RuntimeGuard-AI: Durable Policy-Decision Receipts for AI Audit Evidence
There's a fresh paper on arXiv introducing an innovative AI audit log system called RuntimeGuard-AI, which focuses on improving durability and trust. This system speeds things up by delivering a secure decision before writing data that lasts, plus it keeps evidence safe during crashes with specially chosen synchronization points and Ed25519-signed receipts. When the system restarts, it verifies various aspects like framed records and continuity. The stored records are neatly arranged into linked, signed Merkle epochs, which auditors can check with an outside key. In tests on an Apple M4 Pro using four worker threads and 2,048-byte prompts, the system managed to log 27,193 requests per second with a latency of 141.9 microseconds per request.
Key facts
- The paper is titled 'The Acknowledgment Point Is the System: Durable Policy-Decision Receipts for AI Audit Evidence' and is available on arXiv with ID 2608.17176.
- RuntimeGuard-AI returns a guarded decision before any durable write to minimize latency.
- The system commits a privacy-minimizing record at a caller-selected synchronization boundary.
- An Ed25519-signed receipt indicates whether the synchronization boundary completed.
- Post-restart validation checks framed records, manifests, shard placement, sequence continuity, and replay identity.
- A separate attestation path groups committed records into chained, signed Merkle epochs.
- Auditors verify the Merkle epochs using an externally obtained key.
- Performance testing on an Apple M4 Pro showed 27,193 requests/s with 141.9 microsecond latency.
- The paper was announced as a cross-type on arXiv.
Entities
Institutions
- arXiv
- Apple