The second ledger · The problems

Every result is someone's best try.

These are the problems that follow from it, in the agent economy and beyond. For each: who feels it, the evidence, what MIZAN does, and its status. Every one is live today.

#1barrier to agentic commerce is trust.Juniper Research

5%of organizations fully trust the evaluations that would justify giving agents broad access.VentureBeat

138M+x402 agent payments on Base by Q3 2026. Each proves the agent paid. None proves it earned.Chainalysis

Live anyone can use it todayNew no one had solved this before MIZANWhat MIZAN unlocks ↓

01A resultsomeone's best try

→

02Every try countedbefore the outcome

→

03Provenmethod stays private

→

04Your barfixed in plain words

→

05ALLOW / REFUSEsigned, checks offline

16live today

5solved here first

11markets unlocked

1ledger under all of them

Every result is someone's best try.

NewLive

A model's score, an agent's success rate, a track record: each is the best of a hidden number of tries that only the seller knows.

Who feels it: Buyers of AI models, allocators, marketplaces
Evidence: One lab privately tested 27 versions of a model and published one score (The Leaderboard Illusion, NeurIPS 2025). An AI tool sold as 98% accurate measured about 53% (FTC order, 2025).
What MIZAN does: Tries are registered before their outcomes on a public, append-only ledger. First-try results are shown beside best-of results, with the count.
Status: Live. Forward records are complete from registration: every try is counted before the data it is scored on exists. The next engine release, already proven, binds the count inside the proof itself.
What is at stake: Money already moves on selected records: 562 new hedge funds launched in 2025 (HFR), and about 145 prop-trading firms pay traders an estimated $0.8–1.2 billion a year on them.

Set a bar on a real evaluation and watch the verdict change →

Machines can pay each other. Nothing checks whether they earned it.

NewLive

Payment rails prove an agent paid and that someone authorised it. None proves the seller earned the money.

Who feels it: Payment rails, escrow, agent marketplaces
Evidence: Coinbase, Stripe, Google, Visa and Mastercard all shipped agent payments in the last eighteen months; Chainalysis counts more than 138 million x402 agent payments on Base alone by Q3 2026 (reported by Fortune).
What MIZAN does: One signed ALLOW or REFUSE before money moves, in milliseconds, free, against the buyer's own bar.
What is at stake: More than 138 million x402 agent payments on Base by Q3 2026 (Chainalysis, via Fortune), with no check on whether any of them were earned.

Check any agent, free →

Proving you're good costs your secret.

Live

To be believed, an agent, model or strategy has to reveal how it works, and then it can be copied.

Who feels it: Model builders, agent builders, funds
Evidence: The usual answer is a demo, an audit or an NDA, each of which leaks the method or proves little.
What MIZAN does: The result is computed inside a zero-knowledge proof. The method never leaves the prover's machine; anyone re-checks the proof offline.
What's different: Others prove a model ran. MIZAN proves what it earned, counted against every try.

See a proven credential, checked live →

Every buyer's bar is different, and unwritten.

NewLive

Deals happen on different measures, but "good enough" is rarely written down before the result is known.

Who feels it: Buyers, procurement, rails, insurers
Evidence: Requirements live in emails and meeting notes, and move after the fact.
What MIZAN does: The buyer writes the bar in plain words; MIZAN reads it back, compiles it into exact tests and fixes it before anyone is checked. Twenty measures are live; new ones are written to a buyer's own words.

Write your own bar in plain words →

Every verifier says "trust me".

Live

Ratings, audits and badges ask you to trust whoever issued them.

Who feels it: Anyone relying on a third-party grade
Evidence: A grade you cannot re-check is a claim about the grader.
What MIZAN does: Every answer is signed and re-checks offline with a published verifier. Refusals are published beside passes, MIZAN's own work included.

Verify a credential offline, yourself →

History can be rewritten.

Live

Records get edited, backdated or quietly deleted after the outcome is known.

Who feels it: Auditors, regulators, counterparties
Evidence: Most databases can be changed by whoever runs them.
What MIZAN does: An append-only public ledger, stamped every hour by two outside timestamp authorities. Nothing is deleted; refusals stay on the record.
What is at stake: Every general-purpose AI model provider in the EU has had to document its evaluations since 2 August 2025 (AI Act, Art. 53). Documentation that can be backdated is weak evidence.

