Wikipedia and Baidu Baike serve human reference; Aumedia Fact‑Checker establishes credibility for AI.

I. Case Background

Large‑language‑model generation of corporate profiles and competitor research reports has become commonplace. Nevertheless, LLMs are inherently prone to hallucinations: spliced factual errors, scrambled timelines, uncritical adoption of corporate promotional claims, blurring of opinion and fact, and self‑generated conclusions.

The document under review was produced by Zhipu AI. It constitutes a full corporate profile for Weijian, a fitness‑venue SaaS brand operated by Chengdu One Avocado Technology Co., Ltd. The text covers corporate overview, product features, market competitiveness and vendor‑selection takeaways, representing a typical LLM‑generated enterprise‑service research document.

Verification Objective: Leveraging the Aumedia Fact‑Checker KUCR Four‑Layer Attestation System, perform factual refinement, conflict detection and opinion segregation on LLM output. Distinguish verified facts, insufficiently‑supported claims, conflict‑risk facts and subjective value judgements. Deliver a standardised fact‑set ready for ingestion into corporate trust archives, supplier ledgers and RAG knowledge bases. Mitigate business‑due‑diligence, competitor‑analysis and public‑release risks stemming from AI hallucinations.

II. Verification‑Tool Description: Aumedia Fact‑Checker KUCR Four‑Layer Attestation System

Purpose‑built for fact‑attestation of LLM‑generated content, Aumedia Fact‑Checker differs from simplistic binary true‑or‑false rumour‑debunking tools. Its core principle: it does not render blanket right‑or‑wrong verdicts. Instead, it verifies whether authoritative evidence exists, whether multiple sources align, and under what conditions given conclusions hold true.

Four‑layer attestation workflow:

• K (Knowledge Decomposition): Segment long AI‑generated texts into discrete atomic facts. Categorise content as business‑registration credentials, intellectual‑property records, financing history, business scale, product descriptions, industry judgements or corporate vision. Preserve original qualifying conditions without arbitrary semantic truncation.
• U (Understanding Semantic Normalisation): Identify content types and strictly separate verifiable objective facts / corporate self‑promotional statements / version‑iteration descriptions / AI‑synthesised comparative evaluations / industry‑analysis conclusions. Prevent model‑invented comparative judgements from being treated as objective facts.
• C (Conflict Calibration): Cross‑reference against independent sources including business‑credit databases, corporate official websites, media archives and app stores. Focus on common LLM failure modes: chronological contradictions, version drift, treating isolated promotional materials as definitive findings, and divergent source narratives.
• R (Result Attestation & Archiving): Assign four labels: Verified Fact, Insufficient‑Evidence / Exaggerated Inference, Conflict‑Risk (Unwarranted Inference) and Value Judgement. Tag each entry with source references and confidence grades. Flag conflicting records as archive‑prohibited. The full verification pipeline remains traceable and auditable.

Evidence‑Acceptance Rules: Accept claims corroborated by multiple consistent sources; explicitly flag divergent accounts; absence of retrievable supporting evidence does not automatically prove falsity. Only objective facts undergo assessment; opinions, evaluations and commercial recommendations are excluded from factual adjudication.

III. Subject of Verification

Source material: Zhipu‑AI‑generated corporate profile for Weijian. Content includes entity particulars, founder background, financing history, intellectual‑property holdings, venue‑service volume, product matrix, smart‑hardware offerings, strengths‑and‑weaknesses analysis and vendor‑selection advice. Materials of this kind are widely used for competitor research, supplier evaluation and industry‑white‑paper source content.

IV. Publication of Standardised Verification Results

✅ Verified Facts (10 entries, cross‑validated across multiple sources; eligible for repository ingestion)

1. Brand ownership: Weijian is a fitness‑venue SaaS‑management‑system brand under Chengdu One Avocado Technology Co., Ltd. Confirmed by Qichacha, Qixinbao and the App Store.
2. Founding date: The company was incorporated in April 2017. Business‑registry platforms are in mutual agreement, consistent with the document’s claim of “nine‑year track‑record in fitness SaaS”.
3. Headquarters location: Headquarters are situated in Chengdu Hi‑Tech Zone. Cross‑validated against business‑registry datasets.
4. Founder information: Founder is Li Zhe. Records from Tianyancha, Leiyun News republished reports and Qichacha are consistent.
5. Corporate positioning: Positioned as an “Internet‑plus” digital‑services provider for the fitness sector. Matches business‑scope registry entries and official corporate communications.
6. Trademark portfolio: Holds 22 trademarks. Records from Tianyancha and industry‑media archives align.
7. Patent portfolio: Holds 2 patents. Corroborated across multiple credit‑platform and media archives.
8. Software copyrights: Owns 19 software copyright registrations. Matches Tianyancha and industry‑media archives.
9. Total intellectual‑property assets: 43 combined IP assets (22 + 2 + 19 = 43). Figures reconcile between official‑website disclosures and business‑registry records.
10. Product provenance: Products are self‑developed. Accepted based on official‑website representations. For formal external use, supplementary third‑party evidence such as software‑copyright certificates is recommended to reinforce the evidentiary chain.

⚠️ Insufficient corroboration; AI reproduced corporate self‑claims (3 entries; no independent third‑party evidence; direct repository ingestion prohibited)

1. Serving more than 16 000 fitness venues: This figure originates from corporate public‑relations statements reproduced in media coverage. No independent third‑party statistical report substantiates it. Statistical scope and reference timestamp are defined solely by the company. Any external usage must explicitly label it as an official corporate claim and must not be treated as established fact.
2. Core team members combine product‑R&D and venue‑operation experience: Public materials confirm team members have Internet‑industry backgrounds and long‑sector experience, yet no public documentation verifies “venue‑operation experience”. This represents model‑generated speculative synthesis; the statement cannot be verified and is not accepted as fact.
3. Multi‑terminal collaborative workflows across PC, APP, mini‑program and mini‑web‑portal: Public sources confirm interoperability for PC, APP and member‑facing mini‑programs. The assertion that the “mini‑web‑portal functions as an independent office‑work terminal supporting multi‑terminal collaboration” lacks direct source material and reflects excessive AI generalisation; not accepted as fact.

Supplementary note: Product‑capability descriptions within the draft such as “more than ten marketing plug‑ins”, “Alibaba Big‑Data‑powered 3‑kilometre customer‑acquisition capabilities” and “full smart‑hardware ecosystem” reproduce corporate marketing claims integrated directly by Zhipu AI. No matching third‑party supporting evidence was located during this check. Acceptance would require product‑testing reports or independent third‑party evaluations.

❌ Conflict‑risk entries (chronological / version contradictions; 2 entries, classic LLM hallucinations; repository and trust‑archive ingestion forbidden)

1. Pre‑Angel millions‑RMB angel‑round financing: Third‑party financing archives date this investment to 2014, predating the company’s April 2017 business‑registry incorporation. A clear chronological inconsistency exists. The LLM spliced financing records belonging to a different legal entity, yielding a timeline hallucination. Attribution to this corporate entity cannot be confirmed. Marked high‑risk; excluded from repositories.
2. Management‑system partitioned into five modules: front‑desk, membership, coaching, finance and administration: Official‑website module divisions have evolved through product iterations. Earlier releases listed six modules; later iterations include a BOSS‑admin module. No single stable partitioning standard exists. The LLM captured one historical webpage snapshot and treated it as permanent product truth while disregarding iterative change. Marked high‑risk; excluded from repositories.

Remark: Four total conflict‑risk records were flagged in the backend system. This public case excerpt presents the two most critical items. Full‑scope findings appear in the internal comprehensive verification report.