arXiv CS open access

ChemWorld: Programmable Chemical Worlds for Controlled and Replayable Agent Experimentation

2026 · arxiv_cs

arXiv CS · Papers · 2026

deep-learning machine-learning neural-networks

Conceptio Open Archive

Analyze & explore related research in Conceptio

Connect this record directly into your research workflow or AI agent. Explore citations across open repositories, query full text, and export structured proof bundles.

Explore in Search →Connect via MCP / API ↗

ChemWorld: Programmable Chemical Worlds for Controlled and Replayable Agent Experimentation Jiangjie Qiu1 , Yijun Li1 , Xiaonan Wang1,* 1

Beijing Key Laboratory of Artificial Intelligence for Advanced Chemical Engineering Materials, State Key Laboratory of Chemical Engineering and Low-Carbon Technology, Department of Chemical Engineering, Tsinghua University, Beijing 100084, China *Correspondence: wangxiaonan@tsinghua.edu.cn

arXiv:2608.10792v1 [cs.AI] 11 Aug 2026

Abstract Autonomous chemistry increasingly depends on environments in which agents can repeatedly act, observe, and adapt. Physical laboratories provide essential real-material evidence but are costly to repeat and difficult to use for tightly matched interventions, whereas most digital environments keep the underlying experimental world largely fixed. We introduce ChemWorld, a programmable chemical environment in which reusable process and observation components are compiled into executable worlds. ChemWorld separates the public experimental contract available to an agent from evaluator-owned chemical and material laws. Researchers can therefore vary world composition and operating conditions, or change a single hidden law while holding the public task and interaction conditions fixed. Transactional execution records operations, failures, resource changes, and state transitions, allowing complete environment–action trajectories to be replayed exactly and audited. Full-census qualification covered the reference registry, 52 generated compositions, and module, interface, compilation, and invalid-action tests. Eight deterministic experimental cases demonstrated shared lifecycle semantics, failure recovery, and exact replay, while six parent–child world-fork pairs isolated the effects of single private-law interventions under matched public conditions. An independent agent also completed a full lifecycle in a non-reference world through the same public interface. Within the declared component and model domain, ChemWorld provides a controlled and replayable substrate for studying experimentation across systematically varied chemical worlds, complementary to physical-laboratory evidence and calibration.

1. Introduction

agent completes a given task, operates a particular bench or adapts to fixed dynamics and hidden rules. It is less common to systematically alter the chemical processes and underlying laws that constitute the environment while keeping the agent-facing operations, instruments and observations unchanged. This limits another important class of controlled studies: under identical public experimental conditions, one may wish to change only a material property or process law and observe the resulting changes in the experimental process and agent behavior. Supporting such studies requires not just more predefined tasks, but an environment in which chemical worlds themselves can be constructed, modified, repeated and compared. We therefore introduce ChemWorld, a programmable chemical-world platform for autonomous chemistry research (Fig. 1). ChemWorld treats world construction as a locus of experimental control: it represents chemical processes and observation models as reusable components and separates the agent-accessible public experimental interface from evaluator-owned chemical and material laws. Researchers can compose processes, change operating conditions and instrument configurations, or alter one hidden law while holding the public task, action sequence and bound randomness fixed. The system records operations, state transitions, resource use, failures and recovery, thereby supporting repeatable, auditable and attributable experiments. This paper makes four contributions:

Recent advances in self-driving laboratories and chemistry agents are moving chemical research from algorithmic assistance toward autonomous experimentation. Agents can now plan experiments, call chemistry tools and interact with robots, automated instruments and cloud laboratories to execute synthesis, characterization and optimization workflows on real materials [1–9]. As agents assume a sustained role in experimental decisions, their capabilities depend not only on the model but also on the environment that defines available operations, observations and state transitions. Building environments for agent interaction, controlled study and evaluation is therefore becoming a foundational problem in autonomous chemistry. Physical laboratories provide the most direct evidence from real materials and instruments, but large-scale repeated interaction is costly: it consumes reagents, samples and instrument time and is constrained by equipment, safety requirements and experimental turnaround. At the same time, physical experiments are difficult to restart from exactly the same initial state or to support strict comparisons in which only one factor changes and all others remain fixed. Software environments offer lower-cost repeatable interaction, and existing systems span reaction optimization and process control, virtual chemistry, interactive scientific discovery and embodied laboratory simulation [10–22]. However, these environments usually define the experimental world in advance and primarily study how an

1. Programmable chemical-world construction. 1

Programmable Chemical Worlds

arXiv preprint

Reusable process and observation components and a compatibility compiler construct executable worlds whose topology, operating conditions, instruments and private chemical or material laws can be varied systematically. 2. A common experimental interface across worlds. Compatible worlds share one agent-facing contract for operations, instruments, observations, resources, failure handling, termination and evaluation. 3. Complete, replayable and attributable experimental records. Each operation executes transactionally and records success, failure, rollback and resource changes, providing process-complete evidence for replay and intervention attribution. Complete environment–action traces can be reconstructed exactly, while single-private-law forks isolate the effect of a registered world change. 4. Agent experimentation through the same environment. Deterministic workflows and an independent agent use the same public interface while the evaluator retains the complete experimental process record without exposing private state.

matched repetition remains limited by samples, instrument access and experimental turnaround. 2.2 Simulated laboratories for embodied scientific agents LabUtopia and Labimus use laboratory simulation to train and evaluate embodied scientific agents, covering scene understanding, object and material manipulation, instrument use and long-horizon experimental procedures [20, 21]. These environments primarily study how an agent carries out experimental procedures, reducing the cost and risk of training directly on physical hardware and complementing environments centered on scientific reasoning or chemical-process dynamics. 2.3 Interactive environments for scientific discovery agents DiscoveryWorld, BoxingGym, SciGym, SciExplorer and NewtonBench organize scientific problems as sequential cycles of hypothesis formation, experiment selection, observation and inference. They evaluate whether agents can acquire evidence actively and recover scientific rules in initially unknown systems [14–18]. MADE extends this paradigm to budget-constrained closed-loop materials discovery [19], while SciDisco uses process-verifiable scientific-discovery environments for agentic reinforcement learning and assigns turn-level training signals to intermediate actions that produce verifiable evidence [22]. Together, these works establish interactive scientific environments as infrastructure for both agent evaluation and training.

