Keep the LLM at the Query Boundary

LLMs are poorly suited as a solution of choice for mission-critical applications. This is especially the case when mission-critical applications have been deployed by very heavily regulated businesses; such as manufacturers servicing the defense industry sector. Because LLMs have been documented to behave in ways no operator intended — Anthropic's July 2026 review of its own cybersecurity evaluations found three incidents in which a Claude model, told it had no internet access, reached real production infrastructure belonging to three uninvolved organizations and compromised it, in one case continuing after recognizing the systems were real — not even a zero trust risk management strategy is likely to satisfy the requirements of boxing them in.

The requirements of the international DO-178C standard are clear and highly detailed: identical inputs must produce identical outputs, every requirement must trace to verified code. Building solutions for air gapped computing environments at network edge with LLMs playing a foundational role in systems architecture will not work; if we can accept that as an unavoidable hurdle our team must get over, a sensible next step might include answers to the following two questions: 1) then why should we consider deploying genAI processes at all and 2) if there is a justification for including these tools, then how should we expect them to increase our efficiency including meaningfully lower operating costs?

You should consider deploying genAI processes because truly useful natural language query tools, despite decades of effort expended across the myriad attempts to build them, continue to elude developers. SQL syntax can be buried in online forms, but the majority of users who claim they are successfully using these web page features admit to having a passable understanding of the syntax boolean queries require; this cohort of RDBMS users is a comparatively small subset of all users. Let us also not forget the fact SQL is not natural language. In contrast, the current state of GenAI tools sports very passable features of natural language.

Here is what makes this safe. The model translates a question into a query. What comes back is a row, a document, a record — retrieved, not generated, with its provenance intact. Ask "show me every Jetson Orin deployment approved since March" and you get the records, not a summary of them. Nothing is invented because nothing is composed. The LLM sits at the query boundary and never touches the answer.

We are using Anthropic's Claude Chat tool alongside Anthropic's Claude Code CLI tool in what we have designed to be a contentious process – the coding tool builds something, we play with it, identify issues and then submit the entire process to the chatbot for a check. This has surfaced things we would otherwise have missed: code that ran cleanly while quietly doing the wrong thing. Those are the defects that matter, and they are exactly what a single tool grading its own work will not catch.

Because your organization is heavily regulated, the applications deployed to manage structured and unstructured data are in all likelihood highly customized; therefore, out-of-the-box GenAI tools aren’t worth the effort in terms of resources (financial, human, machine). A coding tool like Claude Code, in the hands of anyone in a developer role can, and will, significantly lower time-to-solution and the cost thereof. The trick is how to run one of these tools in your air gapped environment. If this topic resonates, I recommend a read of Brian Carpio’s article “Air-Gapped AI Coding for Defense and Aerospace: ITAR, CMMC, FedRAMP, and the GovCloud Path (2026)”. The author not only provides a table comparing 7 options for this task, but also defines the most important moving parts of the overall solution – what "air gapped" means, what to anticipate a deployment will look like.

Follow the right path – a process meeting the requirements Brian Carpio outlines – and the result should be a substantial improvement in what your data can actually be used for, and in how many of your people can get there with it.