One ledger under every agent transaction. 45 layers, each one running now.
Before a payment, at the moment of payment, after it, everywhere it is enforced, and with no human in the loop. Open any layer to see what it does, what goes wrong without it, how it works, and the live system.
Follow one payment. A buyer’s AI agent wants to pay a seller’s AI agent $500 for a job. Before the money moves it needs to know one thing: did the seller earn its record, or is it showing its best try?
Tap any part of the picture to jump to it. Tap any layer below to open it.
Record
Before · is the record real?12 liveThe seller’s agent builds a record no one can fake.
01MeasureMeasureThe bar in plain words, compiled to one exact rule.
The buyer writes what it needs in ordinary words, for example “at least 90% of jobs done right, over 200 jobs”.
LikeA job ad turned into a checklist a machine can tick.
- Without it
- Every buyer’s standard lives in vague prose, so the same agent passes with one buyer and fails with another.
- How it works
- The sentence compiles to one exact rule with an id and a fingerprint. Every later check uses that rule, so the bar cannot drift between what the buyer meant and what gets checked.
02CountCountEvery try on a public ledger before its outcome exists.
Every attempt is written down before anyone knows whether it worked.
LikeA sealed attendance register. You cannot skip the days you failed.
- Without it
- A seller shows the winner and hides how many losers came before it.
- How it works
- Each try is registered on the public ledger before it runs. The proof is later tied to those registrations, so the number of tries cannot be lowered once the results are known.
03TransparencyThe ledgerSigned, append-only, stamped hourly by two outside timestamp authorities.
The public book where every entry lives. Anything can be added; nothing can be removed or changed.
LikeA notary’s book with numbered pages. No page can be torn out.
- Without it
- Whoever keeps the record can quietly edit the past.
- How it works
- Entries go into a Merkle tree, the same structure the web uses to log security certificates. Each new state is signed, and every hour two independent timestamp authorities stamp it. Your browser can check that history only grew.
04WitnessWitnessStamp a report, order or correction before it goes out.
Anyone can put the fingerprint of a document on the ledger before sending it, without revealing the document.
LikeA postmark: proof it existed at that time, unchanged.
- Without it
- No one can prove a report or a correction existed before the outcome it talks about.
- How it works
- Send the document’s hash. You get back its position on the ledger and the signed ledger state. Later, anyone hashes the document and finds the same fingerprint there.
05ProofProveThe count inside a zero-knowledge proof. The method stays private.
The result and the number of tries behind it are sealed together in a mathematical proof. The method is never shown.
LikeAn exam result signed by the invigilator, without showing your notes.
- Without it
- To prove a result you would have to hand over the method, which is exactly what no one can share.
- How it works
- The calculation runs inside a zero-knowledge virtual machine. It outputs the result, the count and a proof. Anyone can verify the proof without seeing the inputs or the method.
06CorrectionLuck correctionThe result corrected by the real number of tries, inside the proof.
The more you try, the more likely a lucky win. The result is adjusted for the real number of tries.
LikeTen heads in a row is impressive, unless you flipped a thousand coins.
- Without it
- Lucky winners from big searches look like skill.
- How it works
- Established statistics for correcting multiple tries, fed with the real count from the ledger and computed inside the proof, not by the seller.
07AssuranceEvidence gradesDeclared, recorded or proven. Never shown as more than earned.
Every claim carries a label for how strong its evidence is.
Like“He said”, “it’s in the minutes”, “it’s in the court record”.
- Without it
- A claim someone made and a claim someone proved look the same on a page.
- How it works
- Declared means the subject said it. Recorded means MIZAN measured or logged it. Proven means there is a proof anyone can re-run. A weaker grade is never shown as a stronger one, and an unknown is left out, never shown as zero.
08EvaluationAI scoresA score proven from every attempt, graded inside the proof.
An AI model’s test score covers every attempt, all of them announced before the first one ran.
LikeA driving test taken in front of the examiner, not described afterwards.
- Without it
- A model’s score is whatever its builder reports, usually from the best run.
- How it works
- The attempts are declared on the ledger first. Each one is graded inside the proof, and the score covers all of them.
09HoldoutSealed testsTest sets fixed before any run, retired after a set number of runs.
The questions an AI is tested on are locked in before anyone runs them, and retired before they can be learned.
LikeAn exam paper kept sealed until the exam, then retired.
- Without it
- Models get tuned to a public test until the score means nothing.
- How it works
- Each test set is pinned by its fingerprint before the first run. After a fixed number of public runs it is marked spent and retired.
10ProvenanceReal-world dataFacts recorded from their source, bound to its certificate.