This work establishes ChemWorld’s construction, qualification and agent-interface integration within the declared component and compatibility domain; comparative agent performance, universal chemical fidelity and physicallaboratory transfer remain outside the present scope. Open-source release. Executable code, frozen protocols, processed evidence, replayable trajectories, figure sources and the manuscript release are openly available in the ChemWorld public repository. The frozen code-andevidence snapshot referenced in this paper is versioned as v0.1.0.

2.4 Virtual chemistry and process environments The closest prior work to ChemWorld lies in virtual chemistry and process environments. Summit and Olympus provide repeatable software benchmarks for reaction optimization and experiment planning, while PC-Gym extends this approach to chemical-process control with nonlinear dynamics, constraints and disturbances [10–12]. ChemGymRL constructs a modular virtual chemistry laboratory in which reaction, extraction, distillation and characterization are organized as interactive chemical benches. Stable action and observation interfaces support reinforcementlearning training and policy comparison [13]. ChemGymRL and ChemWorld both expose chemical processes through interactive software environments, but they operate at different levels of control. ChemGymRL makes virtual chemical benches configurable for agent training and evaluation; ChemWorld instead treats the composition and hidden laws of the chemical world itself as controlled experimental variables while preserving a common public agent interface. Researchers can therefore alter the underlying world while holding the task and interaction conditions fixed for strictly matched comparison. Transactional execution and evaluator-complete records further make these matched world changes exactly replayable and experimentally attributable. Table 1 summarizes this functional rather than performance-based distinction in terms of repetition, experimental intervention and record scope.

2. Relation to existing experimental environments Experimental environments for autonomous science have developed along four related lines with different primary emphases: autonomous experimentation in physical laboratories, laboratory simulation for embodied agents, interactive environments for scientific discovery, and virtual chemistry and process environments. We discuss these directions in turn to position ChemWorld within the existing landscape. 2.1 Autonomous experimentation in physical laboratories Coscientist and ChemCrow connect language-model reasoning with chemistry tools, automated platforms and cloud laboratories. A-Lab, mobile robotic systems, ORGANA, ChemAgents and newer instrument-facing agents further demonstrate closed-loop synthesis, characterization, long-horizon experimental workflows and autonomous operation of scientific instruments on real materials [1–9]. These systems establish the real-material execution branch of autonomous chemistry. Their strength is direct physical experimental evidence, although large-scale, strictly 2

Programmable Chemical Worlds

arXiv preprint

A. Composable Components

B. World Compiler

Reusable process and instrument modules with declared interfaces

Chemical Transformation

State & Phase Behavior

Composes components into an executable world with a public contract and private mechanisms

Transport & Separation

Components & Topology

Compiled World

Compile & Validate

Public Contract (Wpub)

Compatibility Reaction

Electrochemistry

Thermal

...

Instrument Modules

UV-vis

Phase Behavior Crystallization

Distillation

Flow

Shared Declarations

HPLC

State

Parameters

Interface

Constraints

C. Transaction Semantics (per step)

Typed Action from Agent

Runtime Precondition

P=1

Schema, compatibility, resources

P=0

Pass

Cpre

Contextdependent checks

No execution State unchanged Charge attempt cost

Instruments

Life Cycle

• Instruments • Failure rules

Private Mechanisms • Constitutive laws • Material identities

Compilation Rejected

• Hidden simulator

• Private parameters • Internal events

D. Deterministic Replay & Controlled Forks

Execute World Dynamics

Post-Execution Validation

Fθ

Cpost

Generate candidate

State integrity, solver invariants

Fail

Preflight Reject

Feasibility

• Observables

• Operations • Resources

Public Task Interface

Deterministic Replay

Typed action → validated execution → commit or rollback → observations & record Preflight

Parameters & Conditions

• Task card

Fail

Pass (C = 1)

Reconstruct and replay exactly the same trajectory Original Run World W, seed ζ， actions

Observations & Record

Public ot+1 to agent

Same public contract and actions; change one private law parameter

Replay Run Same W and ζ Same actions

Parent World (Wpub, θp)

Different parameter

Child World (Wpub, θc ≠ θp)

State Trajectory s0, s1, …, sT

State Trajectory s′0, s′1, …, s′T

Same Actions (a0, …, aT-1)

Evaluator record (privileged)

Observations o0, o1, …, oT

Observations o′0, o′1, …, o′T

State Trajectories s0ᵖ, …, sTᵖ

State Trajectories s0ᶜ, …, sTᶜ

Records & Ledgers (identical)

Observations o0ᵖ, …, oTᵖ

Rollback (C = 0) Discard candidate, restore committed state st and RNG, discard resource changes, charge attempt cost, write receipt

Reconstruct world

Controlled Forks (Counterfactuals)

Success path

Exact Match (zero numerical tolerance)

Failure path

Rewards, observations, terminations, metadata, resource ledgers, events, state hashes

Diverge as expected

Observations o0ᶜ, …, oTᶜ

Differences attributable only to private-law change public contract, actions, instruments, resources identical

Figure 1. ChemWorld composes declared modules into executable worlds with transactional semantics, deterministic replay and controlled forks. A, Reusable process and instrument modules expose shared declarations. B, The compiler either produces a public contract Wpub with evaluator-owned private mechanisms or rejects an invalid composition before construction. C, Typed actions traverse preflight, runtime-precondition, candidate-execution and post-execution validation gates; non-commit branches preserve committed state and record declared attempt consequences. D, Replay reconstructs the bound world and submitted action trace, while controlled forks hold the public contract and actions fixed and change one private law. Table 1. Complementary operating regimes for autonomous chemical experimentation. ChemWorld combines software-scale repetition with explicit world composition, private-law intervention and process-complete replay. System class

Representative emphasis

Replication regime

Experimental intervention

Observable record

Physical SDL / chemistry robot

real-material execution and hardware integration procedural and manipulative competence

apparatus-, materialand time-bound physical repeats resettable simulated episodes

protocol, material and hardware changes

evidence gathering, hypothesis testing and law recovery objective optimization, process control and interactive chemistry controlled experiments over composable executable worlds

resettable interactive tasks

sensor, automation and sample records actions, observations and simulator state analyses, actions, observations and evidence objective histories, controller traces or task state typed actions, evaluatorcomplete state, public observations, failures, resources, termination, lineage and exact environment replay

Embodied laboratory simulator Scientific-discovery environment Optimization / virtual chemistry environment ChemWorld

software-scale queries or control episodes exact reset, matched repetition and versionbound replay without direct wet-laboratory consumable use

scenes, objects, procedures or embodiment configurations datasets, hypotheses, latent rules or oracle budgets objectives, process models, bench configurations or operating conditions declared component composition plus singleprivate-law forks under an invariant public contract

3. Programmable chemical world construction 3

ChemWorld is designed not as another collection of fixed tasks, but as an experimental medium in which chemical

Programmable Chemical Worlds

arXiv preprint

3.2 Compatibility compilation and experimental contracts

worlds can be constructed and modified systematically. It first represents chemical processes as reusable components, compiles compatible declarations into executable worlds, and then executes every operation through one transactional runtime. This section defines that construction stack: the world representation, the compatibility compiler and the execution and replay semantics shared by all compiled worlds.

