Free PDF · ~12 pages

The Source Verification Playbook

Four deterministic checks plus an LLM judgment tiebreak. The 0.8 confidence floor that keeps the bundle from shipping with a fabricated stat. The exact playbook we run on every Qyndex content campaign.

What’s inside

  • HEAD checks — the 30-second filter. How a 5ms HTTP HEAD call kills the most common LLM citation failure mode (fabricated URLs) before any LLM judgement runs.
  • Domain reputation, hand-curated. Why we run a 200-row YAML file instead of a domain-rep API. The exact tier scoring we use, and what to put on the banned-domains list.
  • Freshness signals that are not just dates. HTTP Last-Modified header + visible-date regex. The age-decay formula that handles "stable references" (RFCs) and "fast-moving research" (arxiv preprints) on the same axis.
  • LLM judgment as the tiebreak — not the primary. The narrow prompt we use, why same-model self-check is unreliable, and when to escalate Sonnet → Opus on the verifier.
  • The 0.8 confidence floor. Why every cited source must score ≥0.8 (not the average), and how the floor was empirically calibrated against our internal eval set.

Who it’s for

  • Engineers shipping AI features that quote external sources (RAG, research agents, summary tools)
  • Editorial / content leads who need a defensible "no hallucinations" gate before publish
  • Founders staring at a citation in a draft and wondering if the URL is real