graphaudit v1.14.0

New tools/boilerplate.py: finds lines repeated across a GraphRAG index's text units and drops sources where a pair only meets inside them, without re-indexing; --find WORD prints the text units that mention a word. On the banking-regulation graph, the top 'beyond chance' edge (Santo Domingo - Republica Dominicana, 161 units) falls to 15 units within chance; 517 relationships were only ever seen in the letterhead. Also confirmed: 4 entities (Aurelia, Tiruzia, Alhamia Prison) copied from GraphRAG's prompt example.

graphaudit v1.13.0

Blind test 5: a GLM-4-9B GraphRAG index of A Christmas Carol contains Firuzabad, Aurelia, Cashion and Quintara, copied from the worked example in GraphRAG's default extraction prompt; Scrooge has no type. graphaudit now flags entities named after that example and counts entities with no type in a typed graph.

graphaudit v1.12.0

Blind test 4: Microsoft's GraphRAG test index and the Neo4j blog index are both A Christmas Carol; they share 171 of 587 distinct entities (Jaccard 0.29). New tools/compare_entities.py. A change confined to words of three letters or fewer is never a typo (Bob / Boy Cratchit); a title alone is never a short name (Lord / Lord Mayor).

graphaudit v1.11.0

Blind test 3 on itext2kg (entity resolution): no spelling duplicates, but 16 self-loops show different entities merged (gpt4.1 is_faster_than gpt4.1). The report now lists self-loops, entities whose type is their own name, and rows with empty endpoints; two different one- or two-letter words are never a typo (uk/us government); lower-case graphs get partial names between people only (musk / elon musk). New converter tools/itext2kg_to_csv.py with a restricted unpickler.

graphaudit v1.10.0

Blind test 2 on ai-knowledge-graph's five samples: 17-30% of edges are inferred, and on sourced edges alone each graph falls into 2-16 pieces. The report now counts edges with no provenance and the connected pieces without them; all-lower-case graphs are reported as 'partial names not assessed'; a different last letter on the same stem is not a typo. New converter tools/akg_to_csv.py.

graphaudit v1.9.0

Fixes from the blind test on a 7,945-entity graph: numbered series never pair (Catarey / Catarey I), OCR numerals (ILL, IL, LL), markup stripped from types (<ORGANIZATION), judge prompt names series, things named after an organisation and a word inside a longer name as different. Sharing the judge budget between sources was measured on the same graph (more wrong, fewer right verdicts) and reverted.

graphaudit v1.8.0

--judge: a node judged SAME as two entities the model judged DIFFERENT from each other is a generic node; its SAME verdicts become uncertain, so verdicts cannot chain different entities together. A Christmas Carol with Qwen3-4B: 33 right, 6 arguable, 7 wrong after the rule (was 14 wrong).

graphaudit v1.7.0

--judge now judges named things only (people, organisations, places, events) unless --judge-all; a SAME whose reason describes a hierarchy counts as uncertain. Qwen3-4B on the Wars of the Roses graph: 8 SAME verdicts, all correct, 0 wrong.

graphaudit v1.6.0

Experimental --judge: a local instruction model reads names, types and descriptions and decides whether candidate pairs are the same entity (asked in both orders). On the Wars of the Roses graph with Gemma 4 E2B: 9 of 11 'same' verdicts right, 1 wrong; separates related entities far better than embeddings.

graphaudit v1.5.0

kg-gen support and an Apple Inc. check; experimental --embed with a real negative result on bge-m3; new guards for model numbers, '+', single letters, whole-word swaps; legal-form pairs.