Retrospective coverage for the October 7, 2026 issue.
A refund router can make a consequential mistake without misreading a customer: its software can treat a declined question as “no refund requested.” The distinction matters when structured model output becomes an input to another program. A refusal is a different kind of result, not another point on a yes/no scale.
As of the October 7, 2026 issue date, the pinned llm-openai-decisions README documents an LLM plugin for OpenAI’s Decisions API, illustrated with openai-decisions/gpt-6-luna. It exposes predicates, choices and ordered scores. Crucially, the README says that when the API declines a question, the plugin preserves its answer as {"name":"...","type":"refusal"}. That gives an application an explicit branch to handle rather than requiring it to infer refusal from prose.
Consider a hypothetical support inbox that asks two named questions about each message: whether it requests a refund, and which department should receive it. The plugin documents a multiple-question mode returning answers in question order. Application code could inspect each answer’s type before reading its probability or selected choice. If the refund question returns a refusal, the workflow could preserve that unresolved status for review rather than writing a negative refund flag. This is a proposed design, not a demonstrated customer result.
An “other” department does not solve the same problem. In the README’s category example, “other” is an offered choice; refusal is a separate answer type. Collapsing them would erase whether the system selected a business category or declined the question. Preserving that distinction could let a support team keep useful routing information without silently closing an unresolved request. The snapshot does not establish refusal frequency, causes, or the accuracy of the remaining answers.
The opportunity is therefore narrower than automatic case resolution: build a decision interface that can leave one question unanswered without converting that absence into a business judgment. Structured output makes this branch straightforward to represent; the product still has to decide what work waits there.
