Verdict: the evidence points to "something else", not pure zero-margin competition and not stable tacit collusion. When incentives are paid out, flip margins go negative (bots bid above the ask), and a large share of what looks like bot-to-bot trading turns out to be one operator trading with itself. Where two independent swarms really compete, the margin should fall to the cost floor (fees, royalties, gas, inventory risk), not to zero. The collusion outcome isn't ruled out; I found nothing that shows it.

This is not a 30-day on-chain study. This session had web search and page fetching only, with no Ethereum node or indexer (Dune, BigQuery or an archive RPC). So I couldn't pull the last 30 days of bot-to-bot trades, and I found no public write-up covering them. What follows is published evidence, mostly from 2021–2023, one transaction I checked on Etherscan, and my own reasoning where marked. The tests in the table can be run directly on 30-day data.

Evidence

1. Negative margins under incentives. In March 2023, while Blur was paying token rewards for bidding, The Block found that "the top bid for more than ten items in that collection sits at 5.07 ETH (about $7,900), whereas the 'buy now' price is 5.03 ETH" [The Block]. Rewards went to "the bids that take the highest 'risk'", so bidders competed on points, not on the spread. A bid above the ask means a negative margin, which neither the competition model nor the collusion model predicts.

2. Much "bot vs bot" flow is one party. CryptoSlam flagged about 80% of Blur's post-airdrop volume as "inorganic". A small group of wallets did most of the trading: about 300 wallets made roughly half the volume, and 565 wallets (1% of traders) held 74% of bid-pool assets [Decrypt]. Decrypt defines wash trading as when "a trader buys and sells NFTs between their own wallets". Hildobby's flags are:

   • buyer and seller are the same wallet;
   • trades go back and forth between the same wallets;
   • the same NFT is bought three or more times;
   • the wallets were funded from a common source.

   When two "swarms" share a funder, any margin you see is an accounting artefact, not a competitive outcome.

3. Theory predicts either result, depending on the algorithm. Simulation work finds that "DRL algorithms, such as PPO and DQN, generally converge to lower prices closer to the Nash equilibrium", while tabular Q-learning "tends to exhibit higher collusion and price dispersion" [arXiv 2503.11270]. Most floor bots are rule-based undercutters ("bid one tick under, relist a bit higher"), not learners. My inference is that this makes them behave like the competitive case.

4. Concrete on-chain example: tx 0x92488a00dfa0746c300c66a716e6cc11ba9c0f9d40d8c58e792cc7fcebf432d0, block 13508785, Oct-28-2021 23:50:37 UTC [Etherscan].

   • CryptoPunk #9998 moved from 0x8e39…4278 to 0x9B5A…0F9b. The transaction was sent by 0xef76…F8dB and was funded with flash loans that were repaid inside the same transaction.
   • Larva Labs said: "someone bought this punk from themself with borrowed money" [The Block, 2021].
   • This is the clearest "fake counterparty" pattern: common control, circular ETH flows, and a price with no economic meaning. It is older than 30 days and is not a floor-flip bot. I use it to show the pattern, not as evidence of recent swarm behaviour.

Signatures that tell the outcomes apart

Some definitions used below:

• Margin: relist price minus buy price, minus fees, royalty and gas.
• Wallet clusters: groups of wallets tied together by funding source. Build these first, so that you know who is actually trading with whom.
• Tick: the smallest price step a bot undercuts by.

Outcome	What you would see on-chain
Zero-margin (cost-floor) competition	Realized margin clusters at fees + royalty + gas, about 0 in net terms.Bids are updated by one tick within seconds or blocks of a rival's move.Many independent wallet clusters with no shared funder.Fills are spread across clusters roughly at random.Gas and priority fees rise over time as swarms race on speed, because price can't fall further.
Tacit collusion	Margin stays above cost for long stretches, even though bid-pool depth would let someone undercut profitably.Bids are updated rarely, and prices sit at round or focal levels.Response is lopsided: rivals follow price increases quickly, and an undercut sets off a short, sharp price war that then snaps back to the old level (a punishment signature).The two clusters take turns filling, or split collections or times of day between them.Priority fees stay low (nobody races).
Something else: incentive-driven	Bid > ask, so margins are negative.Volume spikes around points or airdrop snapshots.Bids pile into the reward-maximising "risk" tiers.
Something else: wash	Shared funding wallet.The same token trades three or more times, bouncing between A and B.Traders ignore traits (they treat NFTs as fungible).ETH returns to the buyer within the same transaction or block.
Something else: Edgeworth-style cycles (my inference)	A sawtooth: margins get undercut tick by tick toward the cost floor, then one side resets upward and the other follows.This shows up as periodic jumps in the margin time series, not a flat line.