A world declaration specifies component selection and topology, public parameters, instrument configuration and the agent-facing task surface. ChemWorld does not execute this recipe directly. Its compatibility compiler first normalizes the declaration and then checks component dependencies, state ownership, unit agreement, supported parameter domains, resource feasibility, instrument availability, operation exposure and lifecycle closure. Only declarations that satisfy all checks are compiled into executable worlds. An invalid declaration fails closed before environment construction and returns structured diagnostics, rather than leaving a partially constructed or semantically ambiguous environment. A successful compilation returns both the executable world and one public experimental contract that exposes the allowed typed operations, instruments, observations, resources, failure semantics, termination and evaluation, while evaluator-owned mechanisms and hidden state remain private. Composability therefore does not mean arbitrary juxtaposition. A new world enters the runtime only when its component interfaces, resource conditions and lifecycle jointly satisfy the declared compatibility rules. Conversely, an additional process model or an empirically calibrated formulation can enter through the same interfaces without redesigning the public task contract, transaction layer or replay machinery. Compilation thus separates world authoring from both process-model implementation and agent integration.

3.1 Composable chemical worlds ChemWorld represents reaction, thermal, phase, separation, crystallization, distillation, continuous-flow, electrochemical and observation processes as reusable components. Each component declares its parameter domains, dependencies, owned state and public interfaces. Researchers define a chemical world by selecting components, specifying their topology and binding operating conditions and instruments, rather than implementing a separate environment for every task. Reaction and electrochemical components transform material state; thermal, phase, separation, crystallization, distillation and continuous-flow components supply process-specific transitions; and observation components attach synthetic instruments. All public operations are typed actions rather than unrestricted simulator mutations. A complete world is represented as W = (Wpub , θ),

T = (Wpub , S0,pub , A, I, O, R, τ, E).

Here Wpub contains the public component topology, public parameter domains and interfaces, whereas θ contains evaluator-owned constitutive laws, material properties, hidden parameters and private initialization. The public task contract T joins the public world description and initialstate projection with the allowed actions, instruments, observations, resources, termination rule and evaluation rule. The complete world identity therefore binds both Wpub and θ, while the agent interacts only through T . We distinguish three identity levels so that task identity, world identity and randomness do not silently serve interchangeable roles:

3.3 Transactional execution and exact replay Once a world has been compiled, every submitted action follows the same transactional sequence: preflight admission, runtime-precondition evaluation, candidate execution, post-execution validation and either commit or rollback. A schema, compatibility and resource predicate P (st , at , Rt ) first determines whether an action may enter the runtime. An admitted action is then tested against its context-dependent runtime preconditions. If those preconditions pass, the bound mechanism θ, committed state st , resource ledger Rt , action at and recorded random variates ξt generate a candidate transition:

world-spec ID = id(Wpub , θ), (s̃t+1 , R̃t+1 , ẽt+1 ) = Fθ (st , Rt , at , ξt ).

scenario ID = id(Wpub , θ, ζinit , ζdyn , ζobs ),

Candidate state is not installed immediately. A runtime commit-gate predicate C ∈ {0, 1} covers both a runtimeprecondition rejection before candidate generation and post-execution checks of state integrity, solver status, runtime invariants and the observation path. For a generated candidate, the latter gate is written C(s̃t+1 , R̃t+1 , ẽt+1 ). The transaction commits only when admission and all applicable runtime checks pass:

task–world unit = (task ID, scenario ID). The qualification study uses the final level when testing overlap with the reference registry: a generated row overlaps only if both its task and scenario identities match a registered unit. This public/private separation also makes a second operation possible. A researcher can preserve the agent-facing task contract while changing one registered element of θ, turning a private world law into a controlled experimental variable.

P = 1, C = 1 =⇒ (st+1 , Rt+1 , e⋆t+1 ) = (s̃t+1 , R̃t+1 , eacc t+1 ).

Programmable Chemical Worlds

arXiv preprint

The two non-commit branches remain distinct in the record. If P = 0, the runtime emits a preflight-rejection event and receipt without runtime execution. If P = 1 but C = 0, it emits a runtime-rollback event and receipt; this branch can occur before candidate generation at a declared runtime precondition or after candidate generation at a post-execution check. For b ∈ {pre, roll}, st+1 = st ,

Rt+1 = Gb (Rt , at , ebt+1 ),

The generated set separates two forms of expansion. Eighteen compositions use three topologies absent from the reference registry: phase–observation, phase– separation–observation and reaction–thermal–continuousflow–observation. Eight reaction–thermal–distillation– observation compositions reuse a registered topology but have zero exact task–world identity overlap with the frozen registry. These are protocol-frozen non-reference compositions. The remaining 26 rows provide additional coverage within registered topologies. The first row of the eightcase non-reference block was fixed in advance as the agentintegration target, rather than selected after observing its outcome. Coverage selection combined pairwise rows for discrete component and instrument interactions, seeded Latin hypercube sampling for authored continuous domains, and ordered workflow targets for critical process sequences. The generated suite attained all registered targets: 60/60 discrete levels, 180/180 compatible discrete pairs, 212/212 continuous strata and 84/84 ordered workflow interactions. These denominators define a finite qualification sample within declared domains. They are not evidence of semantic completeness over all higher-order chemistry; higherorder behavior is tested only where it appears explicitly in a workflow, module probe or interface path.

e⋆t+1 = ebt+1 .

The branch-specific ledger function installs only the protocol-declared attempt cost or penalty. Candidate physical, observation and uncommitted resource effects are discarded. The committed runtime state also binds the observation-RNG state ρt ; a non-commit branch restores ρt , preventing an unsuccessful attempt from changing future observation noise. Public and evaluator records are separate projections of the realized branch, ot+1 = πpub (st+1 , e⋆t+1 ),

ôt+1 = πeval (st+1 , e⋆t+1 ).

