General-purpose LLMs, despite their impressive capabilities, often falter in the high-stakes, jargon-heavy world of Model-Based Systems Engineering. Their outputs can be syntactically plausible but semantically vacuous, lacking the specific structural and relational understanding required for formal systems modeling. The path to reliable automation lies in specialization through domain-specific fine-tuning.

Consider the development of a hypothetical model, 'CurieLM.' The process begins not with scraping the web, but with meticulously curating a dataset. This involves compiling thousands of high-quality examples: legacy system specifications, validated requirement documents, interface control documents (ICDs), and corresponding, human-verified SysML models. The objective is to create a corpus that reflects the unique linguistic patterns and logical structures of your engineering domain.

The core of this process is instruction tuning, a supervised fine-tuning technique that explicitly teaches the model to map unstructured or semi-structured inputs to a precise, structured output format. The goal isn't just to paraphrase, but to translate. The model learns to parse a sentence like "The telemetry subsystem shall transmit health and status data every 500 milliseconds" and restructure it into a formal requirement with properties like ID, text, type (performance), and verification method.

While fine-tuning embeds implicit knowledge of a domain's structure, Retrieval-Augmented Generation (RAG) provides explicit, verifiable context at runtime. This is indispensable for ensuring compliance with dynamic standards like the INCOSE Systems Engineering Handbook or ISO/IEC/IEEE 15288. Fine-tuning on a standard is a snapshot in time; RAG allows the model to reference the latest version continuously.

In this architecture, the retrieval corpus is a vector database populated with the parsed, chunked text of all relevant standards, design patterns, and organizational best-practice documents. When an engineer provides a piece of unstructured text, the system first retrieves the most relevant sections of these standards. The original text, along with this retrieved context, is then fed to the fine-tuned LLM. The prompt effectively becomes: "Given these rules from ISO 15288, reformulate the following text as a compliant system requirement."

Extracting and structuring requirements is only the first step. For complex relationships and nested logic, more advanced prompting techniques are necessary. Chain-of-Thought (CoT) prompting guides the model to "show its work," forcing it to externalize its reasoning process. Instead of asking directly for a SysML diagram, you prompt it to first identify entities, then their relationships, then their attributes, and finally to assemble the SysML representation. This stepwise process drastically reduces errors in complex scenarios.

Ultimately, every AI-generated requirement must be considered a recommendation, not a final artifact. This is where the human-in-the-loop and confidence filtering become critical, especially in safety-critical systems. The model should not only output the structured requirement but also a confidence score. This score is not a measure of factual correctness, but rather a proxy for the model's internal statistical certainty in its generation.

A robust workflow integrates these scores directly into the MBSE tool. Requirements generated with a confidence score below a certain threshold (e.g., 95%) are automatically flagged for mandatory human review. Those above the threshold might be provisionally accepted but are still subject to standard engineering review cycles. This approach balances automation's speed with the non-negotiable need for expert oversight, filtering out low-quality or hallucinated outputs before they can corrupt the system model.