On 14 March, at 09:41 UTC, a solver address ending c914 won 58% of all winning bids across the two largest intent-based routing networks in a single four-hour window. It paid 0.41 ETH in gas to do so. Nine days earlier, that same address held a balance of zero.
That is the anomaly. Not the concentration — solver markets are oligopolistic by construction and always have been in their first eighteen months of life. The anomaly is the gas bill. A 0.41 ETH spend to capture 58% of auction share does not describe a competitive bidder. It describes a bidder with prior access. Either the bids were placed with advance knowledge of the batch contents, or the auction's own latency floor was being gamed at the boundary where the oracle feed hands off to the solver.
Every transaction leaves a scar on the blockchain. Most analysts read the scar and stop. The scar is the gas line. The wound is upstream, in the price feed.
Methodology & Data Source
Everything that follows is reproducible, and I am naming the inputs so the reader can rebuild the work rather than trust it. First, trace-level data on 1,183,412 intent fills settled across the 90 days ending 31 March, drawn from Nansen's solver-attribution tables. Second, oracle update timestamps from three push-based feeds, correlated against settlement blocks at one-second granularity. Third, the public bid logs of both auctions.
No founder interviews. No Discord. No roadmap documents. Data is the only witness that cannot be bribed.
Context: The Auction Nobody Audits
Intent-based architecture is the dominant design pattern in 2026. The pitch is clean. A user signs a declarative statement — "sell 500 ETH for at least 1.8 million USDC within 30 seconds" — and a competitive network of solvers races to fill it. The user gets speed and price improvement. The protocol gets flow. The MEV that once extracted value from the user's transaction is, in theory, rebated back to that user as surplus.
That theory has one large hole in it. The auction is not on-chain. The bid ranking, the winner selection, and the price quotes all resolve in an off-chain matching engine, and only the net settlement hits the EVM. What lands on-chain is the outcome, not the process. The blockchain does not forget — but it also does not record what it never saw.
I have spent my career on this exact asymmetry. In 2017 I spent three weeks auditing a staking-reward distribution algorithm against its own whitepaper and found that the math quietly favored early depositors. The contract was public. The bias was hidden in the ordering of operations. Intent auctions in 2026 have the same property at a larger scale: the mechanics are public, the sequencing is not.
The scale matters. Across the 90-day window I examined, intent-based routing settled roughly 61% of all DEX volume I could attribute to retail-sized swaps — trades under $250,000 notional. Retail flow is now overwhelmingly mediated by solvers. Whatever happens inside those matching engines is, functionally, the market.
And the matching engine's single most important input is the price feed. Not the user. Not the pool. The feed.
Core: The Evidence Chain
Start with the timestamp correlation. I aligned every oracle update from the three feeds I tracked against the settlement block of every fill in the dataset, at one-second resolution. The distribution is not uniform.
Across 1,183,412 fills, 34.7% settled in the same second as a feed update. That is expected — solvers act on fresh prices. What is not expected is the shape of the tail. Fills settling in the 200-to-900-millisecond window before a feed update — that is, fills priced off a stale quote — lost an average of 41 basis points relative to the mid-price across the following ten seconds. Fills settling within 300 milliseconds after a feed update lost 6 basis points.
The population of fills that settled on stale quotes was not random. I clustered the winning solver addresses behind them. Three addresses accounted for 71% of all stale-quote fills in the sample, and those three addresses shared infrastructure: identical gas-price bidding curves, sequential nonce patterns on the same settlement contracts, and overlapping funding sources traced back to two exchange deposit addresses.
That is the scar. Three solvers, one operational entity, and a consistent ability to fill against prices that were about to move.
The second piece of evidence is the auction's own latency floor. Both networks publish a bid deadline of roughly 150 milliseconds before the batch is sealed. I measured actual winner determination using the delta between the last accepted bid timestamp and the settlement block's parent hash timestamp. In 12 of the 90 days examined, the effective window collapsed to under 40 milliseconds — and on each of those days, the c914 address cluster's win rate exceeded 50%.
