Skip to content

replit.md

Latest commit

 

History

History
144 lines (113 loc) · 7.32 KB

replit.md

File metadata and controls

144 lines (113 loc) · 7.32 KB

Skyla - Cryptographically Verified Symbolic AI Agent

Overview

Skyla is an AI platform that implements a context-aware epistemic gate system with 2-model consensus architecture. The platform combines symbolic reasoning, advanced mathematical model selection, and cryptographic verification to create transparent, auditable AI behavior with mathematically verifiable uncertainty quantification.

Key Innovation: Unlike traditional AI systems that use duplicate models, Skyla implements true architectural diversity by comparing Claude Haiku (efficiency-focused) against Claude Sonnet (quality-focused), providing authentic consensus measurement and quality-based selection.

User Preferences

Preferred communication style: Simple, everyday language.

Core System Architecture

Consensus Engine

  • True 2-Model Consensus: Claude Haiku vs Sonnet with genuine architectural differences
  • Mathematical Quality Scoring: 5-factor analysis (length, nuance, efficiency, richness, coherence)
  • Context-Aware Epistemic Gate: Analyzes conversation history before triggering clarification requests
  • Dynamic Model Selection: Quality-based selection with architectural bonuses for query complexity

Context-Aware Epistemic Gate System

  • Context Analysis: Epistemic gate that analyzes conversation history before requesting clarification
  • Pronoun Resolution: Resolves "this", "it", "that" references against recent conversation topics
  • Topic Continuity Detection: Prevents false positive clarifications for contextual queries
  • Refined Uncertainty Detection: Uses maximum divergence + contextual patterns for genuine ambiguity detection

Advanced Model Selection Logic

  • Quality Analysis Algorithm: Mathematical scoring across 5 sophisticated metrics
  • Architectural Bonuses: +0.1 bonus for Haiku on simple queries, +0.1 for Sonnet on complex queries
  • Dynamic Token Allocation: 400-800 tokens based on input complexity
  • Transparent Logging: Detailed quality breakdowns and selection reasoning

Frontend Architecture

  • EJS Template Engine: Dynamic rendering with shared partials
  • Interactive Console: Terminal-style UI with real-time state transitions
  • Minimal Design System: Beige background (#f5f4f0) with monospace typography
  • Animated Logging: Console output with syntax highlighting and integrity indicators

Backend Architecture

  • Express.js Server: ES modules with robust error handling
  • Conversation Memory: Session-based context storage (10 exchanges per session)
  • Multi-Model Integrity System: Simultaneous model calls with divergence analysis
  • Cryptographic Proofs: ZK proof generation for all state transitions

Symbolic Processing Engine

Three-Layer Processing System

  1. Exact Symbolic Layer: Direct pattern matching for commands (spiral, daemon, build, analyze)
  2. Semantic Pattern Matching: Regex-based categorization with contextual vector adjustments
  3. Deterministic Hash Fallback: Consistent micro-adjustments for unknown inputs

Identity Vector System

  • 4-Dimensional State Space: [cognitive, emotional, adaptive, coherence]
  • Deterministic State Transitions: All changes produce verifiable proofs
  • Mode Transitions: analytical ↔ adaptive ↔ creative ↔ coherent based on dominant dimension

Context-Aware Features

Session Memory Management

  • 10-Exchange History: Maintains conversational context across interactions
  • Smart Context Extraction: Identifies recent topics, problems, and technologies
  • Ambiguity Resolution: Resolves pronouns and contextual references automatically

Epistemic Uncertainty Detection

  • High Topic Divergence Detection: Flags genuine ambiguity when topic divergence >0.8
  • Context Gate: Prevents false positives by checking conversation history first
  • Pattern Recognition: Identifies ambiguous phrases like "How do I fix this?" without context

Mathematical Quality Assessment

5-Factor Quality Scoring

  1. Adaptive Length Scoring: Optimal response length based on query complexity
  2. Nuance Detection: Philosophical and analytical concept identification
  3. Efficiency Indicators: Directness and clarity measurement
  4. Semantic Richness: Vocabulary diversity assessment
  5. Structural Coherence: Sentence quality and organization analysis

Query Complexity Analysis

  • Complex Terms: philosophy, consciousness, analysis, theoretical, abstract, nuanced, sophisticated
  • Simple Terms: hi, hello, what, how, yes, no, ok, spiral, basic
  • Dynamic Weighting: High complexity favors nuance (35% weight), low complexity favors efficiency (30% weight)

Implementation Status

Currently Functional

✅ True 2-Model Consensus System - Claude Haiku vs Sonnet with real architectural diversity
✅ Context-Aware Epistemic Gate - Revolutionary context analysis before clarification requests
✅ Mathematical Model Selection - 5-factor quality scoring with architectural bonuses
✅ Session Memory System - 10-exchange conversation history with context extraction
✅ Symbolic State Management - Identity vector transitions with cryptographic proofs
✅ Integrity Measurement - 4-dimension divergence analysis (length, sentiment, topic, tone)
✅ Interactive Console Demo - Real-time state transitions with animated logging
✅ Comprehensive Documentation - Technical implementation details and API reference

In Development

🔄 Full Cryptographic Verification - Production-grade ZK proof validation
🔄 Multi-Agent Coordination - Agent-to-agent communication protocols
🔄 Advanced API Endpoints - Production-ready REST API with authentication

Technical Dependencies

Runtime Environment

  • Node.js (>=18.0.0): JavaScript runtime with ES module support
  • Express.js (^4.21.2): Web framework with middleware support
  • EJS (^3.1.10): Server-side templating engine
  • @anthropic-ai/sdk (^0.61.0): Official Claude API integration

Environment Configuration

  • CLAUDE_API_KEY: Required for multi-model consensus system (graceful fallback if missing)
  • PORT: Server port (defaults to 5000)

Deployment Architecture

  • Build Process: Automated dependency installation via npm install
  • Autoscale Deployment: Stateless web application with on-demand scaling
  • Enhanced Error Handling: Comprehensive environment validation and graceful API fallbacks
  • Static Asset Serving: Public directory with brand assets
  • Template-Based Routing: Clean URLs without file extensions
  • Production Ready: Full deployment configuration with proper build commands

API Interface

Primary Endpoint

POST /api/claude - Multi-model consensus processing with context-aware epistemic analysis

Response Format

{
  "success": true,
  "response": "Selected model response text",
  "integrity": "high|medium|low|epistemic_gate",
  "clarificationNeeded": false,
  "epistemicReason": "AMBIGUITY_RESOLVED_BY_CONTEXT",
  "metadata": {
    "model": "claude-3-haiku-20240307",
    "integrityScore": 0.789,
    "consensusStrength": 0.789,
    "divergenceMetrics": { "topicDivergence": 0.234 },
    "epistemicAnalysis": { "contextResolution": {...} },
    "contextMemory": "4 exchanges remembered"
  }
}

This implements context-aware epistemic uncertainty detection with multi-model consensus measurement.