Convergence is not enough

The strongest discussion signal was not “CRDTs are bad.” It was that serious systems split operations into classes. Pure field updates can often use simple last-writer-wins or similar merge rules. Topology-changing or permission-sensitive actions need to be revalidated against current state during replay, and sometimes rejected even if they were valid when created offline. One concrete example came from a team building a multiplayer docs and planning tool with tree and graph structure. They keep guarded operations in the log, replay them deterministically, and reject ones that would create cycles or violate permissions. That landed as the practical model: convergence for safe ops, constraint checks for dangerous ones, and explicit refusal when the data model would otherwise be corrupted. From there, the conversation sharpened the boundary conditions. People pointed to research such as BloomL and to systems like Coln and Pijul that try to encode constraints or semantic merges rather than rely on raw operation replay. Others pushed back on using CRDTs for domains like MMO inventories or financial transfers, where correctness depends on a single clear order of operations. A side complaint was that the article buried its point under repetitive prose, but nobody really disputed the core claim. The takeaway was straightforward: sync technology solves replica agreement, not application correctness, and teams still need domain-specific merge policy.