Answers are grounded in this paper and cite their sources.
As robots increasingly operate under the control of autonomous AI agents in safety-critical and socially sensitive environments, the ability to audit and verify past decisions becomes essential for trust, liability, and post-incident analysis. However, conventional logging mechanisms do not provide guarantees against post hoc manipulation of recorded agent behavior. This paper proposes a lightweight, reproducible framework for tamper-evident logging of autonomous multi-agent decisions, based on cryptographic commitments and blockchain anchoring, as an integrity building block for AI accountability. In the proposed approach, agents record selected decision events off-chain, aggregate them using Merkle trees, and periodically anchor resulting root hashes to an immutable ledger. This design enables efficient storage while ensuring that any modification of committed decision logs after anchoring becomes cryptographically detectable. We implement the framework in an open-source multi-agent robotic simulation and evaluate it under honest and adversarial logging scenarios. Across twelve attack types with 1,000 trials each, every modification of committed decision records was detected against the on-chain commitment, with no false positives in 1,000 honest runs, whereas a local check without the anchor missed every attack once the attacker also rewrote the local commitments. Verification took about 7 µs per recorded decision, a single decision could be verified in about 20 µs with a proof below 1 kB, and each run required one anchoring transaction of 188 virtual bytes, regardless of the number of agents and decisions. The framework guarantees the integrity of what has been committed while keeping the underlying decision records off-chain until an audit requires them. The main contribution is a practical architecture for verifiable decision records of AI agents. It combines selective logging, efficient cryptographic verification, and reproducible experimentation, and lays a foundation for post hoc auditing of autonomous robotic systems.
No takes yet. Share an insight, caveat, or question.
Gürpinar et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: