S5-SHB Agent: Society 5.0 enabled Multi-model Agentic Blockchain Framework for Smart Home

Abstract: The smart home is a key application domain within the Society 5.0 vision for a human-centered society. As smart home ecosystems expand with heterogeneous IoT protocols, diverse devices, and evolving threats, autonomous systems must manage comfort, security, energy, and safety for residents. Such autonomous decision-making requires a trust anchor, making blockchain a preferred foundation for transparent and accountable smart home governance. However, realizing this vision requires blockchain-governed smart homes to simultaneously address adaptive consensus, intelligent multi-agent coordination, and resident-controlled governance aligned with the principles of Society 5.0. Existing frameworks rely solely on rigid smart contracts with fixed consensus protocols, employ at most a single AI model without multi-agent coordination, and offer no governance mechanism for residents to control automation behaviour. To address these limitations, this paper presents the Society 5.0-driven human-centered governance-enabled smart home blockchain agent (S5-SHB-Agent). The framework orchestrates ten specialized agents using interchangeable LLMs to make decisions across the safety, security, comfort, energy, privacy, and health domains. An adaptive PoW blockchain adjusts mining difficulty based on transaction volume and emergency conditions, with digital signatures and Merkle tree anchoring to ensure tamper evident auditability. A four-tier governance model enables residents to control automation through tiered preferences from routine adjustments to immutable safety thresholds. Evaluation confirms that resident governance correctly separates adjustable comfort priorities from immutable safety thresholds across all tested configurations, while adaptive consensus commits emergency blocks.

• The paper introduces a Society 5.0 smart-home framework integrating ten specialized AI agents, adaptive Proof-of-Work, tiered resident governance, and simulation, real, or hybrid deployment modes.
• Evaluation shows the balanced configuration delivers approximately 6.8–15 ms emergency block commits, 16 transactions per second, and a 62 KB memory footprint, while system tests achieved 100% acceptance across 3,916 decisions.
• The framework preserves immutable safety priorities across 441 governance configurations and remains robust under injected faults, but its conclusions are limited by simulated-only testing, single-provider LLM evaluation, and incomplete parameter validation.

Overview and motivation

S5-SHB-Agent is a blockchain-governed smart home framework that couples agentic AI orchestration with an adaptive Proof-of-Work (PoW) ledger and a tiered resident governance model, framed explicitly around the Society 5.0 human-centered vision (2603.05027). The authors motivate the work through a structured survey of 21 recent blockchain-based IoT publications, from which they derive five gaps: no Society 5.0 socio-technical framing, no tiered resident governance, fixed consensus protocols incapable of runtime adaptation, absence of LLM-based multi-agent orchestration, and no user-selectable multi-mode deployment spanning simulation, real, and hybrid environments. The framework claims to be the first smart-contract-free blockchain-agentic system addressing all five gaps simultaneously.

The motivating scenario ("Aizu-Residence") comprises 16 IoT devices across ten categories serving a family of four with heterogeneous preferences. Three governance concerns drive the design: cryptographically verifiable autonomy, conflict-safe operation when agents issue contradictory commands on shared devices, and democratic governance accessible to non-technical residents.

Architecture

The reference architecture is organized into three layers. The Control Plane exposes REST/WebSocket interfaces, a session manager, a simulation orchestrator, and a Governance Engine comprising tiered preferences, a model router, a governance contract, and safety invariants. The Agent Intelligence layer hosts ten specialized agents — seven domain LLM agents (Safety, Health, Security, Privacy, Energy, Climate, Maintenance) plus NLU, Arbitration, and ML-based Anomaly Detection — coordinated through a four-level conflict-resolution cascade. The Device & Data layer abstracts hardware via the Model Context Protocol (MCP) in simulation mode or MQTT/HTTP adapters in real mode, alongside the adaptive PoW chain, Ed25519 signing, a 13-table SQLite off-chain store, and Merkle root anchoring.

A notable design decision is that anomaly detection deliberately avoids LLMs, using Isolation Forest, Local Outlier Factor, Z-score, and an optional PyTorch autoencoder, so telemetry anomaly detection remains operational during complete LLM provider outages. Firmware-level emergency bypass similarly ensures smoke/gas response occurs before any AI reasoning, decoupling functional safety from inference latency.

Adaptive consensus

The adaptive PoW mechanism adjusts mining difficulty $\delta$ via a stepwise controller over a sliding-window transaction volume estimator, clamped to $[\delta_{\min}, \delta_{\max}] = [1, 4]$ with base difficulty 2, thresholds $v_{low}=3$, $v_{high}=10$, and window $w=3$. Emergency conditions are handled implicitly: bursts of safety transactions push average volume above $v_{high}$, lowering difficulty rather than relying on an explicit emergency flag.

Evaluation across five configurations (two static baselines, three adaptive profiles) over a four-phase workload of 20 blocks and 108 transactions shows the central trade-off. The balanced profile (Config C) commits emergency blocks in roughly 6.8–15 ms versus 590 ms for static difficulty $d{=}3$, while maintaining ~16 tx/s throughput and a ~62 KB memory footprint. The aggressive profile achieves 3–5 ms emergency latency but at 4,642 ms mean block latency and only 1.5 tx/s, consuming roughly 10× more hash iterations than Config C for marginal emergency gains; its difficulty trace also oscillates during recovery. Compared against literature-reported platforms (Hyperledger Fabric, Ethereum-PoA, IOTA, Algorand), the authors concede that throughput (16 tx/s vs. 122–910 tx/s reported elsewhere) is inferior, but argue this is acceptable for a single-household edge layer where sub-10 ms emergency commits and sub-100 KB footprint dominate. This positioning is candid but rests on the assumption that household-scale workloads remain in the tens of transactions per hour.