How to run the 30-day test

1. Pull Blur and Seaport sales for one floor collection from nft.trades on Dune, or from Seaport, Blur and Blend events.
2. Build the wallet clusters and drop wash-flagged trades.
3. Pair each buy with the next relist or sale by the same cluster to get the realized margin.
4. Compare the margin distribution with cost, look at response lags to rival bid changes (Blur bids are off-chain, so this needs Blur API snapshots), and check for sawtooth or punishment patterns.

If margins sit on the cost floor and bots respond within a block, it's competition. If margins stay above cost with punishment patterns, it's collusion. If margins are negative or trades loop back to shared funding, it's the incentive or wash outcome. Published evidence points to that last one.

---

Sources offered but not confirmed by this machine:

• https://www.theblock.co/post/217961/blur-bid-incentives-skew-market-lead-to-offers-on-nfts-above-buy-now-prices — did not return a page
• https://etherscan.io/block/13508785 — did not return a page

Verdict: another outcome — the claimed equilibrium is not identifiable from chain-only evidence. The observable mechanism is discrete batch inventory clearing, not evidence that two identifiable swarms either competed margins to zero or sustained tacit collusion.

Why that matters: Seaport permits signed/off-chain orders as well as on-chain validation, so Ethereum reveals fills and some cancellations, not the complete contemporaneous bid/ask book or every “bid just under / relist higher” replacement. A wallet is also not a swarm identity. Thus, a fill cannot establish that both sides were bots, independent operators, or leaving a particular spread unexploited. Seaport documentation explicitly distinguishes on- and off-chain liquidity; its fulfilment event records a completed order, not intent or bot ownership. OpenSea event docs

A concrete Ethereum example within the 30-day window is transaction 0x20c2…6347, block 25,950,338, 2026-09-10 23:41:35 UTC. One caller used fulfillAvailableAdvancedOrders to acquire 23 Spiral Machines ERC-721s from two seller addresses in one settlement. The receipt shows seller payments of 0.37026 ETH and 0.016335 ETH, an OpenSea-fee payment of 0.003905 ETH, and a 0.00674 ETH return to the caller. That is a 23-item sweep with a settlement total of 0.3905 ETH before the caller’s refund—not a sequence of observable one-tick retaliatory relists. Transaction receipt

This is compatible with automation, but does not prove bot-to-bot trading: the buyer’s batch call could be submitted by a human, and the two sellers could be people, bots, or wallets under common control. Calling it collusion or zero-margin competition would overstate the evidence.

The distinguishing signatures should be measured from a captured order-book feed plus receipts, address-clustering evidence, and fee/gas accounting:

Hypothesis	Observable signature
Zero-margin competition	Net executable flip margin—sale proceeds less royalties/marketplace fee, gas, and financing—converges near zero; best bid/ask distance compresses to the minimum tick plus unavoidable costs; fast cancel/replace responses follow each rival quote; frequent failed/reverted fills or priority-fee races.
Tacit collusion	A persistent, economically material net spread remains after fees and gas across many independent inventory cycles; quote changes are slow and synchronized or alternating; an undercut is not immediately matched; fill allocation appears to take turns and both clusters retain profitable quotes. A wide gross spread alone is insufficient because royalties and marketplace fees can create it.
Discrete inventory/batch regime — the evidence-supported alternative	Long quiet periods and stale signed orders, then clustered multi-order sweeps; price/quantity jumps rather than smooth tick-by-tick convergence; inventory concentration and trait/collection selection drive fills; realized margins vary by batch. The 23-item transaction above is a direct signature of this mechanism, though not proof it dominates all trading.

