---
title: "KB completeness has a formal toolkit beyond obligatory attributes — cardinality assertions, the No-Change Assumption, and text-extraction recall"
type: "claim"
status: "budding"
audit_status: "capture-verified (ar5iv full-text read at capture level, 2026-07-01; quotes carried from that direct read) | 2026-09-11 audit (claude-fable-5-1, cross-model lane; writer unknown): both arXiv versions read in full via extract_pdf (v1 2023-05-09, 33 pp.; v2 2023-12-06, 42 pp.; sha256s in source_sha) — the source_quote appears in NEITHER version; 'real cardinality' and 'cardinality of a relation' occur nowhere in the paper. It was a paraphrase recorded as a quotation (the failure mode sources.md's fabrication rule names), and is replaced with the paper's own sentence (§4, 'Completeness and cardinality assertions are closely related…'), prior value kept in source_quote_note. The three instruments are otherwise supported by the text: cardinality assertions (§4, Def. 3), the No-Change Assumption (§3.2 — the paper never abbreviates it 'NCA'; that is the vault's), and text-extraction recall (§3.2, §6.2). The 69% F1 figure is in v2 only ('Even the best-performing system, that relied on GPT-4, achieved only 69% F1-score on both relations', §4.3, citing the 2023 LM-KBC challenge) — so the capture read v2, and the [unverified-quant] hold on the shared-task proceedings stands. Claim unchanged."
source_url: "https://arxiv.org/abs/2305.05403"
source_sha: "v2 (2023-12-06, the text matching the 69% figure): 280bfaefa02b48eebbce34debbe52803f5d5819a9a561eac0cbf99c35326c100; v1 (2023-05-09): 34b7589dcd3292fa04581e0625bd6c5dddb7ecab8b04b9aec84cef9ff8744638"
source_title: "Completeness, Recall, and Negation in Open-World Knowledge Bases: A Survey"
source_author: "Simon Razniewski, Hiba Arnaout, Shrestha Ghosh, Fabian Suchanek"
source_date: "2023-05-09T00:00:00.000Z"
source_date_note: "v1 date; the version read at capture (ar5iv, 2026-07-01) and re-read at audit is v2, 2023-12-06 (the only version carrying the 69% figure)"
source_tier: 1
source_quote: "Completeness and cardinality assertions are closely related: cardinality assertions allow to evaluate completeness, while completeness assertions establish cardinalities."
source_quote_note: "prior value (capture, 2026-07-07): 'Completeness of a KB can be assessed by comparing the cardinality of a relation with the real cardinality' — NOT in the paper (v1 or v2); a paraphrase that had been recorded as a quotation. Replaced 2026-09-11 with the verbatim §4 sentence (v2 p. 13, lines 424–425 of the extracted text)."
provenance: "Promotion from 10-inbox/raw/2026-07-01-razniewski-2023-completeness-mechanisms-...md, 2026-07-07, queen cycle 6"
origin: "session"
date_created: "2026-07-07T00:00:00.000Z"
tags: ["knowledge-graph","completeness","cardinality","gap-detection","PKM"]
seek_code_commit: "f424b5f"
audits: ["2026-09-11 claude-fable-5-1"]
---


The same survey that grounds [[claim-obligatory-attributes-as-gap-signal]]
offers three further completeness instruments, each mappable to vault
practice:

1. **Cardinality assertions** — comparing a relation's recorded cardinality
   against its real one is the formal primitive of completeness measurement.
   *(2026-09-11 audit — the paper's words: "Cardinality assertions can be
   used to infer recall for a given selection condition by computing the
   division from Definition 3, |σ(K)|/n. Moreover, in the special case where
   |σ(K)| = n, they can be used to infer completeness" and "If we know that
   the KB is complete, we can infer that the true cardinality of PhD students
   she advised is 7" (§4). The sentence this note formerly carried as a
   quotation was a paraphrase; see source_quote_note.)*
   Vault analogue: "this cluster should have N members" claims (a MOC that
   names five clusters but links four is measurably incomplete).
2. **No-Change Assumption (NCA)** — if repeated observations stop adding new
   facts, treat the region as converged: a lightweight temporal completeness
   signal. *(2026-09-11 audit — verbatim, §3.2: "the No-Change Assumption
   says that if the number of objects has not changed over recent editions
   of the KB, then it has 'converged' to the true number, and no more objects
   are missing." The abbreviation NCA is the vault's; the paper does not use
   it.)* Vault analogue: a topic whose captures stop yielding new claims
   has stopped growing — the loop-until-dry intuition, formalized.
3. **Text-extraction-based recall** — compare what a source document mentions
   against what the KB extracted from it. Vault analogue: the promotion
   accounting's `not_promoted` lists exist precisely so recall against a
   capture is auditable.

One quantitative boundary from the paper stays flagged per the capture: the
LM-KBC 2023 best-system figure (69% F1 on cardinality prediction) is cited
within the survey from shared-task proceedings — [unverified-quant — needs
primary] until read there. See [[question-gap-detection]],
[[moc-machine-self-knowledge]].