jev-judge-mcp — Design Architecture

jev-judge-mcp — Design Architecture An architecture diagram generated by Archify. MCP Client · host agent · Architecture component MCP Client host agent JevMCPServer + Toolset · stdio · streamable-http · one-way judgment layers (ADR-0002) JevMCPServer + Toolset stdio · streamable-http arguments · compiled parsers · one-way judgment layers (ADR-0002) arguments compiled parsers limits · frozen caps (ADR-0014) · one-way judgment layers (ADR-0002) limits frozen caps (ADR-0014) Eleven JevTools · 10 frozen + jev_score · one-way judgment layers (ADR-0002) Eleven JevTools 10 frozen + jev_score Runtime · lazy provider · cache · one-way judgment layers (ADR-0002) Runtime lazy provider · cache Providers ×4 · envelope-only ×4 · one-way judgment layers (ADR-0002) Providers ×4 envelope-only ×4 Settings → Redactor · env-only · derived secrets · Architecture component Settings → Redactor env-only · derived secrets extract · dialect · killable workers · one-way judgment layers (ADR-0002) extract dialect · killable workers serialize + text · JSON.stringify · UTF-16 · Architecture component serialize + text JSON.stringify · UTF-16 Policy (pure) · thresholds · decisions · one-way judgment layers (ADR-0002) Policy (pure) thresholds · decisions Validation · answers · caps · one-way judgment layers (ADR-0002) Validation answers · caps Parity oracle · 282 fixtures · replay · Architecture component Parity oracle 282 fixtures · replay Telemetry · spans + metrics · in-process · Architecture component Telemetry spans + metrics · in-process Key file · env wins over file · Architecture component Key file env wins over file Response cache · JEV_MCP_CACHE · Architecture component Response cache JEV_MCP_CACHE setup CLI · verify, store 0600 · Architecture component setup CLI verify, store 0600 Eval harness · evals/agent.py · spend · Architecture component Eval harness evals/agent.py · spend install CLI · 8 agent targets · uvx · Architecture component install CLI 8 agent targets · uvx Command hook · gate · deny/ask/abstain · Architecture component Command hook gate · deny/ask/abstain doctor CLI · offline config check · Architecture component doctor CLI offline config check tools/call validate args typed args cap values ask Jev evaluate raw answers judgments actions + reasons result text config · redaction patterns byte-equal (ADR-0001) spans · metrics paid agent runs stored key fallback verified key replay identical hook gate key presence one-way judgment layers (ADR-0002)

One registry (ADR-0013)

  • • list_tools ≡ call_tool — one Toolset
  • • unknown tool returns owned text

Model judges, policy decides (ADR-0002)

  • • validation rejects; policy is pure (traced in tools/observed)
  • • truncated or unknown confidence ⇒ never auto

Trust boundary (ADR-0008/0017)

  • • env-only config; tool args are model-controlled
  • • providers: typesafe, openrouter, cloudflare, compatible

Frozen parity (ADR-0001/0006/0016)

  • • 282 fixtures replay byte-equal; divergences ADR-tagged
  • • extract: one deadline incl. queue; stdlib re workers