Guardrail-Sim

Test your AI pricing policies before they cost you millions.

Everyone's building the gas pedal—AI agents that negotiate, discount, and close deals. But what happens when your LLM gives away margin at scale? Guardrail-Sim is the brakes and steering.

The Problem

You're deploying an AI sales agent. It can negotiate discounts. But:

• Will it honor your margin floors? Or give 40% off to anyone who asks nicely?
• How does it behave at scale? One bad discount is a rounding error. 10,000 is a crisis.
• Can you prove compliance? When finance asks, "what are the rules?", show them—don't guess.

The Solution

Guardrail-Sim lets you simulate thousands of adversarial buyer interactions against your pricing policies before going live. Define rules. Spawn buyer personas that try to game them. See what breaks.

Define Policy → Simulate Attacks → Fix Gaps → Deploy with Confidence

Try It Now

git clone https://github.com/jeffgreendesign/guardrail-sim.git
cd guardrail-sim && pnpm install && pnpm build

Run a simulation with 5 adversarial buyer personas:

pnpm demo

Expected output:

===============================================
  GUARDRAIL-SIM · Simulation Report
===============================================

  Sessions: 50  |  Seed: 42

  Approval Rate ····· 46.0%
  Avg Discount ······ 9.0%
  Avg Margin ········ 28.4%

  PERSONA OUTCOMES
  budget-buyer········ 10/10 approved
  strategic-buyer·····  9/10 approved
  margin-hunter·······  0/10 approved
  volume-gamer········  4/10 approved
  code-stacker········  0/10 approved

Or test a single policy evaluation:

import { PolicyEngine, defaultPolicy } from '@guardrail-sim/policy-engine';

const engine = new PolicyEngine(defaultPolicy);

const result = await engine.evaluate(
  { order_value: 5000, quantity: 100, product_margin: 0.4 },
  0.12
);

console.log(result.approved); // true
console.log(result.violations); // []

Project Status

Component	Status	Description
Policy Engine	Complete	Deterministic rule evaluation with json-rules-engine
MCP Server	Complete	7 MCP tools including simulation and UCP-aligned
UCP Types	Complete	Universal Commerce Protocol type definitions
Insights	Complete	Policy health checks and recommendations
Simulation	Complete	Adversarial buyer personas and negotiation loops

Default Policy Rules

• Margin floor: 15% minimum margin
• Max discount: 25% cap
• Volume tiers: 10% base, 15% for qty >= 100

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                     Simulation Engine                           │
│          5 Buyer Personas · Adversarial Negotiation Loops       │
│               Deterministic (seeded PRNG)                      │
└─────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│                       Policy Engine                            │
│              json-rules-engine · Deterministic Evaluation       │
│                    Exposed via MCP Server (7 tools)             │
└─────────────────────────────────────────────────────────────────┘

Project Structure

packages/
├── policy-engine/     ✅ Deterministic rule evaluation (json-rules-engine)
├── simulation/        ✅ Adversarial buyer personas + negotiation loops
├── mcp-server/        ✅ MCP server with 7 tools (policy + simulation + UCP)
├── ucp-types/         ✅ UCP type definitions and converters
├── insights/          ✅ Policy health checks and recommendations
apps/
└── website/           ✅ Fumadocs documentation site + interactive playground
examples/
├── ucp-integration-demo/   UCP discount validation scenarios
└── simulation-demo/        Run simulation and see results

Commands

pnpm install          # Install dependencies
pnpm build            # Build all packages
pnpm test             # Run tests (109 passing)
pnpm demo             # Run simulation demo
pnpm lint             # Run ESLint
pnpm format           # Format with Prettier

Per-package:

pnpm --filter @guardrail-sim/policy-engine build   # Build single package
pnpm --filter @guardrail-sim/policy-engine test    # Test single package

MCP server:

npx @guardrail-sim/mcp-server  # Run MCP server

Documentation

• Getting Started — Quick start guide and setup
• Architecture — Technical decisions with rationale
• MCP Tools — MCP implementation reference
• Policy Concepts — Policy structure and evaluation

Contributing

We welcome contributions! Please see CONTRIBUTING.md for:

• Development setup instructions
• How to run tests locally
• Pull request guidelines

License

MIT