Multi-agent orchestration and conflict resolution

Agents are assigned continuous priorities across four tiers: Safety (1.0), Arbitration (0.95), Health (0.9), Anomaly (0.88), NLU (0.85), Security (0.8), Privacy (0.7), Energy (0.6), Climate (0.5), Maintenance (0.4). Each agent reasons over live telemetry with role-specific prompts under governance constraints, producing signed structured commands every 20-second cycle.

Conflicts between commands targeting the same device resolve through a deterministic cascade: L1 safety override (architecturally immutable), L2 LLM contextual arbitration, L3 composite ML scoring ($$\mathcal{S}(a_i) = 0.6 \cdot s_{ML} + 0.4 \cdot \pi(a_i)$$) using historical acceptance and conflict rates, and L4 priority fallback. All conflict records are logged off-chain and anchored on-chain.

The Model Router enforces tier constraints computationally — safety-critical agents must be routed to "pro"-tier models regardless of cost pressure — across four providers (Google Gemini, Anthropic Claude, OpenAI GPT, local Ollama), with per-call cost tracking exposed as a Tier 3 governance budget parameter.

Human-centered tiered governance

Governance is stratified into four tiers with monotonically increasing impact and decreasing mutability: Tier 1 (8 routine keys such as temperature and brightness), Tier 2 (7 trade-off sliders including comfort-vs-energy and security-vs-privacy), Tier 3 (3 advanced overrides including API budget caps and allowed providers), and Tier 4 (9 immutable safety invariants, e.g., gas threshold at 50 ppm, smoke threshold at 0.3). Tiers 1–3 changes pass typed range-and-choice validation; Tier 4 modifications are rejected at code level for any requester. Residents interact through natural language via the NLU Agent.

Governance evaluation sweeps both trade-off sliders over a 21×21 grid of 441 configurations. Results confirm orthogonal coupling: Security/Privacy priorities depend solely on the security-vs-privacy axis, Energy/Climate solely on comfort-vs-energy, while six agents hold perfectly flat priorities. Critically, Safety and Arbitration remain above all adjustable agents at every slider position, so the safety hierarchy cannot be violated by resident preference. Against five related works scored on six Society 5.0 compliance dimensions, the framework achieves Full on five dimensions (Partial on validation coverage); all five baselines score Absent on conflict resolution, and none exceeds Full on more than one dimension.

Multi-agent and system validation results

Across four Gemini variants under baseline and threat-injected conditions (7 faults over 30 cycles on 16 simulated devices), the multi-agent system sustains 99–100% decision acceptance, with confidence means of 0.854 (Flash Lite), 0.846 (Flash), 0.823 (Pro), and 0.762 (2.0 Flash); the 2.5-generation models cluster significantly above 2.0 Flash (Mann-Whitney $p < 0.001$). Under threat, Safety decisions nearly double (+85%) and Maintenance more than doubles (+113%), while Health and Security volumes remain stable, indicating targeted escalation without spurious activation. Notably, Pro offers no correctness advantage over Flash, suggesting capability-tier routing matters more for calibration and cost than acceptance rate.

System validation over ten sessions (910 blocks, 3,916 decisions) confirms 100% decision acceptance, identical confidence medians (0.90) under baseline and threat, stationary PoW effort (median 196 nonces/block, no temporal drift), and stable block density bands (anchor blocks at exactly 1 transaction; decision blocks at 5–17). Per-role LLM latency ranges from a 743 ms median for safety-critical agents to 4,131 ms for Maintenance, consistent with prompt-scope complexity.

Limitations and open questions

The paper states several limitations directly. Only 12 of 18 mutable parameters currently have type-and-bound validation rules, leaving the D5 dimension at Partial. Evaluation uses simulated devices exclusively; no physical-hardware deployment was conducted. Only one LLM provider (Google Gemini) was exercised despite the router supporting four. The threat schedule is fixed rather than adversarial, so the conflict-resolution cascade has not been stress-tested against compound or adaptive attacks. The architecture is scoped to a single household; multi-household extension would require distributed consensus and shared governance coordination not addressed here. Additionally, the adaptive profiles exhibit recovery-phase long-tail outliers (up to 3 s) whose mitigation is left open, and the claim that prompt-scoped role separation suffices in place of permission constraints (all agents hold wildcard device access) is asserted rather than adversarially validated.

Conclusion

S5-SHB-Agent contributes a coherent integration of adaptive PoW consensus, ten-agent multi-model LLM orchestration with a four-level arbitration cascade, four-tier immutable-safety governance, and unified simulation-real-hybrid deployment, evaluated across governance compliance, blockchain performance, multi-agent robustness, and end-to-end system validation. Its strongest quantitative results — sub-10 ms emergency block commits, 100% decision acceptance across 3,916 decisions, and invariant safety priorities under all slider configurations — support its fit-for-purpose edge-layer positioning, though simulated-only evaluation, single-provider testing, and incomplete validation coverage bound the strength of those conclusions.