The Synthetic Simulation Stack

Human simulation, stage by stage. If you squint, what we used to call

"synthetic data", "synthetic rubrics", "AI researcher", and "end-to-end RL

environments" is one trend: increasingly ambitious human simulation —

10% worse, but 100x cheaper and 10,000x faster.

Each stage below replaced one more human role in the training loop with a

model. The remaining human roles narrow as the loop closes on itself.

• Stage 1 — The reward signal — the judge goes synthetic (2022)
• Stage 2 — The training data — the corpus goes synthetic (2023)
• Stage 3 — The teacher — distillation from models (2023)
• Stage 4 — The curriculum — models choose what to learn (2024)
• Stage 5 — The researcher — agents run experiments (2026)
• Stage 6 — The environment — RL worlds get synthesized (2026)
• Stage 7 — The human subject — simulated people, simulated panels (2025)
• Stage 8 — The physical world — the part that can't be synthesized (2026, in progress)
• Where simulation breaks — the 2026 skeptical correction (variance, tails, lived experience)
• From papers to procurement — NIH grand challenges, buyer's guides, vendor directories (Sept 2026)

*Updated October 2026: stage 6 extended with co-evolution; new page on

institutionalization; the playground gained a Synthesizer/Solver

co-evolution mode, a gossip/social layer, tail-fidelity sampling, a

feature-gated real LLM adapter, and a web UI served by a small Rust

server.*

This tracker ships with Simulacra, a Rust workspace that prototypes the

pipeline: synthesized environments, synthesized binary verifiers with

oracle/no-op/unsolved stress tests, heuristic agent personas behind an LLM

provider trait, and a simulated-town playground demo.

FAQ

What is synthetic human simulation?

The trend of replacing each human role in the AI training loop — judge,

data labeler, teacher, curriculum designer, researcher, environment

builder, and research subject — with a model. Synthetic data, synthetic

rubrics, AI researchers, and end-to-end RL environments are all instances

of the same idea: simulation that is roughly 10% worse than the human

original but 100x cheaper and 10,000x faster.

What are the stages of the synthetic simulation stack?

Eight stages, in the order they were automated: (1) the reward signal —

RLHF reward models and LLM judges; (2) the training data — Phi-style

synthetic corpora; (3) the teacher — model distillation; (4) the

curriculum — self-instructing models; (5) the researcher — autonomous

experiment loops like Karpathy's autoresearch; (6) the environment —

synthesized RL worlds with synthesized verifiers; (7) the human subject —

digital twins of real people for surveys and A/B tests; (8) the physical

world — the one layer that resists full synthesis.

Where does synthetic simulation break?

Independent evaluations converge on one finding: simulated populations

track average responses well but fail on variance, price sensitivity, and

distribution tails, and they break on lived experience and emotional

nuance. Simulation quality is a measured quantity — like the synthesized

verifiers it depends on, a simulator must be stress-tested against ground

truth before anyone acts on its output.

What is Simulacra?

Simulacra is an open-source Rust workspace that prototypes the pipeline:

seeded environment synthesis, constraint-based binary verifier synthesis

with oracle/no-op/unsolved stress tests, heuristic agent personas behind

an LLM provider trait, a Synthesizer/Solver co-evolution loop, a

gossip-based social layer, and distributional-fidelity experiments that

check simulated populations against theory.