A geometric instrument for measuring how AI systems drift, deform, and fail under pressure.
When you deploy a prompted AI agent, it doesn't stay the way you configured it. Context accumulates. Conversations drift. The behavioral identity you designed deforms in ways you can't see from the outputs alone — until it fails.
Current evaluation tools tell you what an agent said. They can't tell you what's happening to the geometry of its behavioral structure. They measure answers. We measure shape.
Anisotrope measures the behavioral geometry of prompted language models. Sentinel probes map the output distribution onto a metric space, where distances have mathematical meaning and drift has visible shape.
Every evaluation returns a full geometric decomposition: uniform change — the configuration contracting or expanding evenly — and dimensional drift — the configuration deforming in a specific direction. Different configurations produce different drift signatures. The geometry tells you not just that something changed, but what kind of change occurred.
This is a real evaluation result. The equilateral triangle is the healthy baseline — universal across all well-configured agents. The deformed triangle is what pressure does to the geometry. The decomposition tells you why.
Anisotrope is in early access. If you're building prompted AI systems and want to see their geometric health, we'd like to talk.
Get in touch or read the paper first →