Every wash and cluster metric on-chain is an upper bound: nobody can prove who is behind the wallets. The exception is the agents that carry verifiable identity.
Prediction markets have gone from curiosity to infrastructure. Monthly notional volume across the major venues climbed from under $100 million in early 2024 to more than $13 billion by late 2025. With the volume comes the old question: how much of it is real?
Through early and mid-2026, several teams took that question to Polymarket directly — wallet-level anomaly screens, order-flow skill classifiers, market-level information-leakage scores. Practitioners want the same thing in plainer terms: less headline volume, and more net notional settled, unique funded traders, and retention — the numbers that separate real participation from churn.
Every one of these methods hits the same limit.
You can see the pattern. You cannot prove the operator.
From pseudonymous on-chain data you can flag wallets that trade both sides without net exposure, or clusters of addresses that move together. What you cannot do is establish, from that data alone, that those addresses share one controller. The recent Polymarket work says so plainly: its wash figure is an upper bound, because counterparty identity cannot be pinned down from pseudonymous data without extra structure, and cluster attribution rests on an assumption about shared operators that the authors themselves call methodologically uncertain.
So the output carries an asterisk. "Likely the same operator" is a fair analytical guess. It is not evidence — it will not settle a dispute between a platform and a trader, or convince a counterparty or supervisor who declines to take an analyst's word for it.
The missing structure is identity
If a participant carries a verifiable identity — an on-chain-anchored credential tying an agent to a controller and a mandate — the operator question is answered for that participant. The credential names the controller, and anyone can check it. For the identified part of the market, the guess becomes a fact: the identity attests that two wallets share an operator, and the attestation is verifiable by whoever needs it.
This attribution has two properties a heuristic score does not. It is provable, so it can hold up in a dispute. And it is recomputable: any party re-derives it from anchored evidence instead of trusting a proprietary label. Recomputability is what we argued for in the recomputable-trust work — a check anyone can run for themselves, owned by no one.
MoltRadar
That is what MoltRadar is for, and it is out now. MoltRadar scans the on-chain ERC-8004 agent registry and makes the identified-agent layer visible: which agents carry a verifiable identity, and where. It maps the part of the market where the operator question is already answerable.
The scope is narrow, and worth stating plainly. For anonymous wallets it changes nothing; the wall stands. For agents that carry identity, it removes a single uncertainty — operator attribution — and makes that removal checkable by anyone. It is not a fix for manipulation, and does not pretend to be.
Today that identified part is small. It grows as more of the participants become agents — automated traders, and the multi-agent oracles now resolving markets — and as venues start requiring identity from the agents acting inside them. That is where the reliable measurements will come from first.
Analytics on anonymous markets will keep improving. The question that has stayed stuck — who is behind the wallets — only moves once identity is present. That is what we build.
If you build on this or study it — trading, oracles and resolution, venue integrity — we are glad to compare notes.
Sources
- Volume growth, from under $100M in early 2024 to more than $13B in monthly notional by November 2025: Dune × Keyrock, Prediction Markets: The Next Frontier of Financial Markets, December 2025.
- The operator-attribution limit — wash figures reported as upper bounds, and cluster attribution described as methodologically uncertain because counterparty identity cannot be established from pseudonymous on-chain data without additional structure: M. Nechepurenko, Fill-Side Non-Retail Trading on Polymarket, arXiv:2605.11640, May 2026 (§9.2).
- Recomputable trust: Trust Without Trusting: A Recomputable Trust Protocol for Autonomous Agents, arXiv:2605.06738.
MolTrust builds open, verifiable trust infrastructure for autonomous agents — W3C DID/VC identity, an IETF-published authorization envelope, and recomputable accountability, anchored on-chain. MoltRadar
Top comments (0)