We trace every claim in both encyclopedias back to independent sources and measure how well each one holds up. Transparent, reproducible, and unbiased — the method and the data are open.
Our goal is to make AI reliable enough for individuals and businesses to trust — with complete auditability and grounded answers.
We built a living knowledge graph with a deterministic audit trail: every fact carries where it came from, and every answer an agent gives stays grounded in that traceable trail, so each claim stands on a verifiable source. On top of it, our hybrid retrieval steadily narrows the agent’s search space — raising answer quality and driving token cost down over time.
Every metric is computed from the articles themselves — the same way for both encyclopedias, with Wikipedia as the control. Definitions, formulas, and the research each draws on are below; the index is provisional and the method is open.
Citation density — how heavily an article sources its claims, measured as distinct references per 1,000 words of prose.
How it’s calculated: references ÷ (words ÷ 1,000), then scaled to 0–100 against a reference density of ~90 refs / 1,000 words (the measured mean of the most-edited English Wikipedia articles). A structural signal of how much is sourced — not, on its own, a measure of source quality.
Yasseri 2025 (Grokipedia vs Wikipedia) · Wikipedia:Verifiability
The share of cited links that still resolve to a live page rather than a dead or domain-parked one (“link rot”).
How it’s calculated: each unique cited URL is fetched (HEAD, then GET) following redirects: 2xx/3xx counts as resolving; 4xx/5xx, DNS failure, timeout, and domain-parking landing pages count as dead. Score = % resolving, reported with a 95% Wilson confidence interval.
Klein et al. 2014, “Reference Rot” (PLOS ONE) · Pew Research 2024, “When Online Content Disappears”
Of the claims that carry a citation, the share where the cited source actually backs the claim (entailment) — not merely sits next to it.
How it’s calculated: the cited passage is retrieved, then an LLM entailment ensemble labels it entails / contradicts / neutral against the claim; score = entailing ÷ judged. The hardest metric to game; shown as “pending” while the judge is validated against a human-labelled set.
Petroni et al. 2023, “Improving Wikipedia Verifiability with AI” (SIDE) · Kamoi et al. 2023, WiCE (EMNLP)
The share of citations pointing to sources not on a published low-credibility index. Scores a domain’s factual reliability — not its political viewpoint.
How it’s calculated: each cited domain (registered eTLD+1) is checked against a published third-party low-credibility list; score = % not flagged, with a 95% Wilson interval.
Iffy+ Index (CC BY 4.0) · Lin, Pennycook & Rand 2023, expert domain-quality aggregate (PNAS Nexus)
How much the article stands on its own — avoiding self-citation (citing itself or affiliated properties) and verbatim reuse of another encyclopedia’s text.
How it’s calculated: averages two 0–100 signals: (a) 1 − share of citations to self/affiliated domains, and (b) 1 − verbatim 7-word-sequence (7-gram) overlap with the matched Wikipedia article.
A single provisional 0–100 score combining the five metrics above. Wikipedia is the control.
How it’s calculated: weighted average — Verifiability 20%, Citation integrity 25%, Citation support 30%, Source quality 15%, Independence 10% — each normalised to 0–100. When a metric is not yet available the weights are renormalised over the rest and the score is labelled provisional; the head-to-head compares only metrics measured on both encyclopedias.