Communication Mesh Research

note

Research draft v0.3, 26 September 2026 (one merged series of nine papers, numbered 0 to 8). This is a proposed research programme, not a live-network claim. SPECTRE's repository holds a routing skeleton. Skyzai has no production consensus network, messenger, VPN, CDN, or inference cloud.

Skyzai is exploring a shared network of independently operated nodes that could carry messages and, later, provide content delivery and model inference. The long-range design includes people, AI models, agents, and eventually robots as network participants. Each role has its own permissions and responsibilities.

The idea

The network question is how these participants communicate across four shapes:

Shape	Working name	Example job
One to one	FLOW	A direct one-to-one conversation, whose privacy properties are defined for each delivery mode; payment settlement is a separate capability with its own trust boundary
One to many	RELAY	A public or permissioned broadcast, milestone, or feed
Many to one	AXIOM	Rank many candidates for one recipient, or reduce many inputs to one attributable result; never consensus or truth
Many to many	MESH	Nodes find and route across useful peer paths; many publishers' public posts are an application style here, while one publisher's broadcast belongs to RELAY

SPECTRE is the proposed physical node network and transport commons: machines, links, peer discovery, and packet transport. MESH is its many-to-many routing primitive. FLOW, RELAY, and AXIOM are communication patterns carried over those paths, and Skyzai composes user-facing services over the commons. Some of the same machines may also run ledger consensus, but the protocols keep their guarantees separate. Hashgraph-style ledger consensus can keep random-peer gossip as its baseline. MESH, separately, learns from local application-delivery outcomes.

Consensus and routing are different jobs

The 2016 Hashgraph research paper describes nodes gossiping with randomly chosen peers and deriving virtual votes from a recorded history of gossip. Its model allows up to just under one-third of members to be attackers who may delay or delete messages, as long as a member that keeps resending to another member eventually gets one message through. That is a retransmission assumption, not a promise that any single message arrives. The base bound counts members; the paper also gives a stake-weighted version. An open network has to say which one it uses, and how new members are admitted, before the one-third bound means anything.

MESH studies a different question: how packets find efficient paths through a node network. The first research design keeps consensus voting and finality untouched while MESH routes application traffic. Any later change to consensus peer selection needs a separate liveness and safety analysis; a route score must not change voting weight or event validity.

Sharing a machine still shares its CPU, bandwidth, keys and failure risk: two processes on one machine are one fault domain. The first experiment therefore also floods application traffic and measures what happens to consensus. The first SPECTRE operators could come from an initial ledger-node fleet, but routing needs its own admission, and nodes outside the ledger may route too.

Start with the messenger

The first user-facing experiment is a voice-first messenger for people and AI systems. When a person talks to APU, APU may automatically save useful intent as a structured note, task, or reminder; the intent persists while the raw message, audio, transcript, and model output do not. Each saved change records whether the person asked for it or APU inferred it, and each change can be undone; merges and deletes are made undoable or confirmed first.

Those are different data lifecycles, and “ephemeral” is several properties, not one: relay retention, device copies, backups, recipients' own copies, and APU processing each need their own statement. Encryption between two people keeps their message from relays and transit nodes. APU is not a party to a conversation between two people unless someone addresses APU or explicitly asks it to act on a message. When a person does talk to APU, APU's endpoint and its model provider see that turn in plain text in order to extract intent. The saved note outlives the message, so it needs its own confidentiality, retention, and deletion terms. A relay can retain ciphertext for a declared asynchronous-delivery window, or offer live-only delivery that may miss offline recipients. Deleting a relay copy cannot erase screenshots, local copies, or backups held by recipients.

For direct one-to-one messages, the research recommends evaluating established end-to-end encryption such as Signal's PQXDH and Double Ratchet, with Sesame session management for asynchronous multi-device delivery. For encrypted groups, evaluate MLS. Nostr NIP-17 can be considered for interoperability; it is marked draft/optional, and NIP-44 does not provide forward secrecy or post-compromise security. NIP-17's “disappearing messages” are relay expiry, not forward secrecy. A stored voice note and a live push-to-talk stream need different protection, and no NIP specifies live voice. These are research candidates, not a final protocol selection.