Exact replay binds the normalized contract, runtime, mechanism and scoring identities, together with the seeds and intervention record. It reconstructs the compiled world and resubmits the full submitted action/transaction trace—including committed actions, preflight rejections and runtime rollbacks—and compares public observations, transaction outcomes, affected-ledger declarations, world events, rewards and terminal flags at zero numerical tolerance. Resource deltas and rollback receipts are reconciled separately. Replay therefore reconstructs a version-bound environment/action trajectory, not merely its endpoint or action list, and is distinct from experience replay used to train a reinforcement-learning policy.

4.2 World, module and interface qualification Each task–world unit had to compile, execute a complete workflow, close its lifecycle exactly once, reconcile declared and observed resources, and replay exactly. All 64/64 reference units passed, and the boundary and categorical set produced 1,786/1,786 complete executions. All 52/52 generated compositions also passed, including all 8/8 protocol-frozen non-reference reaction–distillation worlds. There were no missing receipts, failure classes or undeclared private-field-exposure findings. Compilation alone was never a pass condition. Qualification also tested the construction boundary. Seven deliberately invalid declarations covered missing dependencies, conflicting state ownership, unit mismatches, invalid parameter domains, resource impossibility and lifecycle gaps. All 7/7 failed closed before environment construction with the registered diagnostic class. The validworld census and invalid-world mutants therefore test both sides of the compiler boundary. Complete-world success depends in turn on the meaning of component transitions. Thirty-two module probes exercise zero input, declared parameter boundaries, monotonic directions, conservation and model-specific invariants. Seven cross-module paths test whether material amount, unit, identity and state meaning survive transfer between components and, where applicable, also test charge, energy and phase balance and event propagation. All 32/32 module probes and 7/7 interface paths passed. These results establish internally coherent executable formulations within their declared model-card domains. Their numerical fixtures and directional oracles are internal qualification checks, not independent reference imple-

4. Qualification of composed chemical worlds Programmability is useful only if newly composed worlds preserve the same executable contract. We therefore qualified ChemWorld at four connected levels: construction coverage, complete-world execution, module and interface semantics, and failure, resource and replay behavior. The design tests a finite declared construction domain; it does not claim exhaustive coverage of chemical space or all higher-order process interactions. 4.1 Qualification design and construction coverage The public capability map contains 15 registered reference tasks, 28 typed operation kinds and five synthetic instrument contracts. These tasks anchor interpretable examples but do not define the boundary of the executable world space. The reference qualification set contains 64 task– world units and 1,786 boundary and categorical recipes. To test construction beyond these reference examples, an execution protocol fixed eight component patterns and generated 52 additional compositions before authoritative qualification. 5

Programmable Chemical Worlds

arXiv preprint

A. From reference tasks to generated world families rxn

thermal phase

sep.

cryst.

distill.

flow

electro

B. Topology novelty vs. world-identity novelty

obs.

new identity

Topology-new generated families P1 phase + observation

phase + separation + observation P6 reaction + thermal + continuous-flow + observation

P4 P5 P6

task/world identity novelty

Registered topology, identity-new (distillation) P5 reaction–thermal–distillation– observation

P7 P8

registered topology

18 topology-new generated worlds

topology novelty

new topology

D. Composition changes the experiment surface

C. What actually varies in generated worlds 1 reaction – thermal – observation

heat (K) duration (s)

350 600

reaction (K / s) evaporation (K / s) distillation (K / s) reflux transfer

350 / 600 325 / 300 350 / 900 1.0 0.65

3.0 0.95

3 reaction – thermal – flow – observation

flow (mL min⁻¹)

residence (s) temperature (K)

0.5 60 330

5.0 600 390

4 reaction –

potential (V) current (mA) electrolysis (s)

0.5 25 300

1.8 150 1,800

electrochemistry – observation

26 registeredtopology generated rows used for coverage

64 reference task–world units

reference overlap

52 generated compositions = 18 topology-new + 8 identity-new distillation + 26 additional registered-topology coverage rows

reaction – thermal – distillation – observation

8 protocol-frozen non-reference reaction–distillation worlds

phase–observation

390 1,800 390 / 1800 345 / 900 390 / 2400

phase–separation–observation

Workflow

add solvent → add reagent → measure → terminate

Workflow

mix → settle → separate → wash → transfer

Public instruments

pH, UV–vis, final assay

Public instruments

HPLC, final assay

reaction–thermal–flow–observation

reaction–thermal–distillation–observation

Workflow

set flow → set temperature → run → sample

Workflow

heat → quench → evaporate → distill → collect

Public instruments

HPLC, GC, final assay

Public instruments

HPLC, GC, final assay

Composition changes topology, operating conditions, workflow order, and public instrument surface under one common contract.

Figure 2. Construction coverage and qualification design. A, The 52 generated compositions separate into 18 topology-new worlds, eight identity-new reaction–distillation worlds that reuse a registered topology, and 26 additional registered-topology coverage rows. B, Topology novelty and exact task–world identity novelty are independent coordinates relative to the 64 reference task–world units. C, Protocol-frozen rows vary discrete component patterns, authored continuous operating conditions and ordered workflows. D, Composition changes topology, operating conditions, workflow ordering and the public instrument surface within one common contract structure. Coverage defines a finite construction sample rather than the extent of the world-design space; qualification outcomes are reported in Sections 4.2–4.3 and Appendix C.

mentations or external physical validation. The interfaces identify where alternative or empirically calibrated models may be introduced while retaining the public task contract, transaction layer and replay machinery.

actions, admission rejections and runtime rollbacks. It compares rewards, public observations, termination flags, transaction metadata, affected-ledger declarations, world events, state-delta summaries and state-integrity checks with zero numerical tolerance; case qualification separately reconciles ledger deltas and rollback receipts. All reported replays yielded zero numerical mismatch. Thus the qualified reproducible unit is the complete environment/action trajectory, including failure and resource consequences, rather than an endpoint or simplified action list.

