A new paper proposes something narrower and more concrete than most agent-governance frameworks: a way to verify, without trusting anyone, whether the operator of a closed agent marketplace or consortium actually applied the admission rules it published. The paper's framing starts from scale. Its author notes autonomous agents already transact heavily on open platforms where signed credentials can be checked without a central service — the abstract cites a single marketplace running 69,000 bots and 165 million transactions worth $50 million. The problem isn't the open web of agents; it's the closed spaces layered on top of it. A marketplace, platform, or consortium that sets its own house rules also holds sole authority to apply them, and can do so "behind closed doors," with no outside party able to check that the stated rules match the rules actually enforced. The proposed fix is the Combined Evidence Protocol (CEP), described as a five-condition predicate that any party can recompute from anchored data, turning the question of rule-following into something verified rather than asserted. The design borrows its trust model directly from optimistic rollups: correctness comes from the fact that anyone can redo the computation and check it, so no oracle or central auditor is needed. The paper positions CEP alongside other 'trustless' verification schemes — optimistic and zero-knowledge rollups, verifiable machine learning, self-sovereign-identity predicates — and targets it specifically at consortiums of mutually distrusting peers operating under a shared charter, where each member wants proof, not a promise, that jointly agreed rules are being applied uniformly. What distinguishes this from adjacent identity and credential work is the object being verified. It isn't proving who an agent is or what action it took; it's proving that the gatekeeper enforced its own admission policy consistently, using data anchored on a public ledger — the deployment description points to a W3C verifiable-credential and DID trust layer running since March 2026, anchored on Base L2. That's a narrower claim than general 'agent accountability,' but it's the kind of narrow, falsifiable claim that's rare in a field mostly producing frameworks and taxonomies rather than working, recomputable checks. The paper is a single-author arXiv submission (v2, revised mid-June 2026), not yet peer-reviewed, and its production numbers and deployment claims come from the author's own description rather than independent audit. Worth tracking as a concrete technical proposal, not as settled infrastructure.