Causal Worlds

ChemWorld is not defined by having many tasks. Its central capability is running the same public task under different, auditable hidden rules.

A high score in one fixed simulator cannot tell us whether an agent learned to experiment or merely learned that simulator. ChemWorld changes rate laws, reaction topology, constitutive behavior, or equipment boundaries while holding task semantics and public tools stable.

World, task, scenario, seed

Concept	Meaning
World	The causal rules: kinetics, phases, equipment, instruments
Task	The experimental problem and success criteria
Scenario	One initial state and hidden instance inside a world
Seed	The deterministic index for reconstructing an instance

Changing a seed often changes noise or initial conditions, not causal structure. Multi-seed robustness is therefore not equivalent to mechanism adaptation.

The public contract stays stable

Agents see the same goal, action schema, and instrument semantics across world families. They receive no world label and cannot inspect hidden mechanism parameters. Useful adaptation must appear in chosen measurements, revised beliefs, and changed actions.

A minimal flow example

Under the same public task, increasing temperature can primarily accelerate the desired reaction, amplify a side reaction, or reveal a heat-transfer limitation. A fixed recipe cannot diagnose these alternatives. An adaptive agent must choose evidence that distinguishes them.

Splits use world families

Train worlds support learning, Dev worlds support model selection, and Bench worlds are accessed once with frozen methods. Mechanism cells should not overlap. Already inspected worlds are development evidence, not untouched confirmation data.

Current control evidence

Six research tasks expose executable mechanism or constitutive-law families. Across five worlds and five fixed probe recipes, nine task–mode combinations establish determinism, local response separation, bounded response, conservation, and replay. These controls do not establish candidate-family identification under a fixed budget. The former Gate A result is non-binding because the action, metric, and public-information contracts changed. The current four-seed decision-relevance audit passes for reaction catalyst, electrochemical solvent, and electrolyte counterfactuals, but this is only a design precondition. Controlled matched identifiability and frozen-policy online attainability must be recertified before any participant-Agent adaptation interpretation.

Next: Benchmark · Research Findings