Skip to main content
Rubric scoring is the method VectorLint uses when reviewing content on a quality spectrum: clarity, tone, technical accuracy, or completeness. The rule defines the criteria and rubric levels; the LLM rates the content against each criterion, and VectorLint calculates the final score.

​
How it works

The rule defines criteria, each rated by the LLM on a 1–4 scale. VectorLint normalizes each rating to the 1–10 range, then combines them into a weighted average. The normalization maps the full 1–4 rating range onto 1–10: 
normalized = 1 + ((score - 1) / 3) * 9
RatingLabelNormalized score
4Excellent10.0
3Good7.0
2Fair4.0
1Poor1.0
The normalized range is 1–10, not 0–10: a rating of 1 (Poor) floors at 1.0, so even a fully failing criterion still contributes a small positive score. Each criterion carries a weight (default 1 when unset). A higher weight gives that criterion more influence over the final score: 
finalScore = sum(normalized_i * weight_i) / sum(weight_i)
Criteria are defined in the rule’s frontmatter: a human-readable name, an id, and an optional weight. The LLM rates the content against each criterion using the rubric described in the rule’s body. See Rule frontmatter.

​
Key properties

A rule with two criteria:
CriterionWeightLLM ratingNormalized
Technical Accuracy403 (Good)7.0
Readability304 (Excellent)10.0
weighted points = (7.0 * 40) + (10.0 * 30) = 280 + 300 = 580
total weight    = 40 + 30 = 70
finalScore      = 580 / 70 ≈ 8.29
VectorLint reports scores to one decimal place, so this document scores 8.3.

​
Constraints

  • Strictness has no effect on rubric scoring. Strictness is a multiplier on error density and applies only to density scoring. See Strictness.
  • For documents above the chunking threshold, VectorLint reviews each chunk independently. Judge scores are averaged across chunks, weighted by chunk size. This is an implementation detail. No configuration is required.
  • Density scoring: the method used when reviewing for discrete issues instead of rating quality
  • How it works: the full review pipeline, including chunking