Scoring

July 3, 2026 ยท View on GitHub

The v0 scorer is deterministic and intentionally plain. It is a gate, not an oracle.

The scorer includes English and Chinese cues for rule language, examples, transferability, and conversational noise. Chinese source length is estimated with CJK characters as effective tokens rather than only whitespace-delimited words.

Signals

SignalMaxMeaning
Source depth25More words usually provide more room for structure and evidence
Structure20Headings, bullets, and numbered lists make extraction safer
Actionability20Rule-like language suggests reusable behavior
Transferability15Reuse, workflow, checklist, and decision cues suggest future use
Evidence10Examples and cases reduce unsupported claims
Signal cleanliness10Conversational noise lowers confidence

Output-Level Mapping

Default score mapping:

ScoreLevel
0-19Discard
20-39Note
40-59Skill Seed
60-79Mini Skill
80-100Full Skill

Length caps prevent short sources from being promoted too far:

  • under 120 words: at most Skill Seed
  • under 600 words: at most Mini Skill
  • under 2,500 words: at most Mini Skill

This is conservative on purpose. Dense short sources are often useful, but they rarely deserve a full standalone skill.

Why No LLM Scoring In v0

LLMs can be useful later, especially for evidence classification and claim checking. The first version keeps scoring transparent so users can understand why a source was promoted or held back.