Study verifies a portable protocol for AI-agent execution traces, highlighting an innovative approach to offline verification.
AEGIS specifies a portable evidence-bundle protocol for offline verification of AI-agent execution traces. It normalizes captured events, commits them through domain-separated named length-prefixed frames, signs a six-field certificate binding chain identity, agent identity, genesis state, event count, and chain head, and exports the chain, certificate, signature, and public key as one verifier-independent object. The bounded contribution is an evidence contract rather than a new cryptographic primitive. A Node.js implementation passes 206 conformance tests, while an independently implemented Python verifier provides six cross-language acceptance and mutation-rejection tests. Evidence boundary. A valid bundle establishes signer commitment to the captured sequence under the stated assumptions. It does not establish capture completeness, source truth, absolute time, split-view consistency, legal responsibility, or regulatory compliance. Version note. Version 3.0 restructures the prior omnibus AEGIS manuscript into a self-contained technical main article and two separately versioned companions. Stable artifact citations use concept DOIs: Supplemental Material A (Technical Appendix) is at 10.5281/zenodo.21348560; Supplemental Material B (Structured-Review and Evidence Appendix) is at 10.5281/zenodo.21348562. The main article remains under concept DOI 10.5281/zenodo.18955102. Earlier versions remain permanently available in this version chain.
No takes yet. Share an insight, caveat, or question.
Alex Chengyu Li (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: