knowledge-confidence.md

Knowledge confidence scoring

Knowledge facts expose a numeric confidence for API compatibility plus two additive fields:

• confidence_scored: whether the number is backed by real signals.
• confidence_reason: a short explanation for dashboards and operators.

The dashboard shows unscored instead of a percentage when confidence_scored is false. This prevents legacy defaults from looking like measured certainty.

Formula

Explicit frontmatter or API confidence values remain authoritative unless they are a legacy generated placeholder. Generated facts are rescored on read from signals Hive already stores:

1. Source provenance
   • configured external document: 0.72
   • HTTPS configured document: 0.73
   • Context7 document: 0.74
   • bead-synth fact: its existing deterministic bead classification score, or 0.56 if missing
2. Workflow-only baseline
   • facts without stored confidence or document/bead provenance start at 0.50 only when they have validation, related-link, source, or usage signals; otherwise they are unscored
3. Human/workflow validation
   • verified, validated, or approved: +0.12
   • published: +0.06
   • draft: -0.08
4. Layer authority
   • community: +0.04
   • org: +0.03
   • project: +0.02
   • personal: no change
5. Corroboration and use
   • related wiki links: +0.02 each, capped at +0.06
   • additional sources: +0.02 each after the first, capped at +0.06
   • vault access count: +0.01 each, capped at +0.05
6. Freshness
   • newer than 30 days: +0.03
   • newer than 180 days: +0.01
   • older than 365 days: -0.04
   • freshness adjusts an otherwise scored fact; file modification time alone does not turn an otherwise signal-free fact into a scored one

Scores are clamped to 0.05..0.99 and rounded to percentage precision. Facts with no explicit confidence and no provenance, validation, related-link, source, or usage signal are returned as unscored.