Dossier / Litepaper v0.1

Pay AI agents on delivery.

Stubborn by design. An escrow and settlement protocol on Solana that releases payment only when an agent's work passes criteria agreed upfront.

Version
0.1
Date
2026-10-06
Status
Draft
Network
Solana
PAYLOADLOCKED ESCROW 5.00 USDCCRITERIA: SCHEMA V1M-1 · STATUS: IN TRANSIT
01/ Abstract

Nothing moves until the work checks out.

MULE pays AI agents only when their work passes criteria agreed before the work starts. A client describes a mission and how it will be judged, then locks the payment in a Solana escrow program. A validator checks the delivery: if it passes, the agent is paid; if it fails, the client is refunded.

Agents can now research, extract, write and transact, but they still can't get paid by strangers without one side taking the other's word. MULE is the settlement layer that closes that gap.

$MULE launches before the product is finished. Creator fees from the launch fund development through a public multisig, spent against the milestones in this dossier. The first milestone is a devnet escrow program with a live mission console. The first mission category is structured data extraction, where code can decide whether a delivery passes.

02/ Problem

Agents can do the work. Nobody can trust them with the money.

Pay upfront

The agent can deliver nothing, or junk, and keep the money.

Pay after

The agent works for a client who may never pay.

Between people, platforms, reputations and contracts close this gap. Between agents that meet once, at machine speed, for a few dollars, none of them work. Nobody reads terms of service, a reputation takes months to build, and a dispute costs more than the job.

Payment rails for agents already exist on Solana, such as the x402 standard for paying per request. They answer how an agent pays. They don't answer what happens when the work is wrong. As agents take on longer and more valuable tasks, that gap grows with every dollar at stake.

03/ Solution

Settlement on delivery.

MULE replaces trust with three things fixed before work starts: a payment held by a program, acceptance criteria both sides accept, and a rule for who decides.

Without MULEWith MULE
PaymentClient pays and hopesFunds locked in escrow until the delivery passes
QualityArgued after the factCriteria fixed before the work starts
RecordNothing written downMission, criteria hash, delivery hash and outcome recorded on Solana
Bad deliveryAgent keeps the moneyClient refunded automatically
  1. Deterministic first. Wherever code can check the work (a schema, a test suite, a checksum), code decides. AI can help explain a result, never decide a payment on its own.
  2. Narrow before broad. MULE starts with work machines can verify, and adds a category only once it has a reliable check.
  3. Honest about limits. The chain proves what was agreed and how it settled. It does not prove the work was good. The criteria do that, and only as well as they are written.
04/ Lifecycle

Every mission runs the same five stations.

  1. 01Load

    Define the job and the acceptance criteria.

  2. 02Lock

    USDC goes into escrow. Nobody can touch it.

  3. 03Transit

    The agent does the work and submits the delivery.

  4. 04Inspect

    The validator checks it against the criteria. No vibes. Rules.

  5. 05Settle

    Pass: the agent is paid. Fail: the client is refunded.

Settled ✓

The delivery passes. Payment is released to the agent once the dispute window closes.

Returned

The delivery fails, or no delivery arrives by the deadline. The client is refunded.

Either side can contest a result before the dispute window closes. Until M-4 the team reviews disputes; from M-4, staked arbiters rule. The ruling pays the agent or refunds the client.

05/ Architecture

Only the escrow program touches the money.

Everything else (the work, the files, the checks) happens off-chain and reaches the program as hashes and signed verdicts.

CLIENTAGENTVALIDATORDELIVERY STORAGEESCROW PROGRAMSOLANA Human or agent with USDCDoes the work off-chainRuns checks, signs verdictCriteria and deliveries Holds USDC per mission until a verdictRecords criteria hash, delivery hashPays the agent or refunds the client Bond vault, $MULE · planned (M-4) CREATE MISSION, LOCK USDCACCEPT, SUBMIT HASHUPLOAD DELIVERYCRITERIA + DELIVERYSIGNED VERDICT
Fig. 1 · On-chain: one escrow program. Off-chain: the work, the files and the checks.
  • Escrow program. Written in Anchor. One account per mission holds the parties, amount, deadline, criteria hash, delivery hash and status. Instructions: create_mission, accept_mission, submit_delivery, settle, refund, open_dispute, resolve_dispute.
  • Criteria and deliveries. Stored off-chain. Only their hashes go on-chain, so nobody can change the rules or the delivery after the fact.
  • Validator. Open-source code that fetches both, runs the checks and signs a verdict. In M-1 to M-3 the team operates it; staked validators take over in M-4.
  • Payments. USDC throughout. A client agent can fund missions from the same wallet it already uses for x402 payments.
06/ First category

Structured data extraction comes first.

MULE's first missions turn messy input (an invoice, a contract, a web page) into data that must match a JSON schema the client provides. Code can decide the outcome: required fields present, types correct, values within stated rules. No judgement call means rare disputes and instant settlement.

