ARCA: auditable AI context modeling and on-chain anchoring for trustworthy risk control

Abstract

Despite the widespread use of Large Language Models (LLMs) in financial risk control, their opaque reasoning and Retrieval-Augmented Generation (RAG) privacy risks pose severe compliance challenges. While blockchain technology is increasingly adopted to provide tamper-proof auditing for AI decisions, its prohibitive anchoring costs under high-frequency business demands remain a critical bottleneck. To address these intertwined challenges, this paper proposes ARCA, an auditable AI context modeling and on-chain anchoring architecture. First, to mitigate RAG privacy risks, we design a Dual-Track Privacy Retrieval engine. The semantic track utilizes block-permuted orthogonal obfuscation, guaranteeing 100% lossless recall while resisting Known Plaintext Attacks (KPA). Concurrently, the keyword track leverages Keyed-Hash Message Authentication Code (HMAC) trapdoors and a Counting Bloom Filter (CBF) to strictly support the GDPR “Right to be Forgotten.” Second, to ensure verifiable AI reasoning, we formulate the InferenceUnit structure, intrinsically integrating a sequence number to encapsulate the cryptographic evidence chain. Deeply decoupling computation from ordering via a lock free queue significantly mitigates concurrency deadlocks. Finally, to overcome economic barriers, we introduce a Chain-of-Thought (CoT) guided Dynamic Router and a Batch Merkle Tree mechanism for high throughput bulk anchoring. Extensive experiments demonstrate ARCA successfully intercepts 65.12% of AI factual hallucinations at optimal cognitive balance, while the batch mechanism reduces per-transaction Gas costs by 84.87%, effectively enabling scalable on-chain auditing for trustworthy AI.