Research
Enreef treats a world as a system that should be testable, replayable, and explainable. Research connects behavioral modeling, geospatial data, scalable execution, synthetic populations, and selective agent intelligence to measurable questions about how a world behaves.
Research foundations
Enreef draws from established travel modeling, synthetic-population methods, agent research, and city-scale simulation. These foundations inform how Enreef designs, validates, and scales living worlds.
Memory, reflection, retrieval, and planning show how language models can support believable individual behavior. Enreef uses deeper reasoning selectively rather than turning every resident into a permanent natural-language loop.
Park et al. · Generative Agents ↗ActivitySim and PopulationSim demonstrate the discipline behind activity-based travel models and synthetic populations: explicit person and household state, calibrated behavior, and reproducible model inputs.
ActivitySim · PopulationSim ↗GenWorld demonstrates a practical scaling direction for city-scale agent studies: ground a synthetic city in real data, reason selectively, and compile expensive decision signals into reusable policies for rollout.
GenWorld · arXiv 2606.27650 ↗SF-LIFE shows how quickly dense trajectory logging becomes enormous: 500,000 simulated Bay Area agents over 70 days at 1 Hz generate trillions of location records. Enreef treats transitions and events as first-class data instead of storing every coordinate at every second.
SF-LIFE · arXiv 2606.00430 ↗Grounded worlds
Real-world branches are compared with observations, baselines, and known history rather than judged by whether the output merely looks convincing.
Created worlds
Created worlds answer to their authored constraints, continuity, relationships, resources, and the consequences accumulated through time.
Evaluation
Enreef uses reproducible evaluation across execution, routing, behavior, historical replay, branch storage, and world compilation.
Research and runtime