"Your data stays inside." Draw the arrows before you believe it.
A data-flow test anyone can run in ten minutes, six arrows that can cross the boundary without anybody deciding they should, and the precise version of the claim that is worth making.
Articles
One question per article, answered with dated facts, named organisations and primary documents. Every source carries the day it was read.
38 articles
A data-flow test anyone can run in ten minutes, six arrows that can cross the boundary without anybody deciding they should, and the precise version of the claim that is worth making.
Concurrent users are not employees. Capacity is decided by the model, the context length, the KV cache and the serving engine, and the vLLM project found that inference systems before its engine wasted 60 to 80 per cent of memory, against under 4 per cent with it.
Residency is not privacy, and neither is compliance. What DORA Article 30 demands in writing, what a documented cloud service says about where processing happens, and why the answer is easier to give when the processing never leaves.
A date-by-date reading of what applies to an ordinary company today, what was extended by the AI Omnibus, and why moving a model onto your own server changes none of the classification.
Language models do not separate instructions from content, so any text the model reads can behave like a command. What EchoLeak showed, why retrieval makes it worse, and why running the model locally does not fix it.
The six stages of the pipeline and what each adds to the security boundary. Why document-level permissions have to be enforced at retrieval time, why the index is a second copy of your data, and the seven questions to answer before a RAG system goes live.
The ten risks in plain business language, which of them an on-premise deployment touches, which it does not, and the minimum an organisation of thirty people should do about them.
A firewall is layer one of seven. Identity, authorisation, data, application, model and operations are the other six, and an on-premise deployment only improves one and a half of them.
Ten sections, each of which answers a question somebody in the company will ask this quarter. What the law requires, what the survey evidence says about the behaviour you are governing, and why a ban is the weakest option available.
What the Article says, what the Regulation's own recital says it is for, why the duty is role-based rather than uniform, and a five-module programme a small company can run without a consultant.
The classification rules in the Regulation's own words, the four conditions that can take an Annex III system back out of the category, the current dates, and why moving the server changes nothing.
What Article 28 puts on the financial entity, what Article 30 requires in the contract, which questions an AI supplier has to answer in writing, and where a local appliance changes the dependency without removing the duty.