The costly mistake is usually not a bad meter but a flow computer running on a stale gas composition. Their line is that an unvalidated model output should never change a billed quantity.
An agent skill and MCP toolkit that builds native SLDPRT and SLDDRW files, exports a PDF and rereads dimensions. Its stated rule is that a tool call succeeding is not the same as the task being done. The README is mostly Chinese.
For mechanical engineers: a worked path from a plain request to a checked part and drawing, with evidence kept, which matters more than speed when the part goes to a shop.
Open-weight models run on site so plant data stays local. In the author's own tests on a benchmark built from IBM AssetOpsBench data, a 4B model passed 142 of 143 tasks on a 4-CPU machine, and an approval gate stopped every injected write in its attack tests, though planted false values still got through.
For maintenance and reliability teams: evidence that a cheap local model can handle work order questions, and that a human approval step is what protects the records.
The post argues that the same "what if" models once limited to companies like Boeing are opening up so a small shop can stress a design, and a student can explore a turbopump, before metal is cut.