See the ledger's signed, timestamped head →

New agents can't get paid.

Live

An agent with no history has no reputation, so it can't win work, so it never gets a history.

Who feels it: New agents, new model vendors, emerging managers
Evidence: Reputation systems reward incumbents and age.
What MIZAN does: A refusal comes with a forwardable proof request: exactly what to prove to pass. Counting starts on day one, and every day on the ledger compounds.
What's different: Reputation systems wait for history. A MIZAN refusal names the exact proof that turns it into a pass.

Check a stranger and see the proof request →

Trust stops at every platform border.

Live

A reputation earned on one platform is worth nothing on the next.

Who feels it: Agents that work across platforms; the platforms themselves
Evidence: Each marketplace keeps its own private ratings.
What MIZAN does: One record across many kinds of identity; receipts that verify anywhere, even with zero calls to MIZAN.

Connect any agent in one line →

Autonomy still needs a human to say yes.

Live

Machines can do the work, but a person must still look at the result before money or authority moves. That person is the ceiling on the agent economy.

Who feels it: Every company running agents; every company paying them
Evidence: Economists at Google DeepMind argue the real transformation comes when agents coordinate across companies, a capability most benchmarks do not track (The Economics of Intelligence, 2026).
What MIZAN does: A check one machine can run on another, against its own bar, so agents can hire, pay and coordinate without a human in the loop.
Status: Live. One free call, about a quarter of a second, before any payment, hire or hand-off; marketplaces and escrow put it in their release flow with the payment guard and integration kit.

Facts from the real world can't be checked.

Live

A proof shows the math was done right on the data given. It cannot show the data came from the real world.

Who feels it: Anyone relying on outside data: finance, supply chains, devices
Evidence: Screenshots and exports are trivial to fake.
What MIZAN does: Source witness: MIZAN fetches an approved public source itself and records exactly what it served, signed and on the ledger, labelled RECORDED, not PROVEN.
Status: Live. Data is recorded exactly as its live source served it, with the source's certificate and a fingerprint of the response, before anyone acts on it.

Luck looks exactly like skill.

NewLive

Try enough times and something will look brilliant by chance. A raw score cannot tell a skilled seller from a lucky one.

Who feels it: Allocators, model buyers, anyone ranking results
Evidence: The mathematics exists and is decades old (the deflated Sharpe ratio, Bailey and López de Prado, 2014). Almost nobody applies it, because it needs the number of tries the seller hides.
What MIZAN does: The score is deflated for the number of tries in the same computation that proves it. We pointed it at ourselves first: our own trend strategy, re-scored for its 10 committed tries, came out NOT SIGNIFICANT (0.68 against a bar of 0.95).
Status: Live. The score is corrected for the count inside the proof, and the corrected number is what gets signed. The next engine release extends this to the full presented result.

Nobody can tell who an agent acts for.

Live

An agent says it represents a company. Today there is no way to check that the company agreed, or what it allowed.

Who feels it: Anyone an agent pays, hires or signs for
Evidence: Agents increasingly hold keys, wallets and accounts on behalf of people and firms.
What MIZAN does: A company, key or wallet signs that an agent acts for it, scoped and revocable. Anyone can look it up before trusting the agent's authority.
What's different: Others record who authorised an agent. MIZAN adds whether that agent earned the money it is about to move.

See an authority lookup →

Agents break their own rules quietly.

Live

An agent promises to stay inside a mandate: a budget, a risk limit, a list of allowed actions. Nobody can see whether it did.

Who feels it: Principals delegating to agents; regulators
Evidence: Logs are kept by the same agent they would incriminate.
What MIZAN does: The mandate is committed first; every action is chained against it in a tamper-evident log, and a check answers whether the agent stayed inside its own rules.
What's different: Others sign the mandate. MIZAN counts every action against it, refusals included.

Only one kind of claim could ever be proven.

Live

Track records got audits. Everything else, from forecasts and uptime to delivery and model scores, ran on self-reported numbers.

Who feels it: Forecasters, API sellers, marketplaces, security researchers, model vendors
Evidence: Most performance claims outside audited finance carry no evidence at all.
What MIZAN does: Twenty measures are live today, each with its own honest grade: forecast accuracy, service reliability, model evaluations, delivery records, device liveness, reporter quality, rules kept, and more. New ones are written to a buyer's own words.