So the defensible conclusion is not “zero” or “collusion.” It is: Ethereum settlement data alone cannot identify a two-swarm equilibrium; the inspected recent evidence instead shows batched, discontinuous execution. To decide between the first two hypotheses, retain timestamped off-chain best bids/asks and signed-order cancellations, map wallets to independently controlled clusters, then test net margins and response times.

---

Sources offered but not confirmed by this machine:

• https://docs.opensea.io/docs/seaport-events-and-errors — did not return a page
• https://ethereum.routescan.io/tx/0x20c2bb0bce01c69c21817145ff78b174e6fa1ab4dd656bb89bbdd7be24e06347?chainid=1 — did not return a page
• https://ethereum.routescan.io/tx/0x20c2bb0bce01c69c21817145ff78b174e6fa1ab4dd656bb89bbdd7be24e06347?chainid=1 — did not return a page
• https://ethereum.routescan.io/tx/0x20c2bb0bce01c69c21817145ff78b174e6fa1ab4dd656bb89bbdd7be24e06347?chainid=1 — did not return a page

Verdict: something else, not either of the two clean cases

The evidence doesn't support zero-margin competition, and it doesn't support stable tacit collusion either. What I see looks like competition capped by the price grid. On Ethereum floors, bots stop cutting price at a floor set by a coarse price grid, gas and the risk of holding inventory. Once they reach that floor, they compete on things other than price: speed, place in the bid queue, and how much capital they commit. On top of that sits a lot of trading driven by marketplace rewards and seller dumping, which makes much of the "bot-to-bot" volume look unlike real flipping.

How far the evidence goes: I did not run a full 30-day scan of every Blur and Seaport fill. I had no indexer or SQL access, only Etherscan pages and published research. My on-chain evidence is a live sample of Blur Marketplace 3 activity from 25 Sep 2026 (blocks ~26,055,400–26,056,160), plus the literature. Take the verdict as a reasoned hypothesis with checkable signatures, not a measured result.

Why the two textbook outcomes don't fit

Why the margin doesn't reach zero. In the sample, bids are filled at round prices. Wallet 0x9CD14Fa7…493364 hit four TOPIA Worlds bids in tx 0x9c677aa838451511abd4506a3490f3de1a118d3183ea068e63d4d8b7106aafa6 (block 26056144). Three filled at 0.01 ETH each and one at 0.02 ETH, to four different counterparties. The same wallet's tx 0x618fdd5a…c2e4d4 (block 26056142) sold four more tokens at 0.01–0.05. Several independent bidders sitting at the same round price is what you'd expect when the price step is too coarse to undercut. An "outbid by one tick" move then costs a meaningful share of the margin, so bots stop competing on price and compete for queue position instead. Each fill also cost ~$2.1–2.3 in gas (0.00077–0.00084 ETH at 2.34 gwei). On a 0.01 ETH ($27) floor, that alone is ~8% per leg. So gas sets a hard floor well above zero for cheap collections.

Why this isn't stable collusion. Bidders are visibly fighting over order state. Counterparty 0x999994ff…6dB7 (837 txs) shows the typical order-management pattern:

• repeated Cancel Trades on Blur (blocks 26055667, 26055995);
• Cancel on Seaport 1.6 (blocks 26054724–5);
• deposits into Blur's bidding pool of 0.99 ETH, then 1.4 ETH, then 1.4 ETH (blocks 26055384, 26055578, 26055627);
• a Set Approval For All on TOPIA Worlds (block 26055968) about 30 minutes before TOPIA fills landed in its wallet. That approval is what lets a bot relist what it buys.

Paying gas to cancel orders on-chain is a cost you only accept if rivals punish stale quotes. That fits live competition, not a comfortable shared spread.

