Skip to content

axiomdriftai/Variance_Register

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

3 Commits
 
 
 
 
 
 

Repository files navigation

Variance Register

Variance Register is a privacy-preserving interaction analysis instrument for measuring stability, volatility, and regime transitions in long-form conversational exchanges. Rather than evaluating individual responses in isolation, Variance Register treats conversation as a temporal system. It analyzes turn-level behavioral signals — specifically response length and variance — across time to reveal how interaction dynamics evolve, stabilize, or fail under extended context.

Variance Register does not analyze semantic content, sentiment, or meaning. All measurements are derived from non-semantic structural properties of the interaction itself, making the instrument suitable for sensitive or private conversational data.

What Variance Register Measures

Variance Register computes rolling statistics over conversational turns to identify:

  1. Interaction volatility (response length variance)
  2. Boundedness (whether variability remains constrained)
  3. Persistence (whether stability is sustained across time)
  4. Time-to-stability (the point at which a stable regime emerges, if at all)

Stability is defined operationally as a period in which rolling variance remains below a configurable threshold for a minimum duration. This allows interaction regimes to be detected without assumptions about model internals, task correctness, or semantic coherence.

What Variance Register Is and Is Not

Variance Register is an instrument, not a benchmark or evaluation score. It is designed to:

  1. Observe interaction-level behavior across scale
  2. Detect regime transitions (exploratory → convergent → unstable)
  3. Compare interaction dynamics across threads or models using a shared metric space

It does not claim to measure: understanding, alignment, intelligence, intent, or agency.

Datasets

This instrument is demonstrated on two curated example archives derived from long-form human-LLM interaction data.

Basalt — a 1,027-turn archive exhibiting a measurable transition to a stable response regime over the course of the exchange.

Chalk — a 590-turn archive exhibiting sustained volatility with periodic structural disruptions and no convergent regime under default parameters.

Both archives consist of derived, non-semantic metrics only. No message content is stored or displayed.

Current Scope

This release demonstrates Variance Register on fixed example archives. External dataset ingestion and streaming analysis are intentionally excluded to preserve scope clarity and privacy boundaries. Variance Register is part of a broader suite of interaction analysis instruments focused on understanding how stability forms and persists in human-LLM exchanges.

Implementation

Variance Register is implemented as a static single-file HTML application. All computation — rolling statistics, stability detection, and visualization — runs client-side. No data is transmitted or logged.

Notes

This project was developed by Axiom Drift AI. AI-assisted coding tools were used during implementation, consistent with modern software development practice. All conceptual framing, measurement logic, and evaluation criteria are human-directed.

Related Work

An exploratory preprint describing the temporal analysis methods and findings presented here is available at:

Beyond Content: Temporal Dynamics of Long-Form Human-LLM Interaction https://doi.org/10.5281/zenodo.18273459

Axiom Drift AI

Releases

No releases published

Packages

 
 
 

Contributors

Languages