Skip to main content
Coming soon. Not in this release. This page says what the feature is, so that the shape is public before the code is.
A run already asks for numbers in three places. A branch reads a route and a score to choose a path. A confidence gate decides whether a result is good enough to pass. An advisor is consulted part-way through and its answer changes what happens next. Today each of those is answered by a model that writes prose, and the runtime reads a number back out of the words. That works, and it costs a full turn of inference every time. A decision model answers directly: one choice from a set declared in advance, with a confidence attached. The same three places, the same contract, without the round trip through prose. See Conditions and gates.