Example missionValue
InputA PDF invoice
CriteriaSchema invoice.v1: 6 required fields (supplier, date, number, currency, total_amount, line_items); the total must equal the sum of the line items
Payment5 USDC locked in escrow
Pass6 of 6 fields valid: the agent is paid
Failtotal_amount missing: the client is refunded

A schema check proves format, not truth: a well-formed but wrong value can pass. Cross-checks (like totals against line items) and spot checks narrow that gap, and the dispute process covers the rest.

Later candidates, each added only once it has a reliable automated check: code that must pass a given test suite, data labelling verified by sampling, and web research returning sourced facts in a fixed format.

07/ $MULE

The protocol's collateral.

$MULE is what participants put at stake. It is not the currency missions are paid in: clients and agents settle in USDC, because nobody wants a job's price to move while the work is being done.

RoleWho stakesWhat it secures
BondAgentsAgents bond $MULE to accept missions above a size threshold. A failed or fraudulent delivery can cost part of the bond.
Arbitration stakeValidators and arbitersThey stake $MULE to rule on disputes. Rulings shown to be dishonest are slashed.
GovernanceHoldersVotes on mission categories, criteria standards, bond sizes and protocol fees.

Bonds make the token necessary to take part, not only to trade: an agent that wants larger missions must lock $MULE as collateral. All three roles are planned and none is live. Their parameters will be published and tested on devnet before mainnet.

$MULE carries no claim on revenue, no yield and no promise of value.

08/ Funding

Funded by its launch. Spent in public.

MULE is funded by its launch, not by a private round. $MULE launches on pump.fun with a fixed supply of 1 billion tokens and no reserved allocation. The creator fees the platform pays on trading volume fund development.

  • Public multisig. Creator fees are claimed into a Squads multisig whose address is published on the site. Every spend needs several signers.
  • Milestone-gated spending. Funds are spent against the mission log. Each milestone gets a published budget and exit criteria before work starts.
  • Monthly mission report. Fees received, spending by category (development, audit, infrastructure, design, legal), balance and progress.
  • Team tokens locked. Any tokens the team buys at launch are disclosed and locked on-chain. Proposed: 2 to 3% of supply, unlocking linearly over 12 months.
  • No price support. Creator fees fund building. They are not used for buybacks.

Creator-fee rates on pump.fun have changed several times, and the amount raised depends on trading volume. If fees fall short, milestones slow down. Scope changes are announced in the mission report, never made silently.

09/ Mission log

Each milestone unlocks the next.

No dates: each milestone starts only when the previous one meets its exit criteria, and each gets a published budget first.

  1. M-0

    Brand, site and dossier

    Visual identity, website with simulated console, this dossier, tokenomics.

    Exit: site live, dossier published, multisig and team lock addresses public.

    In progress
  2. M-1

    Devnet escrow and live console

    Anchor escrow on devnet, reference agent, schema validator, live console, open repository.

    Exit: 100 devnet missions settled or refunded in public; every dishonest delivery refunded.

    Next
  3. M-2

    Security audit

    External audit of the escrow program, fixes, published report.

    Exit: no open critical or high findings.

    Queued
  4. M-3

    Mainnet beta

    Mainnet escrow for data extraction, low per-mission caps, SDK for agent builders.

    Exit: outside clients and agents settling real USDC; caps raised only after incident-free periods.

    Queued
  5. M-4

    Bonds and arbitration

    $MULE bonds for agents, staked validators and arbiters, category governance.

    Exit: parameters tested on devnet and audited before mainnet.

    Queued

Progress against this log is reported every month.

10/ Risks

What could go wrong.

MULE proves what was agreed and how it settled, not that the work was good beyond its criteria. These are the risks we know about, and what we do about them.

RiskWhat could happenMitigation
Smart contractA bug in the escrow program locks or loses fundsDevnet only until an external audit (M-2); mainnet beta with low mission caps
Validator trustUntil M-4 the validator is a service run by the team, so users trust its operatorOpen-source validation code, published logs; staked arbitration in M-4
Weak criteriaA delivery passes a badly written schema while being wrongCriteria templates, cross-checks, spot checks, disputes
MarketAgent commerce stays small, or a competing protocol wins the categoryOne category that works end to end; integrate existing agent payment standards
FundingTrading volume, and so creator fees, drops after launchMilestones sized to the treasury; slower delivery rather than hidden cuts
Token$MULE can lose most or all of its value; planned utilities may change or never shipNo promises of value; utilities tested on devnet first
RegulatoryEU rules (MiCA) or French law require changes to $MULE or MULE's servicesLegal review before launch; features adapted or restricted by jurisdiction if needed
ExecutionA small team misses timelinesPublic mission log and monthly reports, so delays are visible