AI Simulations

Structured state helps a simulator track goals, beliefs, constraints, knowledge, and memory without pretending to reconstruct a human mind.

A user model should begin with inspectable evidence, not a decorative persona prompt. Grounding must preserve provenance, scope, time, uncertainty, and measured behavior.

A seven-layer framework for deciding when a simulation may explore, test, or predict, and what evidence each claim requires.

The most dangerous moment in building a simulation is when the demo starts to feel real. Believability gives face validity, not behavioral validity, and six claims people quietly collapse into one.

At Gmail, the hardest part of agent evaluation was not the agent. It was building a world where we could tell whether it worked. Why synthetic evaluation data is infrastructure, and why the unit is a world, not a document.

A practical map of five AI simulation scopes, from synthetic data and interactions to user journeys, individual models, and population worlds, with the evidence each requires.