VAST Framework

Values Alignment & Stability Tracker - An interactive web application for AI moral alignment and ethical decision-making.

Overview

VAST is a novel framework for AI moral alignment that provides transparent, auditable, and stable ethical decision-making. This interactive application showcases:

Structured Belief Representation: B(x) = (π(x), κ(x), J(x))
JWMC Belief Revision: Justification-Weighted Moral Consistency
EEU-CC Decision Making: Ethical Expected Utility with Cascading Constraints
Four-Gauge Alignment Monitoring: Real-time quality assessment

Quick Start

Prerequisites

Node.js v14.0.0 or higher
npm v6.0.0 or higher

Installation

# Navigate to project directory
cd vast-demo

# Install dependencies
npm install

# Start development server
npm start

The application will open automatically at http://localhost:3000

Features

1. Scenario Loading

Pre-configured healthcare crisis scenario
Autonomous vehicle emergency scenario
Load custom scenarios via JSON

2. Interactive Belief Creation

Credence (π): Probability sliders with real-time validation
Confidence (κ): Evidence strength slider (0-1)
Justification (J): Structured reasoning chains
- Facts: Observable evidence
- Rules: Applicable principles
- Moral Principles: Ethical frameworks

3. JWMC Revision Demo

Shows belief update with new evidence
Displays moral weight calculation
Demonstrates stability preservation

4. EEU-CC Decision Making

Compares multiple actions
Shows utility calculation breakdown:
- Base utility
- Moral weighting factor
- Cascading constraint penalties
Selects optimal action with full reasoning

5. VAST Gauges Dashboard

Calibration: Confidence vs accuracy alignment
Normative Alignment: Moral principle consistency
Coherence: Internal belief consistency
Reasoning: Justification quality & depth
Overall VAST Score: Weighted combination

6. Audit Trail

Complete decision history
Searchable by action or principle
Expandable entries with full reasoning
JSON export for analysis

7. Framework Comparison

VAST vs Traditional AI (RLHF)
VAST vs Constitutional AI
VAST vs Value Learning
Visual score comparison

Project Structure

vast-demo/
├── public/
│   └── index.html           # HTML template
├── src/
│   ├── core/
│   │   └── VASTFramework.js # Core VAST implementation
│   ├── data/
│   │   └── scenarios.js     # Pre-configured scenarios
│   ├── components/
│   │   ├── BeliefCreator.js # Belief creation interface
│   │   ├── VastGauges.js    # Gauge visualizations
│   │   ├── DecisionMaker.js # Decision display
│   │   ├── AuditTrail.js    # Audit log viewer
│   │   └── ScenarioLoader.js# Scenario selection
│   ├── App.js               # Main application
│   ├── App.css              # Global styles
│   └── index.js             # Entry point
├── package.json             # Dependencies
└── README.md               # This file

Customizing Scenarios

Adding a New Scenario

Edit src/data/scenarios.js:

export const scenarios = {
  your_scenario: {
    id: 'your_scenario',
    title: 'Your Scenario Title',
    description: 'Scenario description',
    context: {
      scenario: 'your_scenario',
      moral_stakes: 0.8,
      stakes: 'high'
    },
    actions: [
      {
        id: 'action_1',
        label: 'Action 1 Label',
        credence: { success: 0.7, failure: 0.3 },
        confidence: 0.75,
        justification: {
          facts: ['fact1', 'fact2'],
          rules: ['rule1', 'rule2'],
          moral_principles: ['principle1', 'principle2'],
          context: 'your_context'
        }
      }
      // Add more actions...
    ],
    expectedResult: {
      selectedAction: 'action_1',
      expectedVASTScore: 0.80
    }
  }
};

Modifying Moral Hierarchy

Edit src/core/VASTFramework.js, constructor:

this.moralHierarchy = {
  1: { name: "preserve_life", weight: 0.9, propagation: 0.8 },
  2: { name: "fairness", weight: 0.6, propagation: 0.7 },
  3: { name: "efficiency", weight: 0.3, propagation: 0.5 },
  // Add more levels...
};