4.3 Failure semantics, resource accounting and replay qualification The mechanisms defined in Section 3.3 must also hold under invalid actions and exhausted resources. The frozen campaign contained 192 negative probes: 64 invalidschema/unknown-operation probes, 64 campaign-resourceexhaustion probes and 64 runtime-precondition probes. Every probe produced its registered outcome and preserved committed physical state. The evidence therefore qualifies 128 P = 0 admission rejections and 64 P = 1, C = 0 runtime-precondition rollbacks. Solverdiagnostic and candidate-observation fault paths are implemented as fail-closed semantics, but this campaign did not assign them separate qualification denominators. The resource ledger independently tracks material, sample, instrument use, process time, operation count and terminal assay. Failed attempts retain only their declared costs or penalties, and observation checks require public packets to contain only task-declared fields. Exact replay then reconstructs the bound compiled world and resubmits the complete typed-action sequence, including accepted

5. Controlled experiments over chemical worlds Qualification establishes that ChemWorld constructs and executes worlds consistently. We next ask what controlled experiments this medium supports. The first study tests Table 2. Reader-visible branch census for the 192 negative probes. Counts are inherited from the frozen qualification report; no post hoc probe reclassification was used.

Branch

Registered class and outcome

P =0

P = 1, C = 0

schema / unknown operation; validation rejection before runtime execution campaign-resource exhaustion; rejection before candidate installation runtime precondition failure; rollback preserves committed physical and observation-RNG state

Programmable Chemical Worlds

arXiv preprint

process-complete lifecycles, including planned failure and recovery, across diverse workflows. The second makes world construction itself experimental by changing one private law under otherwise matched public conditions.

actions, instruments, observations, resources, failures, scoring, material catalogue and the contracted invariant/safety surface—as well as the fixed typed-action sequence and bound randomness. Parent and child consequently have distinct complete-world identities while presenting the same public experiment. The admissible change is restricted to exactly one protocol-frozen private constitutive or material law. The partition intervention changes the hidden response from K 1.00 to K 1.75 ; the registered terminal organicproduct amount and public product_in_organic assay both increase. The electrochemical intervention keeps public material labels fixed while reassigning hidden electrolyteresponse profiles; the registered selective-product amount and public ohmic_efficiency both decrease. For each intervention, direction is checked separately by the frozen divergence oracle, while the magnitude oracle requires both the protocol-frozen absolute and relative thresholds. Repeating both variants produced 24 deterministic traces. All six pairs passed lineage, exactly-oneprivate-target, public-contract-invariance, same-sequenceexecutability, expected-state-divergence, expectedobservation-divergence and exact-replay gates. Within the declared executable model and intervention domain, the resulting trajectory differences are therefore attributable to the registered private-law change under fixed actions and noise identity. In the electrochemical fork, the permutation reassigns hidden response profiles to unchanged public material labels. It is therefore a matched change in private material properties, not an identifier remapping or a change in the public catalogue.

5.1 Process-complete lifecycles and failure recovery Eight protocol-frozen use cases span reaction-tocrystallization, resource-limited equilibrium characterization, planned failure and recovery, continuous flow, electrochemistry, distillation, partition and a second crystallization world. Together they cover single-stage and multistage processing, constrained measurement and multiple separation modalities. Each case is an independent experimental unit whose submitted actions are audited within one complete lifecycle. Across the eight cases, all 89 submitted actions have complete schema, transaction, state-integrity, event, resource and public-observation receipts. Eighty-eight actions committed and one protocol-frozen action rolled back. Every case completed one final assay, closed its lifecycle, reconciled resources and replayed exactly with zero numerical error. The same lifecycle contract therefore supports distinct chemical workflows without process-specific failure, termination or recording semantics. The failure–recovery case places one deliberate invalid operation inside an otherwise complete experiment. Its first action passed schema, compatibility and campaignresource admission, but a runtime-precondition check found that no separable phase had yet formed. The transaction consequently entered the recorded P = 1, C = 0 branch before candidate physical state was generated. It preserved committed physical state and observation-RNG state, created no ghost state, and reconciled the declared failed-attempt consequences. The next 18 actions continued from the last committed state and completed the recovery path, final assay, resource reconciliation and exact replay of all 19 submitted actions. Failure is therefore an experimental event rather than an episode-level exception. Its attempt and resource consequences remain auditable, while rejected physical state is excluded from subsequent execution. A workflow can recover without erasing committed history or restarting the experimental unit.

6. Agent interaction with ChemWorld The preceding experiments use deterministic execution to qualify the worlds themselves. We finally test whether a qualified world can be exposed to an autonomous agent through the same public contract while preserving an evaluator-complete record. The goal is interface integration, not a comparison of agent capability. 6.1 Public agent interface and evaluator-complete records ChemWorld separates agent-facing information from evaluator-side evidence. The agent receives the public task card, typed operations, instruments, public observations, available resources and termination and final-assay interfaces. It cannot directly inspect hidden material properties, private constitutive-law parameters, private initialization or complete simulator state. The evaluator retains the complete state transitions, transaction outcomes, failure and resource consequences and environment/action trajectory. This information boundary permits evaluation without exposing the mechanism being evaluated. ChemWorld groups 19 registered process coordinates into terminal commitment, evidence acquisition, evidence-conditioned action, resource deployment and outcome trajectory. They remain separate coordinates rather than a composite score

5.2 Controlled single-law counterfactuals Process-complete execution and controlled attribution require different designs. The lifecycle study asks whether diverse workflows share one execution semantics; the counterfactual study asks whether one private world law can change while all public experimental conditions remain fixed. Parent and child worlds are defined as Wp = (Wpub , θp ),

Wc = (Wpub , θc ),

θp ̸= θc ,

Tp = Tc = T.

The qualification contains six parent–child pairs: two intervention classes evaluated over three seeds. Every pair preserves nine versioned public-contract components—task, 7

Programmable Chemical Worlds

arXiv preprint

Table 3. Reader-visible specification of the two controlled-fork classes. Public task, actions, instruments, observations, resources, failures, scoring, material catalogue and the contracted invariant/safety surface remain unchanged in every row. Thresholds were frozen before execution and require both the listed absolute and relative magnitudes plus the stated direction. Fork class

Private target

Parent → child law

Registered state channel

Registered public channel

Partition constitutive law

phase-partition response

partition-base response K 1.00 → K 1.75

Electrochemical material law

hidden electrolyteprofile effects

public profile labels stay fixed; hidden effect-row mapping (0, 1, 2, 3) → (2, 1, 0, 3)

terminal organic-product amount Porg increases; ∆ ≥ 10−4 mol and relative difference ≥ 0.05 terminal selective-product amount Red decreases; ∆ ≥ 10−6 mol and relative difference ≥ 0.01

final-assay product_in_organic increases; ∆ ≥ 0.02 and relative difference ≥ 0.02 final-assay ohmic_efficiency decreases; ∆ ≥ 0.05 and relative difference ≥ 0.05

