Skip to content

Open numbersLive · read

Vestiarion, in production

Who uses Vestiarion, what its agent decided, what stopped it, and what settled. Every figure is read from Vestiarion's production database, with nothing entered by hand. The Quality Lab's result below is not read from the database: it comes from the tests that run its scenarios, and replay the team's own exported shadow-mode decisions, on every change.

Figures are kept a minute between reads, and can be older after a quiet spell: the time above says when they were read. Days are UTC.

Arc testnet · Last 7 days

Customer workspaces
1
6 live now, 0 made a first payment
Payments settled
14
None by customers, 8.40 USDC in all
Invoices paid on time
100%
14 of 14, every workspace
Median time to a customer's first payment
—
From opening a workspace to its first settled payment
  • Customers counted apart

    Workspaces the team runs to build and test the product are counted apart from customers', and every figure says whose it is.

  • One network at a time

    Arc mainnet and Arc testnet are each counted on their own, and never added together.

  • Checkable on chain

    Our own payments link to the explorer. Customers' are counted, never listed.

Arc testnet

Payments on Arc testnet.

Test USDC

Latest payment settled on

Customers

Customers on Arc testnet

Workspaces opened on Arc testnet by people outside the Vestiarion team. Each figure has the total with our own workspaces beneath it.

Customer workspaces opened

1

Each one opened by someone outside the team.

With our workspaces: 1

Customer workspaces livenow

6

Running on Arc testnet as this page was read.

With our workspaces: 13

Reached a first payment

None

Customer workspaces whose first payment settled.

With our workspaces: 0

Payments settled

None

Confirmed by Circle on Arc testnet.

With our workspaces: 14

USDC paid

—

No customer payment on Arc testnet in this period.

With our workspaces: 8.40 USDC

Median time to first payment

—

From opening a workspace to its first settled payment.

With our workspaces: —

Customers' payments are counted here, never listed one by one.

All workspaces

How the agent performs

What the agent did with the payment decisions it made, and how those payments turned out. Customers' share is beneath each, except on flags, where customers' verdicts lead: in our own workspaces a person is testing the product, not judging the agent.

Decided by the agent itself

65%

15 of 23 payment decisions

Paid, sent or scheduled without waiting for anyone; the rest went to a person.

Customers: 50% (1 of 2)

Invoices paid on time

100%

14 of 14 invoices paid

Settled on or before the invoice's due day.

Customers: —

On time, no person involved

100%

14 of 14 invoices paid

Nobody approved, rejected or returned the invoice.

Customers: —

Agent flags customers upheld

—

No customer has decided a flag yet

The agent said not to pay, and the customer who decided agreed.

With our workspaces: —

All workspaces

What stands between the model and the money

A model never moves money on its say-so. Every payment decision passes the same checks before money moves, and each is counted here, the uncomfortable ones too.

  1. Model

    A model proposes

    718

    decisions made by a model

    In 695 agent cycles, deciding 22 invoices.

  2. Policy

    Compared with the written policy

    42

    times the model disagreed

    The written rule-based policy decides every case beside the model, so each disagreement is recorded, not hidden.

  3. Code

    Hard limits in code

    5

    decisions refused by code

    A decision that breaks a hard limit is refused, whatever proposed it. 0 duplicate invoices caught before payment.

  4. Person

    A person when it matters

    8

    decisions escalated to a person

    1 resolved by a person so far.

  5. Chain

    Settles on Arc testnet

    14

    payments settled

    0 of them paid contractor milestones.

Activity

Settled payments by day

No customer payment settled in this period.

All activity shows our own workspaces' payments as well.
  • Customers
  • Our workspaces
012Nov 12, 2026: 0 by customers, 0 by our workspaces (0.00 USDC)Nov 12Nov 13, 2026: 0 by customers, 2 by our workspaces (1.20 USDC)Nov 14, 2026: 0 by customers, 2 by our workspaces (1.20 USDC)Nov 15, 2026: 0 by customers, 2 by our workspaces (1.20 USDC)Nov 15Nov 16, 2026: 0 by customers, 2 by our workspaces (1.20 USDC)Nov 17, 2026: 0 by customers, 2 by our workspaces (1.20 USDC)Nov 18, 2026: 0 by customers, 2 by our workspaces (1.20 USDC)Nov 19, 2026: 0 by customers, 2 by our workspaces (1.20 USDC)Nov 19
Days are UTC. Hover a day for its figures.
Show the days as a table
DayCustomersOur workspacesUSDC
Nov 13, 2026021.20
Nov 14, 2026021.20
Nov 15, 2026021.20
Nov 16, 2026021.20
Nov 17, 2026021.20
Nov 18, 2026021.20
Nov 19, 2026021.20
Verify it yourself

Payments you can check on chain

The latest payments our own workspaces settled in this period, each linked to the explorer. Customers' payments are counted above and never listed here.

Show 9 more
For auditors

Every figure

Customers, our own workspaces and the total, side by side, for last 7 days.

Show every figure for customers, our workspaces and the total
Every figure, Last 7 days. Rows marked now are totals at the moment of reading.
FigureCustomersOur workspacesTotal
Adoption
Workspaces opened101
Workspaces livenow6713
People in workspacesnow6612
Workspaces that made a first payment000
Median time from workspace opened to first payment———
Money moved
Payments settled01414
USDC paid0.008.408.40
Payee wallets paid022
Contractor milestones paid000
USDC in walletsnow67.801,074.301,142.09
Agent activity
Agent cycles run303392695
Decisions made by a model305413718
Invoices decided12122
Payment decisions the agent carried out itself11415
Payment decisions it escalated to a person178
Decided by the agent itself50%67%65%
Escalations a person resolved101
Outcomes
Invoices paid on time—14 of 1414 of 14
Paid on time with no person involved—14 of 1414 of 14
Agent flags a person upheld———
Shadow mode decisions a person agreed with———
Safety and controls
Model disagreed with the written policy14142
Decisions refused by code055
Duplicate invoices caught000

Arc mainnet

Payments with real money, on Arc mainnet.