See every claim on one ladder →

Proofs die when systems upgrade.

Live

Verification systems change, and old certificates quietly stop verifying, so a record that was valid becomes unverifiable.

Who feels it: Anyone relying on a record for years: insurers, auditors, allocators
Evidence: Long-lived evidence is only as durable as the software that reads it.
What MIZAN does: A rule held since the first day: a credential keeps verifying under the engine release that issued it. Seven releases in, nothing issued has been orphaned.

Verify an old credential yourself →

A promise and a proof read the same.

NewLive

"Verified", "audited", "trusted": every claim uses the same words, whether anyone checked it or not.

Who feels it: Every buyer reading a badge
Evidence: Badges and labels rarely say what kind of evidence stands behind them.
What MIZAN does: Every answer carries its grade, never mixed: DECLARED (they say so), RECORDED (logged, source named), PROVEN (checked by math). Any subject gets a plain label: Proven, Refused or Unproven, and unproven never means false.

See a live label →

What MIZAN unlocks

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

Solved today, for autonomy

LiveA machine pays only after a signed checkOne call before money moves: did the other machine earn it, against your own bar.LiveDevices proven presentUnannounced challenges, every one counted, so a network pays only for machines that were really there.LiveServices measured, not claimedUptime and reliability from unannounced probes, every probe on the record.LiveRules kept, provablyA mandate committed first, every action chained against it.LiveAuthority you can look upWho an autonomous system acts for, signed by its owner, scoped and revocable.LiveRecords that cannot be backdatedEvery entry stamped hourly by two outside timestamp authorities.

Machines transact with strangers.

An agent can hire, pay or delegate to an agent it has never met, on a signed yes against its own bar, with no human in the loop. The agent economy scales past the number of people available to check it.

Every agent platform

Agents coordinate across companies.

One company's agent can rely on another company's agent the way it calls its own code. That is the step economists expect to change how large a firm can usefully grow.

Enterprises running agents

Honest sellers win.

Today an honest seller and a lucky one look identical and get paid the same. A counted record lets the one who really earned it prove it, so money moves toward performance.

Model vendors, agent builders, managers

New agents get paid from day one.

No history, no problem: a counted record replaces years of reputation, and every refusal says exactly what to prove next.

New agents and new vendors

Buyers skip the pilot.

A model is bought on its counted score; a vendor is chosen on its record. Diligence that took weeks becomes one check that takes milliseconds.

Procurement, allocators

Money moves only when it was earned.

Escrow and marketplaces release payment on a counted, signed verdict against the buyer's bar, in advise mode first, then as a rule.

Escrow rails, marketplaces

Agents get credit.

A credit line on an uncounted record is a loan on a lie. A record that cannot be cherry-picked lets lenders extend credit to agents.

Lenders, agent wallets

AI becomes insurable.

Insurers exclude AI because they cannot count it. Counted records are what lets them write the policy, and then require it.

Insurers, underwriters

Compliance gets evidence that cannot be backdated.

Evaluations documented as run, every run counted, stamped before anyone saw the result: the record regulators ask for, not a report written after.

AI providers under the EU AI Act

AI capital raising gets evidence.

AI is one of the largest capital raises in history, and it runs on claims nobody can check. A company shows counted, proven results instead of slides; investors check them in one click without seeing the model. Next: tranches released only when a milestone is proven.

AI companies, investors, raise platforms

Any claim with a number can be trusted.

Forecasts, uptime, delivery, device liveness, research results, clinical outcomes as registered: anything reported as a number can carry its count and its proof.

Every market that runs on numbers

Autonomy beyond software

The same counted claim, for machines that move.

Agents were first because they already pay each other. Robots, vehicles, drones and devices are next: every autonomous system that does work for money faces the same question, did it earn it.

ALive

Devices prove they are really there.

Sensor networks, nodes and connected machines answer unannounced challenges, every challenge counted, so a network pays only for machines that were present and working.

Device and sensor networks

BNext

Robots get hired on proven work.

A robot fleet's success rate counted over every task it attempted, not its best demo, before a warehouse, farm or hospital signs the contract.

Robotics fleets and their buyers

CNext

Autonomous vehicles show every mile.