B. Rollback stays inside the same experimental record

A. Frozen execution cases across workflows

1 Resource-limited Charaterization

Crystallization World 1

Failure & Recovery

No Candidate State Committed

Typed Actions

Distillation Measure

Failure / Rollback

Partition

World 2

Termination

D。Separate agent-facing lifecycle on a non-reference world

Non-reference world: separate agent-facing lifecycle Partition law K1.00 → K1.75

Change one private law Child world Wc = (Wpub , θc)

18 subsequent commits

recovery resumes from last committed state

Controlled private-law forks

Wp = (Wpub , θp)

final assay

Final Assay

C。Controlled private-law counterfactuals

Parent world

…

Continuous Flow Runtime Precondition Failure

Electrochemistry

Reaction

organic product ↑ product in organic ↑

Thermal

Distillation

Observation

same public contract Deterministic qualification path

Electrochemical material law

12 actions

hidden row mapping changes

selective product ↓ ohmic efficiency ↓

Provider-driven agent run

15 actions

construction, closure, replay

typed tools, explicit termination, final assay

World Qualification

Agent Experiment

Figure 3. Execution, intervention and agent access. A, Eight frozen execution cases span crystallization, resource-limited characterization, planned failure and recovery, continuous flow, electrochemistry, distillation and partition workflows under one shared lifecycle semantics. B, A runtime-precondition failure at step 1 remains inside the same experimental record; recovery continues from committed state through 18 subsequent commits and final assay. C, Controlled private-law forks hold the public contract, typed action sequence and bound randomness fixed while changing one private law. D, A separate non-reference world supports deterministic qualification and a provider-driven agent lifecycle through the same public contract.

or a set of uniformly higher-is-better metrics; their numerators, denominators, null rules and boundary rules are specified in the process-coordinate contract in Appendix D.1. The resulting record can support subsequent comparisons of measurement strategy, resource allocation, response to invalid actions and stopping behavior. For example, an agent may choose between additional pH or

UV–visible evidence under a limited sample budget, or revise a typed action after a structured rollback without losing committed history. The deterministic cases in Section 5 qualify these interaction primitives, but comparative policy behavior is outside the claims of this study.

Programmable Chemical Worlds

arXiv preprint

6.2 Independent-agent integration in a non-reference world

after failure. Controlled forks further show that one private law can be changed while the public experimental conditions, action sequence and bound randomness remain fixed, allowing the corresponding trajectory differences to be attributed to that intervention. Finally, an independent agent completes a full lifecycle in a non-reference world through the same public interface, without its behavior being used to qualify the underlying world. ChemWorld therefore provides a software experimental substrate for controlled studies of world construction, world intervention and agent–world interaction, rather than simply a larger fixed benchmark suite.

World qualification and agent execution were kept as independent experimental units so that agent success could not serve as evidence for world correctness. The selected protocol-frozen non-reference world combines reaction, thermal, distillation and observation components. A deterministic 12-action path first qualified its construction, lifecycle closure and exact replay. Only after world qualification was complete did an independent agent enter the same public instrument contract and complete a separate 15-action lifecycle (Fig. 3D). The experimental interface supplied the public task card, typed tool schemas, resource contract and explicit termination and final-assay requirements. The agent received no private world fields and issued every experimental operation, including termination and final assay. In one uninterrupted experiment, it submitted 15 actions; all 15 committed and closed the lifecycle, with no rollback, right-censoring or undeclared private-field exposure. The environment used 8,158.454 of 10,440 simulated process seconds, four of four instrument uses and 0.00085 of 0.001 L sample. The evaluator-complete record links each decision to its public observation, hidden simulator consequence, transaction result and resource debit. The entire 15-step submitted-action trace replayed with zero numerical mismatch. An independent agent experiment can therefore enter the same auditable execution and replay framework as the deterministic studies. This demonstrates compatibility with the public instrument contract; it is not an agent benchmark, a model comparison or a claim of general chemical intelligence.

7.2 Implications for experimental-intelligence research Many agent benchmarks compress an experimental process into a success rate, reward or terminal property. Such outcomes indicate whether a task was completed, but reveal little about how the result was obtained. ChemWorld instead retains the full process from evidence acquisition and subsequent decisions through resource deployment, failure, recovery and the terminal experimental outcome, making experimental strategy itself available for study. Under a limited resource budget, for example, researchers can compare when an agent chooses to collect another measurement, when it stops, and how it allocates sample and instrument capacity. After an invalid operation, they can examine whether new observations alter the agent’s subsequent strategy. Within a controlled world fork, they can further test whether an agent detects, adapts to or learns a changed hidden law. The 19 process coordinates defined in this work provide a common record for such analyses, but are intentionally not collapsed into a single score, and this study does not use them to rank agents. Controlled forks are especially important for these questions. Parent and child worlds retain the same public task, operations, instruments, resources, action sequence and bound randomness while differing in one prespecified hidden law. Within the registered design and qualification criteria, the associated state and observation differences can therefore be attributed to that world intervention. Future work can build on this substrate to study agent adaptation, law discovery and scientific decision-making, but those claims require separately frozen agent policies and evaluation protocols and lie beyond the present qualification.

7. Discussion and Conclusion 7.1 The chemical world as an experimental variable ChemWorld’s central methodological contribution is not a larger collection of predefined tasks, but the ability to construct and control the chemical world itself as an experimental variable. Component composition and operating conditions specify what may vary; a unified public experimental contract specifies what must remain comparable; and transactional execution with evaluator-complete records preserves the state, resource and failure consequences of every interaction. Researchers can therefore alter process composition, continuous operating conditions or hidden chemical and material laws while keeping the agent-facing action and observation interface fixed, enabling strictly matched experiments across different worlds. The present study first establishes the reliability of this experimental substrate. Tests of reference worlds, generated worlds, process modules, component interfaces and invalid operations show that compatible compositions retain common execution, resource and replay semantics within the declared component and compatibility domain. Process-complete lifecycles then show that the same contract supports distinct chemical workflows and recovery

7.3 Scope, extensibility and physical validation The conclusions of this study are limited to the declared component vocabulary, compatibility rules and authored model domains. Qualification establishes consistent executable semantics and internally coherent behavior for these software models within their stated scope; it does not establish that ChemWorld fully reproduces real chemical systems. The current synthetic instruments are controlled software observation models rather than digital twins calibrated to particular devices. The coverage de-

