AEGIS：连接 AI 智能体日志义务与独立审计的可移植证据接口

Third-party audit of AI-agent actions requires a portable handoff from heterogeneous execution records to independently retained evidence, auditor-owned procedure and report, and forum judgment. AEGIS defines that handoff through an ARC1–ARC6 conformance boundary joining declared capture, portable sequence commitment, producer-independent acceptance, independently governed custody, auditor-owned procedure and report, and competent-forum judgment. Cryptographic algorithms, encodings, authenticated structures, and anchors implement replaceable profiles within this stable architecture. AEGIS v1 provides a byte-specified signed bundle and V1–V4 acceptance procedure that a verifier can reperform without a platform API or private key. IBM Client Engineering's public DFAH-Bench repository expressly attributes its hash-chain and certificate modules to AEGIS and uses them to construct and verify benchmark evidence bundles, providing direct downstream engineering reuse. The AEGIS capture–commit–third-party-verification handoff was publicly deposited on 11 March 2026, establishing the chronology examined in this article.