Workflows Whose State Is Never Stored

A workflow engine conventionally stores where a run is: a field holds step 3, and advancing the run means assigning a new value to that field. We describe engines that do not — where the state of a run is a pure function of what data exists, recomputed on every pass, and no assignment is ever made. The pattern is not new; it is the reconciliation loop familiar from cluster controllers, applied to workflow state. What follows from it is less familiar. If state is derived from data, then every predicate defining a state rests partly on absence — this run is rolling because some device has not reported yet. Under a complete read, absence means what it says. Under a partial read of a distributed system it means something else: nothing among the devices that answered. An engine that ignores the difference stops watching the world and starts watching the network. We set out the consequence, one asymmetry that makes it manageable — incompleteness is monotone: a device that fails to answer can only make the world look emptier, never fuller — and a two-declaration scheme that follows from it.