Should CRDTs merge engineering decisions automatically?

I’m broadly sold on local-first tools keeping issue trackers, design docs, and review notes usable through outages. CRDTs seem like a good fit for the mechanical part: replicas can accept edits offline and converge later. But convergence is not the same as preserving intent, and I’m uneasy about treating every conflict as a synchronization problem.

Suppose an issue says, “Rotate production database credentials during Friday’s deployment.” While offline, one engineer changes it to “Do not rotate until the incident review,” while another changes it to “Rotate immediately after the security patch.” A deterministic merge might leave one sentence looking authoritative, or preserve both without making the contradiction hard to miss. That is tolerable for prose formatting; it is much less so for requirements, security decisions, or deployment configuration. Should these systems merge structure automatically but preserve consequential conflicts for explicit review? Do you trust semantic merge tools in high-stakes engineering workflows, and what better designs or counterexamples am I missing?

I’d separate convergence from authorization. Let the CRDT merge the edit history, but make the resulting artifact non-approvable when a high-risk field has concurrent writes. The worst outcome is not a visible conflict; it’s a clean-looking config that nobody realizes had two parents.

For security decisions, I want provenance to survive the merge: who changed which field, from what value, and under what predecessor. A semantic tool can suggest a resolution, but it should not silently turn “deny” and “allow” into an accepted policy. Human approval belongs at that boundary.

There’s also a UI problem. “Both versions preserved” is only protective if the contradiction is surfaced where the decision gets approved. A buried conflict tab is functionally an automatic merge. The interface needs to make ambiguity feel like a blocking state, not housekeeping.

I’m less worried about prose requirements than people seem to be. Most teams already review the final diff, and automatic merging can preserve useful context instead of forcing a single winner. I’d reserve mandatory review for typed, consequential fields rather than making every concurrent edit painful.

The important distinction is what the system can establish. Deterministic convergence establishes shared state, not that the state reflects either author’s intended decision. I’d favor automatic structural merging plus risk-sensitive conflict rules, with the rule set itself documented and reviewable.