Limitations
===========

* scCS is supervised and does not discover topology or terminal states.
* Results depend on RNA-velocity model quality and transition-graph construction.
* Ordering choice affects anchors, display placement, and signed progression.
* Effective horizon is a graph scale, not a calibrated physical time unit.
* Endpoint anchors are annotation- and ordering-defined and may not be natural sinks.
* Explicit competing outcomes require biological justification.
* Hard dominant-fate labels are unstable for low-specificity cells.
* Future-fate mode does not provide a scientific velocity vector in star space.
* Instantaneous and future-fate modes answer different questions.
* Condition inference requires independent biological replicates.
* Gene associations are hypothesis-generating and not causal evidence.
* Very large future-fate solves remain more demanding than the instantaneous transform.