AIMP — AI Mesh Protocol

An experimental, serverless networking protocol for resilient state synchronization between autonomous agents in fragmented, low-bandwidth networks. Built on Merkle-CRDTs and cryptographic identity — no central authority, no global DNS, always writeable.

---

Papers

#	Version	Title	Link
1	v0.1.0	Merkle-CRDT Protocol (L1/L2)	ResearchGate
2	v0.2.0	Epistemic Layer (L3)	ResearchGate
3	v0.3.0	Correlation-Aware Aggregation (L3)	ResearchGate
4	v0.4.0	Deterministic Semantic Topologies (L3)	ResearchGate

Author profiles: ResearchGate · ResearchHub

---

Protocol Stack

Layer	Version	Purpose
L1/L2	v0.1.0	Merkle-DAG CRDT, Ed25519 signing, Noise Protocol transport, BFT quorum
L3	v0.2.0	Epistemic Layer: integer log-odds, two-pass trust propagation, Sybil-resistant reputation
L3	v0.3.0	Correlation-Aware Aggregation: geometric discounting for correlated sensors/LLMs
L3	v0.4.0	Deterministic Semantic Topologies: autonomous edge generation via 256-bit SimHash

---

What's New

v0.4.0 — Deterministic Semantic Topologies

L3 v0.3.0 required applications to manually construct the knowledge graph (Supports/Contradicts edges). v0.4.0 eliminates this bottleneck with autonomous edge generation:

• Claims carry an optional QuantizedEmbedding([u64; 4]) — a 256-bit SimHash computed application-side from a canonical embedding model.
• At each epoch boundary, the protocol computes pairwise Hamming distances (XOR + popcount, ~1 ns per pair) and emits Supports edges for close pairs (d <= 30 bits) and Contradicts edges for distant pairs (d >= 200 bits).
• Edge strength scales linearly with distance in basis points (d=0 → 10000 bps, d=30 → 1000 bps).
• A max_k_nearest cap bounds edge density to O(N), preventing trust propagation explosion.
• embedding_version: u32 isolates disjoint latent spaces for protocol-level model upgrades.
• Auto-edges are materialized via L2 gossip, surviving GC via Holographic Routing.
• Dead zone (31-199 bits) orphans ambiguous claims — epistemologically correct isolation.

// Autonomous truth discovery: 10,000 claims → 50 ms scan, O(N) edges
// No floats. No coordination. No central authority.
let edges = auto_edge_generator.generate_edges(&epoch_claims);

v0.3.0 — Correlation-Aware Belief Aggregation

L3 v0.2.0 assumes all evidence sources are statistically independent (Naive Bayes). This produces pathological hyper-confidence when physically correlated sensors (e.g., 100 IoT devices on the same rooftop) or semantically correlated agents (e.g., LLMs fine-tuned on the same dataset) report concordant observations.

v0.3.0 introduces Grid-Cell Correlation Discounting:

• Each claim carries an optional CorrelationCell(u64) — a discrete coordinate for spatial, semantic, or temporal proximity.
• Within each cell, evidence is ranked by strength and geometrically discounted: the strongest source retains 100% weight; each subsequent source receives discount_bps^rank / 10000^rank (default 30%).
• With 30% discount, N correlated sensors converge to ~1.42x the evidence of a single sensor — regardless of N. The naive approach would produce Nx amplification.
• The CRDT associativity challenge (geometric decay is non-associative across partial merges) is solved architecturally: epoch reduction buckets by (temporal_grid, fingerprint, correlation_cell), guaranteeing atomic computation on the complete set.
• Claims with correlation_cell: None behave identically to v0.2.0 (zero regression).
• All arithmetic is integer-only (i32/i64, basis points). No floats. ZK-ready.

// 100 co-located sensors, 70% confidence each:
// v0.2.0 (naive):  100 × 847 = 84,700 milli-log-odds → ~100% (hyper-confident)
// v0.3.0 (30%):    847 × Σ(0.3^i) ≈ 1,207 milli-log-odds → ~77% (realistic)

---

Architecture

aimp_node/          Rust reference implementation (Cargo workspace member)
  src/
    crdt/           Merkle-DAG engine, actor model, arena allocator, quorum consensus
    crypto/         Ed25519 identity, BLAKE3 hashing, zero-trust firewall
    network/        UDP gossip, Noise Protocol XX sessions, per-peer rate limiting
    protocol/       Wire format (MessagePack), typed payload enum
    epistemic.rs    L3 Epistemic Layer (v0.3.0): log-odds, trust propagation, correlation discounting
    semantic_topology.rs  L3 Semantic Topology (v0.4.0): SimHash embeddings, auto-edge generation
    decision_engine.rs  Pluggable deterministic decision engine (trait + rule engine + hot-reload)
    error.rs        Unified AimpError type hierarchy
    dashboard/      Ratatui TUI
    config.rs       Dynamic configuration with validation
    event/          Structured logging + Prometheus metrics (counters + histograms)
  tests/            Integration tests
  benches/          Criterion benchmarks
aimp_testbed/       Python SDK (aimp-client) + CLI tool + chaos testing
deploy/             Systemd service, Firecracker microVM, install script
formal/             TLA+ convergence + quorum safety + belief convergence specification
docs/               Paper 1 (Typst source + PDF)
v0.2.0/             Paper 2: Epistemic Layer (Typst source + PDF)
v0.3.0/             Paper 3: Correlation-Aware Aggregation (Typst source + PDF)

---

Strategic Advantages

Feature	AIMP (Merkle-CRDT)	Traditional (Raft/Paxos)
Topology	P2P Mesh / Decentralized	Leader / Quorum
Availability	AP (Always Writeable)	CP (Requires Majority)
Ordering	Causal (Vector Clocks)	Total (Sequential)
Integrity	Cryptographic (Merkle-DAG)	Log-based
Hardware	Edge/IoT Optimized	Data Center Grade

---

Key Features

Core Engine (v0.1.0)

• Actor Model with zero-shared-state CRDT via tokio::mpsc
• Slab/Arena allocation with O(1) insertion and SoA layout
• Durable persistence via redb with ChaCha20Poly1305 encryption at rest
• HKDF-SHA256 key derivation with domain separation
• Cached merkle root with invalidation-on-write
• Real mark-and-sweep GC with slab memory reclamation
• Epoch-based GC tracking integrated into the CRDT actor