Action
The feasible action set 𝒳(s) as an OBJECT rather than as scattered members of the decision context (§4.4.6.5).
An earlier form of §4.4.6 put feasibleBounds, isFeasible and violations directly on DecisionContext and left feasibleActions unbuilt. That correctly identified that the set was missing and then failed to name it, so every rule that wanted to search 𝒳(s) rebuilt the loop, the filter and the argmin for itself. Naming it makes those reusable and testable without a simulation.
All members are pure and epoch-scoped: an instance is valid only during the action call that received the context carrying it.
Properties
Accepted draws over attempted draws, since this set was created. NaN before any.
How many levers an action vector has.
How many actions the set contains, or null when that is not a useful question — any CONTINUOUS lever, or a discrete product beyond ENUMERATION_CEILING.
Functions
Every action in the set. Throws when size is null, because a rule that reaches here without checking has made an error the library should not paper over.
Bounds in force for this lever at this epoch: envelope ∩ narrowed ∩ 𝒳(s) (§4.3.3).
Membership. The same predicate the element applies when the action arrives.
Up to count feasible actions drawn at random — the only way to search a set that asSequence cannot walk, which on a model with several state-dependent levers is the usual case rather than the exception (§8.2.10: 97% of epochs).
Why not. Empty exactly when contains is true.