ChemWorld System Model¶

This page is the normative source for architectural terminology, trajectory fields, and manuscript boundaries. The authoritative evidence state is configs/current.json; implementation does not imply empirical closure.

ChemWorld unifies a Physical Causal World Substrate, an Experimental Interaction Runtime, and a Task and Evaluation Contract so that search, execution, identification, and adaptation can be studied as one experimental- intelligence problem. The agent, trainer, model weights, and private agent memory remain outside these three layers.

trainer / pretraining / fine-tuning
                │
                ▼
              Agent
                │ public task, actions, observations, history
                ▼
┌────────────────────────────────────┐
│ Task and Evaluation Contract       │
├────────────────────────────────────┤
│ Experimental Interaction Runtime   │
├────────────────────────────────────┤
│ Physical Causal World Substrate    │
└────────────────────────────────────┘

ChemWorld can generate data for an external trainer or host training interactions. It does not update agent weights, maintain agent beliefs, or select the next experiment for the agent.

Layer responsibilities¶

Layer	Owns	Does not own
Physical Causal World Substrate	State, dynamics, constitutive laws, equipment, observation generation, constraints, controlled interventions	Task objectives, leaderboard weights, agent beliefs, training
Experimental Interaction Runtime	Action validity, transactions, lifecycle, measurements, visible feedback, failures, resource ledgers, trajectories	Agent action selection, silent closeout assistance, cross-task ranking
Task and Evaluation Contract	Public goal, action/observation permissions, budgets, termination, scoring, scenario distribution	Runtime physics or hidden-truth disclosure

The current scoring compiler lives under chemworld.world.scoring, but the task layer has authority over objectives and weights. The runtime executes a task-supplied scoring contract; the physical world does not define the research objective.

Physical causal worlds¶

A world family can be written as

[ W=(\mathcal X,\mathcal U,T_\omega,O_\omega,C_\omega,\Delta_\omega), ]

where the terms denote typed physical state, operations, hidden transitions, observation generation, constraints, and controlled world interventions. Implemented intervention axes include parameters, rate-law form, topology, constitutive relations, material mappings, and selected apparatus boundaries. Observation noise is configurable, but a general independently validated sensor-law family is not a current claim.

Experimental runtime¶

An Experiment begins from an explicitly initialized sample or process state, contains a sequence of operations and measurements, and ends in final assay, explicit termination, failure, or budget truncation. Only a contract-valid final assay is a comparable formal outcome; failures and incomplete runs remain in autonomy denominators.

Campaign
└── Experiment
    └── Operation / Measurement

The same runtime exposes three action abstractions:

Campaign Design: select a complete recipe or experiment;
Procedure Execution: select one operation at a time;
Process Control: select bounded equipment setpoints or process-control actions.

The current control interface is a bounded setpoint/process-control abstraction, not a claim of universal high- frequency continuous control. The runtime validates and records lifecycle semantics but does not choose terminate or final-assay actions for an autonomous agent.

Task and evaluation contract¶

A task specifies the public goal, action and observation permissions, budgets, scoring rules, and world/intervention distribution. Task endpoints, online shaping, constraints, information efficiency, resources, and procedural autonomy remain separate; unlike physical quantities are not averaged into one cross-domain intelligence score.

Canonical entities and identifiers¶

Entity	Definition	Trajectory identifier
Task	Stable public goal, permissions, budget, and scoring contract	`task_id`, `task_contract_hash`
World	One hidden physical-law instance	`world_id`, `mechanism_hash`
Scenario	Task–World composition plus initial state, interventions, reset/feedback condition, and seed	`scenario_id`
Campaign	One agent run on one Task × Scenario × Seed	`campaign_id`
Experiment	One initialized physical run within a campaign	`experiment_index`
Operation	One state-changing, measurement, or lifecycle action	`operation_id`, `action`
Run	Traceable execution instance	`run_id`

The canonical benchmark cell is Task × Scenario × Agent × Seed.

Trajectory v0.1 stored a seed-specific run label in task_id and the stable task in benchmark_task_id. In v0.2, task_id is the stable task-contract identifier and run_id carries the execution label. benchmark_task_id remains a compatibility alias. The v0.2 writer stops materializing all five v0.1 aliases in chemworld-trajectory-0.3; read-only support for old trajectories is a separate policy.

Three outcome layers¶

Trajectory v0.2 fixes three top-level fields:

Field	Meaning
`environment_outcome`	The transaction, physical observation, resource consequence, and lifecycle state actually produced by the world/runtime
`agent_visible_observation`	The observation and feedback actually released under this information condition
`evaluation_outcome`	The true evaluation endpoint/reward and scoring-contract binding

Feedback ablations may change only agent_visible_observation; they must not rewrite the environment or evaluation outcomes. The v0.1 aliases observation, reward, agent_view, and leaderboard_score remain during migration, but new analysis must use the three outcome layers. Their writer compatibility, together with benchmark_task_id, ends at chemworld-trajectory-0.3.

Bounded completeness¶

ChemWorld distinguishes:

structural completeness: the hidden-world → action → transition → measurement → feedback → next-experiment interaction chain is closed by design;
evaluation completeness: outcomes, constraints, resources, adaptation, and autonomy have explicit contracts, but fixed-world formal evidence exists on two tasks while mechanism-adaptation cross-method evidence remains incomplete;
attribution completeness: diagnostics are designed to separate under-identifiability, experiment choice, feedback use, recovery, and lifecycle failure. Historical RC28 Gate A passed on its frozen source; its current binding is stale, and participant-Agent Gates B–E remain to be frozen and executed.

The normative boundary is:

ChemWorld targets structural completeness of the experimental-interaction stack across selected physical- chemistry archetypes; chemical coverage and numerical fidelity are bounded and explicitly declared, not exhaustive.

Core, Diagnostic, and Extended¶

Core is the six-serious-task agent-comparison environment. Its environment contract is ready; two flagship tasks have frozen formal descriptive Participant results, while the remaining tasks and cross-task method matrix remain unfrozen.
Diagnostic contains identifiability, no-change, feedback branching, counterfactual, adaptation-decomposition, and autonomy protocols. Mechanism v0.3 starts with crystallization and electrochemistry and separates static, calibrated-change, uncalibrated-stress, and recovery tracks.
Extended contains the remaining registered tasks, training uses, and demonstrations. Coverage does not grant a formal ranking claim.

These are evaluation roles, not separate engines. The current evaluation contract defines Core; changing its scope requires a new protocol version and revalidation, and prior results cannot be carried forward silently.

Mechanism-understanding evidence¶

Declared: an auditable mechanism distribution or change probability;
Predictive: a testable prediction for an unexecuted intervention;
Actionable: a belief that changes experiment choice and improves recovery or regret under a fixed budget.

The mechanism protocol requires explicit old-world reference acquisition before calibrated change attribution. It makes never first-class, uses relation closure and within-campaign pre-change cross-fitting for reference sufficiency, blinds the policy to prefix and changepoint support, and separates environment, Participant, and private-confirmation cohorts. Static current-world identification and early uncalibrated nonstationarity remain separate, non-gating diagnostics. An independent Agent counterfactual-prediction probe is still a future result rather than completed evidence.

Current boundary¶

The candidate backend and replay controls are operational, and two flagship tasks have formal descriptive Participant results. Historical environment Gate A has a stale current-source binding. Participant Gates B–E, private confirmation, and external bridge evidence remain incomplete, so current benchmark_ready=false. Documentation and scientific claims must follow configs/current.json.