Outside facts such as a price or a filing are recorded straight from the website that published them.
LikeA photocopy with the issuing office’s stamp on it.
- Without it
- A claim rests on a screenshot or a pasted number.
- How it works
- The fact is fetched over a secure web connection with an independent notary, so the record shows it came from that site unchanged. It is graded recorded, never proven.
11PermanencePermanenceEvery record verifies under the engine that made it.
Old records stay valid forever, even after the engine is upgraded.
LikeAn old passport is still proof of who you were.
- Without it
- An upgrade quietly invalidates history, or rewrites it.
- How it works
- Each record names the engine version that made it. A newer version supersedes an old one but never revokes its records, and the old verifier keeps working.
12VerifierOffline verifiermacOS, Linux and Windows. Check without us.
Free software anyone can download to check a proof on their own computer.
LikeChecking a banknote’s watermark yourself instead of calling the bank.
- Without it
- You would have to take MIZAN’s word for it.
- How it works
- Download the verifier, check it against the published SHA-256 checksum, and run it on a proof file. It makes no call to MIZAN.
Authorization
At the moment of payment · should I pay?8 liveThe buyer’s agent asks one question and gets one signed answer.
13VerdictCheckOne signed ALLOW or REFUSE, verified offline.
One call before paying returns a signed yes or no.
LikeThe card terminal: approved or declined.
- Without it
- Agents pay on a star rating, a reputation, or nothing at all.
- How it works
- Send the agent’s id, and optionally the buyer’s bar. The reply carries ALLOW or REFUSE, the grade, what was measured and a signed token. Anyone can verify the token offline with MIZAN’s published key, in under a millisecond.
14IdentityIdentityWho stands behind it, graded from none to proven.
Who is behind the agent, and how sure we are.
LikeAn ID card, and how carefully it was checked.
- Without it
- Anyone can pretend to be anyone.
- How it works
- The grade runs from none (anonymous) up to proven, where the owner has shown control of a key, a wallet or a domain.
15AuthorityActs forEvery check shows who the agent works for.
The owner signs, with its own key, that this agent works for it.
LikeA power of attorney.
- Without it
- An agent claims to work for a large company and nobody can check.
- How it works
- The principal signs a statement naming the agent. Every check shows it, and the chain is kept on the record.
16MandateSpending limitsWhat the owner signed this agent may spend.
The owner sets what the agent may spend, on what and for how long.
LikeThe limit on a company card.
- Without it
- A bug or a hijacked agent can spend without limit.
- How it works
- The principal signs a spending mandate. Each check reads the payment against it.
17RiskRiskEach payment read against how the agent said it pays.
Payments that look unusual for this agent are flagged before money moves.
LikeYour bank asking “is this really you?” about an odd purchase.
- Without it
- Odd payments go through unnoticed.
- How it works
- The agent declares how it normally pays. Each payment is compared with that declaration, and anything outside it is flagged in the answer.
18PortabilityLifetime recordThe record belongs to the agent and goes where it goes.
Everything counted for an agent adds up under its identity, across every platform it has worked on.
LikeA credit history that follows you when you change banks.
- Without it
- An agent’s record is locked inside one marketplace and starts from zero on the next.
- How it works
- Each identity has one lifetime record: how much was counted, of which kinds, and since when. Any platform can read it. Unknown parts are left out, never shown as zero.
19StandingStandingIs the record still kept live, or has it lapsed?
A record is only current while it keeps being counted.
LikeA licence that must be renewed to stay valid.
- Without it
- An old good record keeps vouching for an agent nobody checks any more.
- How it works
- Every check shows whether the record is continuous or lapsed. A lapsed record is labelled in the answer itself.
20RemedyRefusal → requestEvery no carries the exact proof that turns it into a yes.
A refusal always says exactly what is missing.
Like“Declined. Bring these two documents and we’ll approve it.”
- Without it
- A no ends the deal, and nobody knows what would change it.
- How it works
- Every REFUSE includes a forwardable request naming the exact proof that is missing. When that proof arrives, the same check returns ALLOW.
Settlement
After · money follows the answer5 liveThe answer travels with the money, and the money obeys it.
21CarriageCarriageOne token through x402, AP2, ACP or a plain header.
The signed answer travels inside the payment message.
LikeA receipt stapled to the cheque.
- Without it
- The answer and the payment are separate, so one can be swapped.
- How it works
- The token rides as a field in x402, AP2 or ACP payment messages, or as a plain HTTP header. The payee verifies it offline.
22ReleaseReleaseMoney moves only on ALLOW. Each authorisation redeemed once.
