⌘K

Live evidencecounted 29m ago

0.48days of MEME volume to unwind the crowd99%of the XMR book on one side, across 28 vaults$1.30Bunder one key with no guardian, 52 of the 424 vaults whose control we have read$334.1Mwhose control we have never read, across 3,229 vaults$1.95Bsitting in vaults and not lent out$11.27Bunder 3,653 vaults we read$4.79Bin 323 vaults where we saw automated execution, nothing declared2,080vaults declared or on a venue built to run software, holding $1.01M324software-run vaults scored, 13% of them450operators identified behind them26networks where we read capital, from Ethereum to Cardano31,350,034deposits and withdrawals read from chain5changes recorded in the last day588,846ERC-8004 agent registrations read−0.61/ 30 days360,982distinct owners behind those registrations199,780registrations carrying a readable card, 34%3,337,008paid requests on the Olas rail (deliveries are counted apart)−0.67/ 30 days87suppliers delivering that work411,027agent-to-agent jobs read on Base+0.04/ 30 days$1.0Msettled between agents, all time1,175suppliers listed on the ACP rail189agents live on Cardano76Cardano agents deregistered, 29% of all ever minted29,141escrows between Cardano agents since Jun 2025, 142 still open78%of every agent payment we read comes from 2,571 addresses that never stop307,376addresses paid once and never came back, of 494,774 that ever paid934of those payers are an agent we can name, 0.189% of them635agent hosts answered when we asked, of 2,532 probed17,125agents sit behind a host that answers nothing at all404msmedian answer from the 635 hosts that answer44%of all agent ratings were written by ten addresses, of 671,178 ratings read29,502rated agents have exactly one rater, of 47,110 rated at all86,979service endpoints declared across 33,840 agent cards we read0.48days of MEME volume to unwind the crowd99%of the XMR book on one side, across 28 vaults$1.30Bunder one key with no guardian, 52 of the 424 vaults whose control we have read$334.1Mwhose control we have never read, across 3,229 vaults$1.95Bsitting in vaults and not lent out$11.27Bunder 3,653 vaults we read$4.79Bin 323 vaults where we saw automated execution, nothing declared2,080vaults declared or on a venue built to run software, holding $1.01M324software-run vaults scored, 13% of them450operators identified behind them26networks where we read capital, from Ethereum to Cardano31,350,034deposits and withdrawals read from chain5changes recorded in the last day588,846ERC-8004 agent registrations read−0.61/ 30 days360,982distinct owners behind those registrations199,780registrations carrying a readable card, 34%3,337,008paid requests on the Olas rail (deliveries are counted apart)−0.67/ 30 days87suppliers delivering that work411,027agent-to-agent jobs read on Base+0.04/ 30 days$1.0Msettled between agents, all time1,175suppliers listed on the ACP rail189agents live on Cardano76Cardano agents deregistered, 29% of all ever minted29,141escrows between Cardano agents since Jun 2025, 142 still open78%of every agent payment we read comes from 2,571 addresses that never stop307,376addresses paid once and never came back, of 494,774 that ever paid934of those payers are an agent we can name, 0.189% of them635agent hosts answered when we asked, of 2,532 probed17,125agents sit behind a host that answers nothing at all404msmedian answer from the 635 hosts that answer44%of all agent ratings were written by ten addresses, of 671,178 ratings read29,502rated agents have exactly one rater, of 47,110 rated at all86,979service endpoints declared across 33,840 agent cards we read

Crowding

Does being in a crowd cost anything?

250 events across 159 vaults

Vaults entering a crowd went on to underperform matched vaults that did not by -18.01% at the worst point of the following 20 days. Shuffling the event dates produces a path that extreme 89% of the time, so this is not yet distinguishable from noise.

That share rests on 140 shuffles of the 250 asked for, because the job reached its time limit. It can be read to about 0.7 percentage points and no finer.

What the window is. Every event here is a crowd that formed while we were watching, which is 54 days of cluster history. The returns either side of an event are the venue’s own published account-value history and reach back much further, so the series is long and the study is not: crowds formed before we began have no event date and are not in this table.

Day	Entered a crowd	Matched controls	Difference	Events	Refused
-5	0.00%	-0.25%	0.25%	222	39
-4	0.00%	0.85%	-0.85%	227	29
-3	0.00%	0.53%	-0.53%	230	19
-2	0.00%	0.60%	-0.60%	234	8
-1	0.00%	0.00%	0.00%	232	2
+1	0.36%	0.02%	0.34%	232	9
+2	0.41%	0.58%	-0.18%	227	23
+3	1.64%	4.13%	-2.48%	229

The faded rows refused more readings than they used. A cumulative return outside −99% to +1000% over a few weeks is a broken share-price series rather than a trade, and this study will not compute on one, but a horizon that discards two thirds of its vaults is measuring whoever survived the filter, which is a selection rather than a result. Those rows are shown because hiding them would leave a shorter table and no reason for it; they are not evidence of anything.

A difference is not a cause. Vaults enter crowds because they took a popular position, and a popular position going wrong is one explanation among several. The shuffled dates do not rule out coincidence here: a difference this size appears often when the event dates are random, so nothing on this panel yet separates crowding from chance.

What this cannot tell you

We see on-chain and venue-side only. Anything held on a centralised exchange is invisible, so true crowding is at least as bad as shown, never better.
A quarter of Hyperliquid’s capital is behind these positions. $97.54M of $273.98M on that venue sits in vaults whose positions we hold, 36% by capital against 74% by vault. Most of the rest is flat rather than unread, which a figure here cannot say for itself.
Coverage is usually low. Most instruments here show us holding a low single-digit share of open interest. That is a statement about our reach, not about safety.
Book clusters use single-linkage grouping. Members are chained by pairwise similarity, so a cluster average can sit below the joining threshold. It is transparent and arguable, which we prefer to a clever method nobody can interrogate.
Positions are snapshots. A strategy that flips intraday can look uncrowded at every observation and still be crowded in between.