Why an infrastructure artificial intelligence must not decide
A language model generates plausible text. It does not solve a state equation. That distinction determines whether a value shown to an operator is computed or invented.
The orchestration architecture
Identify what needs computing, call the tool, validate that the result falls inside the admissible physical domain, then explain. The model never produces the figure: it produces the question and the account of the answer.
Why this separation makes the system auditable
Every recommendation traces back to the equation that produced it and to the measurements that fed it. A monolithic architecture, where the same component interprets and computes, does not allow that traceability: you cannot replay a generation.
The decision stays human
This is not a precautionary formula. It is the only tenable position once software is a product and liability is assessed without fault.
