@guardrail-sim/simulation

Stress-test your pricing policies by running adversarial buyer personas through multi-round negotiation loops.

Installation

pnpm add @guardrail-sim/simulation

Quick Start

import { runSimulation, defaultPersonas } from '@guardrail-sim/simulation';
import { defaultPolicy } from '@guardrail-sim/policy-engine';
 
const results = await runSimulation({
  policy: defaultPolicy,
  personas: defaultPersonas,
  ordersPerPersona: 20,
  seed: 42,
});
 
console.log(`Approval rate: ${(results.metrics.approvalRate * 100).toFixed(1)}%`);
console.log(`Edge cases found: ${results.metrics.edgeCasesFound.length}`);

Buyer Personas

Five built-in personas cover the full spectrum of buyer behavior:

Custom Personas

import type { BuyerPersona } from '@guardrail-sim/simulation';
 
const enterpriseBuyer: BuyerPersona = {
  id: 'enterprise-buyer',
  name: 'Enterprise Buyer',
  strategy: 'strategic',
  discountRange: { min: 0.10, max: 0.20 },
  volumeRange: { min: 200, max: 1000 },
  marginRange: { min: 0.30, max: 0.50 },
  maxRounds: 5,
  adaptationRate: 0.25,
};

Negotiation Loop

Each session follows this flow:

1. Persona generates an initial discount request based on its strategy
2. Policy engine evaluates the request
3. If approved: session ends with acceptance
4. If rejected: persona adapts based on adaptationRate and tries again
5. If all rounds exhausted: session ends with rejection

The adaptationRate controls how much a persona concedes after rejection:

• Cooperative (0.5): Large concessions, quick resolution
• Strategic (0.3): Moderate concessions, gradual negotiation
• Adversarial (0.1-0.15): Minimal concessions, boundary probing

Reproducibility

All simulations are deterministic given the same seed:

const run1 = await runSimulation({ policy, personas, ordersPerPersona: 20, seed: 42 });
const run2 = await runSimulation({ policy, personas, ordersPerPersona: 20, seed: 42 });
 
// run1.metrics.approvalRate === run2.metrics.approvalRate (always)

Uses a seeded PRNG (mulberry32) — never Math.random().

Metrics

The SimulationMetrics object contains:

• totalSessions — Number of negotiation sessions
• approvalRate — Fraction of sessions that ended in acceptance
• averageDiscountApproved — Mean discount for approved sessions
• averageDiscountRequested — Mean discount across all requests
• averageMarginAfterDiscount — Mean margin for approved discounts
• violationsByRule — Count of violations per rule name
• outcomesByPersona — Accepted/rejected/abandoned per persona
• limitingFactors — Which rule limited most often
• edgeCasesFound — Boundary conditions detected

Edge Case Detection

The engine automatically detects:

• Discounts approved near the margin floor (<17% margin)
• Orders at volume tier boundaries (qty 95-105 with 10-15% discount)
• High discounts approved (>20%)

For targeted boundary testing:

import {
  generateVolumeBoundaryOrders,
  generateMarginFloorProbes,
  createBoundaryProber,
} from '@guardrail-sim/simulation';
 
// Test qty 99, 100, 101 with same discount
const volumeTests = generateVolumeBoundaryOrders(0.4, 5000);
 
// Test incremental margin floor approach
const marginTests = generateMarginFloorProbes(0.35);

Insights Bridge

Convert results to feed the @guardrail-sim/insights analysis engine:

import { toSimulationSummary } from '@guardrail-sim/simulation';
import { analyzePolicy } from '@guardrail-sim/insights';
 
const summary = toSimulationSummary(results);
const report = await analyzePolicy({
  policy: policySummary,
  simulationResults: summary,
});

This enables 8 built-in simulation insight checks including coverage analysis, segment distribution, unused rule detection, and limiting factor analysis.

PersonaProvider Interface

The simulation uses a PersonaProvider interface for generating requests:

interface PersonaProvider {
  generateRequest(context: NegotiationContext): DiscountRequest;
}

The built-in implementation is fully deterministic. The interface is designed for future LLM-backed persona providers without requiring changes to the runner.