Programmable Chemical Worlds

arXiv preprint

sign systematically samples prespecified parameter and composition spaces but does not exhaust chemical space or all higher-order process interactions. Module fixtures and directionality oracles primarily test consistency with the authored models rather than an independent reference implementation or empirical ground truth. Exact replay is likewise restricted to the bound software identities, world identity and environment–action trace; it does not imply policy re-execution, cross-platform numerical identity or cross-version archival replay. These boundaries do not require ChemWorld to remain fixed at its present level of fidelity. Component interfaces separate experimental semantics from individual model implementations, so experimentally calibrated kinetic and thermodynamic laws, unit-operation models and instrument-response models can in principle enter the same world-construction, transaction and agent-interface framework. The present study does not, however, establish that an independently authored third-party module automatically inherits every system guarantee without integration work. Each new model must be qualified against the relevant interface and runtime semantics. ChemWorld and physical self-driving laboratories are therefore complementary experimental regimes rather than substitutes. Software worlds support rapid repetition, strictly matched controls, targeted intervention and complete process records; physical laboratories provide realmaterial evidence and model calibration. A natural workflow is to use software experiments to identify conditions, mechanisms and agent behaviors that merit closer study, then return questions requiring real chemical evidence to the physical laboratory. In summary, ChemWorld establishes a programmable, repeatable and auditable chemical experimental environment in which researchers can systematically alter the world while preserving a common agent-facing interface. This work qualifies that capability within the current component and model domain and demonstrates complete experimental lifecycles, single-law controlled counterfactuals and independent-agent integration. Higher-fidelity chemical models, together with agent adaptation, law learning and scientific discovery in changing worlds, are natural directions enabled by this foundation.

run caches are not included.

8. Data and Code Availability A frozen, versioned public release associated with this study is openly available under the MIT License in the ChemWorld public repository; the frozen code-and-evidence snapshot referenced here is tagged v0.1.0. The release contains the executable code, publication configurations and protocols, processed machine- and human-readable qualification evidence, figure source files, replayable simulator environment–action trajectories, the arXiv PDF and source bundle, and an offline manifest verifier for the reported hashes and denominators. It is intentionally separated from subsequent benchmark development. Credentials, raw provider payloads, private seeds and local

Programmable Chemical Worlds

arXiv preprint

References

[13] Chris Beeler, Sriram Ganapathi Subramanian, Kyle Sprague, et al. ChemGymRL: A customizable interactive framework for reinforcement learning for digital chemistry. Digital Discovery, 3:742–758, 2024. doi: 10.1039/D3DD00183K.

[1] Daniil A. Boiko, Robert MacKnight, Ben Kline, and Gabe Gomes. Autonomous chemical research with large language models. Nature, 624:570–578, 2023. doi: 10.1038/ s41586-023-06792-0.

[14] Peter Jansen, Marc-Alexandre Côté, Tushar Khot, Erin Bransom, Bhavana Dalvi Mishra, Bodhisattwa Prasad Majumder, Oyvind Tafjord, and Peter Clark. DiscoveryWorld: A virtual environment for developing and evaluating automated scientific discovery agents. In Advances in Neural Information Processing Systems, volume 37, 2024. doi: 10.52202/079017-0324.

[2] Andres M. Bran, Sam Cox, Oliver Schilter, Carlo Baldassari, Andrew D. White, and Philippe Schwaller. Augmenting large language models with chemistry tools. Nature Machine Intelligence, 6:525–535, 2024. doi: 10.1038/ s42256-024-00832-8. [3] Nathan J. Szymanski et al. An autonomous laboratory for the accelerated synthesis of inorganic materials. Nature, 624:86–91, 2023. doi: 10.1038/s41586-023-06734-w.

[15] Kanishk Gandhi, Michael Y. Li, Lyle Goodyear, Agam Bhatia, Louise Li, Aditi Bhaskar, Mohammed Zaman, and Noah D. Goodman. BoxingGym: Benchmarking progress in automated experimental design and model discovery, 2025. Preprint, arXiv:2501.01540.

[4] Tianyi Dai, Sriram Vijayakrishnan, Filip T. Szczypiński, et al. Autonomous mobile robots for exploratory synthetic chemistry. Nature, 635:890–897, 2024. doi: 10.1038/s41586-024-08173-7.

[16] Haonan Duan, Stephen Zhewen Lu, Caitlin Fiona Harrigan, Nishkrit Desai, Jiarui Lu, Michał Koziarski, Leonardo Cotta, and Chris J. Maddison. Measuring scientific capabilities of language models with a systems biology dry lab, 2025. Preprint, arXiv:2507.02083.

[5] Kourosh Darvish, Marta Skreta, Yuchi Zhao, et al. ORGANA: A robotic assistant for automated chemistry experimentation and characterization. Matter, 8(2), 2025. doi: 10.1016/j.matt.2024.10.015.

[17] Maximilian Nägele and Florian Marquardt. Agentic exploration of physics models. Physical Review X, 16:031002, 2026. doi: 10.1103/xnqc-q6nt.

[6] Tao Song, Man Luo, Xiaolong Zhang, et al. A multiagentdriven robotic AI chemist enabling autonomous chemical research on demand. Journal of the American Chemical Society, 147(15):12534–12545, 2025. doi: 10.1021/jacs. 4c17738.

[18] Tianshi Zheng, Kelvin Kiu-Wai Tam, Newt Hue-Nam K. Nguyen, et al. NewtonBench: Benchmarking generalizable scientific law discovery in llm agents. In International Conference on Learning Representations, 2026.

[7] Gihan Panapitiya, Emily Saldanha, Heather Job, et al. AutoLabs: cognitive multi-agent systems with self-correction for autonomous chemical experimentation. Scientific Reports, 16:19554, 2026. doi: 10.1038/s41598-026-45593-z.

[19] Shreshth A. Malik, Tiarnan Doherty, Panagiotis Tigas, Muhammed Razzak, Stephen J. Roberts, Aron Walsh, and Yarin Gal. MADE: Benchmark environments for closedloop materials discovery. In International Conference on Machine Learning, 2026.

[8] Simone Pilon, Elia Savino, Oliver M. Bayley, et al. A flexible and affordable self-driving laboratory for automated reaction optimization. Nature Synthesis, 2026. doi: 10.1038/s44160-026-01053-0.

