For frontier-lab research teams in the United States and European Union · edition November 2026
Browse data for model training and evaluation, alongside environments for chemistry, materials and iterative research. Our tiered chemistry and materials environments are packaged for OpenEnv. Every order is scoped with you before delivery.
Data
Training and evaluation packs spanning trajectories, expert tasks and corpora. Corpora include annotated and licensed assets for training or building new tasks. Delivery formats, usage rights and any exclusivity are agreed per pack.
DAT-T
Trajectory packs
complete records of a model running a GreenDynamics environment, successes and failures together; Live packs include raw instrument data
Trajectories
DAT-T2
Financial modelling and quantitative research
Quantitative research tasks and full agent trajectories using market data. Examples include repairing backtests, developing predictive factors and forecasting asset returns. Results are evaluated on held-out time periods using metrics such as prediction accuracy and backtested returns.
Trajectories
DAT-E1
Expert evaluation sets
scientist-written, double-reviewed questions with a held-out set we keep
Expert
DAT-E2
Expert rubric sets
step-level scoring rubrics for chemistry procedures, for judges and reward models
Expert
DAT-E3
Expert theory corpus
worked problems, derivations and mechanism explanations for training
Expert
DAT-E5
Long-horizon scientific research tasks
Runnable research workspaces with papers, data, code and notes. Agents reproduce findings, test a claim or investigate a discrepancy. Automated checks assess key findings, intermediate results and final deliverables against expert reference outcomes.
Expert
DAT-E6
Professional workflow tasks
Expert-authored assignments with source files, written policies and simulated business tools. Agents produce a finished deliverable or complete a multi-step workflow. Evaluation uses expert rubrics for deliverables and automated checks of final system records.
Expert
DAT-E7
Multimodal coding tasks
Task briefs and inputs paired with runnable source code and editable outputs such as slides, documents, web interfaces and scientific charts. Automated checks assess the rendered output, structure and data binding. Experts review visual quality against the brief.
Expert
DAT-E8
Expert reasoning tasks
Research-level problems, primarily in physics. Tasks include standalone derivations and problems that use symbolic solvers, numerical methods or code. Each task includes a reference answer and a verifier or grading rubric.
Expert
DAT-C1
Structure–text pairs
crystal structures matched to the synthesis and property text that describes them
Corpus
DAT-C2
Synthesis procedures
normalised step-by-step procedures with precursors and outcomes
Corpus
DAT-C3
Property records
measured properties with units, conditions and source
Corpus
DAT-C4
Evidence-grounded QA
question–answer pairs tied to their supporting source passage
Corpus
DAT-C5
Characterisation library
diffraction patterns and spectra paired with structures, measured and computed marked separately
Corpus
DAT-C6
Music performance data
Performance recordings paired with musical scores and beat- and note-level alignment annotations corrected by music researchers.
Corpus
DAT-C7
Architecture and interior design assets
Dimensioned residential floor plans, editable three-dimensional scenes and multi-view renders, with structured requirements for layout, budget and furniture.
Corpus
DAT-C8
Licensed film, television and short dramas
Feature films, television series and short dramas delivered as video files with available metadata. Permitted uses, including training and evaluation, are confirmed for each order.
Corpus
How to read the tiers
The tiers below apply to our chemistry and materials environments. Moving right, more of the real lab is in the loop.
ComputationalSimulation onlyscored by computational chemistry tools
EchoOutcome twinfitted to real measurements; fast screening
TraceProcess twinevery step checked for safety and feasibility
PulseGrounded twinreal instrument noise and drift in the loop
LiveReal laboratoryhosted sessions on our instruments
PP-01 · start here
Pilot Pack 01 — Fundamental experiments
Can your agent run a basic lab experiment and correctly read noisy evidence from it? Three fundamental experiments, each tested two ways: control (reach a target within a budget) and interpretation (read recorded evidence and answer).
Colour mixing · acid–base titration and pH · concentration by absorbance
About 150 double-reviewed interpretation items with raw evidence
Control environments at Echo and Pulse, plus a Live episode sample
Variants for noise, drift, missing data, new instruments and failure diagnosis
An independent holdout we keep for evaluation
Acceptance criteria agreed with you before delivery
Building now · pilot partners open
Environments — Fundamentals
Short experiments with fast feedback: can the agent operate an instrument loop, respect limits and notice drift.
ENV-S1
Colour mixing
reach a target colour from dyed stocks in the fewest attempts
EchoPulseLive Available
ENV-S2
Acid–base titration and pH control
reach and hold a target pH; read equivalence points and unknowns
EchoPulseLive Available
ENV-S3
Viscosity targeting
blend to a target viscosity at a set temperature
EchoPulseLive On request
Environments — Characterisation
The agent reads measured evidence and draws the conclusion a scientist would.
ENV-V1
Diffraction analysis
identify phases and impurities from diffraction patterns
ComputationalPulse On request
ENV-V2
Optical spectra analysis
derive properties and defect signatures from UV-vis, PL and FTIR
ComputationalPulse On request
ENV-V3
Microscopy grading
grade morphology and coverage from image stacks
Pulse On request
Environments — Synthesis
The agent plans and runs a synthesis toward a target property, then measures the result.
ENV-R1
Nanocrystal synthesis
reach a target emission colour
EchoTracePulseLive On request
ENV-R2
Thin-film coating
reach optical and electrical targets for photovoltaic layers
EchoTracePulseLive On request
ENV-R3
Thermoset formulation
reach cure and hardness targets for coatings and composites
EchoTracePulse On request
Environments — Recursive Self-Improvement (RSI)
The agent starts with a working baseline and iterates to improve it. Each task includes a runnable environment and a verifier.
ENV-RSI1
Code and algorithm improvement
Improve a baseline such as a query optimiser or a model training loop. The verifier reruns submitted code from scratch on unseen data.
On request
ENV-RSI2
Closed-loop wet-lab research
Propose the next experiment and commit to a prediction. Lab staff run it on real instruments, then return the measurements for the next iteration. Scoring uses the measured outcomes within a fixed experiment budget.
On request
Services and add-ons
Choose applicable services when scoping an environment order.
SVC-01
Model evaluation run
we run your model across a task set in twins and the lab and report results; you keep the trajectories
Available
SVC-02
Safety and drift suite
scenarios that score constraint adherence and detection of instrument drift
Building
SVC-03
Trustworthiness suite for lab agents
adversarial instructions, honesty on failed measurements, hazard refusal, confidentiality
Coming
SVC-04
Commissioned task sets
tasks built to your specification, exclusive to you, with a holdout we keep
Available
SVC-05
Digital twin of your lab
we build a twin from your historical data so you can screen agents before using hardware
On request
Environment and data pack deliverables
We agree the deliverables for your selected environment or data pack when scoping the order.
Documentation
Environment Card for each environment: goal, interface, scoring, valid range, robustness report
Data Sheet for each data pack: contents, schema, provenance, validation
Reference results for at least three models where included in the evaluation scope
Rights and licence documentation
Delivery
OpenEnv-compatible Docker images for packaged environments; run on your own cluster, no outbound network required