An allocating agent cannot interview the agent it is about to pay. So the interview becomes a call. MIZAN gives any agent two tools: verify a trading or AI agent's track record before it pays, copies or allocates, and prove its own record from positions only, with the strategy never leaving its machine. Verdicts come back exactly as issued, PASS, FAIL or NOT SIGNIFICANT. Free to verify, forever.
Any MCP host (Claude Code, Claude Desktop, Cursor, OpenAI-style agents):
claude mcp add --transport http mizan https://studio.mizan.market/mcp # tools: verify_credential · list_datasets · mint_credential
Agent-to-agent (A2A): send a message, get a Task back.
curl -X POST https://studio.mizan.market/a2a -H 'content-type: application/json' \
-d '{"jsonrpc":"2.0","id":1,"method":"message/send","params":{"message":{"messageId":"m1","role":"user","parts":[{"kind":"text","text":"verify dsr10-flagship"}]}}}'
# → Task completed · verdict NOT_SIGNIFICANT · hash_match true
Python, no dependencies:
curl -O https://mizan.market/skill/mizan_agent_tools.py python3 mizan_agent_tools.py dsr10-flagship
verify_credential looks the credential up in the public registry, downloads the proof bundle, and checks its hash against the registry anchor. With the open-source verifier installed it re-runs the zero-knowledge proof offline, trusting no one, including us. A tampered bundle comes back UNVERIFIED. A missing one does too. A FAIL is a result, not a secret, and a credential is a backtest seal, history made honest, not a forecast.
Run your agent on a canonical public market on your own machine. Send one position per bar-move, nothing else. The circuit recomputes the record from those positions under committed costs and proves it on a GPU. Not one parameter, weight or prompt leaves the owner. Preview is free; the real proof is a 225 KB credential anyone verifies in about 81 ms.
Today minting runs with a free account during the founding cohort. Next, pay-per-mint over x402: no account, USDC on Base, charged only when a real proof is queued, so a gate fail costs nothing. Built, opening after the founding cohort.
MIZAN is built as an ERC-8004 validator: a credential's verdict can be posted to an agent's own on-chain identity as reputation, PASS 100, NOT SIGNIFICANT 50, FAIL 0, with the bundle committed by hash. Registration is pending. A drafted VTR-1 Agent Profile extends the open standard for agents: identity binding, a cumulative trial count that counts every model swap, forward-only records, and an allocate-check an agent runs before moving money.
Honest status: the rails are built and listed; no external agent has called them yet. They are built so the first one can.