ChemWorld Compiles Chemical Processes Into Replayable Agent Environments, Preserving Transaction Receipts Even for Failed Actions

Evaluating autonomous chemistry agents typically involves two extremes: physical laboratories can provide evidence involving real materials, but are difficult to reset at low cost; general-purpose simulators are easy to rerun, but their underlying worlds are usually fixed, making it difficult for researchers to change only one chemical rule while keeping every other condition identical. Released on August 11, ChemWorld attempts to turn the “chemical world itself” into a programmable, versionable experimental variable.

ChemWorld decomposes reactions, heat, phase behavior, separation, crystallization, distillation, continuous flow, electrochemistry, and observational instruments into components. After researchers submit a world declaration, a compatibility compiler checks dependencies, state ownership, units, parameter ranges, resources, and lifecycles before generating the public contract visible to the agent. Material properties, constitutive laws, and hidden parameters remain under the evaluator’s control. The agent can change state only through typed operations and instrument interfaces; it cannot directly modify the simulator’s internal values.

The execution layer uses transactional semantics: every action proceeds through preflight checks, runtime preconditions, candidate-state generation, and postcondition validation. Failed actions roll back the physical state but still retain the request, failure reason, and resource impact, allowing the full trajectory to be reconstructed. The authors performed exhaustive checks on the reference registry and 52 generated combinations. Across eight frozen workflows, 89 actions were submitted: 88 committed and one rolled back as planned. All were replayed with zero numerical error. Six parent–child world pairs also replaced only one private law each, producing 24 attributable trajectories under the same public contract, action sequence, and randomness.

The code, tests, evidence, and offline verifier have been released under the MIT License, and the local Student Lab can run without a model API key. However, the paper demonstrates only that component compatibility, transaction logging, and controlled forks work as intended; it does not compare multiple agents or calibrate the system against a wet lab. Engineering teams should next focus on external validation of the simulation laws, whether agents can exploit simulator shortcuts, and version consistency between the paper’s frozen snapshot and the continuously updated library.