Methodology

Retrieval

Three independent layers run per query: a lexical BM25 index, a substring (trigram) index, and a vector index. Their ranked lists are combined with Reciprocal Rank Fusion, which needs no training, no learned reranker, and no randomness — a ranking can be replayed from the retrieval trace we return with every search result.

Lexical evidence is weighted above semantic evidence, because an exact term match is a fact and a nearest neighbour is a guess. Below three query tokens the semantic layer is dropped entirely: absent evidence contributes nothing.

Contradiction detection

Detection is rule-based and therefore replayable. We report only classes of disagreement that are decidable without a model: two dates for the same event, two quantities for the same measurement, or an assertion and its negation.

Findings are advisory. We state that claims conflict and show both with their sources. We never declare which claim is true. A year-only record and a full-date record for the same event are treated as compatible, not as a conflict.

Detection is measured against a hand-labelled evaluation set on every change, with precision as the release gate. A false alarm costs more credibility than a missed conflict ever costs coverage.

Provenance

Every article is sealed into an append-only SHA3-256 hash chain. Each entry commits to a content root over its own files, and to the digest of the entry before it. The chain is re-derived from its genesis point on every verification rather than trusted from storage.

A hash chain alone is self-consistent under a total rewrite, so entries are signed and the verification key identifier is pinned out of band. A verifier that accepts whatever key the ledger presents has proved nothing about authorship.

Licensing

Structured factual content derives from Wikidata, released under CC0. Prose is our own synthesis under the Anticloud licence. Where content is derived from Wikipedia article text it is served under CC BY-SA 4.0 with attribution and an indication of modification, as that licence requires.

What this is not

This system is not a certification and claims nothing. It produces evidence that an independent assessor may examine. Nothing here has been audited by a third party.

encyclopegia.org is operated by Anticloud FZ LLE (UAE), founded by Lois-Kleinner Alpasan. Anticloud also operates 0-1.gg.