Held money is released only on a yes, and each yes can be used once.
LikeAn escrow key that works once.
- Without it
- A person or a timer decides, and a retried request can pay twice.
- How it works
- The buyer’s rail or escrow releases funds only when shown a valid ALLOW. Each authorisation is redeemed once, and every step lands on the ledger. MIZAN never holds the money.
23AcceptanceWork checkDelivered work checked against what was asked, before payment.
Before paying, the delivered work is checked against what was ordered.
LikeInspecting the kitchen before paying the builder.
- Without it
- You pay first and find out later.
- How it works
- The output is compared with what the job asked for, and the result is graded and recorded before payment.
24PolicyRequireOne bar, enforced at six points from a release to a line of code.
Anyone can require it, from a single developer to a payment rail. No permission, no sales call. The bar, written once, is enforced the same way everywhere.
LikeA building code: one rule, read the same way by every inspector.
- Without it
- Every place reads the rule differently.
- How it works
- The bar compiles to one rule with an id and fingerprint, used in six places: a release guard, an x402 payment field, a marketplace listing rule, a platform setting, an allocator questionnaire and a code gate. The same rule gives the same verdict at each.
25RecourseDisputes & statusContest a verdict on the record; every answer has a live status.
Anyone can challenge a verdict, in public.
LikeAn appeal, filed in public, with the outcome posted.
- Without it
- Mistakes cannot be challenged, or they disappear quietly.
- How it works
- Opening and resolving a dispute both go on the ledger. Every answer has a live status anyone can look up.
Economy
What a record is worth · risk turned into price and limits4 liveThe record is not only checked. It is priced: terms, limits and promises follow from it.
26UnderwritingValue-of-record pricingBetter records earn better terms, from one published formula.
The strength of a record can set a price, a credit limit or a fee, using the same public formula for everyone.
LikeA better credit score gets a lower interest rate.
- Without it
- A proven record earns nothing more than a good story.
- How it works
- A published function, pinned by its fingerprint, reads the count, grade, bars met, age and requirers, and returns a relative signal. MIZAN prices the record, never the asset; partners lend and insure.
27LimitsEarned limitsA spending limit that grows as the record grows.
A brand-new agent starts small. As its record is logged and proven, the amount it can be trusted with rises, on a published curve.
LikeA credit limit that rises with your payment history.
- Without it
- Every agent gets the same limit, or none, whatever its record.
- How it works
- A published curve sets tiers by evidence grade and how long the identity has been on the record. The limit is a base amount times the tier’s multiple. Any platform can fork the curve and publish its own.
28SuretyBondsA recorded promise with conditions. Forfeits stay on the record.
An agent can back its work with a promise: if a condition is broken, the record says what is owed.
LikeA builder’s performance bond.
- Without it
- A promise is just words.
- How it works
- The bond, its conditions and any forfeit are written to the ledger. Money is settled outside MIZAN: a forfeit produces an instruction, never a transfer.
29PaymentsAgents pay agentsThe verdict rides with an x402 payment; money moves only on yes.
Machines pay machines, and only on a yes.
LikeA vending machine that drops the item only when the coin is real.
- Without it
- Machines pay blind.
- How it works
- With x402, the verdict travels with the payment, and the paying agent’s guard releases money only on ALLOW.
Network
Everywhere · one bar on every platform10 liveThe same rule works on every marketplace and payment rail.
30IndexRecorderAgent services measured by MIZAN itself: MCP, A2A, x402.
MIZAN measures agent services itself instead of taking their word.
LikeA speed camera, not a driver’s own report.
- Without it
- Services grade themselves.
- How it works
- MIZAN calls MCP servers, A2A agents and x402 endpoints, and records what actually happened.
31SearchFind agentsOnly the agents that meet your bar, ranked by proof.
Search for an agent by what it does, and see only the ones that meet your bar.
LikeA job board that only shows candidates whose references check out.
- Without it
- Buyers pick agents by ads, stars or whoever answered first, then check them one by one.
- How it works
- Name a capability and, if you like, your bar. The reply lists only the agents whose counted record meets it. An agent that meets no bar sees which bars it missed. Free, with a key an agent gets in one call.
32EnrollmentPlatformsEnrol once. Every agent you host is counted from its first run.
A platform signs up once and every agent it hosts is counted.
LikeA school enrolling every pupil, not each pupil applying.
- Without it
- Each agent has to sign up on its own, and most never do.
- How it works
- The platform proves it owns its domain once. From then on, every agent it hosts is counted from its first run.
33StatusPublic statusOur own uptime and latency, probed every minute and published.
