crdt-merge — Federation

Gossip Convergence Simulation

Each node starts with its own weight tensor. Nodes exchange CRDTMergeState via merge() following the selected topology. Convergence is measured as average pairwise L2 distance. When all nodes reach the same state_hash, the system has converged — no coordinator, no locking, no message ordering required.

This simulates real-world federated model merging where:

• Nodes are geographically distributed (edge devices, data centers)
• Network connectivity is unreliable (partitions, late joiners)
• There is no central server dictating merge order

New in v0.9.5 -- E4 Trust-Delta: Federation now carries trust metadata via the Symbiotic Lattice Trust (SLT) protocol. Every gossip exchange propagates typed trust scores (accuracy, consistency, recency, provenance) as first-class CRDT dimensions. Byzantine peers are detected and isolated with 34% fault tolerance and zero coordinator overhead. The convergence chart now includes a secondary y-axis showing real-time trust evolution for each node.

OR-Set State Trace

Traces the internal provenance() data at each gossip step. Shows wire protocol payload sizes, tombstone counts, and round-trip serialization proof.

Round-trip proof: from_dict(to_dict(state)).state_hash == state.state_hash must hold for all nodes.

How to Read the Results

• Wire Protocol Chart: Shows the serialized state size (bytes) for each node over gossip rounds. As nodes absorb more contributions via merge(), their wire payload grows. After convergence, all nodes have the same payload size (they hold the same set of contributions).
• Provenance Trace Table: Each row shows a contributor inside a node's state at a given round. merkle_hash is the content-addressed hash of that contribution's tensor. weight shows how much influence this contribution has in the final resolve. As gossip proceeds, nodes accumulate contributions from all peers.
• Round-trip Serialization Proof: For each node, to_dict() serializes the state to JSON, then from_dict() reconstructs it. PASS means the reconstructed state has an identical state_hash — serialization is lossless. This proves states can be safely transmitted over the wire without data corruption.
• Tombstone Count: OR-Set remove semantics. When a model contribution is removed (e.g., a node retracts its weights), a tombstone is added. Tombstones ensure the removal is propagated to all replicas — even those that haven't seen the remove yet. High tombstone counts may indicate excessive churn.

New in v0.9.5 -- E4 Trust-Delta: State mutations are now wrapped in CausalTrustClock operations providing Lamport-style causal ordering. Each round's state is tracked via TrustBoundMerkle for tamper-evident provenance. See the E4 section below the tombstone summary for full causal clock trace and Merkle root evolution.