minuteiTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Yes: in one writing linter, adding unrelated prose lowered the score enough to change a draft from review to pass. The problem was not that the tool found fewer issues. Its single score blended deduplicated word-level checks with punctuation and formatting rates that shrink as the word count grows.
Frank Chu’s October 1, 2026 account describes the test, the input changes he made, and the remaining weakness: actual prose can still dilute the body score. The results are his measurements, not an independent benchmark of writing linters generally. Read Chu’s account on DEV Community.
How much did adding words change the score?
In Chu’s example, a draft scored 9 at 78 words. After text was added that did not trigger the same objections, it scored 3 at 648 words—a reduction by a factor of three in the reported score. The issue count rose from one to three because the added text introduced mild flags. The lower score therefore did not mean the linter found fewer issues. Chu’s October 2026 account reports these figures.
A separate input test showed how much the material sent to the checker mattered. The same published post received a PASS when analyzed as a 930-word whole file, but REVIEW / REWRITE as 746 words of body text. Both runs reported the same three issues. Chu attributed the difference to the whole-file word-count denominator and dashes in the front matter. These are the author’s measurements of his checker.
#1 Best Overall
Why can more writing lower a score?
The linter combined two kinds of measurements in one scalar score:
- Set-based word checks: A flagged phrase contributes as a match; repeating it does not keep adding to the score because matches are deduplicated.
- Density-based checks: Em-dash and bold-formatting checks are expressed as ratios against the number of words.
Those measures respond differently to added text. Repeating an already-detected phrase may not change its set-based contribution, while adding words that do not attract the same checks lowers a density ratio’s share of the total. If both families are combined and judged against one threshold, the falling density contribution can offset set-based terms. The verdict can then pass even when the findings remain or increase.
This describes the particular scoring design Chu tested; it is not evidence that all writing linters work this way. The key issue is the single combined score, not simply the presence of a word count.
What text did Chu remove from analysis?
Chu changed the input so the body analyzer would focus on prose rather than surrounding material. Before calling the analysis function, the checker now removes an initial HTML comment note, front matter, fenced and indented code, and inline code. Each exclusion addresses a different source of distortion:
Rank #3
- Notes and metadata are not the article body, but can affect counts and ratios.
- Code can trigger word-level matches for text that is not ordinary prose. Chu gives
struct.unpackmatching the English wordunpackas an example.
The title and description are assessed together in their own region. Chu’s detector has a ten-word minimum, so a short title evaluated by itself could fall below the minimum. Grouping title and description lets the checker assess that reader-facing text without mixing it into the body score.
What does this change fix—and what does it not?
Excluding notes, metadata, and code makes the body input more representative of the text being judged and reduces denominator effects from non-body material. Separate title-and-description analysis also avoids treating a short title as if it were a sufficiently large sample on its own.
But the body is still judged with one scalar against one threshold. Chu notes that adding actual prose can still dilute the density component. Removing irrelevant input does not solve the underlying design problem: unlike measurements are still being combined into one number.
Free tools Windows power users keep installed
One-click scans. No signup required.
What would make this kind of quality gate harder to fool?
Chu’s proposed fuller repair is to score the two rule families separately and let either family fail the post. That would prevent a better result on a density measure from compensating for a failure on set-based checks. More broadly, a quality gate should make clear what each score measures and what text it evaluates.
- Separate binary or set-membership findings from rate-based checks instead of averaging them into one verdict.
- Keep metadata and code out of body analysis when they are not part of the prose being evaluated.
- Check that short regions meet a detector’s minimum sample size before interpreting their scores.
- Test whether adding neutral text changes a verdict even though the underlying findings have not improved.
How widespread was the effect in the archive?
Chu compared whole-file and body-only scores across 100 archive posts. He reports higher whole-file scores in 40 cases and one verdict flip. He characterized the practical damage in that archive as limited, while pointing out that the gate had produced a number trusted for the wrong reason. These figures are his archive review, not an independent estimate of how often similar failures occur in other tools.
His account says the experiment took four minutes once he thought to try it. The useful practice is not to assume every linter is vulnerable, but to deliberately challenge a gate: add material that should not improve the writing and see whether the score or verdict changes without a meaningful change in findings.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
