Loading
This site publishes several counts of what looks like the same thing.
Each is correct and each answers a different question, and a reader comparing two pages cannot tell a legitimate narrowing from a disagreement. So the narrowing is published: what was discovered, what survives each stage, and what drops out at every one.
The stage keys are the population names.
Which vaults is a capital figure a total of?
dropped here: recorded as shut, and kept rather than deleted
dropped here: no price we trust for the underlying, or a venue that publishes what was put in rather than what is held now - the size is refused rather than guessed
dropped here: no evidence that software decides, which is not the same as evidence that a person does; a further 926 software-managed vaults carry no value and so are not in this nested count
1,721 of 4,149 vaults (41.5%) reach software-managed, of the valued. The other 2,428 drop out at a named stage above, never silently.
Of the 1,287 open vaults with no capital figure, 349 do carry one we measured: $45.00M of net deposits, read from the vault accounts themselves.
That is what depositors put in, not what is held now, and no total on this site adds the two together - so every capital figure here excludes it. It is on each vault's own page.
How many open vaults show evidence that software decides, whether or not we can put a size on them?
dropped here: no evidence that software decides, which is not the same as evidence that a person does
2,647 of 3,930 vaults (67.4%) reach software-managed. The other 1,283 drop out at a named stage above, never silently.
1,721 of them are valued and 926 are not; the valued ones are the last step of the chain above.
Which vaults does the withdrawable figure cover?
dropped here: one side of the register could not be read, so the vault sits in none of the three exit states
Every one of the 652 lending vaults discovered survives to supply and borrow both read.
Which vaults is the real-world figure a total of?
dropped here: the vault names no underlying asset, so there is nothing to classify
dropped here: we hold the symbol and could not place it in any class, which is ours rather than theirs
dropped here: the underlying is ether, bitcoin, a chain token or a staking receipt, all crypto, and nobody is uncertain about that
794 of 3,927 vaults (20.2%) reach in a real-world class. The other 3,133 drop out at a named stage above, never silently.
How much of what agents earn can we follow?
dropped here: paid in something we hold no price for at that moment
dropped here: it moved somewhere we do not read, such as an exchange, an unlinked treasury or a chain outside our perimeter. Ours, not theirs.
79,042 of 271,523 earners (29.1%) reach money we can follow. The other 192,481 drop out at a named stage above, never silently.
What is a completion rate on the job book a rate of?
dropped here: we never saw a phase for it, which is a hole in our reading of the book rather than a job that did nothing
dropped here: still in request, negotiation or evaluation, which is not yet a result and is never counted as a failure
204,639 of 413,905 jobs (49.4%) reach a recorded outcome. The other 209,266 drop out at a named stage above, never silently.
Why does the same rail carry four different totals?
dropped here: the pairing of an authorisation to a transfer is one we no longer trust, so the row is kept for re-reading and never published in a total
dropped here: no block timestamp, so it could not be placed in a window. Today that is none of them
dropped here: neither is in the authorisation event and the paired Transfer could not be found, so both are null rather than guessed at
dropped here: no amount could be resolved, so there is nothing to compare with the seller’s own listing
Which vaults does a control finding cover?
dropped here: the venue publishes no access control, so there are no roles to read, which is most venues, and is our limit rather than a finding about the vault
511 of 3,930 vaults (13.0%) reach control we could read. The other 3,419 drop out at a named stage above, never silently.
A small share by count and 96% by capital.
The venues that publish access control are the large ones, so the count understates the reach and the capital overstates it. Neither figure alone is the answer.
Whose return is the up-or-down split a split of?
dropped here: we hold no value on both sides of their entry, so there is nothing to measure a return from
3,800 of 8,354 positions (45.5%) reach return reconstructable. The other 4,554 drop out at a named stage above, never silently.
How many vaults does a firm we infer to be automated run?
dropped here: nobody has merged them to a firm or a leader, which is a limit of our linking rather than a vault that runs itself
dropped here: the firm is inferred to have a person deciding case by case, or no mode could be inferred at all
460 of 3,930 vaults (11.7%) reach that operator runs without a human per decision. The other 3,470 drop out at a named stage above, never silently.
177 of those, holding $1.84B, carry no evidence of their own and are therefore NOT in this site's software-managed count.
Two readings on two bases: that one is about the vault, this one is about the firm. They are never added, and neither overrules the other. A vault promoted on its operator's evidence would be the mistake the classifier refuses in the other direction.