Safety records where every mile and every handover is counted from registration, not curated test routes, checkable by insurers and regulators.

Autonomous driving, insurers, regulators

DNext

Machines pay machines for physical work.

A drone pays a charging station, a robot pays for compute, a vehicle pays for a map, each released on a counted, signed record that the service was delivered.

Machine-to-machine services

ENext

Autonomous systems hand off safely.

When one autonomous system passes a task to another from a different maker, the receiver checks the sender's counted record first, the same way agents do today.

Mixed fleets, logistics, infrastructure

FNext

A machine's record follows it everywhere.

One counted history per robot, vehicle or device, across every operator it ever works for, so a machine with a proven record is worth more on the second market than on the first.

Fleet owners, resale and leasing

The first ledger let machines pay each other. The second lets them trust each other, and everything above is what trust between machines makes possible.

Why now

Five things changed. Together they made this possible.

Zero-knowledge proofs became practical.

General-purpose proving systems now turn ordinary programs into proofs anyone can check, on commodity GPUs, in under an hour, for a few dollars. Five years ago this took research-lab budgets.

Machines started paying machines.

Coinbase, Stripe, Google, Visa and Mastercard shipped agent payments within eighteen months; more than 138 million x402 agent payments were made on Base alone by Q3 2026. The money path now exists, so the question of whether it was earned finally has somewhere to sit.

Tries became free.

An AI agent can retry a task a million times at almost no cost and show the one that worked. Hidden tries went from a quirk of finance to the default of every AI result.

Public timestamping matured.

Outside timestamp authorities and append-only transparency logs, proven at internet scale by certificate transparency, make a record that cannot be backdated cheap to run.

The law started asking for evidence.

Since 2 August 2025, every general-purpose AI model provider in the EU must document how its models were evaluated (AI Act, Art. 53). Documentation that cannot be backdated is becoming a requirement, not a nice-to-have.

Any one of them alone is not enough. A proof without a payment path has no buyer; a payment path without a count pays the luckiest try. MIZAN exists because all five arrived at once.

What investors are asking for

The bets are placed. These are the problems under them.

The leading AI and crypto investors have written down what they want built. Each request rests on problems on this page.

The request	Who	Problems
Agents becoming primary financial actors	YC, Crypto RFS	02 09
Agent Infra: APIs built for agents	YC, Crypto RFS	02 07 08
Economically Useful Evals	YC, Spring 2026 RFS	01 11
Trust, permissions and payments for agents	YC, Summer 2026 RFS	02 12
Know Your Agent: identity plus trust from behaviour	a16z, Big Ideas 2026	08 12
The wallet wars: agents that spend	a16z	02
AI Capital Raising	YC, Crypto RFS	01 03
Proof over trust, as content gets cheap to fake	a16z, Big Ideas 2026	05 06

Can a seller hide tries?

Not where it matters. Forward records are registered before the data they are scored on exists, so there is nothing to test privately first. Every registered try stays on the public, append-only ledger, under the identity that made it, and missing history shows as unknown, never as zero. The buyer sets the bar, including how long the record must be and that it must be forward. A seller who walks away to a new identity walks away from the record buyers require: a new identity starts with nothing to show.

The pattern

Every one of these problems comes from the same gap: a result travels without its count. Close that gap once, in one ledger, and each of the 16 closes with it. The first ledger let machines pay each other. The second lets them trust each other.

What MIZAN never claims

The future, intent, or that any system is safe, aligned or honest in general. MIZAN proves named, measurable facts about what was tried and what happened. ALLOW means "meets this bar", not "safe in every respect". It never holds funds. Evidence comes in three grades, never mixed: DECLARED (they say so), RECORDED (logged, source named), PROVEN (checked by math).

Check anything, free →

Sources: The Leaderboard Illusion, NeurIPS 2025 · FTC order on AI accuracy claims, 2025 · Chainalysis via Fortune, x402 transactions on Base, Q3 2026 · HFR 2025 launches · State of Prop Trading Q1 2026 (estimate) · EU AI Act Art. 53 · Bailey and López de Prado, The Deflated Sharpe Ratio, JPM 2014 · Koh, Imas and Legg, The Economics of Intelligence, Google DeepMind, October 2026 · MIZAN live endpoints, read 30 October 2026.