MODULE 2.2: LIVE WIRE

Status: Active Prototype (v3.0)

Stack: Python (Pandas, NumPy, scikit-learn), Gradio, OpenAI/Anthropic/Google APIs))

Research Content: Communicative Alignment Framework (CAF)  

Live Wire is a Python-based orchestration system designed to observe how large language models stabilize, diverge, and align across synchronized multi-turn interactions. A single user input is broadcast to multiple LLMs under a shared system context, while each model maintains an isolated rolling conversational state. On each turn, Live Wire computes telemetry including Dynamic Communicative Alignment (DCA), cosine similarity matrices, fingerprint overlap, and inter-model consensus measures. The instrument is intended to support evaluation and interpretability research by exposing interaction-level dynamics without modifying model weights or introducing persistent memory. 

What to watch for in the demo below: 

• Turn-level changes in user-model vs. model-model alignment
• Shifts in inter-model coherence under consistent boundary conditions
• Stable stylistic differentiation alongside semantic convergence

Orchestration: Custom Python controller coordinating concurrent API calls to GPT-4o, Claude, and Gemini

State Management: Isolated rolling context windows per model with controlled summarization ("backpack" pattern)

Telemetry

• Dynamic Communicative Alignment (DCA) via cosine similarity of turn deltas
• Fingerprint vectors for semantic overlap
• Consensus detection across model outputs 

Interface: Gradio Blocks API for live state and metric rendering

• Models receive identical user inputs and maintain isolated local histories
• No cross-model communication or shared memory occurs
• Observed patterns emerge without explicit instruction to converge