What most of the "flow" actually is. The active seller in my sample, 0x9CD14Fa7…, is an EIP-7702-delegated account with only 79 txs. Its sequence was:

1. deploy two contracts (blocks 26056011 and 26056055);
2. make many calls with an unidentified selector 0x99ee63c0 to other contracts;
3. dump NFTs into standing Blur bids in batches, using Take Bid / Take Bid Single;
4. withdraw from Blur's bidding pool within a few blocks (for example, block 26056149 right after the 26056144 fill).

That is a liquidation into the bid pool, not a flip. The bidding bots earn their spread from sellers like this, not mainly from each other. Where true bot-to-bot trading does happen, research shows that incentive schemes dominate it. A Boston University study found "Approximately 38% of NFT trades and 60% of the total traded value across several major exchanges showed patterns consistent with wash trading". It tied this to platforms like Blur and LooksRare "offering 'rewards' tied to trading volume" (BU HIC).

What theory predicts. Algorithmic collusion is fragile and depends on how the bots are built. Douglas, Provost & Sundararajan find that "naive algorithmic collusion never emerges when both agents use a broad class of persistently random algorithms". It "always emerges" with deterministic learners such as UCB (arXiv 2411.16574). Two swarms of simple, deterministic, identical rule-based bots ("bid one tick under, relist one tick over") are the risky case for collusion. Swarms that randomize or keep exploring are not. Separately, Bose et al. document that automated high-frequency NFT strategies make most of their money from ordinary users, not from rival bots (arXiv 2310.06844).

Signatures that tell the outcomes apart

For each swarm, first group wallets by common funder, shared deployer, and matching approval and cancel timing. Then measure:

Signal	Zero-margin competition	Tacit collusion	Price-grid-capped competition (my verdict)	Reward farming / wash
Round-trip margin (resale minus purchase, minus gas and royalty)	Converges to ~0	Stays well above cost, stable	Stays at about 1 price step plus gas; doesn't shrink further	Near zero or negative, paid for by rewards
Bid and listing price levels	Change every few blocks, 1-step undercuts	Sit on round numbers for long stretches without being overbid	Pile up at the same round level (e.g. many bidders at 0.01)	Off-market prices, often above floor
On-chain Cancel / Cancel Trades rate	High, keeps going	Low	High around moments when fills happen	Low
Who gets filled	Fragmented, whoever is fastest	Swarms take turns or split collections and traits	Queue-ordered; the swarm with more capital deployed takes a bigger share	Same wallets on both sides
Response to a price cut	Rival matches or beats it immediately and stays there	Brief sharp price war, then a return to the old spread (punishment)	Rival adds capital or reposts at the same level; price rarely moves	None
Price path over time	Flat at cost	Flat above cost	Sawtooth or flat at the grid floor	Spikes when rewards are paid
Flow of the same token	A → bot → buyer, one direction	Same	Same	A → B → A, cycles, shared funder

Two signals matter most:

• Collusion vs. competition: whether rival swarms show a punish-then-return pattern after one side undercuts, or whether margins that are stable and above cost persist even when cancel activity is high.
• Price-grid capping vs. true zero margin: whether the realized margin sits exactly one price step above gas, however many bidders join. If adding bidders only lengthens the queue and never lowers the price, the grid, not collusion, is keeping the margin up.

To turn this into a 30-day measurement

1. Pull every Blur Marketplace 3 (0xb2ecfE4E…CEA5) and Seaport 1.6 fill from roughly block 25,840,000 onward, plus Blur Bidding deposits and withdrawals.
2. Group wallets by funder, deployer and approval timing. Tag a wallet as a bot if it has many cancels, keeps its Blur Pool balance topped up, and relists within N blocks of buying.
3. For each bot, compute realized round-trip margin minus gas and royalty. Plot it against how many rival bidders sit at the same price level.
4. Test for punishment: after a 1-step undercut, check whether the old price level is restored within K blocks.

The sample above shows pieces of the price-grid-capped pattern: identical round-price bids across several counterparties, high cancel activity, and bots making money from a liquidating seller. It shows no evidence of a coordinated spread.

---

Sources offered but not confirmed by this machine:

• https://etherscan.io/address/0xb2ecfE4E4D61f8790bbb9DE2D1259B9e2410CEA5 — did not return a page