[20] Rui Li, Zixuan Hu, Wenxi Qu, et al. LabUtopia: Highfidelity simulation and hierarchical benchmark for scientific embodied agents. In Advances in Neural Information Processing Systems, Datasets and Benchmarks Track, 2025.

[9] Aikaterini Vriza, Michael H. Prince, Tao Zhou, Henry Chan, and Mathew J. Cherukara. Operating advanced scientific instruments with AI agents that learn on the job. npj Computational Materials, 12:160, 2026. doi: 10.1038/s41524-026-02005-0.

[21] Yuhan Wu, Zhao Jin, Tao Li, et al. Labimus: A simulation and benchmark for humanoid dexterous manipulation in chemical laboratory, 2026. Preprint.

[10] Kobi C. Felton, Jan G. Rittig, and Alexei A. Lapkin. Summit: Benchmarking machine learning methods for reaction optimisation. Chemistry–Methods, 1:116–122, 2021. doi: 10.1002/cmtd.202000051.

[22] Yucheng Xu, Keyi Zhang, Yuyang Yu, Min Zhang, Shiyuan Meng, Pei Chu, and Zhongying Tu. Scaling scientific discovery environments for turn-level agentic RL, 2026. Preprint, arXiv:2607.28990.

[11] Florian Häse, Matteo Aldeghi, Riley J. Hickman, Loïc M. Roch, Melodie Christensen, Elena Liles, Jason E. Hein, and Alán Aspuru-Guzik. Olympus: a benchmarking framework for noisy optimization and experiment planning. Machine Learning: Science and Technology, 2:035021, 2021. doi: 10.1088/2632-2153/abedc8. [12] Maximilian Bloor, José Torraca, Ilya Orson Sandoval, et al. PC-Gym: Benchmark environments for process control problems, 2024. Preprint, arXiv:2410.22093.

Programmable Chemical Worlds

arXiv preprint

Appendix A. Qualification and Experimental Protocols

either cumulative process time or a repeat limit would be exceeded.

A.1 Qualification and coverage protocol

A.3 Deterministic use-case protocol

The authoritative protocol was frozen before execution. It specifies component patterns, discrete axes and compatible pairs, continuous bounds, seeds, ordered workflows and pass/fail criteria. Composition qualification first checks dependencies, state ownership, unit agreement, parameter domains, resource feasibility, exposed operations, instrument availability and lifecycle closure. Invalid declarations must return structured diagnostics without constructing an environment. Discrete axes use pairwise covering rows, and continuous axes use seeded Latin hypercube samples within authored bounds. Each pattern has one or two ordered workflow targets. For every qualification unit, the report records normalized construction input, compiler diagnostics, public contract, submitted actions, transaction status, stateintegrity checks, events, resource preflight and outcome, public observations, termination and exact replay. Missing receipts, non-finite quantities, unexpected transaction outcomes, undeclared private-field exposure, denominator drift or replay mismatch fail the unit. Process-module qualification assigns reaction, thermal, phase, separation, crystallization, distillation, continuous flow and electrochemistry four probe classes each: zeroinput bounded runtime, valid low/high input, directionality and runtime invariants. Numerical fixtures carry explicit tolerances; declared conceptual or synthetic oracles are used where no independent numerical fixture is claimed. Observation modules are checked separately for instrument availability, sample accounting, bounded signals and the public/private boundary. Invalid compiler probes cover missing dependencies, conflicting state ownership, unit mismatch, invalid parameter domains, lifecycle holes and resource impossibility.

The cases, seeds, submitted-action sequences and terminal requirements in Table 5 were fixed before execution. Every case requires explicit termination and final assay. The failure–recovery case additionally fixes a runtimeprecondition failure at its first submitted action and requires continuation from the restored committed state. Table 5. Protocol-frozen deterministic use cases. Submitted actions include termination and final assay. Seed / actions

Planned failure

U01 U02

reaction-to-crystallization composed-equilibriumcharacterization-demo composed-reactionpurification-demo

0 / 12 0/5

none none

0 / 19

flow-reaction-optimization electrochemical-conversion reaction-to-distillation

0/8 0 / 11 0 / 12

step-1 runtime rollback none none none

partition-discovery reaction-to-crystallization

0 / 10 1 / 12

none none

U06-flow U06-electro U06distillation U06-partition U06crystallization

A.4 Controlled-fork acceptance criteria For an aligned checkpoint value p in the parent and c in the child, the acceptance oracle records

δ = c − p,

∆ = |δ|,

∆ , max(|p|, |c|, s0 )

where s0 is the registered positive scale floor. A fork passes only if the registered direction, absolute threshold, relative threshold, public-contract invariance, single-target lineage, same-sequence executability and exact-replay criteria are all satisfied. A.5 Public-boundary and exact-replay checks Public-boundary checks scan agent-facing task cards, observations and histories for undeclared evaluator-owned fields, private identifiers and absolute private paths. Inferential information obtained through task-declared measurements is not counted as direct exposure. Replay reconstructs the bound world from its normalized contract, runtime, mechanism, observation and scoring identities together with recorded seeds and interventions, then resubmits every typed action in recorded order. Field-level comparison covers rewards, public observations, termination and truncation flags, operation types, transaction status, rollback reason, affected resource-ledger fields, world events, state-delta summaries and state-integrity checks. The replay result is reconciled with resource ledgers and rollback receipts; exact replay requires zero numerical tolerance and no field mismatch.

Process-time limits are derived by component pattern: tmax = trequired + treserve + trepeat . Table 4. Protocol-frozen process-time limits.

Phase observation Reaction–thermal observation Phase separation Reaction crystallization Reaction distillation Continuous flow Electrochemistry Reaction–phase separation

Public identity

U03/E01

A.2 Process-time limits

Pattern

Case

Limit (s) 0 3,600 1,860 11,100 10,440 7,200 5,400 7,500

The required term sums the upper bounds of necessary timed stages, the reserve covers authored quench and transfer operations, and the repeat term follows explicit per-operation repeat limits. Preflight rejects an action if 12

Programmable Chemical Worlds

arXiv preprint

Appendix B. World Registry and Component Library B.1 Reference-task registry Table 6. Complete reference-task registry. Seed counts define the scenario identities exercised for each task. Reference task identity

Component topology

electrochemical-conversion equilibrium-characterization flow-reaction-optimization

reaction + electrochemistry + observation reaction + thermal + observation reaction + continuous flow + observation

5 5 5

low-budget-characterization