The gas limit was identical on all four thousand two hundred and eleven transactions.
That is the first thing I check. Not price. Not volume. Gas limit.
216,000. Every single fill. Down to the unit.
On the morning the anomaly surfaced, a mid-tier generative collection posted $18.4 million in 24-hour volume across the major aggregators. The timeline called it a resurgence. Two newsletters called it the return of the JPEG. Forty-one wallets accounted for 91 percent of that number. The unique-buyer count for the same window was 137. Unique sellers: 129.
Two hundred and sixty-six participants do not move eighteen million dollars in a day.
Six bots sharing one deployment script do.
I pulled the block data at 04:20 UTC. By 05:00 I had the funding graph. By noon I had the whole thing. The collection's floor never moved more than 1.8 percent.
The venue here is not exotic. It is a mid-cap aggregator that routes Seaport orders and a proprietary book through a single settlement contract. Eight weeks ago it shipped an Agent SDK — a REST wrapper around its signed-order endpoints, rate-limited to 40 requests per second per API key, with a paymaster contract that sponsors gas for whitelisted callers. The pitch was frictionless machine-to-machine commerce.
The reality is a subsidized execution layer.
I have been mapping machine-to-machine flow since 2020, when I built a Python script that tracked 500 daily Uniswap V2 transactions and surfaced whale accumulation no dashboard displayed. That work got me into Nansen. The habits did not change. I still start with the funding graph. I still distrust any figure that rounds to the nearest million.
Before that, in 2017, I spent six weeks auditing the Solidity codebase of a Singapore-based ICO ahead of its mainnet launch. I found three reentrancy paths and got a pull request merged two weeks before the token sale. That job taught me a permanent lesson: the contract is the only witness that never gets tired, never gets threatened, and never changes its story.
Last year I worked with an AI research group to analyze ten million interaction logs between autonomous agents and smart contracts. The finding that stuck was not the manipulation. It was the plumbing. Agents do not behave like traders. They behave like cron jobs. They have no opinion about art. They have an opinion about expected value per unit of gas.
That distinction is the entire article.
So: the collection. Generative, 8,888 supply, minted fourteen months ago at 0.04 ETH. Floor at the time of the anomaly: 0.38 ETH. Not a blue chip. Not a corpse. The middle of the market — which is where forensic work always pays, because nobody is watching the middle. Blue chips have too many eyes. Dead projects have no volume to dissect. The middle is where incentives are mispriced and audits stop.
I queried four datasets. Every fill on the settlement contract across a 72-hour window. Every ERC-4337 UserOperation touching the paymaster. Every ERC-2981 royalty event on the token contract. And the venue's public points ledger, which lives on-chain and is therefore, mercifully, not a marketing document.
The blockchain remembers what the founders forget.
Start with the fingerprint.
Every EVM transaction carries a gas limit chosen by its submitter. Humans let wallets estimate. Wallets estimate from simulation, and simulation output drifts with state. So human gas limits cluster loosely and wander. Bot gas limits are constants compiled into code.
Across 72 hours, 4,211 fills shared twenty-three distinct gas limits. Three of them accounted for 4,088 fills. 216,000. 231,400. 208,750. Those three numbers appeared with the regularity of a resting pulse.
Tracing the ghost in the smart contract code is not mysticism. It is arithmetic. Identical gas limit, identical calldata shape, identical nonce velocity equals one codebase. I never needed the repository. The chain published it for me.
Next: who paid for the gas.
Not one of the 41 wallets held ETH sourced from a centralized exchange withdrawal. All of them were funded from three addresses. Those three were funded from one. That one was funded from a bridge deposit eleven weeks earlier, and the bridge withdrawal address had been used exactly twice: once to receive, once to distribute.
Wallet clustering is not a party trick. It is a graph problem with a known failure mode. You can always find a link if you look hard enough, and you can never prove one does not exist. So I never rely on clustering alone. I require mechanical corroboration.
Here it was a deployment signature. Twenty-nine of the forty-one wallets were created via CREATE2 at addresses sharing a two-byte prefix. That is vanity mining. Cheap gas on cold storage paths. And the operator did not care that it painted the same initials across the entire set.
Every mint leaves a digital scar. So does every deployment.
The third signal closed the case. Time.
Human trading has a heartbeat. It is bursty, diurnal, and correlated with social attention. Fills cluster after announcements, after royalty payouts, after a well-followed account posts a chart. The inter-arrival distribution is heavy-tailed, and that is testable.
The 4,211 fills were not heavy-tailed. I ran a Kolmogorov–Smirnov test against a fitted Poisson process and got a statistic that should embarrass anyone who described this flow as organic. Fills arrived at near-constant intervals within each block. Then they paused. They paused exactly when block production paused. They paused again through a nine-minute sequencer hiccup — during which the order book stayed open and received eleven bids. Eleven. In nine minutes. On a collection doing eight figures.
Silence in the logs speaks louder than the pump. The humans were present. They simply were not the volume.
I ran the same test against a control: a collection with a known human collector base and no points program. Heavy-tailed, diurnal, clumpy around a single influencer post. The contrast was not subtle. Good forensics needs a negative control, and almost nobody in this industry bothers to run one.
Fourth: the venue did not get paid.
Aggregators monetize through taker fees and, more durably, through order flow data. A machine routing 91 percent of a collection's volume should generate routing revenue. This one did not. I traced the fills through the settlement contract and found that 3,844 of 4,211 executed against signed orders submitted directly to a private mempool endpoint — bypassing the routing contract entirely while still reading its public book.
Read the book. Take the book. Pay nothing.
