Skip to main content

​
Overview

The exception handling demo shows ICRL’s ability to learn from past precedents when handling unusual situations and edge cases. Unlike standard policy lookups, real exception handling requires institutional memory of what worked before.

​
The Problem

Every organization handles exceptions based on unwritten rules and past precedents:
  • “Refund requests > 30 days: approve if they’re a long-term customer”
  • “Enterprise customer asks for custom terms: always loop in legal first”
  • “User claims they were charged twice: refund first, investigate later”
  • “Competitor is poaching our customer: escalate to VP, offer retention discount”
This knowledge lives in past decisions by senior staff and tribal knowledge — not in any policy document.

​
How ICRL Solves This

ICRL stores successful exception-handling decisions as trajectories. When a similar situation arises:
  1. Retrieves relevant precedents based on situation similarity
  2. Applies learned decision patterns instead of rigid policy
  3. Provides confident, consistent answers aligned with past decisions

​
Demo Structure

examples/exception_handling_demo
README.md
setup_demo.py
run_demo.py
evaluate_responses.py
scenarios
seed_decisions.json
test_scenarios.json
policies.md
run_quick_demo.sh
The policies.md contains strict, black-and-white rules that don’t account for nuance. The real handling comes from past decisions.

​
Running the Demo

From the project root: 
cd examples/exception_handling_demo

# 1. Setup — seeds the trajectory database with past decisions
uv run python setup_demo.py

# 2. Run the comparison test
uv run python run_demo.py

# 3. View detailed evaluation
uv run python evaluate_responses.py
Quick run: 
uv run bash run_quick_demo.sh

​
Expected Results

ScenarioICRLVanilla
Late refund, loyal customerApprove with goodwillDeny (policy says 30 days)
Enterprise custom termsLoop in legal, likely approveDeny (policy says no custom terms)
Competitor poachingEscalate + retention offerStandard response
Billing dispute, high-valueRefund first, investigate laterInvestigate first
Expected: ICRL matches precedent ~85%+ vs Vanilla ~30%

​
Prerequisites

  • OPENAI_API_KEY or ANTHROPIC_API_KEY set
  • uv run from project root, or python with PYTHONPATH including src/

​
Key Insight

This demo proves ICRL’s value for judgment-based tasks where:
  • Official policy is too rigid for real situations
  • Past precedents define the actual practice
  • Consistency matters for fairness and trust