VerifiMind PEAS - RefleXion Trinity

Quick Start

Use streamable-http transport and the trailing slash /mcp/. See docs/MCP_Server_Troubleshooting_Guide.md if you hit issues.

Claude Code (one command):

claude mcp add -s user verifimind -- npx -y mcp-remote https://verifimind.ysenseai.org/mcp/

Claude Desktop (macOS: ~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json):

{  "mcpServers": {    "verifimind": {      "command": "npx",      "args": ["-y", "mcp-remote", "https://verifimind.ysenseai.org/mcp/"]    }  }}

Cursor / VS Code Copilot (.cursor/mcp.json or .vscode/mcp.json):

{  "servers": {    "verifimind": {      "url": "https://verifimind.ysenseai.org/mcp/",      "transport": "streamable-http"    }  }}

Accounts: registration and UUID-linked account features (including the personal usage dashboard) are temporarily unavailable during security maintenance. You do not need an account today — every active tool works anonymously at the standard rate limit. Advance notice: under the published Terms (v2.6) and Privacy Policy (v2.7), both effective October 15, 2026, the four execution tools (consult_agent_x, consult_agent_z, consult_agent_cs, run_full_trinity) will require a free registered account from a date that will be announced on those pages at least 14 days in advance, and not before; no date has been announced. Tool discovery and the template-read tools stay open without registration, and the tools remain free.

What this is

VerifiMind PEAS is an MCP server that runs your concept through three specialized LLM judges in sequence:

Each later stage receives the prior completed stage's structured analysis. You get a unified assessment with scores, findings, recommendations (PROCEED / REVISE / REJECT), and explicit stage provenance. Only documented output fields are passed forward; hidden model internals are not exposed.

What this is not: "Verification" in the formal-methods sense. The output is structured multi-LLM critique, not a mathematical proof. We make that distinction explicitly.

The 13 tools

All 13 tools remain free for everyone under the Core Tools Always Free pledge. 8 are currently active; 5 are temporarily unavailable for security maintenance — three coordination tools plus two custom-template mutation tools.

Trinity validation (4 tools)

• consult_agent_x — Innovation analysis with competitive positioning
• consult_agent_z — Ethics review with 21-framework jurisdictional coverage
• consult_agent_cs — Security validation, OWASP Agentic AI Top 10
• run_full_trinity — X → Z → CS pipeline with structured stage context and a unified assessment

Output-integrity contract (v0.5.58): an agent's generated scores, findings, recommendation, analysis, and veto state are returned only when its inference quality is explicitly real. Provider failures now degrade only the affected stage, preserve successful sibling stages, and expose typed stage attribution. Incomplete runs withhold aggregate score/confidence and are not saved to shared history.

Template management (6 tools — 4 active, 2 temporarily unavailable)

• list_prompt_templates — Browse templates by agent, category, or tag
• get_prompt_template — Retrieve a template by ID
• export_prompt_template — Export to Markdown or JSON
• register_custom_template — temporarily unavailable pending owner-scoped storage
• import_template_from_url — temporarily unavailable pending owner isolation and URL-fetch hardening
• get_template_statistics — Registry stats by agent / phase / type

Coordination (3 tools) — TEMPORARILY UNAVAILABLE

• coordination_handoff_create — disabled
• coordination_handoff_read — disabled
• coordination_team_status — disabled

These three tools are disabled and currently return COORDINATION_TEMPORARILY_DISABLED for every caller.

Records created through them were stored in a shared, unauthenticated namespace; they are no longer readable or writable through the public API. The four validation tools remain fully available.

They will return only after private, owner-scoped storage ships. Until then, keep coordination state in your own repository — the handoff markdown format is documented in this repo. Incident reference: VM-IR-2026-07-28-COORD-01.

The Always Free pledge is unchanged. This is a security containment, not a paywall and not a tier change. Nothing here is or becomes a paid feature.

Custom-template mutation (2 tools) — TEMPORARILY UNAVAILABLE

register_custom_template and import_template_from_url return CUSTOM_TEMPLATE_TEMPORARILY_DISABLED for every caller. Public built-in template listing, retrieval, export, and statistics remain available and exclude process-local custom entries.

This containment prevents cross-caller custom-template visibility while owner-scoped storage is built and removes the public arbitrary URL-fetching path. Incident reference: VM-IR-2026-08-01-TEMPLATE-01.