A collapsed window is not proof of collusion. It is proof that the auction's timing is operator-controlled, and that a participant with better infrastructure sits closer to the clock. Whether that advantage is earned through engineering or granted through access is not something the on-chain record can tell me. But it is a question the on-chain record can raise, and nobody in the current discourse is raising it.
The third piece is the gas accounting. The c914 cluster spent 0.41 ETH to win 58% of a four-hour window. I reconstructed the implied cost per winning bid and compared it to the rebate schedule published by both networks. The math does not close. At published rebate tiers, the cluster's four-hour gross revenue would have been roughly 2.1 ETH. Net of gas and the settlement contract's fee share, that is a thin margin — unless a meaningful portion of the captured value arrived through a channel that is not in the rebate accounting at all.
Which is the point. The surplus rebated to users is measured against the auction's own reference price, not against an independent mid-price. If the reference price is biased, the rebate can look generous while the extraction continues invisibly. This is the same structural flaw I found in 2020 when I scripted a comparison of Compound's distribution against organic deposit growth and discovered that 40% of deposits came from bot farms farming new-account bonuses. The headline number was real. The underlying demand was not.
I want to be precise about what I am and am not claiming. I am not claiming the networks are fraudulent. I am claiming that the verification surface is inadequate for the volume they now carry. A 1.18-million-fill dataset is large enough to show a pattern. It is not large enough to assign intent. Intent lives off-chain.
There is a second-order problem that compounds this. Oracle feeds themselves are not neutral infrastructure. They are operated by a small set of node operators with published update thresholds, and every operator knows the threshold. A feed that updates on a 0.5% deviation band is not a random process. It is a schedule. Anyone who can model the schedule can position ahead of it, and push-based architectures make that modeling trivial compared to pull-based alternatives. The decentralization of the operator set is irrelevant if the update cadence is deterministic.
Contrarian: The Wrong Scapegoat
The reflexive conclusion is that intents are a scam and that MEV has simply moved from the mempool to the solver network. Half right, and the useful half is the half nobody says out loud: MEV did not move. It was reclassified.
Under a classic AMM, extraction shows up as an on-chain sandwich. It is legible. Anyone with a block explorer can count it, and the community's reaction is immediate and loud. Under intent routing, the same economic loss appears as a slightly worse fill price. It is invisible to the user, invisible to the explorer, and it is disclosed — if at all — inside a rebate metric that the protocol itself defines.
That is not a technical improvement. It is an accounting change. The extraction did not shrink; the reporting window did.
The contrarian corollary is that pure on-chain MEV — the kind that DeFi spent five years building defenses against — is now a shrinking minority of total extraction. The defenses worked. The volume moved. And the industry is applying 2021-era scrutiny to a 2026-era problem, which is why the concentration metrics above generate no headlines despite being observable from public bid logs.
There is a version of this critique that I do not accept, and I want to mark it clearly: the argument that solvers are simply more efficient and that users are better off. Measured against the auction's reference price, users are better off. Measured against an independent mid-price reconstructed from multiple venues, the improvement narrows considerably — in my sample, from an apparent 22 basis points of surplus down to 7. The remaining 15 basis points went somewhere. Correlation is not causation, but a persistent, addressable, non-random 15 basis points is not noise either.
I have watched this pattern before. In 2021 I mapped wallet clusters on a PFP collection and found that 60% of high-value sales were transfers between wallets under common control. The floor price corrected 20% once the dataset was public. Nothing about the collection changed. Only the visibility did.
Takeaway
The signal to watch next week is not the solver leaderboard. It is the auction's published latency floor.
If either network widens its bid window, or commits to publishing per-batch bid distributions after settlement, the concentration in my dataset should decay within two weeks. That is a falsifiable prediction and I will publish against it. If the window stays narrow and the distributions stay private, then the 15 basis points I cannot attribute will keep flowing somewhere, and the scar will simply relocate to a surface nobody is tracing.
Every one of those 1,183,412 fills is sitting on-chain right now, immutable and waiting. The question is whether anyone is going to read them before the next four-hour window opens.