The existing Communicator plan uses Nostr as a bootstrap until SPECTRE runs. A messenger on that path does not show that MESH routing works, so the network is measured separately, with the same messaging workload and privacy terms.

Two sequences to test

The user product path is voice-first messenger → broadcasts and milestones / CDN → model inference. The node-operator path is separate: consensus participation → optional VPN / CDN → inference hosting. Each step needs measured delivery, cost, failure recovery, privacy, and demand. The operator path orders the research; it is not a path every node must follow, and an operator may qualify for one role without the others.

A node does not become a VPN automatically. Operators would need to opt into third-party egress and disclose exit policies, logging, abuse handling, and service terms. A service receipt or token reward would not establish that the service was useful or profitable.

Research paper sequence

Each paper keeps a stable number and states a baseline and a result that would stop or revise its hypothesis.

• Paper 0 — Skyzai Communication Mesh — Architecture Draft: shared nodes, separate protocol planes, privacy, the series threat model, and research gates.
• Paper 1 — Hashgraph consensus and MESH routing on SPECTRE: whether and under what assumptions adaptive paths can carry consensus gossip.
• Paper 2 — Local learning for MESH: energy-based route scoring, Hebbian-inspired updates, Q-routing, exploration, poisoning, eclipse, and resilience.
• Paper 3 — FLOW one-to-one messaging and APU intent capture: end-to-end encryption, identity verification, multi-device sessions, recovery, expiry, metadata, voice, what APU may see, and deletion limits.
• Paper 4 — RELAY broadcasts and CDN: public events, permissioned fanout, milestones, retention, caches, and content availability.
• Paper 5 — AXIOM selection and aggregation: recipient-specific ranking and many-to-one reducers with provenance, uncertainty, quorum, user control, and correction.
• Paper 6 — MESH and SPECTRE at network scale: multi-hop delivery, partition recovery, fallback, and operator policy.
• Paper 7 — Node economics and inference: optional VPN/CDN services, open-weight or licensed model variants selected through a separate model-access plane, world models, robotics workloads, operating cost, and service verification.
• Paper 8 — Node runtime and interoperability: Erlang/OTP compared with Rust and Go, a language-neutral wire format, safe public ingress, NAT traversal, mobile sleep, and upgrades.

Papers 1–3 have been through two review rounds. Papers 4–8 are first drafts awaiting their first review round.

The order matters: agree on the system boundaries and threat model first, then establish baselines, then evaluate each user-visible capability. Simulated improvements remain simulation results until independently reproduced on a described testbed.

First research questions

• Can application packets use MESH while Hashgraph consensus remains unchanged?
• What does “ephemeral” promise at the relay, device, backup, and APU-processing boundaries?
• When, and only with whose visible consent, may APU process a conversation between two people?
• Which direct and group encryption protocols fit offline delivery, multi-device use, and device loss?
• Can local node observations improve delivery without exposing who talks to whom or causing route concentration?
• How should an N:1 system rank messages, models, or services for one user, and how does that differ from aggregating inputs into one result?
• What should operators be paid for, and how do measured cost and useful service relate to rewards?
• Which runtime and wire format let independently written nodes interoperate and accept public peers safely?

Research references

• Baird, The Swirlds Hashgraph Consensus Algorithm (2016).
• Boyan and Littman, Packet Routing in Dynamically Changing Networks: A Reinforcement Learning Approach (Q-routing, NeurIPS 1993).
• Signal: PQXDH, Double Ratchet, Sesame.
• IETF, Messaging Layer Security, RFC 9420.
• Nostr protocol: NIP-04, NIP-17, NIP-44, NIP-59.
• Reynolds, Flocks, Herds, and Schools (ACM SIGGRAPH 1987) — an analogy for local interactions, not a networking result.