Simulation

A research direction, and the least built of the four. Reach out if this is the problem you have.

Once you have a model that predicts what a person does next, you can run it forward. Point it at a population, change one thing about the world, and watch what the population does differently.

That turns a lot of expensive, slow, one-shot decisions into something you can rehearse:

Decision	What you would rehearse
Product	Demand, feature tradeoffs, pricing, and switching behavior, before committing engineering to any of them.
Messaging	Positioning, creative, and offers, evaluated against a population rather than a focus group of twelve.
Segmentation	The groups that actually behave differently, instead of the demographic buckets you inherited.
Scenario planning	A strategy stress-tested against competitive moves and economic shifts that have not happened yet.
Communications	How an announcement reads to the people who will actually receive it.
Policy	The second-order response to a rule change, before the rule ships.

Why we think this belongs on a behavioral backbone

The companies doing this well right now are worth studying. Simile built generative agents grounded in interviews and survey responses. Aaru frames the pitch precisely: ask people and they answer, simulate them and they act. Demosyne runs long-lived worlds where the consequences compound over months rather than resolving in a single prompt.

The shared insight is that stated preference is a poor substitute for revealed preference. People misremember, rationalize, and tell you what they think you want.

Our angle is on the substrate. Most simulation today puts a language model in a persona and asks it to behave like a person, which inherits everything the language model knows about how people describe themselves. A foundation model of behavior is trained on sequences of what people actually did. If that model is good, it should be the better engine to simulate from, because it never had to route through self-report.

That is a claim, not a result. It is the one we most want to test.

How it would work

1. Ground a population. Construct a population from real behavioral sequences, not from written personas. Sampled to match the distribution you care about.
2. Change a condition. Introduce the new price, the new feature, the new message, the new competitor.
3. Run it forward. Each simulated actor produces a sequence of actions, not an opinion. Aggregate over the population.
4. Validate against holdout. The only thing that makes this useful is correlation with a real outcome you withheld. No engagement runs without that check.

On validation

A simulation that cannot be scored against reality is a very expensive way to produce a confident number. If we work on this with you, agreeing the holdout comes before agreeing the scope.

The rest of the downstream surface

Simulation is one thing a behavioral backbone unlocks. The others are closer in and mostly follow the same pattern of adapting one pretrained model rather than building a new system:

Task	What the backbone provides
Recommendation	Shipping first. See Foundation models.
Matching	Compatibility between two people, scored on outcomes rather than stated preferences.
Personalization	Assistants that adapt to how you work without being told each session.
Memory	Deciding what context about a person is worth keeping and surfacing.

---

Talk to the team: jonathan@jeantechnologies.com