A world model that predicts how an organization and its external world co-evolve — users, market, competition, regulation, partners, public — as one dynamic system.
A latent forecaster with scope-controlled output. Ask it about a team and a customer segment over the next thirty days; it returns only that slice. Predict-first → slow plan / fast act.
No PMs, no engineering tickets. In an AI-native org, humans contribute three things — taste, domain, vision — and agents do the rest. The coordination layer is the fabric routing work between them.
Three handoff types — agent ↔ agent · agent ↔ human · human ↔ human — over one event-sourced runtime. Role-type to concrete person resolved at tick time.
Organizations make their hardest moves once. You restructure a team, ship a price change, post a public statement — and the world responds. Most of that response is unrehearsable. We are building the two capabilities that make it rehearsable, on one model:
Social Simulation runs the organization, its agents, and its outside population forward under a candidate decision. Market Prediction forecasts how the outside world — customers, competitors, regulators, platforms — moves in response. Both are scope-controlled: you ask about a team and a segment, you get only that slice.
Spin up populations of agents — employees, internal agents, customers, competitors, regulators, wild AI — and let them interact under a candidate decision. The simulator runs forward; you read the trajectory and the macro behavior that emerges.
A counterfactual rehearsal layer for choices that are too expensive to rehearse in the real world. What if we restructure this team. What if this policy passes. What if the competitor cuts price 30%.
A latent forecaster over the joint state of the firm and its outside world — customers, competitors, regulators, platforms, public. Scope-controlled: ask about one team and one segment over the next thirty days, get only that slice.
Building the right thing beats building twice as fast. The forecast is what makes "right" computable, before commitment.
"Cut entry-tier pricing 30% next Tuesday." Wrap as ScopeQuery with the affected team, customer segment, and competitor set; horizon 30 days; focal metrics = adoption, churn, competitor response.
Social Simulation runs the org × segment forward under the candidate decision. Market Prediction folds in competitor + regulatory + ecosystem signals. Returns a trajectory bundle — point estimate plus K sampled paths (optimistic, base, pessimistic, tail).
Outputs are scope-controlled — only the slice asked for, with uncertainty band. Compare against a control rehearsal (no cut). Commit, hedge, or refine. Real outcomes later check the rehearsal — surprises are where the model wants more attention.
Both capabilities sit on one model. The eight-layer frame below is what the model carries inside: same primitives, two scopes — the firm's inside, and the world the firm is embedded in. Org × world as one evolving system.
Read across each row.
Surveyed seventeen LLM-driven social-simulation papers. All sit at one of two extremes — abstract agents on abstract networks, or a single closed community. None models an organization jointly with the world it operates in.
Existing world-model work hands the planner the whole latent state. Useless when the planner asks "what about this team and this customer segment over the next thirty days." There is no scope retrieval head in the literature.
Social-simulation work still predicts surface text — agent utterance by agent utterance. We predict in latent space: encode population state into z, forecast zfuture, decode only what a head asks for.
境瞳 is building the prediction + coordination substrate for AI-native organizations. We are in the design-doc + early-implementation phase, talking with select operators, early customers, and investors.