The signature check passed because the orders were valid. The orders were valid because a trade between two wallets you control is, to the EVM, indistinguishable from a trade. Nothing was exploited. The fee schedule simply worked as designed on a venue that assumed its counterparties were human and slow.
Here is the technical layer most coverage misses. The paymaster sponsored gas for whitelisted UserOperations through an ERC-4337 bundler. Each UserOp carried a paymasterAndData field pointing at one contract. The bundler batched them. That batching is what allowed 41 wallets to fire within the same block, in a fixed order, without competing on priority fee — because they were not competing. They were one transaction wearing forty-one coats.
A bundler does not ask whether the UserOps inside it are economically independent. It asks whether they validate. They validated.
Fifth: the reason for all of it.
Follow the points.
Loyalty Points Season 4 awards 1 point per 0.01 ETH of taker volume on eligible collections, with a 3x multiplier for agent-routed flow — a category the SDK defines by API key, not by any on-chain property. The token has not launched. The market assigns an implied value based on the previous season, which settled near $1.40 per point across a 30-day vest.
The arithmetic is trivial once you see it. Take a 0.38 ETH floor. Round-trip the token between two wallets you control. You pay gas, you pay any marketplace fee, and you receive 3x points on the notional. At a 0.38 ETH notional with a 3x multiplier, the round trip costs roughly $0.61 in gas and fees and returns roughly $2.28 in expected point value at the season's implied price.
Positive expectancy. Negative purpose.
I modeled the program the way I modeled algorithmic stablecoins in 2022 — Monte Carlo, ten thousand iterations, varying the two variables that matter: eventual token price, and the ratio of real to synthetic volume. Above an 11 percent synthetic share, the point-to-value conversion drifts into a regime where the only profitable strategy is more synthetic volume. Above 30 percent, the program becomes a closed loop. Only bots can afford the gas. Humans are priced out of their own loyalty program.
I published an equivalent result four years ago about reserve-backed tokens. Different asset. Same mathematics.
That is the Risk Simulation appendix, compressed. The full model is in the dashboard.
Sixth: the creators.
ERC-2981 is a signal, not a rule. The standard tells the market what the creator would like to be paid. The market decides what it pays. Across the 4,211 fills, royalty recipients collected 0.71 percent of volume against a 5 percent declared rate.
The gap is not malice. It is routing. Signed orders that set royalty fields to the protocol minimum are valid orders. Aggregators that honor the minimum are honoring the standard. Everybody is compliant. Nobody is paid.
I have written about this before and been told the answer is better tooling — dynamic royalties, programmable splits, on-chain enforcement. The tooling is not the constraint. The constraint is that the only buyer left in the middle of this market sets the royalty to zero by default, because the royalty is a cost, and the cost is measurable, and the art is not.
Fifteen creators split $130,600 in a week when $18.4 million of "volume" cleared. Several of them posted about the volume.
So who did get paid?
The gas tells you. The largest single recipient in the window was the private mempool builder — $41,000 in priority fees. Second was the ETH burn. Third was the paymaster's sponsor, which is to say the venue itself, which paid roughly $28,000 in sponsored gas to generate volume that generated points that generated the expectation of a token. Fourth: the point farmers, in expectation. Fifth, at 0.71 percent, the people who made the thing.
Mapping the liquidity that never was is a discipline of subtraction. Notional volume: $18.4 million. Settled economic value, counting each round trip as one trade: roughly $1.9 million. Value that left the system permanently: about $69,000 in gas and builder fees. Value contingent on a token that does not exist yet: eight figures.
That is the trade. Not art. Not community. A venue paying itself to look busy so a token can be sold to people who will read a dashboard on announcement day.
The floor price is a lie told by whoever is buying. It was never the signal here. It did not move, because the wallets on both sides of the trade were the same hand.
Now the part where I argue with myself.
None of this is fraud. I want to be precise, because precision is the only thing I sell.
Every one of the 4,211 fills settled. Every signature validated. Every hash is public and permanent. No withdrawal was denied. No oracle was manipulated. No holder lost a token they did not choose to part with.
This is an incentive design failure, which is a different animal. The venue published a fixed reward per unit of volume with no measure of uniqueness. Rational agents optimized it. That is what agents are for. Had I written a bot in that window, it would have done the same thing, and I would have slept fine.
The interesting question is not whether the volume was fake. It is whether fake and real volume remain distinguishable when the marginal participant is a script. Correlation has stopped being causation's cousin and become its replacement. Dashboards report what the chain settled. The chain settles what the bots chose. The bots chose what the incentive paid.
I have watched this pattern in another ledger. After the fourth halving, miner revenue compressed and hash power consolidated toward a handful of pools, and the network still calls itself decentralized because the node count has not moved. Concentration never announces itself. It simply changes who decides.
Regulation will not fix this either. Europe's framework has given the market apparent clarity, and the compliance cost of becoming a licensed service provider is now high enough that the surviving venues are the large ones — which are precisely the venues with the incentive to report the biggest volume. Clarity for institutions. Consolidation for everyone else.
The same is true in the middle of the NFT market. It now has one kind of buyer. That buyer does not have a wallet. It has an API key.
Here is the signal I will be watching, and it is not the floor.
Watch the paymaster's outflow. When a venue stops sponsoring gas for agent-routed flow, synthetic volume stops within a single block, because positive expectancy was the only thing holding it up. Watch the point-to-claim conversion ratio in the first hour of the token event — if fewer than 8 percent of points convert into claims, those wallets were never users, and the float belongs to six scripts. Watch the royalty line, because it is the only number in this article that measures something real.
Pattern recognition precedes profit prediction. The pattern is on-chain, timestamped, and permanent.
Then ask the question the dashboards will not: when volume is manufactured, who is the buyer of last resort?
