Research
World models over latent human states.
We study how to predict and steer human trajectories over time — from a current state A to a chosen state B — through proactive interaction.
The problem: consequence-blindness
Modern AI systems optimize the quality of a single response at a single moment. They are consequence-blind: they do not model what happens to the person over the following hours, weeks, and months. Benchmarks measure answers. Nobody measures trajectories.
The missing data
Trajectory models cannot be trained on existing data. Clicks and conversation logs capture engagement, not behavior change. The data that matters — plan versus fact, interaction followed by action followed by outcome, across every area of life, with consent — is not collected anywhere. It can only be produced by a product designed for it from the first line of code.
Every interaction inside Performance is a labeled training example.
The loop
Interaction → Action → Event → Performance
Performance runs on a weekly 168-hour cycle. The system initiates interactions, the person acts, events land on the timeline, and the delta between plan and fact feeds the next cycle. This loop is both the product mechanic and the data engine.
01
Interaction
02
Action
03
Event
04
Performance
Context
World models are an established research direction: predicting the next latent state of an environment rather than the next token. Existing work targets the physical world — video, robotics, simulation. Our training objective is different: the human timeline. The latent states we model are states of a person's life, structured by a weekly cycle.
Mathematical foundations
Our modeling stack builds on a synthesis of established mathematics — ultrametric and p-adic structures, multiary algebras, information-theoretic methods — applied to cyclic behavioral trajectories. The synthesis and the application domain are ours; the components are rigorous, known mathematics.
For research inquiries: research@holisticintelligence.co