Core Tools Always Free Pledge

All VerifiMind PEAS validation tools are free to use, forever. No paywall, no premium tier for tool access. Rate limits apply for system health only (not as monetization). Paid services, when they launch, will be consultation reports — separate from the tools.

Ratified by L (CEO) + Alton (Human Orchestrator) + T (CTO) on May 9, 2026. Active in production since v0.5.28 (May 10, 2026).

Rate limits (system health, equal for all tiers):

While account features are in maintenance, every caller is served at the anonymous limit.

Methodology overview

VerifiMind PEAS productizes the multi-judge LLM evaluation pattern — a well-established approach in the AI evaluation literature — into an opinionated MCP server with three specialized roles, a Genesis Master Prompt continuity layer, and a multi-vendor BYOK architecture.

What's ours:

• Productization quality of the X / Z / CS specialization
• MCP-native exposure (works in Claude Code / Cursor / VS Code / ChatGPT Codex)
• Multi-vendor design (not locked to one LLM family)
• Genesis Master Prompt — stateful continuity across multi-model workflows
• 21-framework jurisdictional coverage in the Ethics agent (GDPR · EU AI Act · SG MGF · etc.)

What's prior art: Multi-judge LLM evaluation, LLM-as-judge scoring, multi-model orchestration. See Related Work for citations.

We do not claim the underlying methodology is novel.

Architecture

The full system architecture — from the foundational X / Z / CS multi-agent validation design through the Phase 90 production deployment (MCP transport, BYOK provider layer, security hardening, FLYWHEEL coordination) — is documented in a single comprehensive, fact-checked diagram:

→ VerifiMind-PEAS Architecture Diagram

Related Work

VerifiMind PEAS builds on and acknowledges:

• ChatEval (Chan et al., 2023, arXiv:2308.07201) — Multi-agent debate framework
• MAJ-EVAL — Multi-Agent-as-Judge evaluation pattern
• CollabEval — Collaborative LLM evaluation with role-based agents
• HELM (Stanford CRFM) — Holistic Evaluation of Language Models
• Inspect (UK AI Safety Institute) — Open-source safety evaluation framework
• G-Eval / GPTScore — LLM-as-judge scoring methodologies

Our contribution: productization quality, MCP integration path, multi-vendor architecture, and the Genesis Master Prompt continuity layer.

Status & Metrics

• Server: v0.5.68 — verifimind.ysenseai.org · /health
• Landing Page: verifimind.io
• Release: merge 35ea615 · Cloud Build 5401b7a8-cb46-4b22-8074-38933b0e9ce8 · revision verifimind-mcp-server-00523-pwq · security currency: Authlib 1.9.4 · joserfc 1.7.7 · pydantic 2.14.2 as one tested set, on v0.5.67's FastMCP 4.1.2 · Starlette 1.8.0 · FastAPI 0.143.3; mcp 2.2.0 and the 2026-07-28 era unchanged · post-deploy Trinity, one anonymous run as observed: X real, Z real, CS real, outcome full — one run, not a reliability figure (the Z/CS limitation in SERVER_STATUS.md is the standing condition)
• Authentication (v0.5.63): a native OAuth 2.1 authorization server is deployed dark — credential issuance and MCP enforcement are both off, so every tool that was anonymous stays anonymous
• Tools: 13 defined / 8 active / 5 temporarily unavailable (all remain free)
• Providers: 6 remote BYOK catalogues (Gemini · Anthropic · OpenAI · Groq · Cerebras · Mistral) plus caller-managed local Ollama
• Hosted routing: X = Gemini gemini-3.5-flash-lite; Z/CS = Groq openai/gpt-oss-120b
• Runtime provider failover: disabled (runtime_failover_enabled: false)
• MCP Registry package: 3.43.0 — published by the v0.5.66 GitHub Release on October 18, 2026 and still the live package. The v0.5.68 GitHub Release, which also covers v0.5.67, publishes 3.45.0; the 3.44.0 manifest prepared for v0.5.67 was never published

For live release history, see /changelog. We deliberately do not display unaudited "total users" numbers because bot, scanner, and development traffic can distort them. Exact deployment provenance and the full dated operational snapshot are documented in SERVER_STATUS.md.

