Silent Updates: Post-Deployment Disclosure Gap

Silent Updates paper appears in the harvested arXiv AI safety item set.

Post-deployment custody gaps are directly relevant to OECD accountability and incident-tracking agendas.

Both stress post-training deployment controls and verifiable linkage between evaluated and served agent/model behavior.

Output watermarking is one provenance tool, but does not by itself prove which model version was served post-deployment.

Both target verifiable authorization/custody: ARF for action authority, silent-updates work for served-model identity.

Silent post-deployment model changes undermine the external verifiability required for accountable agentic decisions.

FinTech agent governance needs provider-exposed inference controls and reproducibility that silent updates currently deny external verifiers.

Both identify missing external verifiability between claimed evaluated systems and systems that actually act.

Missing post-deployment chain-of-custody worsens the verifiability gap for high-stakes agentic financial decisions.