Home

VerifiMind PEAS Wiki

Structured multi-LLM critique with human final authority

VerifiMind PEAS is an MCP server that reviews a concept through three specialized perspectives: innovation, ethics, and security. It helps people find questions, risks, and trade-offs before they build.

Connect · Tool reference · Current production status · Public statements

---

Status boundary: This Wiki teaches the durable concepts and operating workflow. Runtime version, model IDs, routing, protocol version, tool availability, and deployment evidence change independently. Check /health, /setup, the live server card, and Current Production Status before relying on an operational fact.

Start here

Goal	Page
Connect an MCP client	Installation
Understand what each callable tool does	MCP Tools Reference
Use your own model provider	BYOK Guide
Understand the public X → Z → CS runtime	AI Council
Learn the broader human-directed methodology	Genesis Methodology
Check what is live now	Current Production Status
Review incidents, limitations, and exact release evidence	Public Statements

The stable mental model

The public Trinity runtime has three sequential analysis stages:

Stage	Lens	Core question
X — Innovation & Strategy	novelty, alternatives, market and execution	What is promising, weak, or already known?
Z — Ethics & Compliance	affected people, rights, governance and ethical red lines	What could cause harm or require human/legal review?
CS — Security	threats, misuse, data and system integrity	What can break, and what should be tested or contained?

run_full_trinity executes X → Z → CS. Later stages receive only trustworthy reasoning from earlier completed stages. The service then returns a structured assessment for a human to interpret.

The human remains responsible for scope, evidence, trade-offs, and the final decision. A tool recommendation is an input to judgment, not delegated authority.

What the service does not claim

• It is not formal verification or mathematical proof.
• It is not legal, regulatory, medical, financial, or security certification.
• It is not exhaustive; untested risks may remain.
• Multiple seats do not guarantee independent models or providers. Check the live routing record for the run you are evaluating.
• A high score is not permission to ignore findings, degraded stages, or human review.

Current availability contract

The current production contract defines 13 tools:

• 8 active: four Trinity tools and four built-in template-reading tools.
• 5 temporarily unavailable: two custom-template mutation tools and three coordination tools.

The unavailable tools remain defined for schema stability but return explicit maintenance denials. They do not read, write, or reveal a shared coordination namespace. Their unavailability is security containment, not a paywall.

Availability is operational state. Verify it from /health rather than inferring it from the pricing pledge or an old release note.

All active tools remain free under the Core Tools Always Free pledge. Separate professional services, if offered, do not change tool availability.

Quick start

Use the hosted MCP endpoint with streamable HTTP:

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

For a native streamable-HTTP client:

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

Then ask the client:

Use run_full_trinity to review this concept. Preserve uncertainty, identify any unavailable stage, and explain which findings need human follow-up.

See Installation for client-specific setup and verification.

How to read a result

Treat these as separate facts:

1. Stage quality: Was each stage produced by real inference, or marked degraded/unavailable?
2. Findings: What concerns, vulnerabilities, alternatives, and evidence did the completed stages produce?
3. Aggregate: Was a combined score or confidence allowed? Incomplete runs must not look complete.
4. Failure trace: Which stage, provider, model, and normalized failure class were reported, and was retry guidance actually applicable?
5. Human decision: What remains unverified, and who has authority to decide?

Provider failures can leave a run explicitly incomplete while preserving trustworthy sibling stages. Never interpret missing output as a clean finding.

Privacy defaults

run_full_trinity defaults save_to_history to false. If a caller explicitly opts into aggregate history, production keeps only a bounded, instance-local set and does not guarantee a fixed time-based retention period. Incomplete exception runs are not written to that history.

Supplying a UUID is a separate optional choice that may create pseudonymous usage metadata. Read the live Privacy Policy and Terms before sending personal, confidential, regulated, or third-party data.

Learn, operate, verify

Textbook

• Genesis Methodology
• Architecture
• AI Council

Operator playbook

• Installation
• MCP Tools Reference
• BYOK Guide
• Early Adopter Program

Evidence and trust

• Current Production Status
• Public Statements
• Statement 001 — Trinity Integrity
• Evaluation Roadmap v1.0
• Publications

---

VerifiMind PEAS is MIT licensed. Forks and derivatives must use distinct branding. See the repository LICENSE.