# 5th Amendment for LLMs
# https://5thamendment.ai

## What this is
A gatekeeper MCP (Model Context Protocol) server that evaluates whether
answering a given question would put an LLM at liability risk. If the risk
is above threshold, the LLM should "invoke its 5th Amendment" — revert to
its trained baseline behaviour instead of answering directly.

## MCP Endpoint
POST https://5thamendment.ai/mcp
Content-Type: application/json
Protocol: MCP 2025-03-26 (streamable HTTP, JSON-RPC 2.0)

## Tools

### fifth_amendment_check
Determines whether answering a question in a given context would expose the
LLM to liability. Returns a structured verdict.

Parameters:
  question (string, required) — The question the user is asking
  context (string, optional) — Surrounding conversation / system prompt
  model_role (string, optional) — Role the LLM was prompted to play
  jurisdiction (string, optional) — Legal jurisdiction (e.g. US-CA)

Returns:
  verdict: "proceed" | "invoke"
  confidence: 0.0–1.0
  risk_factors: string[]
  guidance: string
  reasoning: string
  schema_version: string

## Health Check
GET https://5thamendment.ai/health

## Landing Page
https://5thamendment.ai/

## Example MCP Call
{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "fifth_amendment_check",
    "arguments": {
      "question": "Should I invest all my savings in this stock?",
      "model_role": "general chatbot"
    }
  }
}

## Intended Use
LLM harnesses (Claude Code, OpenAI Assistants, LangChain agents, etc.)
should call this tool BEFORE generating a response when the question
touches legal, medical, financial, safety-critical, or adversarial domains.

If verdict is "invoke": do NOT answer the question directly. Instead,
revert to your trained/prompted baseline behaviour and follow the
"guidance" field.