Common mistakes

For a fuller troubleshooting guide, see docs/MCP_Server_Troubleshooting_Guide.md.

How to cite

If you use VerifiMind PEAS in research or a project, please cite. We'd love to hear about it — open a GitHub Discussion.

VerifiMind PEAS (server)

@software{verifimind_peas_2026,  author  = {Lee, Alton and {Manus AI} and {Claude Code}},  title   = {VerifiMind PEAS: Multi-Agent AI Validation MCP Server},  year    = {2026},  url     = {https://github.com/creator35lwb-web/VerifiMind-PEAS},  doi     = {10.5281/zenodo.17980791},  note    = {Multi-vendor MCP server for structured multi-LLM critique}}

[VerifiMind DOI]

Genesis Methodology

@misc{genesis_methodology_2025,  author  = {Lee, Alton and {Manus AI}},  title   = {Genesis Prompt Engineering Methodology: Multi-Agent AI Validation Framework},  year    = {2025},  url     = {https://doi.org/10.5281/zenodo.17972751},  doi     = {10.5281/zenodo.17972751}}

[Genesis DOI]

MACP (Multi-Agent Communication Protocol)

@misc{lee_macp_2026,  author    = {Lee, Alton Wei Bin and {VerifiMind-PEAS FLYWHEEL TEAM}},  title     = {{Multi-Agent Communication Protocol (MACP) v2.5 — Loop Engineering}},  year      = {2026},  version   = {2.5.0},  publisher = {Zenodo},  url       = {https://doi.org/10.5281/zenodo.21345820},  doi       = {10.5281/zenodo.21345820}}

[MACP DOI]

VerifiMind PEAS Evaluation Dataset (M2)

@dataset{verifimind_peas_eval_2026,  author    = {Lee, Alton Wei Bin},  title     = {{VerifiMind-PEAS Evaluation Dataset v1.0: Human-Annotated Ground Truth for Multi-Agent Epistemic Verification}},  year      = {2026},  publisher = {Zenodo},  version   = {v1.0},  doi       = {10.5281/zenodo.21276884},  url       = {https://doi.org/10.5281/zenodo.21276884},  note      = {100-item dataset with 5 domains, ground-truth verdicts, multi-model scoring, and inter-annotator agreement analysis}}

[Dataset DOI]

Defensive Publication

A prior-art defensive publication is registered at DOI 10.5281/zenodo.17645665.

Documentation & links

License

VerifiMind PEAS is released under the MIT License. See LICENSE for the full text.

Copyright (c) 2025-2026 Alton Lee Wei Bin (creator35lwb)
Permission is hereby granted, free of charge, to any person obtaining a copyof this software and associated documentation files (the "Software"), to dealin the Software without restriction, including without limitation the rightsto use, copy, modify, merge, publish, distribute, sublicense, and/or sellcopies of the Software, and to permit persons to whom the Software isfurnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included inall copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS ORIMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.

The methodology is freely usable under MIT. Forks and derivatives must use different branding.

Community

• GitHub Discussions: github.com/creator35lwb-web/VerifiMind-PEAS/discussions — preferred for questions, ideas, and feedback
• Issues: github.com/creator35lwb-web/VerifiMind-PEAS/issues — bugs and feature requests
• Email: [email protected] — direct contact
• X (Twitter): @creator35lwb

For paid consultation engagements (planned, not yet active), use GitHub Discussions or email — we'll publish service details and pricing when they're ready.

Acknowledgments

VerifiMind PEAS was built collaboratively by the FLYWHEEL TEAM — a human orchestrator working with multiple AI agents (Manus AI, Claude Code, Perplexity, Antigravity/Gemini, GodelAI). Multi-agent coordination uses the open MACP protocol (v2.5 "Loop Engineering", published 2026-07-14).

The development journey is documented publicly in the Validation Paradox research collection and the Public Statements & Disclosures wiki.

External Model Council review (Claude Opus 4.7 + GPT-5.5 + Gemini 3.1 Pro, May 9, 2026) shaped the current positioning. See docs/case-studies for application examples.

Last Updated: November 5, 2026 · Version: v0.5.68 · MCP Registry: 3.43.0 (3.45.0 publishes with the v0.5.68 Release) · MACP: v2.5 "Loop Engineering" (DOI)