Spaces:
Runtime error
title: Twin Track
emoji: 🚀
colorFrom: red
colorTo: red
sdk: docker
app_port: 8501
tags:
- streamlit
pinned: false
short_description: 'Visualization of human vs simulated interaction dynamics '
license: mit
TwinTrack
Observed vs Simulated Interation Dynamics
TwinTrack is a lightweight visualization tool for comparing long-form human-LLM interactions against synthetic persona simulations, stripped of semantic content and analyzed purely by temporal shape.
This is not about what is said. It is about how interation unfolds over time.
What TwinTrack Shows
Top: A human long-form interaction (Big Flame) Bottom: A selected synthetic persona (Minimalist, Cheerleader, Poet)
Each band visualizes rolling response variance across normalized interaction time. Same scale. Same normalization. Different dynamics.
Synthetic personas often look "reasonable" in isolation. When placed next to a human thread, their temporal fingerprints tell a different story.
How it Works (High Level)
- Conversation threads are converted to numeric features (length, variance)
- Time is normalized to a 0-1 scale to allow comparison across unequal thread lenghts
- A rolling window computes variance over interaction time
- Tow bands are rendered using a shared color scale for direct visual comparison
No embeddings. No sentiment labels. No semantic interpretation. Just shape.
TwinTrack is NOT
- A benchmark
- A performance metric
- A truth or safety score
- A semantic analyzer
TwinTrack is an instrument - not a verdict
Status
- v1: Working prototype
- Interaction persona selection
- Normalized temporal alignment
- Shared color scale for interpretability
Future iterations may explore:
- Additional perturbation controls
- Alternative variance measures
- Multi-human baselines
Related Work
Beyond Content: Temporal Dynamics in Long-Form Human-LLM Interation DOI: https://doi.org/10.5281/zenodo.18273459