Skip to main content
Density scoring is the method VectorLint uses when reviewing content for discrete issues: banned phrases, structural errors, or missing elements. The rule defines what the LLM should look for; VectorLint counts what the LLM finds and calculates a score.

​
Why it exists

A raw issue count treats every document the same. One banned phrase in a 100-word release note and the same phrase in a 1,000-word spec would lose the same points. Density scoring measures issues per 100 words instead, so an error weighs more in a short document than in a long one. The score reflects proportional quality, not absolute count.

​
How it works

VectorLint calculates error density (violations per 100 words) then subtracts a penalty scaled by that density and the rule’s strictness: 
density    = (violations / wordCount) * 100
rawScore   = clamp(0, 100,  100 - density * strictness)
finalScore = rawScore / 10
The final score is on a 0–10 scale. The full expression in one line: 
score = (100 - (violations / wordCount * 100 * strictness)) / 10
strictness is a multiplier (default 10) set in the rule’s frontmatter. See Strictness.

​
Key properties

With the default strictness = 10, one violation scores differently depending on document length:
ViolationsWord countdensityrawScoreScore
1100(1/100) * 100 = 1.0100 - 1.0 * 10 = 909.0
11,000(1/1000) * 100 = 0.1100 - 0.1 * 10 = 999.9
The same error costs more in the short document because it represents a higher proportion of the content. A score below 10 means VectorLint found at least one issue. It reports the score at the rule’s configured severity; a severity of error causes a non-zero exit code and fails CI.

​
Constraints

A score of 0 means rawScore was clamped at the floor. The error density was high enough to exhaust the entire penalty budget.
  • Rubric scoring: the method used when reviewing quality on a spectrum instead of counting issues
  • Strictness: the multiplier in the formula and how to set it