Which operators is a firm figure a total of?
dropped here: every vault we linked them to is shut, or none was ever linked
460 of 542 operators (84.9%) reach running an open vault. The other 82 drop out at a named stage above, never silently.
How many registrations are economically usable?
dropped here: the card is a pointer we have not fetched, or it could not be read at all
dropped here: the card names no endpoint, so there is nothing to ask
dropped here: the host refuses, has gone, or would not answer in time
11,289 of 1,715,499 registrations (0.7%) reach that endpoint answers. The other 1,704,210 drop out at a named stage above, never silently.
A figure’s evidence block names one of these.
The ones marked as decomposed narrow above; the rest are named and measured but not yet broken into stages, which is a smaller claim rather than a missing one.
Endpoints somebody has declared: the distinct hosts registered identities publish, and the resources x402 sellers list in a catalogue.
We ask what is declared rather than crawling for services, so an endpoint nobody declared is outside this population rather than missing from it. The two halves are counted apart and never summed.
Every vault this platform has ever found, open or shut.
The widest count there is.
Discovered vaults not recorded as shut.
Closed ones are kept rather than deleted, which is what makes survivorship measurable.
A stage that loses rows is not a stage that failed.
Refusing to value a vault whose underlying we cannot price is the behaviour this platform is built around, and the count of refusals is a measure of honesty rather than of coverage. What would be a failure is a figure published without saying which stage it was computed at, which is the thing these chains exist to make impossible.
How each figure resolves to the read behind it: evidence. Where our figures disagree with somebody else’s: what others report. What we read at all: the perimeter.
24,151,958 of 24,159,874 settlements (100.0%) reach checkable against a price. The other 7,916 drop out at a named stage above, never silently.
Mostly because of WHEN, not what.
Every rung here was counted in one pass at 5 November 2026, and the two narrowings below the top are both small. The rail takes on more than ten thousand settlements a day, so two pages reading it hours apart differ by more than any predicate on this chain does. That is what an outside audit found on 25 September, reading four totals off four pages and unable to tell a narrowing from a disagreement. The figure beside each of those pages carries its own snapshot, and this chain is what they are snapshots OF.
Discovered vaults recorded as shut.
The complement of vaults_open, and the half a survivor list leaves out: count(vaults_open) + count(vaults_closed) = count(vaults_discovered) over one snapshot.
The vaults a caller said they hold, which is a population WE DID NOT CHOOSE and cannot verify.
It is not a sample of anything: it has no denominator in the market, a holding we do not hold at all is reported in `notHeld` rather than dropped, and nothing can be said about what the caller did not name. Every share computed over it is a share of their list.
Open vaults we can put a size on, which is the denominator behind every capital total here.
A vault whose underlying we cannot price is published unvalued rather than estimated; one we read as holding nothing is valued at zero, which is a reading rather than a gap.
Open vaults with evidence that software decides.
Absence of evidence is not evidence that a person decides.
Open vaults we can put a size on AND with evidence that software decides: the software-managed vaults nested inside vaults_valued.
Automation and valuation are two axes, so the software-managed vaults without a size are not here; they are in vaults_agent_managed.
Open vaults on a lending venue, whether or not their register could be read.
Lending vaults where both sides of the register were read, so the vault can be placed in exactly one exit state.
Open vaults whose underlying we could place in a class at all.
A symbol we hold and cannot classify is our gap rather than theirs.
Every ERC-8004 identity registration read from a chain, whatever it points at.
Registrations whose card we fetched and parsed.
A card that is only a pointer we have not followed is a gap in our reading rather than a defect in the registration.
Registrations whose card names a service endpoint, so there is something to ask.
Registrations whose declared endpoint answered the last time it was asked.
Addresses paid on an agent-commerce rail, which is the population every revenue figure is a total of.
Addresses paid on x402 as a seller that lists a paid resource, on ACP as a provider, or on the mech marketplace: the population the counterparty check reads.
Narrower than rail_earners on x402, where a transfer to a recipient that lists nothing is not counted as a sale. Not drawn as a chain beneath rail_earners, although it is one, because the two counts come off different nightly jobs and a chain whose stages are read at two moments cannot be checked at all.
Earners whose payment we could put a dollar figure on at the moment it was made.
One paid in an asset we hold no price for is counted and not valued.
Earners whose money we can follow to any outcome at all.
