Expertery

Manufacturing runs on engineering tools (SOLIDWORKS and 3DEXPERIENCE alone serve 8 million users), plus every company's own scripts and in-house software, and general AI doesn't know any of them: it invents functions that don't exist, so engineers spend their time fixing "almost right" answers or don't use AI at all. Most teams try to fix this with longer prompts; Expertery asks whether an open model can learn a tool's real rules from outcomes alone, and prove it. Its training loop has one teacher: reality. Every answer the model writes is actually built in a sandbox and measured by a referee, and only answers that pass become lessons, so Meta's Muse Glimmer 30B learns from proof, not imitation. The findings shaped the design: training Glimmer on a bigger model's answers (Muse Spark Open) made it worse, while its own referee-checked answers made it an expert, so Spark Open became a tutor that gives hints, never answers; studying where the model failed before teaching it mattered more than any training trick; and a test generated by the same machine as the lessons overstated results, so the project relies on a hand-made exam locked before training and run once, with rules written in advance, confidence intervals on every comparison, and a public record of which claims did not hold. On that exam the trained model reached 92% with no manual, tying a manual-fed setup at 2.7× fewer tokens. For a company, this means an AI expert on its own tool that runs in its own cloud, keeps its designs private, costs less per answer than pasting manuals, and has to pass an exam built from the company's own past work before anyone relies on it: AI that earns trust like a new engineer. CAD is where this is proven first; the same factory applies to any tool whose output a machine can check, from machine code in factories to chip design.