Quality vector

Every result in a sift response carries a quality_vector. It is the core transparent artifact: each field is derived from a specific input and logged in signals[] so downstream agents (and you) can trace why a classification landed where it did.

Schema

interface QualityVector {
  schema_version: "1.0",
  tier:
    | "regulated_primary"        // SEC / EDGAR / gov / court records
    | "peer_reviewed"            // arXiv / PubMed / .edu papers
    | "independent_editorial"    // BBC / Reuters / Ars Technica (sans sponsored)
    | "vendor_primary"           // vendor's own product / docs / homepage
    | "vendor_content_marketing" // vendor blog / strategic advice
    | "affiliate"                // "Best X" listicle + commercial intent
    | "content_farm"             // templated / AI / low-effort aggregator
    | "ugc"                      // Reddit / forums / community
    | "unknown",
  editorial_standards: "high" | "medium" | "low" | "unknown",
  self_promoting: boolean,
  third_party: boolean,
  commercial_intent: "high" | "medium" | "low" | "none",
  domain_content_mismatch: boolean,    // parasite-SEO flag
  authoritative_weight: number,        // 0..1
  confidence: number,                  // 0..1
  reason: string,                      // under 80 chars
  signals: Array<{
    origin: "safety" | "authoritative" | "llm",
    type: string,
    match: string,
    weight: number
  }>
}

The 9 tiers

See Tier definitions for the full set of examples and boundary rules. Briefly:

Tier	Example	Typical use by agent
regulated_primary	SEC filings, government publications, court records	Cite as authoritative
peer_reviewed	arXiv, PubMed, SAGE/Elsevier journal articles	Cite as authoritative
independent_editorial	BBC, Reuters, Ars Technica (main articles, not sponsored)	Cite with attribution
vendor_primary	Product homepage, API docs, official changelog	Cite for vendor facts
vendor_content_marketing	HubSpot blog, Stripe blog, VC firm thought leadership	Treat as positioning
affiliate	”Best X 2026” listicles	Commercial — do not use for objective comparison
content_farm	Templated AI / SEO churn	Do not cite
ugc	Reddit, Stack Overflow, Quora	Treat as opinion
unknown	Cannot determine from signals	Flag uncertainty

Orthogonal axes

tier is the primary categorical dimension, but several orthogonal flags give the agent finer-grained context:

• editorial_standards — high / medium / low / unknown. A reputable domain’s /sponsored/ path has medium or low standards even if the parent brand is high.
• self_promoting — true when the publisher has a direct stake in recommending itself (e.g., a hosting company’s “Best Web Hosting” list).
• third_party — true when the publisher is independent of the subject.
• commercial_intent — high / medium / low / none. Orthogonal to tier; a vendor_primary page can be none (docs) or high (pricing).
• domain_content_mismatch — true when the domain’s implied business strongly differs from the content topic. A parasite-SEO flag.

authoritative_weight and confidence

• authoritative_weight is a scalar in 0..1 derived from tier + editorial_standards + confidence + domain_content_mismatch. It’s what you’d multiply source claims by when aggregating. The exact formula lives in src/vectorize.ts#authoritativeWeightFromLlm.
• confidence is the classifier’s own certainty about the whole classification. Low confidence + unusual signals is a good trigger for an agent to caveat the result.

signals[]

Every classification is backed by at least one signal. Three origins:

• safety — Google Safe Browsing (threat type as match)
• authoritative — known-trusted / known-good cache hit (domain as match)
• llm — LLM judge output (tier + reason as match)

A result can have multiple signals. For example, a reddit.com URL that Google Safe Browsing also flagged would carry both authoritative (known-good:ugc_community) and safety signals.

Why expose this much?

Because agents need to explain their reasoning. If an agent writes “According to industry data, the SaaS magic number benchmark is 0.7-1.0”, the user (or a reviewer) should be able to ask “where did ‘industry data’ come from?” — and the agent should be able to say “10 vendor_content_marketing blog posts from VC firms, with mean authoritative_weight 0.23; no peer-reviewed source was in the SERP.” That transparency is what the quality vector enables.