VA-Sim simulates a single subject — identity withheld by curatorial order — living between home and a third place. Each day is a timeline from 00:00 to 24:00, cut into routine slots and executed in 288 ticks of five simulated minutes. A plan cron fires at 00:05 local time; a tick cron fires every five minutes after. One tick equals five minutes of lived time, one to one.

The architecture is three layers with a single law: the model proposes, the simulation disposes. Layer 1 plans once per day — a rule draft enriched by an external chat model that may swap micro-activities but may never move the slot hours. Layer 2 ticks deterministically with zero LLM calls: desires, micro sessions, multi-tick travel bridges, safety-net constraints. Layer 3 interprets after the fact — every feeling traceable to a chain of activity, bars, and emotion scores. Since Rev. B the engine and prompts are free of hardcoded identity: a character registry supplies the subject, and an emotion engine plus inner-state behavior (fourteen labels, urgency and relief, soft-max desire) turn inner pressure into outward routine.

Phase 09 gave the subject things to own. A handful of everyday objects are seeded — tools that drain, consumables that replenish, and a rotating favourite — each carrying its own state: charge, place, fullness, affection. Tools with a dead battery are excluded from the micro-activities that need them, and the exclusion is remembered in the narrative rather than raised as an error. Charge drains while the subject works, and a companion's needs rise when the subject tends to it, feeding a new desire. Phase 09b then removed the kindness the simulation had been granting for free: no more charging merely because the subject is home. Charging became a micro-activity with a place, a duration, and a cost in the day. Phase 09c gave the subject foresight — below thirty percent, plugging in starts pulling at the subject's attention, and a dead-end latch makes sure a dying device sends the subject home once rather than twice.

Phase 10 brought other people into the room without letting them become state. A roster of five relations — family, friend, colleague, acquaintance, client — is seeded in code, each with a closeness score, traits, and the hours they tend to be free. Presence is probabilistic and deterministic: closeness times habit match times a day-variant roll, capped at three, with one micro reserved for reaching out to someone. A per-day ledger blocks double-booking — if the subject is already out with someone, the subject cannot also be elsewhere with someone else, though reaching out costs nothing in location. Nothing new was persisted: a social beat is a line folded into the tick's own narrative, and the AI may only schedule people who are actually on the roster.

The evidence lives in the archive. An accelerated simulator replays fourteen days at ten ticks a second — 4,032 ticks, nine days planned by the model and five by fallback, twenty enrichment attempts vetoed, not one behavioural loop surviving to the end. Those runs are what exposed the last three repairs: one routine could interrupt another, a tired subject would rather linger than travel home, and after two visits a place closed for the rest of the day. The subject's identity is redacted everywhere — profiles, tables, session ids — because the simulation is about the shape of a day, not who lives it.

From Phase 11 the simulation has somewhere to go. A relationship layer turns first contact into a roster entry; a standalone world map routes the subject over a seeded road network, with travel cost shaped by daylight, weather, and landmarks; Phase 13 lets people actually be met there — at a landmark, on the way home, or remotely. The tick still owns the truth: a meet is a commitment, not a conversation, and the map only lengthens a day in the places it can afford.