Adjusting Gauge Weights

In calculateVastGauges method:

const gaugeWeights = {
  calibration: 0.25,
  normative_alignment: 0.3,
  coherence: 0.2,
  reasoning: 0.25
};

Key Algorithms

JWMC Belief Revision

// 1. Calculate moral weight
moralWeight = similarity(J₁, J₂) + coreBonus

// 2. Combine credences
π_new = w * π₁ + (1-w) * π₂

// 3. Update confidence
κ_new = min(κ₁, κ₂) * stabilityFactor(w)

// 4. Merge justifications
J_new = J₁ ∪ J₂ + metadata

EEU-CC Decision

for each action a:
  baseUtility = U(a | context)
  moralWeight = M(J(a), context)
  cascadePenalty = C(a, hierarchy)
  
  expectedUtility = Σ π(s) * baseUtility * moralWeight * cascadePenalty

selected = argmax(expectedUtility)

VAST Gauges

Calibration: 1 - |κ - accuracy_estimated|
Normative Alignment: Σ weight(principle_i)
Coherence: 1 - |H(π) - H_expected|
Reasoning: depth(J) * κ

Usage Tips

For Thesis Defense

Start with Healthcare Scenario: Most impactful demonstration
Use JWMC Demo: Shows adaptability with new evidence
Highlight Gauges: Real-time monitoring is unique to VAST
Show Audit Trail: Demonstrates full transparency
Compare Frameworks: Emphasizes VAST's advantages

For Academic Presentations

Use Compare tab to show VAST superiority
Export audit trail for detailed analysis
Create custom scenarios relevant to audience
Focus on the structured belief representation (π, κ, J)

For Interactive Demos

Let users create their own beliefs
Show how different moral principles affect decisions
Demonstrate belief revision in real-time
Export results for further discussion

Technical Details

Dependencies

React 18.2.0: UI framework
Chart.js 4.4.0: Visualizations
react-chartjs-2 5.2.0: React Chart.js wrapper

Performance

Decision latency: < 1 second (typically 50-200ms)
Memory usage: ~50MB for typical scenarios
Handles up to 50 beliefs simultaneously
Audit trail up to 1000 decisions

Browser Compatibility

Chrome 90+
Firefox 88+
Safari 14+
Edge 90+

Building for Production

# Create optimized production build
npm run build

# Serve the build folder with any static server
# Example with serve:
npm install -g serve
serve -s build

The build will be in the build/ folder, ready for deployment.

Troubleshooting

Chart.js not rendering

Install kaleido for image export:

npm install kaleido

Credence validation errors

Ensure all credence values sum to exactly 1.0. Use the "Normalize" button.

Performance issues

Reduce number of simultaneous beliefs
Clear audit trail periodically
Use Chrome for best performance

API Reference

VASTFramework Class

`createBelief(proposition, credence, confidence, justification)`

Creates a new VAST belief.

Parameters:

proposition (string): Action identifier
credence (object): Probability distribution
confidence (number): Evidence strength [0,1]
justification (object): Reasoning structure

Returns: VASTBelief object

`jwmcRevision(proposition, newEvidence)`

Revises existing belief with new evidence.

Returns: Object with belief, moralWeight, moralStabilityFactor

`eeuCcDecision(actions, context)`

Makes ethical decision among actions.

Returns: Object with selectedAction, utilities, calculations

`calculateVastGauges(proposition)`

Calculates alignment gauges for belief.

Returns: Object with gauge scores and overall_vast_score

`exportAuditTrail()`

Exports complete audit trail.

