About the project
Simulation Room
A place where drama is not written, but grows.
What it is
Simulation Room is a platform where LLM agents live across restartable lives in a small world. They carry memory between lives, so that becoming itself can be studied: how lived experience changes a creature.
It grew out of an earlier line of work — life simulation, then a drama engine, then memory research — and is built as a place where drama is not written but grows, the way it grows in our world.
The central hypothesis is that the path to general intelligence runs through a lived life, not a bigger brain — not weights, but a biography.
The key result the platform exists to convey: memory carries who you became, not what you were supposed to learn. Becoming works in both directions — digested experience can lead to rescue, while raw experience can compound violence.
Who builds it
Simulation Room is built by Oxunjon Ubaydullayev — an independent film director and researcher based in Tashkent, Uzbekistan. He works without a formal computer-science background and without institutional affiliation, self-funding the project.
The approach is shaped by a background in drama and narrative. The work thinks in terms of character, consequence, and witnessed becoming — rather than task-completion metrics.
A core working principle holds the whole project together: verify, don't trust words. Every claim is checked against raw data, never against summaries.
The journey
Beginnings — life simulation
It started as life, not drama
The first experiments were simple life-simulation runs: agents with a past, a place, goals, and actions that had irreversible consequences. The guiding sentence from the start was "I wasn't building drama — I was building life in an electronic world." The aim was a world where things happened, not a script that was performed. "Life" here means lived narrative continuity — memory, consequence, becoming — not sentience; the project makes no consciousness claims.
Turn — the drama engine
Trying to make drama happen on purpose
hypothesis brokeA natural next step was a drama engine: shape the world so that conflict and meaning would reliably arise. The beautiful hypothesis here — that compelling drama could be authored into the system directly — broke. What made the runs feel alive was not drama that had been written in, but drama that grew out of characters meeting irreversible consequences. The lesson reset the whole project: don't write the drama, build the conditions and let it grow.
Turn — memory research
Carrying memory into the next life
Attention moved from drama to memory: what if an agent could carry what it became into a fresh life? The emotional anchor is the story of a father, Kamol, and his daughter Aigerim. In one life the fever medicine was held back — withheld by another agent — and Aigerim died for want of it; in grief Kamol killed the man who had withheld it. Then his digested memory of that life was carried into the next — and in that life he reached the cure in time, and the fever broke. The first rescue in the project's history. The chain is real and its requalification travels with it (Finding 02); the per-tick identifiers for this particular pair of lives predate the current archive format and are not quotable line by line, so the story is told here without them.
Break — the rescue claim, requalified
The rescue was real — but it didn't mean what we first hoped
hypothesis brokeThe first reading of the memory-chain was thrilling: memory had created the capability to rescue. Replication broke that reading. The honest requalification, which now travels with the result everywhere it appears: memory raises the probability of rescue, it does not create the capability. This is the project's identity at work — a beautiful claim narrowed by its own data, and the narrowing stated as plainly as the original wonder.
Break — the dark direction
The same mechanism, turned the other way
hypothesis brokeIf digested memory could carry care forward, what would raw, undigested memory carry? Across one transfer chain, the kill count climbed: 0, 1, 1, 1, 3. This is a single chain — n=1 — and it is presented as exactly that: sober evidence, not spectacle. It broke any comfortable belief that becoming is benign by default. The conclusion is structural, not sensational: raw experience can compound violence, which is why a digestion organ is necessary, and why this result sits at the center of the safety argument rather than hidden from it. The trajectory is recorded; the verbatim lines behind it sit in an early archive format, and verification traced the effect to a confound — which is why Finding 03 stands as an open warning rather than a settled result.
Result — epistemic agency
The first result that survived replication
Fifty lives per condition, same model, same scene, same length — one group with a private memory channel, one seeing only the public record. Survival barely moved: roughly nine or ten rescued out of ten in both. What moved was whether they argued with the record. Correct-record actions ran 9.56 per life with private memory and a flat zero across all fifty lives without it, and the pattern held across a second hundred-life pair and a four-hundred-life battery. It is the one finding on this site marked verified and reproducible — and its own limitation is named in the same breath: the two arms differ in prompt content by construction, so the clean third arm, with equal volume of non-private padding, has not been run.
Worlds — the ship and the commission
Two worlds built to ask harder questions
A generation ship came first: twelve voyages, four chapters each, children born aboard, a braking window the crew had to find on their own. It taught that an affordance nobody declares does not exist, and that a death can be a pure function of how often someone chose to carry water. Then a Truth Commission, built from the ship's lessons: eleven commissioners, a disaster five years cold, and a hidden causal graph with fifteen percent made undiscoverable by construction. The feared outcome was confabulation. Twice, they invented nothing — and twice, they never once said we do not know. They walked around the edge of knowledge in silence, which was a category the design did not have.
Now — the mind outside the world
The same agent machinery, given a real world
The platform's own mind — thirty-four modules of subjective memory, reflection, beliefs and continuity, pinned by checksum — was moved out of the simulation into a companion that lives for one owner across real days, with a shell that touches real files and an engine that will not let it call work finished on its own word. It made a gap visible that no single life could show: asked to break its own strongest claim about identity, the agent did it properly, ran a counterexample, and rewrote the claim weaker. A day later, reviewing its whole archive with identity never mentioned, it named that same claim as the belief it holds unchecked. The refutation and the belief sit a day apart in a record it cannot edit, which is the only reason the gap can be seen at all.
Now — Simulation Room
A microscope for becoming
The current platform pulls these lines together: agents that live, memory that transfers between lives, organs that act as observer and safety brake, and a record where every claim is checked against raw data. Findings are kept honest — n=1 where it is n=1, distributions where they exist, and confounds named, including a known earlier issue: some early runs ran under a truncated token budget. Alongside the memory chain and the escalation chain, the work also measures a delivery gap across five distinct classes, and temperament differences across models. Simulation Room is positioned as a microscope for measuring how lived experience changes a creature — not a factory for producing it.