Provenance Monogamy in an LLM Knowledge-Verification System: A Negative Result and an Audit Method

We report a negative result from a production knowledge-verification pipeline(OmniCore, PARAM BILIM cluster) together with the audit method that producedit. The system maintains a triple store with per-triple epistemic status( unverified / verified / disputed / retracted ) and was designed to movefacts toward verified by cross-checking sources. A provenance audit of the fullstore (N = 1,125 triples, snapshot 2026-07-19) found that 98.5% of triplesdescend from a single source ecosystem (Wikimedia), that 100% of resolvablesource URLs concentrate on one host, that zero triples carry an externalindependent provenance, and that all seven triples marked verified werepromoted by internal executor agents consuming the same ecosystem — self-approval rather than verification, including one closed citation loop in which acached “confirming” snapshot points to the triple’s own source URL. We framethe finding with a six-class taxonomy of witness-failure modes (C1–C6) and astructural impossibility result: against an adversary who owns the substrate, root-independence of witnesses hosted on that substrate is unattainable byconstruction, for reasons adjacent to — but distinct from — the Knight–Levesoncorrelated-failure result, Byzantine fault-tolerance assumptions, and the trusted-computing-base boundary. The main practical contribution is a three-query auditrecipe that any operator of a triple store or RAG corpus can run to measureprovenance concentration. We describe a deliberately modest repair in progress(ingesting a state registry as a second, constitutively distinct root) and list thedesign questions this repair leaves open.