FIELDCRAFT #010 — THE LOCAL MACHINE
Every Query You Send to a Cloud Model Becomes Someone Else’s Data. Some Investigations Cannot Survive That.
SUBJECT: LOCAL AI INFERENCE FOR SENSITIVE RESEARCH // OLLAMA, LM STUDIO, LLAMA.CPP, MODEL SELECTION, HARDWARE TIERS, AIR-GAPPED OPERATION, CONSENSUS OVERRIDE DEPLOYMENT
DATE: MAY 23, 2026
CROSS-REF: FIELDCRAFT SERIES | FIELDCRAFT #001 — THE CONSENSUS OVERRIDE | THE OPERATOR | THE PUBLICATION GAP
DATA CONFIDENCE: VERIFIED (every tool referenced is publicly accessible; every technique is reproducible)
This is the tenth and final installment of FIELDCRAFT. The briefings give you the intelligence. FIELDCRAFT gives you the tradecraft.
We are building the video side of this operation. If you want the briefings in a format you can watch, subscribe to The Sentinel Network on YouTube. New episodes drop weekly. Every FIELDCRAFT installment gets a video companion.
Why Local
Every other installment in this series has assumed you are using cloud AI. ChatGPT, Claude, Gemini, Grok, Perplexity. These are the most capable models available and they make most investigations possible. They have a cost that we have not yet discussed in detail.
Every query you send to a cloud model is logged. The query, your account, your IP address, your session history, and the response. Most providers retain that data for a defined window for “safety review” before deletion. Some retain it longer. All of it is subject to legal process. If an investigation you are running becomes legally interesting to a powerful adversary, your query history is sitting on a corporate server with a subpoena address. The data does not have to be exposed by malice. It can be exposed by court order, by breach, by acquisition, or by the routine compliance response to a federal request.
For most investigations, this risk is acceptable. The records are public, the targets are corporate, the questions are not sensitive. For some investigations, it is not acceptable. When the subject of your investigation has the resources to compel disclosure, the cloud is not your friend.
This installment teaches you how to run AI on your own hardware, where nothing leaves the machine. No queries logged. No accounts tied to your name. No subpoena exposure. No rate limits. No content restrictions imposed by the provider. Total control over what the model sees and what it produces.
The tradeoff is capability. Local models in 2026 are not as powerful as frontier cloud models. They are powerful enough for most investigative work. For the queries that genuinely require the cloud, you still have the cloud. For the queries that cannot survive cloud logging, you now have an alternative.
This is the final installment of the series. It is also the one that takes the longest to set up. Plan for an afternoon. The infrastructure you build here will outlast the rest of the toolkit.
Why Local, Specifically
Six reasons to run a model locally, in order of operational significance.
Privacy. No corporate server sees your prompts. No log records what you asked. No metadata trail connects you to the investigation. The query happens entirely on your hardware, in your network, under your control.
