Introducing Delos

Delos solves an initialization problem for agent simulation. Before a run begins, the system needs to know who exists, where each person can go, which relationships and obligations connect them, and which facts the world can act on. Terrarium does not fill those fields later; an omitted person, place, or possession is absent from the run. Delos turns evidence about a real place into a finite initial world for Terrarium: synthetic people, shared places, reciprocal relationships, routines, and material facts, all assembled before Terrarium starts its clock. Delos owns that starting state. It does not decide what the people do next.

Population marginals cannot supply that state. Separate totals for age, tenure, occupation, or ancestry say nothing about how those properties occur together; multiplying the marginals would manufacture intersections the source never observed. Proportional rounding creates a second problem in a small cast: a consequential group can disappear because its expected count is less than one. Generic personas leave a different gap. They leave relationships, routines, and dispositions to one model prior, so different biographies can still produce the same agreeable, reflective person. Delos treats demographics as constraints on allocation, not as a shortcut from group averages to individual behavior.

Research therefore begins with a target: a resolved place, population, geography, and period. Agents inspect the strongest source that fits each claim, freeze the exact material used, and store task-shaped research records with source IDs and locators. Quantitative records keep their universe, definitions, method, uncertainty, and derivation when those affect interpretation. Approved names retain name-bank record IDs. Human speech retains its speaker and corpus record IDs and remains verbatim in the canonical world; the Terrarium projection strips selected transcriber notation and normalizes horizontal spacing. A profile statement can point to the research records that bound it, while an empty pointer marks a construction choice. No aggregate source is presented as an observation of a synthetic person, and no cast member is a de-identified copy of a resident.

Those records move through a short set of durable handoffs: target, research revision, cast, cast plan, world, and release. Research belongs to the target, so a later cast of the same place can reuse the dossier and add another immutable revision. The cast contributes a requested size, specifications, and seed. Its plan then allocates exact person slots, social units, and a shared map before profile prose is written. The map is a bounded stage for action, not a geographic scale model. People, places, relationships, and place ties are constructed together so a routine reaches places that exist and recurring obligations point to people who are actually in the world.

A finite cast still loses information, and Delos makes that loss reviewable. The casting agent chooses what the sample must preserve, then records rounding choices, omitted populations, unsupported intersections, deliberate overrepresentation, and alternatives in the cast plan. The package validates the parts that code can establish: size and seed, unique identities, exact realization of planned slots, social units and places, valid starting-place ties, resolvable research pointers, and exact name and speech-bank records. Richness, plausible network density, routine closure, and evidence fit remain agent judgments. The casting skill requires explicit walks and whole-cast review for those properties; the schema does not pretend to prove them. Stored artifacts are not automatically rerun through a newer semantic audit, so reuse requires a fresh validation walk.

Joint construction matters because Terrarium does not fill omissions after the run starts. A routine that mentions a workplace needs either that place or an explicit bridge beyond the map. A credential, vehicle, tool, or shared history that physics may test needs a present-tense fact on the person or place that carries it. A relationship is one shared object rather than two private stories that can contradict each other, while unrelated people can still meet through a shared place. These connections give a decision a route back through a neighbour, employer, household, or institution. Whether the consequence actually returns is runtime behavior, not something Delos scripts or guarantees.

When a world is ready, Delos stores it as an immutable revision and releases that revision explicitly. Saving a later world does not change what delivery emits until another release promotes it. Terrarium receives a versioned projection of the global places, each person’s starting location, profile, and visible and private facts; the delivery identifies its source cast and world revision, while detailed evidence stays in Delos. This method does not make the cast a representative survey sample, recreate any real resident, or show that a simulation predicts how the real population will behave. It provides an inspectable starting world whose evidence, constructions, and representation tradeoffs can be reviewed before the simulation begins.