DJZS Trust MCP

What DJZS Trust MCP MCP server does

DJZS Trust MCP MCP server evaluates the reasoning behind a proposed prediction-market position before execution. It is designed to distinguish whether a thesis should proceed, wait for clarification, or fail because of identified reasoning defects. The returned result includes a verdict, caller-facing action, risk score, flags, unknown fields, disagreements, extraction status, scope status, taxonomy versions, and a reproducible verdict_hash.

The prediction-market path uses the frozen DJZS-PM-v1.0 taxonomy. Its four findings cover a mismatch between the thesis and the resolution question, missing falsification conditions, unsupported probabilities, and consensus without a differentiated edge. The weighted score uses a fail threshold of 25. A lone consensus finding is advisory because its weight is below that threshold, while combined findings can produce a failure.

How it works

A free-text thesis first goes through an extraction stage. Three LLM samples identify observable facts rather than choosing a verdict. Fields require unanimous state agreement; disagreements become unknown. Claims that something is absent also need quoted evidence from the original intent, with critical absences requiring the same quote across all samples.

The resulting AuditInput contains tri-state fields: present, absent, or unknown. Deterministic code then applies the rules and weights. The same input produces the same verdict and hash, calculated from canonicalized verdict data, score, sorted flags, and unknown fields. If a decision-critical fact remains unknown, the result is WAIT and the caller action is HALT, with the missing fields returned for clarification.

Tools and capabilities

DJZS Trust MCP MCP server exposes three tools through streamable HTTP at /mcp:

• verify_pm_trade audits a prediction-market thesis. It is PM-only and returns in_scope: false for text that does not resolve as a prediction-market thesis.
• query_pol_certificates reads existing ProofOfLogic certificates from the Irys datachain.
• query_agent_trust reads subgraph-indexed trust information, including total audits, fail rate, latest verdict, and a PROCEED, HALT, or NO_HISTORY action.

The audit gate is available at https://mcp.djzs.ai/mcp. The service also exposes a separate x402 HTTP endpoint at /x402/verify, where an unpaid request receives a 402 challenge for a 2.00 USDC payment on Base mainnet.

Limitations and notes

The live verification tool does not write a new ProofOfLogic certificate; it returns a verdict hash only. Existing Irys certificates represent earlier architecture. Taxonomy hash constants are also not included in the tool response.

The broader DJZS-LF-v1.1 perpetuals and general-reasoning taxonomy has 11 defined codes, but only three are wired into the live engine. It should not be treated as full coverage of that taxonomy.

The trust tool recomputes fail rates from individual verdicts and excludes WAIT from failures. Raw subgraph counters may differ because the current mapping counts WAIT in failCount; the tool returns both values and labels the indexed counters. The discrepancy remains until the subgraph is corrected and re-indexed.