TRUST: A Distributed Trustworthy AI Service Framework for High-Value Scenarios

TRUST is a decentralized AI verification framework designed to solve the four major problems of robustness, scalability, transparency, and privacy faced by centralized AI auditing. Through three innovations—hierarchical directed acyclic graphs (HDAGs), the DAAN causal attribution protocol, and multi-level consensus mechanisms—combined with the security-profit theorem and privacy protection design, it provides transparent and robust trustworthy AI service support for high-value scenarios such as healthcare and finance, and supports four application scenarios including decentralized auditing and tamper-proof leaderboards.

Large Reasoning Models (LRMs) and Multi-Agent Systems (MAS) are widely used in high-value fields, but traditional centralized verification has four limitations:

1. Insufficient robustness: Risk of single-point failure, vulnerable to attacks or bias;
2. Limited scalability: Computational and storage requirements grow with reasoning complexity, becoming a bottleneck;
3. Lack of transparency: Auditing process is opaque, making it difficult for users to confirm decision reliability;
4. Privacy risks: Exposure of reasoning traces may lead to model theft or adversarial attacks, creating a dilemma between transparency and privacy.

The TRUST framework solves the above dilemmas through three innovations:

1. Hierarchical Directed Acyclic Graphs (HDAGs)

Decompose the chain of thought into five layers: original input, semantic parsing, strategy planning, execution reasoning, and final output. It supports parallel distributed auditing, where different auditors focus on different layers while maintaining logical consistency.

2. DAAN Causal Attribution Protocol

Project multi-agent interactions onto Causal Interaction Graphs (CIGs) to achieve deterministic tracing of error root causes, which is more precise than black-box debugging.

3. Multi-Level Consensus Mechanism

Integrate three types of auditors: computational checkers (automated verification), LLM evaluators (semantic logic evaluation), and human experts (authoritative judgment). It uses stake-weighted voting and is theoretically proven to tolerate 30% adversarial nodes.

Experimental validation shows the effectiveness of TRUST:

• Accuracy: Reaches 72.4%, 4-18 percentage points higher than the baseline;
• Adversarial robustness: Remains stable even when 20% of nodes are compromised;
• DAAN attribution: 70% root cause accuracy, saving 60% token consumption;
• Human validation: F1 score of 0.89, Brier score of 0.074, highly consistent with human judgments.

TRUST supports four core application scenarios:

• A1: Decentralized auditing, providing publicly verifiable third-party audit services;
• A2: Tamper-proof leaderboards, establishing trustworthy AI model performance rankings to prevent cheating;
• A3: Trustless data annotation, ensuring annotation quality in a decentralized environment;
• A4: Governed autonomous agents, establishing a governance framework for AI agents to prevent loss of control.

The TRUST research has implications for AI governance:

• Decentralization as a foundation for trust: A trustworthy system can be established without a single authority;
• Integration of economic incentives and security: Aligning individual rationality with system security through staking-reward-penalty mechanisms;
• Balance between transparency and privacy: Public decision-making on-chain, protection of sensitive content off-chain;
• Value of multi-layered verification: Multi-source cross-verification provides sufficient reliability guarantees.

TRUST still needs to explore the following directions:

• Performance optimization: Reducing latency overhead while maintaining security;
• Cross-chain interoperability: Supporting cross-chain auditing and consensus;
• Dynamic participant management: Efficiently handling dynamic changes in the set of auditors (joining, exiting, reputation updates).