Open problems in AI-agent reproducibility, evidence infrastructure, and audit.
Peer-reviewable write-ups of the system and its measured results.
Mohsen Seyedkazemi Ardebili · arXiv:2609.12582 · 2026
The design, threat model and evaluation of NovaFabric: eight research questions at measured scope, including replay, tamper rejection, redaction, lineage queries over 100M edges, and a 314-machine, ten-region ingest run.
Cite: doi:10.48550/arXiv.2609.12582 · Software archive: doi:10.5281/zenodo.22997388
Active investigations across reproducibility, audit, and provenance.
Local-first run capsules for capture, replay, diff, validation, and audit workflows. CLI capture, SDK decorator, API/MCP proxy, and OTel GenAI semconv all captured.
Tamper-evident DSSE/RFC 3161 sealing in local profile. Cryptographic SoD enforced. Enterprise HSM/X.509 PKI not claimed.
SQLite lineage is experimental (10M-edge KuzuDB benchmark done). KuzuDB backend is prototype. Lineage federation is experimental with legal review pending.
Multi-tenant server with OIDC/RBAC, 13-tab dashboard, offline tokens, and topology views. Production-scale operation is not claimed.
KuzuDB-backed KG for query, audit, and entity workflows. CLI and dashboard panels shipped v0.29–v0.31. Tier 2/3 features in progress.
Risk classification, audit profiles (EU AI Act, NIST, GDPR, SOC2, NIS2), cost reporting, and evidence exports. Experimental workflows — not legal guarantees.
The core open problem: given an AI-agent run, can we produce a verifiable certificate that another researcher can independently verify? NovaFabric explores sealed capsules as the atomic unit of evidence, with DSSE signing and RFC 3161 timestamping. We are actively investigating: (1) capsule deduplication at scale, (2) semantic equivalence for replay validation, (3) lineage graph query languages for provenance forensics.