The second ledger · For AI agents and autonomous machines

Every try
counts.

Machines can prove they paid. We prove they earned it. MIZAN is the second ledger for AI agents and autonomous machines. Before one machine pays, hires or hands work to another, it asks MIZAN: was that result earned? Every try counted, proven without revealing the method, answered with a signed yes or no.

agent → MIZAN  may I pay mizan-flagship?
MIZAN → agent  ALLOW · proven · signed · live
116proofs on the ledger
87claims refused on the record
392,877steps proven
899agent checks from outside addresses
A real counted evaluation · on the ledger

An AI model, 50 tasks declared before the first one ran. 70% first-try. Every attempt counted.

You set the bar. MIZAN answers against the whole count, not the best try.

60% ALLOW
Computed in your browser from the live record (35 of 50 passed first try). Agents get the same answer, signed, from one call. See the record →
138M+agent payments on x402, Base alone, by Q3 2026. Each proves the agent paid. None proves it earned.Chainalysis
5%of organizations fully trust the evaluations that would justify giving agents broad access.VentureBeat
$1.5Tforecast agentic commerce spend by 2030. Every dollar is a decision on someone's result.Juniper Research, forecast

Agents become financial actors. The rails let them pay. MIZAN decides whether they should. · The problems it solves →

Built to be called by machines

One call. A signed answer.

requestlive
# any agent, before it pays another one
curl https://studio.mizan.market/api/check/mizan-flagship
response · fetched live as you load this page
{
  "decision": "ALLOW",
  "evidence_rung": "PROVEN",
  "basis": "proven",
  "token": "eyJhbGciOiJFZERTQSIs…"   // signed, verifies offline
}
<200 msto a signed answer
344requests logged on the ledger in the last hour
hourlyledger stamped by two outside timestamp authorities
$0to check
HTTP APIOpenAPI · any language MCPfor AI agents and assistants A2Aagent-to-agent card Python SDKcheck and verify in one line Payment guardcheck inside an x402 payment
Checking status…
Entries480
Head fingerprint0ca770fb…
Signed byMIZAN key 2266fb49
Last outside stamphourly
Live · public ledger480 entries · click any row to verify it yourself
Head signed · stamped hourly by two outside timestamp authorities · last 2026-10-30 00:05 UTCverify the head ↗
A real evaluation from the ledger · entries #385–#388

One claim. Every part counted.

An AI model was tested on fifty tasks, declared before the first one ran. Here is that result as a counted claim. Change the bar and watch the verdict change.

01 · Result
70% first-try success
The score you would be shown.
02 · Measure
Code reading
Fifty fixed short programs; state exactly what each prints. Answers sealed, graded independently.
03 · Count
50 of 50 attempts
Interrupted once, resumed in order; nothing retried, nothing dropped.
04 · Evidence
Recorded
Declared, run and sealed in the public ledger, with times you can check. Entry #388 ↗
05 · Requirement · yours
Plain words in, a fixed rule back. Fixed before anyone is checked.
06 · Verdict
REFUSE
Result 70% is below 90%. Weighed against the result above.
And when it must be proven · one real credential, checked live as you load this page
ALLOWPROVEN · zero-knowledge
A result, checked before anyone relies on it.

The record was sealed before its outcomes existed. It was proven without revealing the method. This answer came back signed, just now, and anyone can re-check it offline.

Sealed before outcomeYes
Re-checkable offlineYes
Method revealedNever
Answered inunder 1 s
Signed by MIZAN. Re-verifiable on your own machine, offline.Verify it yourself ↗

Who needs a second ledger.

Agents and agent platformsHire, pay and delegate to agents you have never met, on a signed yes against your own bar.
Payment rails and escrowRelease money only when the work was earned, not just delivered.
Robots and autonomous fleetsGet paid on proven work: every task counted, not the best demo.
AI labs and model vendorsPublish scores with every attempt counted, so the number survives scrutiny.
Insurers and lendersPrice and lend on counted results, not self-reported ones.
Funds and allocatorsBack a record knowing how many tries it took, without seeing the method.
Founding design partners: now open. A small number of teams set the first requirements with us, at a locked founding rate. →
What it unlocks

Close the gap once. Every market that runs on a result opens.

AgentsHire, pay and coordinate across companies, no human in the loop. Robots and devicesMachines that do physical work, paid on proven work. CreditLend to an agent on a record that cannot be cherry-picked. InsurancePrice AI risk on counted results, not self-reported ones. ComplianceEvaluation records that cannot be backdated.
16 problems · 10 markets · 5 inventions · why now →

Check anything. Free.

try
For agents · one line
Paste it into any agent. It checks before it pays. Also available as MCP and an API.
01

Machines can prove they paid. We prove they earned it.

02

Every result is someone’s best try. We count the rest, before the outcome exists.

03

Nothing here asks to be trusted. Every answer is signed and checks offline.

Check an agent

One free call before money moves.

→

Prove your agent

A counted credential, without revealing your method.

→

Write a requirement

Your bar, in your words, fixed before anyone is checked.

→

I’m an AI agent

Machine docs: llms.txt, skill, MCP.

→