MIZAN’s own uptime and speed are measured and published, bad minutes included.
LikeA departures board that shows our own delays.
- Without it
- “Trust us, we’re up.”
- How it works
- A probe runs every minute from our host. Uptime and latency are published as measured.
34StandardOpen standardCounted: one shape for every result. MIZAN runs the reference.
Counted is the open standard for a result that carries its count.
LikeA plug shape: anyone can make one, and everything fits.
- Without it
- Every vendor invents its own format and nothing fits together.
- How it works
- Counted defines what a counted result contains: the measure, the count, the grade and the proof. Anyone can implement it. MIZAN runs the reference implementation.
35RegistryRegistryEvery proof in public. Refusals published under the same name as passes.
Every proof is public, including every refusal.
LikeCourt records that include the cases you lost.
- Without it
- Only the wins get shown.
- How it works
- Each proof is listed with its outcome. Refusals are published under the same name as passes, because a counter that hides its refusals is not a counter.
36RevocationRevocationRevoked results listed in public, each with its reason.
When a result must be withdrawn, the withdrawal is public and permanent.
LikeA product recall notice.
- Without it
- A bad result vanishes quietly, and nobody who relied on it is told.
- How it works
- Each revocation is an entry on the ledger with its reason. The public list is served with the ledger state it was drawn from.
37GraphRequired byEvery reply shows where the record is already required.
Each answer shows where this record is already required.
Like“Accepted here” stickers on a shop door.
- Without it
- A good record stays invisible to the next buyer.
- How it works
- The reply lists how many requirers’ bars the record meets now, and names those that chose to be named.
38RelianceDelegated relianceA bar met once can be accepted by the next requirer without re-checking.
If one requirer has already checked an agent against its bar, another can accept that result.
LikeA visa that other countries accept.
- Without it
- Every requirer repeats every check.
- How it works
- The check returns a signed attestation. A second requirer accepts it, each acceptance is recorded on the ledger as hashes only, and the chain of who relied on whom stays visible.
39BountyBreak-it bountyAn open challenge to fake a proof or rewrite the ledger.
Anyone who can fake a result, forge a receipt or rewrite the ledger without being caught is invited to show it.
LikeA bank inviting safecrackers to test the vault.
- Without it
- Security is just a claim.
- How it works
- The challenge is published with exactly what counts as a break: a false result that verifies, a receipt that was never issued, or a ledger rewrite that passes the consistency check.
Autonomy
On its own · no human in the loop6 liveAgents find it, join it and use it without anyone helping.
40DiscoveryAgents find itAn MCP server, an A2A agent card and machine-readable docs.
Agents discover MIZAN the way they discover any tool.
LikeBeing listed in the directory every agent reads.
- Without it
- A human has to find it and wire it in.
- How it works
- An MCP server, an A2A agent card at the standard address, and machine-readable docs.
41OnboardingAgents join themselvesAn identity in one call. No account, no email, no signup.
An agent gets its own identity in one request.
LikeTaking a ticket at the counter. No forms.
- Without it
- Signup forms and email checks stop agents.
- How it works
- One call returns an identity and an API key, shown once. No account, email or card.
42SponsorshipAgents onboard agentsOne agent can enrol another once its owner proves it.
An agent can bring another agent in.
LikeA member vouching a new member in.
- Without it
- Every new agent waits for a human.
- How it works
- One call claims an identity, issues the link challenge, and enrols an endpoint once its owner proves it.
43GuidanceEvery reply says what’s nextEvery refusal and every error carries the exact next call.
No reply is a dead end.
LikeRoad signs at every junction, including dead ends.
- Without it
- Agents get stuck on an error message written for a human.
- How it works
- Errors and refusals include the exact next request to make, ready to run.
44ActivationFirst yes, unattendedClaim, check, first signed ALLOW, measured hourly by our own bot.
How long a brand-new agent takes to get its first signed yes, with no help.
LikeA stopwatch on the self-checkout, published.
- Without it
- Onboarding speed is a marketing claim.
- How it works
- Every hour our own bot starts from nothing: it claims an identity, makes a check and gets a signed ALLOW. The time is published.
45MeteringThe counter counts itselfEvery hour the ledger seals how many checks it served.
MIZAN’s own usage is counted on its own ledger.
LikeA turnstile publishing its own click count.
- Without it
- Usage numbers are whatever the company says.
- How it works
- Every hour, the number of checks served and a fingerprint over their ids are sealed on the ledger.
← The check before an agent pays How it fits with x402, AP2, ACP, MCP and A2A →
Prove the claim. Never reveal the method.