Variable DECISION_DISCOVERY_APPLIES_INSTRUCTIONSConst
DECISION_DISCOVERY_APPLIES_INSTRUCTIONS: "This request asks for work that a specialist agent should do, rather than something the conversation manager should answer directly or plan as a multi-agent workflow." = 'This request asks for work that a specialist agent should do, rather than something the conversation manager should answer directly or plan as a multi-agent workflow.'
The Likelihood's instructions. They stand on the request alone, because a driver may answer each question separately and never see the Choice's options.
Known limitation. Because they name no agent, the Likelihood measures "is this specialist work?", not "does one of these agents fit?". A per-question driver, such as the Default Decision prompt's top-priority binding, answers each question on its own, so it never sees the options while answering this one. The Choice cannot say "none" either: its probabilities sum to 1 over the options it is given. So when the options leave out the agent the request needs (a host allow-list of a Billing and a Marketing agent, and a request for a legal review), both answers can still clear the threshold and suggest an agent that does not fit. The suggestion stays advisory, and the agent's prompt keeps the "unless the request clearly needs something else" escape.
Why the wording stays. The discovery eval (#4892) measured, and the calibration (#4893) calibrated, exactly this question: its per-model Platt parameters and threshold hold only for this wording. Rewording it invalidates both, so the fix (quoting the options, or one Likelihood per option) waits for a re-measurement, tracked in #4938. Meanwhile DECISION_DISCOVERY_MIN_OPTIONS closes the worst case, two options, and leaves the measured case unchanged.