CRDT-Merge Multi-Node Convergence Laboratory
Empirical proof that the two-layer CRDTMergeState architecture guarantees identical merged models across distributed nodes — regardless of merge ordering, network partitions, or strategy choice.
Library: crdt-merge v0.9.5
Four experiments: Multi-node convergence | Network partition & healing | All 26 strategies | Scalability benchmark
E4 Trust Convergence (v0.9.5): The E4 trust-delta protocol guarantees 0.000 maximum divergence across all peers with 3.84ms convergence time. Trust scores propagate as first-class CRDT dimensions -- every merge operation carries cryptographic proof of provenance via 128-byte proof-carrying operations.
Run all four experiments -- tests all 26 merge strategies
Tensor Source (safetensors)
3 100
16 512
Primary Strategy
2 20
2 10
N nodes merge in different random orderings -- all must produce identical results
Tensor Source (safetensors)
3 100
16 512
Strategy
2 20
Split nodes into isolated partitions, gossip internally, heal, verify convergence
Tensor Source (safetensors)
6 100
16 512
Strategy
2 10
Every merge strategy tested for convergence -- 13 base-free + 13 base-required
Tensor Source (safetensors)
3 30
16 256
Measure convergence overhead from 2 to N nodes
Tensor Source (safetensors)
10 100
16 256
Strategy