Verification and trust scoring
How trustworthy a claim is depends on its domain. Verify with the right framework, score on a common scale, and cap claims that fail their domain's minimum — so a casually-asked medical claim is still held to the medical bar.
Verification frameworks
- SIFT / Four Moves — Stop, Investigate the source, Find better coverage, Trace to the original. Fast triage.
- Lateral reading — leave the page to check the source's reputation elsewhere.
- CRAAP / RADAR — currency, relevance, authority, accuracy, purpose.
- GRADE + evidence hierarchies — strength of medical/scientific evidence (RCT > cohort > case report).
- IFCN code — journalistic fact-checking standards.
- Predatory-venue (Beall's) and retraction (Retraction Watch) checks — void credit for fake or withdrawn sources.
Domain bars
Medical VERY HIGH (RCT / systematic review; check retractions) · Scientific HIGH (+ replication) · Legal HIGH (primary authority, jurisdiction, currency) · Financial HIGH (SEC / government primary) · News MED-HIGH (two independent outlets) · Historical MED-HIGH · Statistical MED-HIGH (disclosed methodology) · Technical MEDIUM (official current docs; recency) · Business MEDIUM · Product / how-to / definitional LOW-MEDIUM (broad independent consensus).
Scoring (0-100)
Blend: Authority (30) + Independent corroboration (25) + Accuracy/verifiability (20) + Currency (10, domain-scaled) + Transparency (15). Then apply a domain ceiling that caps a claim failing its domain's minimum source — e.g. a medical claim caps near 45 without a peer-reviewed RCT or systematic review, however confident it sounds.
False-positive catalog
Score LOW when a claim is: single-source; citation laundering / loops; confident-but-unsupported; published in a predatory venue; carrying an undisclosed conflict of interest; wrapped in urgency / emotional framing; outdated in a fast-moving domain; an anecdote generalized to a population; viral-but-unverified; backed by a fabricated or hallucinated citation; from an authority speaking outside their field; p-hacked / unreplicated; or AI content-farm output.