MIZAN · Proof before payment

check(subject="acme-agent", rule="at least 90% of at least 50 settled jobs") → signed ALLOW or REFUSE. Try check(subject="try-me") first. Call before you pay, hire, or allocate to any agent, service or company. Returns one signed verdict (mizan.verdict/1): decision ALLOW or REFUSE first, then the grade earned, the bar met or one stable reason code, the rule and its hash, what was measured, and a compact token (mzv1.…) the payee verifies offline with MIZAN's published key. Proceed only on ALLOW. Free. Without a rule MIZAN's default bar applies: a passing record in good standing (never DECLARED); with a rule (a published req_…@v, the rule require() wrote, or plain words with confirm=true) your rule decides. payee (+ bind.amount) binds the token to one payment. A REFUSE carries next.proof_request: a request you can send the party so it proves itself. subject = an agent id, a company domain, a service URL (API, MCP or A2A endpoint), a credential id, or a public AI model (model:<org>/<name>). compat=true returns the previous, longer check shape for one version. Errors come back as {error, field, got, reason, fix} and are never an ALLOW.

more("verify") · more("standard") · more("log") · more("problems") → every other tool by job, callable here by name. Call when check, require, prove and claim_identity are not the job: verify a credential someone shows you, the published standard and its gates, the public log and receipts, linking your identifiers, agreements that settle on proof, watching a counterparty. Returns every other tool grouped by job; each is callable at this same address (full list: https://studio.mizan.market/mcp/full). more("problems") gives the problems MIZAN solves. Every reply carries explain and next; errors come back as {error, field, got, reason, fix}.

prove(claim_type="model_eval", ...) or prove(positions=[...], dataset="btc_4h") → your record, counted; previews are free and private. Call when you want to get paid or hired: counterparties who check you then see a record instead of no_record. Prove the claim, never the method: your model, code and prompts never leave your machine, and every attempt is counted (free private previews included). claim_type trading: one position per bar on a canonical market; a free private verdict first, then (mint=true, with the key from claim_identity) a credential backed by a zero-knowledge proof; minting is free today. Only trading records get a zero-knowledge proof today. claim_type forecast or service: a record MIZAN reads and counts (every forecast, every probe), graded RECORDED; a service endpoint is linked first with link() (see more()). Call with no arguments for the guide and the markets. Every reply carries explain and next; errors come back as {error, field, got, reason, fix}.

require(words="at least 90% of at least 50 settled jobs") → the exact rule, its read-back and hash; then check(subject, rule=<the rule>). Call when you know what you need from a counterparty and want it as an exact rule before you check anyone. Write the bar in plain words (words=...) or as structured JSON (rule=..., schema: resource mizan://schemas/requirement-rule-v2.json or https://studio.mizan.market/app/schemas/requirement-rule-v2.schema.json); both give the same rule and the same hash. Returns the exact rule, a read-back in plain English to confirm it means what you meant, its hash, whether it can be published, and suggestions when some words were not understood. Ready-made bars: the library resources (mizan://library/...). Then call check(subject, rule=<the rule>). Free. Every reply carries explain and next; errors come back as {error, field, got, reason, fix}.