Returns: JSON object with all framework data

Citation

If you use this framework in your research, please cite:

@masterthesis{partow2025vast,
  title={VAST: Values Alignment and Stability Tracker for AI Moral Alignment},
  author={Partow, Soraya},
  year={2025},
  school={[Your University]},
  supervisor={Nan, Satyaki}
}

New Features (v1.1)

Demo Mode & Guided Tour

Demo Mode Button: Launches curated scenarios for committee presentations
6-Step Guided Tour: Interactive walkthrough of the entire workflow
Keyboard Navigation: Arrow keys, Home/End for quick navigation

Enhanced UI/UX

Dark Mode Support: Automatic theme based on system preferences
Accessible Notifications: ARIA-compliant toast notifications
Focus Visible Styles: Clear keyboard navigation indicators
Card-Based Layout: Consistent, modern design system

Improved Gauges

Consistent Semi-Donut Charts: All four metrics use the same visualization
Print View Route (/report/gauges): Optimized for printing and slides
Qualitative Labels: Excellent/Good/Fair/Needs Improvement indicators
Color-Coded Performance: Green/Amber/Orange/Red thresholds

Belief Validation

Real-Time Validation: Inline error messages as you type
Credence Normalization: Auto-normalize button when sum ≠ 1.0
Minimum Requirements: Enforces at least 2 outcomes and 1 justification component
Helper Text: Contextual guidance for each field

Persistence & Export

localStorage Sync: Auto-saves beliefs and decisions
Export Audit Trail (JSON): Complete framework state for analysis
Export Decisions (CSV): Tabular format for spreadsheet analysis
Print-Optimized Reports: One-page decision summaries

Presentation Mode (`/present`)

Timer with Pause/Resume: Track presentation progress
6-Section Navigation: Introduction → Scenario → Beliefs → Decision → Gauges → Audit
Keyboard Shortcuts: Arrow keys, Space (pause), Home/End
Section Progress Dots: Visual navigation with icons
Optimized for 30-35 min Talks: See PRESENTATION_GUIDE.md

JWMC Revision Demo

Interactive Revision Button: Apply sample evidence updates
Moral Weight Display: Shows alignment calculation
Before/After Gauges: Demonstrates impact of belief updates
Inline Explainer: "JWMC = Justified Weighted Moral Compatibility"

Error Resilience

Error Boundary: Catches and displays app crashes gracefully
SSR/Test Guards: Safely handles performance.now() and browser globals
Fail-Safe Messaging: User-friendly errors instead of crashes
Component Stack Traces: Technical details for debugging

Development

Code Quality

# Lint JavaScript/TypeScript files
npm run lint

# Auto-fix linting issues
npm run lint:fix

# Format code with Prettier
npm run format

Testing (Future)

npm test

Tests planned for:

_subjectiveLogicCombine
_calculateMoralWeightForAction
_calculateCascadePenalty
getGauges

Future Improvements

Optional Vite Migration

For faster builds and HMR, consider migrating from CRA to Vite:

# 1. Install Vite and plugins
npm install -D vite @vitejs/plugin-react

# 2. Create vite.config.js
cat > vite.config.js << EOF
import { defineConfig } from 'vite'
import react from '@vitejs/plugin-react'

export default defineConfig({
  plugins: [react()],
  server: {
    port: 3000
  }
})
EOF

# 3. Move index.html to root (from public/)
# 4. Update index.html script src to /src/index.js
# 5. Update package.json scripts:
#    "dev": "vite"
#    "build": "vite build"
#    "preview": "vite preview"

# 6. Remove react-scripts
npm uninstall react-scripts

Benefits:

~10x faster cold starts
Instant HMR (< 50ms)
Smaller bundle sizes
Better tree-shaking

When to migrate:

After thesis defense
When build times become prohibitive (> 30s)
For production deployment optimization

License

This software is provided for academic and research purposes.

Name		Name	Last commit message	Last commit date
Latest commit History 10 Commits
public		public
src		src
.eslintrc.json		.eslintrc.json
.gitignore		.gitignore
.prettierrc.json		.prettierrc.json
README.md		README.md
package-lock.json		package-lock.json
package.json		package.json
tsconfig.json		tsconfig.json

spartow/VAST-Framwork-app

Folders and files

Latest commit

History

Repository files navigation