METHOD 1.0.0 · CHECK SET 2026-08-30
Evidence & Trust methodology
A deterministic map of page claims, nearby external references and visible responsibility signals. It locates the evidence chain; it does not replace fact-checking.
What the auditor measures
The auditor isolates numerical, attributed and—under the high-stakes standard—prescriptive claims. It asks whether an external reference appears in the same heading-bounded section, whether up to eight unique sources resolve, what broad domain class they belong to, and whether the page exposes authorship, review, dates and disclosures.
This differs from AI Citability: citability asks whether a passage is easy for an answer system to reuse and attribute. This tool asks whether a human editor can follow an important claim back through a visible evidence and responsibility chain.
Process
- Fetch public response HTML with redirect, response-size and time limits.
- Prefer main or article content; remove non-editorial interface regions.
- Split content at headings and isolate sentences with specific-claim patterns.
- Associate every external link in the same section with those claims.
- Check up to eight unique source URLs at audit time.
- Read metadata, structured data and visible text for author, dates, reviewer and disclosures.
- Score four dimensions under the selected evidence standard.
Scoring
| Dimension | General / research | High stakes |
|---|---|---|
| Claim support | 40% | 35% |
| Source integrity | 25% | 25% |
| Accountability | 20% | 25% |
| Freshness signals | 15% | 15% |
Claim support combines section-level reference coverage with the proportion of official or academic source classes. Source integrity combines current reachability and the same disclosed source-class heuristic. Accountability rewards a named author, named reviewer, dates and relevant disclosures. No signal proves correctness.
Check definitions
Specific claim detection
Identifies sentences with numbers, years, percentages, attributed research language or mode-specific prescriptive language. Important qualitative assertions can fall outside the detector.
Nearby source support
Counts a claim as linked when its heading-bounded section contains an external reference. It does not assume that the source proves the sentence.
Source reachability
Fetches up to eight unique references with the same public-web safety limits. Reachability is point-in-time; blocked automation and temporary errors can produce an Unable result.
Source class
Uses domain patterns to label official, academic, established-publisher or other sources. This is a routing aid only. A government page can be outdated; an “other” source can contain primary evidence.
Specialist review
High-stakes mode looks for visible “reviewed by” style language. It rewards a declared review trail, not the validity of credentials or depth of review.
Publication and modification dates
Reads article metadata and structured data. A recent date is not proof that the content was substantively reviewed.
Disclosures and limits
Looks for professional-advice limitations, commercial relationship language and editorial-process references. Wording patterns are incomplete and must be interpreted in page context.
Limits
- Source proximity is not entailment: human review must confirm that a source supports the exact claim.
- Only raw response HTML and a capped number of sources are checked.
- Source-class labels are coarse domain heuristics, not quality ratings.
- Authorship, reviewer and date declarations are not independently verified.
- Claim detection is pattern-based and strongest in English.
- Legal, medical and financial accuracy requires qualified expert review.
- The report does not measure rankings, citations, conversions or reputation.