VerifiMind-PEAS

<div align="center">
  <img src="docs/assets/branding/VerifiMind-PEAS-Logo-transparent.png" alt="VerifiMind PEAS" width="200"/>

  # VerifiMind PEAS

  **An opinionated MCP server for structured multi-LLM critique.**

  Three specialized agents — Innovation, Ethics, Security — review your concept before you build it. Multi-vendor (Gemini · Claude · GPT · Groq · Cerebras · Mistral · Ollama). Free, open-source, MCP-native.

  [![Version](https://img.shields.io/badge/version-v0.5.61-blue.svg)](CHANGELOG.md)
  [![License](https://img.shields.io/badge/license-MIT-green.svg)](LICENSE)
  [![Status](https://img.shields.io/badge/status-Operational-success.svg)](SERVER_STATUS.md)
  [![MCP Registry](https://img.shields.io/badge/MCP%20Registry-Listed-purple)](https://registry.modelcontextprotocol.io/?q=verifimind)
  [![Health](https://img.shields.io/badge/Health-LIVE-success)](https://verifimind.ysenseai.org/health)
  [![Genesis DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.17972751.svg)](https://doi.org/10.5281/zenodo.17972751)
  [![MACP DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.21345820.svg)](https://doi.org/10.5281/zenodo.21345820)
  [![Dataset DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.21276884.svg)](https://doi.org/10.5281/zenodo.21276884)
  [![HuggingFace Dataset](https://img.shields.io/badge/HuggingFace-Dataset-yellow)](https://huggingface.co/datasets/YSenseAI/verifimind-peas-eval)
</div>

---

## Quick Start

> Use `streamable-http` transport and the trailing slash `/mcp/`. See [docs/MCP_Server_Troubleshooting_Guide.md](docs/MCP_Server_Troubleshooting_Guide.md) if you hit issues.

**Claude Code** (one command):
```bash
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`):
```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`):
```json
{
  "servers": {
    "verifimind": {
      "url": "https://verifimind.ysenseai.org/mcp/",
      "transport": "streamable-http"
    }
  }
}
```

After [registering](https://verifimind.ysenseai.org/register), add `--header "X-VerifiMind-UUID:${VERIFIMIND_UUID}"` to opt into the personal usage dashboard at `/early-adopters/dashboard/{uuid}`. Registration is optional.

---

## What this is

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

| Agent | Role | Question it answers |
|---|---|---|
| **X (Innovation)** | Innovation & competitive positioning | "Is this novel? What's the prior art? What's the strategic angle?" |
| **Z (Ethics)** | Ethics, compliance, 21-framework jurisdictional check | "What risks does this raise? GDPR, EU AI Act, SG MGF, etc." |
| **CS (Security)** | Security validation, OWASP Agentic AI Top 10 | "What can break? What's the attack surface? What's the reasoning-layer audit say?" |

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](#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):**
| Tier | Identity | Limit |
|---|---|---|
| Anonymous | IP only | 10 req/60s |
| Scholar | UUID (free registration) | 30 req/60s |
| EA / PILOT | UUID + email | 100 req/60s |

---

## 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](#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](docs/architecture/VerifiMind-PEAS-Architecture-Diagram.md)**

---

## Related Work

VerifiMind PEAS builds on and acknowledges:

- **ChatEval** (Chan et al., 2023, [arXiv:2308.07201](https://arxiv.org/abs/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](https://crfm.stanford.edu/helm/)) — Holistic Evaluation of Language Models
- **Inspect** ([UK AI Safety Institute](https://inspect.ai-safety-institute.org.uk/)) — 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:** `v0.5.61` — [verifimind.ysenseai.org](https://verifimind.ysenseai.org) · [/health](https://verifimind.ysenseai.org/health)
- **Landing Page:** [verifimind.io](https://verifimind.io)
- **Release:** merge [`3019f5c4`](https://github.com/creator35lwb-web/VerifiMind-PEAS/commit/3019f5c4889d8334063d4a2d9243e87d96fc93a8) · Cloud Build `be6ed621` · post-deploy smoke **31/0/0**
- **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/gp