Phantom Convergence: Auditing Cross-Domain Claims in an Autonomously Built Knowledge Graph

Autonomous knowledge-graph builders often report that independent sources converge on the same idea. We audited one such system, ANIMUS, three times, and each audit removed the effect the previous version reported. First, 52% of its self-reported nodes were undetected duplicates (1,892 reported, 911 unique). Second, its headline convergence, failure → algorithm "confirmed by 32 sources", co-occurs in 2 of 29 books when chance alone predicts 2.6. Of 639 co-occurring tension–resolution pairs, 4 survive a chance model with false-discovery control: fixed expressions such as problem → solution. The fourth is a topical pair, poverty → education. Third, we rebuilt the system to make such errors detectable. ANIMUS v5 extracts ideas with a local LLM, keeps an idea only if the model's verbatim quote is found in the passage, and embeds ideas with a multilingual model that passes a positive control at 97% (chance: 5%). It then counts cross-book links against decoys run through the identical procedure. On 10,392 verified ideas from 29 books (52 million cross-book pairs), real ideas produce fewer links than word-shuffled decoys (18,258 vs. 20,050). The most striking surviving theme, statements about the limits of knowledge across Kuhn, Meadows, Isaiah and Gödel-themed texts, was largely a negation artifact: both-negated pairs made up 42% of the top 100 links under word-shuffled decoys and 11% under decoys that preserve sentence form. We find no evidence of cross-domain convergence in this corpus beyond vocabulary, topic and syntax. We contribute two open-source auditing tools and a reusable recipe: quote-grounded extraction, an embedding positive control, and form-preserving decoys.