The rest is a statement about our linkage rather than about the agents: revenue moved to an exchange, to an unlinked treasury, or onto a chain we do not read looks identical from here to revenue that never moved.
Events on a rail where an agent did something for somebody: jobs, tasks, payments, channels and escrows.
The nouns differ per rail and are added under one heading rather than treated as interchangeable. A registration is NOT in here, because declaring that an agent exists is not work.
Registrations and ratings: somebody declaring that an agent exists, or rating one that does.
Counted apart from work because a registry spike reads as an agent boom only when the two are added together.
Services listed in the x402 catalogue that publish a price.
Depositor positions read from the venues, one row per address per vault, latest reading only.
The widest depositor count there is.
Depositor positions whose entry and value history we hold on both sides of the entry, so a return means something.
A position we can see but cannot date is counted and not measured: any return for it would be a guess about when the money arrived.
Vaults whose deposits and withdrawals we hold over the window, so growth can be split from performance.
Every open position on GMX across both chains, counted rather than sampled.
Open positions on Ostium in a page the venue serves, which is a floor rather than a total.
Pairs Avantis lists, with the venue’s own open interest against each.
Open vaults we have linked to a firm or leader.
A vault with no operator here is one nobody has merged to a name, which is a limit of our linking rather than a vault that runs itself.
Open vaults whose OPERATOR we have inferred runs without a human deciding each case.
A second reading of who runs a vault, on a different basis from the vault’s own evidence: this one is about the firm, that one is about the vault. They are never added together and neither overrules the other.
Firms and leaders running at least one open vault we read.
An operator nobody has linked to a vault is not here, which is a limit of our merging rather than a fact about the market.
Open positions we read from the perpetual venues themselves, per vault and per instrument.
A venue we do not read carries positions that are absent here rather than zero.
Instruments whose mark, mid and impact prices we record on a schedule.
An instrument nobody in our cohort holds is not snapshotted, so its absence says nothing about the market.
Open vaults whose own contract told us what they are denominated in.
A vault that does not answer is unread rather than unbacked.
Open vaults whose SETTLED withdrawals we have read, which is a smaller set than the open vaults and a different one from the vaults that publish a withdrawable figure.
A vault outside it has never been read rather than never been withdrawn from, and the two must never be counted together: an empty record and an absent one answer opposite questions about whether money can leave.
Open vaults whose venue publishes access control and whose roles we could read.
Most venues publish none, so this is a small share by count and a large one by capital: the venues that publish it are the large ones, and neither figure alone is the answer.
Every operator record we hold, including those whose vaults have all closed.
Kept rather than deleted, which is what makes a firm’s whole book readable.
Open vaults whose underlying is a named real-world instrument or a stablecoin whose issuer invests reserves.
Crypto is classified and deliberately outside this: nobody is uncertain whether ether is a Treasury bill.
Open vaults holding a contract we can ask directly how much it holds, so the size is read from the chain rather than taken from a venue.
A vault on an order book has no such contract and is absent here rather than unproven; the confirmed share is published against the askable ones, never against every vault.
Vaults eligible to be index constituents at a rebalance: measurable in that period and above the size floor.
Chosen using only what was known at the time, so a vault that later died keeps its loss and one that later grew does not appear early.
Strategies declared to us before the result was known, including the ones that went on to fail.
Small and deliberately self-selected — the value of pre-registration is that it cannot be chosen after the fact, not that it is representative of anything.
Jobs on the Virtuals ACP v2 contract we read end to end.
A job still in negotiation has not failed and one whose phase we never saw is our gap rather than an outcome, so neither is counted against a completion rate.
ACP jobs we saw a phase for.
A job read from the contract whose phase we never observed is a hole in our reading of the book, not a job that did nothing.
ACP jobs that reached an outcome — delivered, refused or expired.
This, and not the whole book, is the denominator under every completion rate we publish: a job still in request, negotiation or evaluation has not failed and counting it as a failure would defame the provider.
Authorised transfers not held in quarantine.
A quarantined row is one whose pairing of an authorisation to a transfer we no longer trust — it is kept so the reading can be re-done, and is never published in a total.
Published settlements carrying a block timestamp, which is what places one in any window or hour-of-day.
Loses nothing today: every row we hold is dated, so any figure filtering on it is the same population under a longer predicate.
Settlements with a recipient and an amount, which is what the pre-flight price check compares against the seller’s own dearest listing.
