Originally published at rantum.xyz, which is the canonical source. The full preprint is on Zenodo, DOI 10.5281/zenodo.21513263.
Decentralized exchange volume is one of the most-cited numbers in crypto. Liquidity-mining programs pay against it. Layer-2 foundations report it as adoption. Grant committees read it as evidence a product is being used. It is treated as a proxy for real people doing real trades.
It is a weak proxy, for a structural reason: the trade-execution infrastructure everyone shares exposes no standard way for a frontend to announce who it is on-chain. A wallet-native swap widget, an aggregator, an institutional smart contract, and a meta-router can all produce the same kind of settled trade, and none is required to leave a signature saying which one it was. Once volume is not tied back to a true origin, it is trivially inflated — by wash trading, proxy routing, and MEV-adjacent flow — and any incentive program built on top of it rewards the wrong behavior.
I wrote a preprint on the attribution methodology we built for ClearTrace, a neutral third-party execution-intelligence system covering Ethereum, Base, Arbitrum, and Optimism. This is the plain-language version; the full methodology, validation, and limitations are in the paper.
The problem stated plainly
If you cannot say which frontend produced a trade, you cannot say whether the volume is organic. And you usually cannot say which frontend produced a trade, because that information is not natively recorded anywhere. Attribution is the missing first step every downstream number quietly depends on.
Four independent heuristics
The paper describes four ways to recover frontend origin from raw transaction and trace data. They are deliberately independent — each fails on different trades, so where several agree the signal is strong, and where they disagree you have found something worth a closer look.
- Calldata-suffix trapping — many frontends append an identifying tag to the end of the transaction calldata. Where it exists, it is the cleanest signal available.
- Proxy-router mismatch detection — when the contract that appears to send a trade is not the one that actually originated it, the mismatch itself is evidence of routing.
- Multi-hop origin tracing — following a trade back through intermediary contracts to the interface a user actually touched, rather than stopping at the last hop.
- Fee-recipient clustering — grouping trades by where the fee ends up, which tends to cluster by frontend even when nothing else does.
No single heuristic is trusted on its own. The point is the ensemble.
Attribution first, then execution quality — kept separate
On top of the attribution layer sits an execution-quality layer that benchmarks the price a trade actually got against a volume-weighted average price (VWAP) oracle. That quote-accuracy score is kept deliberately separate from a sandwich/MEV-exposure score. They answer different questions — "did you get a fair price?" and "were you exposed to predatory ordering?" — and blending them into one number hides more than it reveals.
What the paper reports
Rather than assert the method works, the preprint reports what happened running it in production:
- Validation from an in-production quote-accuracy sampler, with 112 of 168 tracked cells reaching statistical significance over a 9.6-day window.
- A worked hypothesis reversal: a major aggregator that looked strong until additional data revealed order-size-dependent quote degradation.
- A structural finding separating genuine sandwich-attack immunity from what was actually a coverage gap on Arbitrum — a distinction that is easy to get backwards.
It also states the limitation plainly: this is single-team validation, and the paper argues for treating cross-frontend attribution as a first-class, auditable input to incentive design rather than an afterthought bolted onto a volume dashboard.
Read the paper
The full preprint — methodology, validation, and limitations — is on Zenodo under DOI 10.5281/zenodo.21513263, licensed CC-BY 4.0. The abstract and citation live at rantum.xyz/research, and the methodology is deployed in production at cleartracedata.com.
Top comments (0)