Simulation Environment

When the goal expands from assessment to capability building, the assessment environment evolves into a simulation one: millions of interactions instead of thousands, randomization versus rote learning, instant feedback. Differences from the assessment environment: reliable reset semantics (each episode is a deterministic clean slate), throughput higher by orders of magnitude. The assessment verifier directly becomes the RLVR reward function. Spectrum of fidelity: from API stubs to the real world; domain randomization narrows the gap.

Related: [Assessment Environments], [Data and Environment are More Important than Algorithm], [Sim2Real]