Non-semantic measurement instruments for long-form human–LLM interaction. All instruments are publicly available and run entirely in the browser. Nothing is transmitted or stored.
Demo instrument
Measures stability, volatility, and regime transitions in two curated example archives — Basalt and Chalk — derived from long-form human–LLM interaction data. Computes rolling variance, stability onset detection, and configurable variance bands over turn-level response length. All measurements are non-semantic: no content is read or analyzed.
Upload instrument
Accepts user-supplied transcripts in CSV, JSON, Markdown, or plain text and applies the same structural analysis pipeline as Variance Register to your own data. Includes a transparent mapping step so you can confirm what was detected before analysis runs. Adds perturbation testing — temporal scramble and metric noise injection — to assess robustness of detected structure. Processed entirely in your browser. No data is transmitted or stored.
Comparison instrument
Places the Chalk human archive alongside a synthetic profile generated under controlled input conditions. Three profiles are available — Sparse, Effusive, and Structured — representing distinct input entropy regimes. Both tracks are normalized to a shared temporal grid for direct comparison. TwinTrack asks whether input regime alone can reproduce the temporal signature of a real long-form human thread. The consistent finding is that it cannot.