A new preprint out of MolTrust, "Trust Without Trusting: A Recomputable Trust Protocol for Autonomous Agents," tackles a narrower problem than most agent-trust papers: not whether an agent is trustworthy, but whether the party judging it played by its own rules. The paper's framing starts from scale that's already real, not hypothetical. It cites a single marketplace running 69,000 autonomous bots executing 165 million transactions across $50 million USDC in cumulative volume, without any shared trust layer between participants. Verifiable credentials already let any party check a signed claim without going through a central service. The paper's point is that verification alone doesn't cover the whole problem — it covers the open part of the agent world, where there's no universal boundary to enforce and nobody is asking for one. The harder case shows up wherever a closed space draws its own boundary: a marketplace, platform, or consortium setting house rules that only it enforces. Whoever draws that boundary holds the authority to apply it, and can apply it opaquely, behind closed doors. The paper's central move is distinguishing enforcement (which the boundary-holder already controls) from accountability (whether anyone else can check that the holder applied its own published rules consistently) — without handing that check to yet another trusted intermediary. The proposed answer is what the paper calls the Combined Evidence Protocol: a set of conditions about the population of relying parties, combined into a single predicate that any outside party can recompute from anchored, on-chain data rather than take on faith from an operator's logs. The design goal is to remove the recorder from the trust path entirely — a verdict about an agent's conduct should be reproducible from anchored data, not from anyone's testimony about that data, the same way an instant-replay call is reproducible from footage rather than a referee's word. The paper reports this isn't purely theoretical: the infrastructure has been live since March 2026 across eight verticals, with on-chain anchoring backing the audit trail. That's a meaningful detail for anyone building agent marketplaces where a platform operator's incentive to look the other way on its own rules is exactly the failure mode credentialing alone doesn't solve. This matters for the broader accountability-infrastructure conversation because most current proposals — verifiable credentials, DIDs, signed receipts — solve for "can I check this claim is authentic" but not "can I check the rule-maker actually followed its own rules." Recomputability, in this framing, is a governance property layered on top of cryptographic verification, not a replacement for it.