POLIS 5,280-Episode Study: Provenance-Aware Guards Block Authority Laundering That Local-State Guards Miss in 22 of 96 Episodes

arXiv 2608.09828·high signalXBlueskyLinkedInCopy link

The first paper from the POLIS research programme, posted 2026-08-10, runs a frozen 5,280-episode suite across four model families to ask which parts of a multi-agent 'institution' actually produce safety. A detailed constitutional prompt and a provenance-aware executable guard both yield 0/384 realized violations, but the guard blocks prohibited attempts in 51/384 episodes with 44/51 later completing safely. In matched laundering scenarios — where an ordinary transformation changes visible policy while originating authority stays fixed — a local-state guard admits violations in 22/96 episodes versus 0/96 for provenance enforcement (p = 4.77e-7). Identical final violation rates can hide completely different mechanisms.