Most agent-trust proposals still lean on someone to vouch: a registry, a certificate authority, an enterprise identity provider. A new paper, "Trust Without Trusting: A Recomputable Trust Protocol for Autonomous Agents," takes a different position: instead of asking a central party to attest that an agent is trustworthy, it makes trust something any party can recompute independently, without relying on a central authority, by extending the recomputation approach across organizational boundaries. The paper's framing of the problem is worth sitting with. It surveys the current landscape — Singapore's IMDA framework requiring verifiable digital identity and audit trails, NIST's new AI Agent Standards Initiative, the EU AI Act's human-oversight provisions, Anthropic's zero-trust guidance, Google's SAIF 2.0 risk map, Microsoft's Entra Agent ID — and notes a shared assumption running through all of it: some party sits above the transaction and certifies who did what. That works inside one company. It gets shakier across organizational lines, where the two agents transacting may have no shared registrar and no reason to trust each other's home platform. The paper's proposed fix is to make the trust claim itself something a third party can independently regenerate from the underlying evidence, rather than something they have to take on faith from an issuer. If a receiving party can recompute the same trust score or authorization result from the same inputs, the central authority's role shrinks to something closer to a data source than a gatekeeper — a meaningfully different trust model than the credential-and-signature approach most current agent-identity specs assume. It's a narrow technical move, but it names something real: nearly every governance framework published this year — from IMDA to NIST to the EU AI Act's Article 14 — puts the accountability burden on the relying party, not on the protocol. The paper is explicit that this is the gap it's trying to close: holding the exercise of authority across a network-wide boundary to account, by recomputation rather than a central authority. Whether recomputable trust holds up outside a paper's threat model is an open question — the approach hasn't been tested against adversarial agents at scale, and the paper itself is recent enough that no independent implementation has surfaced yet. But it's a useful counterpoint to the assumption, baked into most 2026 agent-identity work, that verification always needs a verifier sitting in the middle.