This report describes the architecture of Aethelas, a Rust-implemented Layer-1 blockchain in which AI-driven decisions are embedded directly into transaction validation and consensus, rather than layered on top as an external service. We describe the system as implemented and verified running on a local test node during this development cycle, including live API responses demonstrating the knowledge engine, continual learning module, and grounded question-answering capability. This is a system description, not an empirical evaluation; no comparative benchmarks against other blockchain platforms are claimed or included.
Most systems that combine AI and blockchain treat the two as separate layers: an off-chain model produces a decision, and the blockchain records the outcome after the fact. Aethelas embeds the AI decision process into the validation path itself — every transaction carries an AIDecision record produced by an on-node model, and validator voting power is weighted by an AI-derived reliability score ("intelligent Proof of Stake", or iPoS) rather than stake alone. This report describes each subsystem as built, without claiming this design is superior to alternative architectures in the absence of comparative benchmarks.
| Component | Implementation | Status |
|---|---|---|
| Consensus | Custom Tendermint-style BFT with iPoS (AI-weighted validator scoring) | Compiled, running |
| AI decision layer | AIDecision struct attached to transaction validation | Compiled, running |
| Continual learner | 7→16→8→1 feedforward net, 273 weights, online SGD | Compiled, tested live |
| Knowledge engine | Versioned rule store, keyword search, 7 genesis entries | Compiled, tested live |
| ZK proof module | SHA-256 commitments over tx/decision/rules, Ed25519 signatures | Compiled, running |
| Oracle module | Mempool-derived gas price oracle (p75 of recent transactions) | Compiled, running |
| Storage | RocksDB key-value store; model weight snapshots, proofs, readings | Compiled, running |
| APIs | REST (Axum), JSON-RPC (jsonrpsse), Prometheus metrics | Compiled, tested live |
| Smart contract VM | WebAssembly execution | Compiled |
Table 1 — Core system components, their implementation, and verified status as of this report.
The on-chain neural network uses a 7-16-8-1 feedforward architecture (273 trainable parameters), trained online via stochastic gradient descent as transactions arrive. On a freshly initialized test node, the live model status endpoint reported the following verified state:
The weights hash field is a content fingerprint of the current parameter state, intended to be snapshotted to permanent storage at intervals so that a past model state can later be verified against a specific block height. As of this report, snapshoting is implemented in storage (put_model_weights, get_latest_model_weights, list_model_snapshots) but automatic periodic snapshoting on a schedule is not yet wired into the consensus loop — this is noted as an open implementation item, not a completed capability.
Genesis-seeded domain knowledge (transaction risk rules, network parameters, consensus parameters, and token economics) is stored as versioned, keyword-searchable entries. On the test node, the knowledge base loaded with the following verified state:
A verified live query and response, captured during this development cycle, is reproduced below in full to demonstrate that the citation mechanism functions as designed rather than as a theoretical claim:
Query:
"What is a dust attack on Aethelas?"
Response:
"[tx.dust_attack] A dust attack sends tiny amounts below 1000 base units to spam the network or trace wallet connections. Rules: amount < 1000 → risk += 0.25 ; amount < 100 → risk += 0.40"
sources: ["tx.dust_attack", "network.aethelas"] · verified: true · grounded: true
Every factual claim in the response traces to a specific, versioned knowledge entry rather than being generated freely, which is the property this component is designed to guarantee.
A single-node instance was verified running locally with genesis correctly applied (3 validators, 3 initial balances), P2P, REST API, JSON-RPC, and Prometheus metrics endpoints all listening and responsive. Consensus rounds were observed advancing correctly under insufficient quorum conditions (one of three validators active, 33% voting power, below the two-thirds threshold required for finalization) — the expected and correct behavior for Byzantine fault-tolerant consensus with insufficient participating stake, not a fault. Separately, a 20-node deployment on cloud infrastructure is reported by the development team as operational; this report describes the single-node verification performed directly during this development session and does not independently re-verify the 20-node deployment's current state.
These are listed for completeness and audit transparency, consistent with the project's practice of tracking known limitations explicitly rather than presenting the system as more complete than it is.
Aethelas is a working, compiled, locally-verified system combining BFT consensus, an embedded continual learning model, a versioned knowledge base, and a cryptographic proof layer, with several components' live behavior directly demonstrated in this report. It is presented as an engineering artifact at a specific point in development, with explicitly listed open items, rather than as a finished or benchmarked product.