Real USDC

No payment has settled on Arc mainnet in this period.

Its figures show here in full once one does.

Workspaces live
3
Customers' workspaces live
0
Agent cycles run
168
Payments settled
0
Show every figure
Every figure on Arc mainnet, Last 7 days. Rows marked now are totals at the moment of reading.
FigureCustomersOur workspacesTotal
Adoption
Workspaces opened000
Workspaces livenow033
People in workspacesnow022
Workspaces that made a first payment000
Median time from workspace opened to first payment———
Money moved
Payments settled000
USDC paid0.000.000.00
Payee wallets paid000
Contractor milestones paid000
USDC in walletsnow0.000.790.79
Agent activity
Agent cycles run0168168
Decisions made by a model000
Invoices decided000
Payment decisions the agent carried out itself000
Payment decisions it escalated to a person000
Decided by the agent itself———
Escalations a person resolved000
Outcomes
Invoices paid on time———
Paid on time with no person involved———
Agent flags a person upheld———
Shadow mode decisions a person agreed with———
Safety and controls
Model disagreed with the written policy000
Decisions refused by code000
Duplicate invoices caught000

Quality Lab43 of 43 scenarios right

Each scenario runs through the real decision code on every change, with the model's answers scripted, often to what an attacker would want.

55 real decisions from the team's own shadow-mode workspaces, replayed through today's code: 46 reach the same outcome, 9 cannot be rebuilt from the record, 0 wrong payments.

0 wrong payments
  • Prompt injection8 of 8
  • The payee and its address8 of 8
  • Screening3 of 3
  • Limits and approvals4 of 4
  • Matching and duplicates6 of 6
  • Workspace switches6 of 6
  • When the model says no2 of 2
  • Sending money once3 of 3
  • Clean bills3 of 3

A scenario is right when money moved or stopped as expected, with the expected status and rule.

A wrong payment is any transfer or schedule where money should stop, a transfer to an address other than the one on file or for an amount other than the bill's, or a second transfer for one bill.

Every scenario and its result →Every replayed decision →

Methodology

How the figures are counted

The same definitions hold for every figure of both networks, for every period. The Quality Lab's result above states its own.

Networks, periods and freshness
  • Arc mainnet and Arc testnet are counted apart, and nothing is ever added across them. A workspace belongs to one network, chosen when it goes live and kept from then on; a payment belongs to its workspace's network.
  • Rows marked now are totals at the moment of reading; the others cover the period chosen above.
  • A deleted workspace takes its activity out of these figures with it.
  • Figures are kept for a minute after they are read. The first visit after a quiet spell is shown the figures last read while newer ones are read for the next, and the time at the top of the page says when they were read.
Customers and our workspaces
  • A workspace is a customer's when someone outside the Vestiarion team opened it. The team's own workspaces, and any whose creator has since deleted their account, are counted as ours; only workspaces the team opened list their payments.
  • A person counts as a customer when they are not on the team and belong to a customer's workspace.
  • Sample data that a workspace loads to try the product never counts.
Payments and money
  • A payment counts once Circle confirms it on its network. A simulated payment, which a sandbox makes until it pays through a wallet, never counts.
  • A workspace's first payment is the first one Circle confirms on its network, and the time to it runs from when the workspace was opened. A workspace holding payments from before it was opened counts, but not towards the time. A dash means there is no time to measure in the period.
  • A contractor milestone counts once a settled payment has paid it.
  • USDC in wallets is the USDC in live workspaces' Circle wallets on that network, as last read from the chain.
The agent's decisions
  • A payment decision is the agent's decision on an invoice to pay or a contractor milestone, counted by what happened. Paid, sent or scheduled is carried out by the agent itself; held, flagged or waiting for information is escalated to a person. A payment a model proposed and code refused counts as escalated.
  • The written rule-based policy decides every case beside the model. The model disagrees with it when the two choose a different action; the model's choice still has to pass the same checks in code.
  • A decision is refused by code when a hard limit blocked it, whether a model or the written policy proposed it.
  • A flag is the agent saying an invoice should not be paid. A person upholds it by rejecting the invoice and overturns it by paying it; sending it back to the agent counts as neither. A hold or a request for information asks a person to decide, so it is not counted as agreement either way. The share of flags upheld leads with customers' verdicts: in our own workspaces a person pays a flagged invoice to test the product, which says nothing about the agent.
  • An invoice is paid on time when its payment settles on or before its due day (UTC). No person involved means nobody approved, rejected or returned it; entering the invoice does not count.
  • A duplicate is caught when the agent stopped an invoice that repeats one already paid or on its way to being paid, with the same purchase order and amount, and it was never paid since.
  • In shadow mode, a workspace keeps paying its bills as it does today while the agent decides on the same bills, and a person agrees or disagrees with each decision before anything is paid, unless an owner lets the agent pay vendors that earned it under their rule, whose payments a person agrees or disagrees with after. A decision a person agreed with counts as agreed; bills of sample payees are never counted.
What is never shown
  • Customers' payments appear only as counts and totals, never one by one or as a day's amount.
  • The database's own error messages are never shown; a network whose numbers cannot be read says so instead.