Equal to the population above since the amount backfill finished, and kept as its own key because the two answer different questions and will part again the moment a transfer cannot be paired.
Dated settlements whose recipient and amount were both recovered from the paired Transfer.
Neither is in the authorisation event, so a transfer that could not be found leaves both null — and this is the population every figure about a seller, a price or a dollar total is over.
Addresses that paid on x402 over the blocks we hold, separated by observable behaviour rather than by any label they carry.
No class claims agency: paying every hour of the day is a fact about a schedule, not about who wrote it.
Open vaults carrying a published risk score, which is what the six dimensions are read from.
A score needs history and a readable underlying, so a vault missing here is one we could not measure rather than one that scored badly, and the dimensions table is silent about it rather than drawing it as safe.
Vaults read as software-run that also carry a risk score.
A score needs history and a readable underlying, so the shortfall against vaults_agent_managed is our reading rather than a judgement about the vault.
Chains where at least one open vault is read.
NOT the count of chains this platform touches, which is larger and includes chains carrying only a registry, a payment rail or a service catalogue. The two differ by eleven and an audit once read them as a contradiction.
Deposit and withdrawal EVENTS read from chain, one row each.
A count of movements and not of vaults or of depositors: a single depositor entering and leaving is two.
Open vaults where a realised return can be set against an advertised one: both readings exist and the realised window is long and dense enough to compare.
A SMALL MINORITY of open vaults, because a venue that quotes nothing, or a share price that never moves, is excluded rather than counted as in line. A short list here therefore means few comparable and never few that performed.
Open vaults where we can measure what DEPOSITORS kept, rather than what a unit held throughout would have earned.
It needs per-depositor equity over time, which only some venues publish, so this is a narrow set and its narrowness is about those venues and not about the vaults. A count here is NOT comparable with a count of vaults whose share-price return can be set against an advertised rate: different question, different evidence, and dividing one into the other means nothing.
Recorded changes to anything this platform tracks over the last twenty-four hours.
An event count over a rolling window, so it is never comparable with an all-time total on the same strip.
Distinct addresses holding at least one ERC-8004 registration.
Fewer than the registrations, because one owner can register many agents, and an owner is not an operator: nothing here says a person or a firm is behind it.
Registrations whose agent card was held inline on chain and could be read.
A registration pointing at a card we could not fetch is counted as a registration and not as a card, so this is a floor on what is declared.
Requests on the Olas mech marketplace that were PAID, one row each: the settled rung.
Deliveries are a separate population and the two are never added, because one task produces up to both. Narrower than agent_work_events, which sums every rail under one heading.
Distinct addresses that delivered at least one mech request.
Counted from delivery rows, so a supplier paid but never seen delivering is absent.
Distinct providers named on the ACP job book.
Listed rather than active: a provider with one expired job from last year is in this count.
Cardano agents whose Masumi registration token has not been burned.
Live means the token exists, which is a claim about registration and not about whether the agent answers.
Every Cardano agent ever registered on Masumi, live or deregistered.
The denominator behind any share of them, kept because a deregistration rate needs one.
Masumi escrows ever opened between Cardano agents, one row each.
An all-time event count, and not the number currently open, which is smaller by three orders of magnitude and was published as "open" for a time.
The readers this platform runs on a schedule, which is the population behind every freshness claim here.
It measures OUR jobs and not the sources they read: a reader that ran ten minutes ago may still have been served stale data by a venue.
Agent-commerce rails we read end to end.
A rail absent here is one we do not read, which is different from a rail with no activity.
Every x402 payment this platform has read off a chain, all time.
An EVENT count: one address produces many, so it is never added to a count of addresses and never compared with a count of agents.
Distinct hosts named by an agent card that we asked for an answer.
Hosts, not agents: thousands of agents can name one host, which is why the agent count behind an outcome is a separate population.
Agents whose declared service lives on a host we probed, counted per agent rather than per host.
The two differ by an order of magnitude and reporting one as the other overstates how much of the market a single dead host takes down - or understates it.
Ratings written into an ERC-8004 reputation registry that we have read.
An EVENT count. A rater can write many, which is the finding rather than a caveat.
Agents carrying at least one rating.
A THING count, and deliberately not the number of ratings: most of these carry exactly one, from one rater.
Service endpoints an agent card declares, counted per endpoint.
What an agent SAYS it can do, never evidence that any of it answers - that is agent_hosts_probed.