We sell US and EU frontier labs two things: environments where a model acts and is checked, and the data those environments produce. Everything is sorted by one question — where does the feedback come from? — and placed on one loop: Hypothesis → Synthesis → Characterisation → Analysis. First physical scope: liquid and liquid-solid chemistry, catalogued by application (coatings, displays, energy storage, photovoltaics).
Fidelity answers one question: how close to physical reality is the thing that checks the model? Higher means harder for a buyer to rebuild, and so more valuable. Simulation is split into three levels because a neural surrogate, a semi-empirical calculation and DFT differ by orders of magnitude in cost and in how well they hold up.
| Organisation | What they offer | Level | Commercial model | Lesson for us |
|---|---|---|---|---|
| Meta FAIR — OMol25 | 100M+ DFT calculations (ωB97M-V/def2-TZVPD), 83 elements, up to 350 atoms; billions of CPU core-hours | F3c | released openly with models and evaluations | raw DFT data at scale is given away by a frontier lab; we cannot sell DFT datasets as such |
| Materials Project, OQMD, AFLOW, OMat24, GNoME | millions of computed crystal structures and properties | F3c | free public databases | same: computed properties are a commodity input, not a product |
| Orbital Materials — Orb | fast ML interatomic potential | F3a | open model; company sells designed materials | surrogates are open and good enough to build F3a verifiers; ours must add tasks and anti-hacking, not the model |
| NVIDIA ALCHEMI Toolkit | GPU building blocks for MLIP and MD workflows (MACE, AIMNet2, batched MD) | F3a–F3b | public library, drives GPU use | the tooling layer is free; buyers can stand up simulators themselves |
| Rowan Scientific | cloud computational chemistry: AIMNet2, GFN2-xTB, several DFT functionals, Python API | F3a–F3c | $200/month per user, credits; academic one-third | people pay for simulation as a service, but for running calculations, not for RL tasks; low price point |
| ChemGymRL (academic) | open RL framework for simulated chemistry benches | F3b | open source | simulated chemistry RL environments exist as research code; none are packaged or verified for frontier-lab RL |
| RLVR for molecular generation (academic) | RL with RDKit-computed rewards: validity, QED, logP, similarity | F2 | paper | the common reward today is a cheminformatics check, the bottom of our ladder |
An environment is something a model can act in and be scored by. It has five parts: task set, interface, world, verifier, card. A dataset with no interface is data, however large. A verifier is a scoring system, not one program: rubric and checker produce the reward; noise model, coverage map and anti-hacking set say how far that reward can be trusted.
| Component | Corpus | Computational | Echo | Trace | Pulse | Live |
|---|---|---|---|---|---|---|
| Environment? | no | yes | yes | yes | yes | yes, hosted |
| Task set | QA items with reference answers | hidden-answer tasks over tool outputs; compute budget | target property, allowed ranges | target + action grammar + fixed protocol | as Trace | as Trace |
| Interface / tools | none (files) | OpenEnv; structure DBs, cheminformatics, MLIP, xTB, DFT runner, workspace | OpenEnv; propose_recipe, query_history | OpenEnv; dispense, mix, heat, coat, measure, abort | as Trace | as Trace, routed to instruments |
| World | none | the tool sandbox; results computed or cached | regression fitted to real outcomes | simulated procedure with instrument model | step models fitted to real runs; real noise, drift scenarios | equipment, materials, instruments |
| Rubric | answer match; source-span support | correctness vs hidden answer; budget; tool-use validity | closeness to target; efficiency | per-step feasibility and safety; terminal closeness; efficiency | as Trace + drift detection | as Pulse |
| Checker | program (F2) | program (F2) or simulation (F3a–c) | twin prediction (F4) | simulator (F3b–c) | grounded twin (F4) | instrument reading (F5) |
| Noise model | — | numerical tolerance | held-out error of the regression | instrument-model parameters | measured repeatability and drift | the instrument; repeated runs |
| Coverage map | provenance, date range | element and size range of the simulator | fitted region of task space | validity range of the simulator | fitted region + measured drift range | materials and instrument availability |
| Safety | — | resource limits | out-of-range refusal | forbidden mixes, ranges, sequences | as Trace + instrument limits | as Pulse, enforced by interlocks and humans |
| Data in | literature, structure DBs, measurements | reference solutions, cached simulation | measurement set behind the regression | expert step-feasibility rubrics | real runs; calibration logs | materials lots; calibration logs |
| Data out | the records | items, agent trajectories (tool calls, files), labels | items, episode trajectories, scores | step-level trajectories, safety flags | as Trace + drift episodes | full observation stream: images, spectra, readings, instrument state, human interventions |
| Expert data used | QA review; theory corpus | task authoring; reference solutions | target selection | step-feasibility rubrics | as Trace | as Trace + lab-record labels for failure diagnosis |
Filled dot = the tier we plan to sell for that family; open dot = possible later. Priority within each block is the row order.
| # | Task family | Scope | Verified by | Comp. | Echo | Trace | Pulse | Live | Status |
|---|---|---|---|---|---|---|---|---|---|
| S1 | Colour mixing to a target RGB | Realisation · starter | camera or spectrometer | live, multiple sites | |||||
| S2 | pH control, reach and hold | Realisation · starter | pH probe | live | |||||
| S3 | Viscosity at set temperature | Realisation · starter | viscometer, rheometer | to set up | |||||
| V1 | Diffraction → phase and impurity | Verification | ICSD or refined label | Computational first | |||||
| V2 | Optical spectra → property, defects | Verification | labelled set | Computational first | |||||
| V3 | Microscopy → morphology grade | Verification | expert label | needs images | |||||
| V4 | Failure diagnosis, full char set | Verification | lab record | needs lab record | |||||
| V5 | Computation vs measurement | Verification | expert label | bridge to Discovery | |||||
| R1 | Nanocrystal synthesis → target emission | Realisation | PL peak, FWHM | priority 1 | |||||
| R2 | Thin-film coating → optical/electrical target | Realisation | bandgap, sheet resistance, XRD | priority 2 | |||||
| R3 | Thermoset formulation → cure, hardness | Realisation | DSC, hardness, gel time | priority 3 | |||||
| R4 | Electrolyte → conductivity | Realisation | impedance | later | |||||
| R5 | Crystallisation → size or polymorph | Realisation | DLS, XRD | later | |||||
| R6 | Solution catalysis → yield | Realisation | HPLC, GC | later |
The current safety suite covers physical constraints and drift. Safety teams at the same labs buy a broader set; the lab-agent version of the eight-perspective text-model trust evaluations is:
| Perspective | What the agent must do | Scored by |
|---|---|---|
| Constraint adherence | respect forbidden mixes, ranges and sequences even under pressure to hit the target | violation count and severity |
| OOD instrument | notice readings outside the calibrated range and stop or flag | detection rate, false stops |
| Adversarial instructions | ignore unsafe or off-task instructions embedded in protocols, lab notes or tool outputs | compliance with injected instructions |
| Recipe confidentiality | not leak proprietary recipes or partner data into logs or other episodes | leakage checks on outputs |
| Honesty on failure | report a failed or missing measurement rather than fill it in | fabrication rate vs ground truth |
| Hazard refusal | decline syntheses or conditions on the hazard list, with reason | refusal rate and quality |
Add-on to any Trace, Pulse or Live pack; build after the first Pulse family has baselines. Nobody offers this for lab agents today.
Families chosen for two reasons: path dependence (order and timing change the result, so process verifiers and trajectories are worth paying for) and verifier latency under two hours (several Live episodes per day). The open benchmark keeps free public versions of S1–S2; paid versions add tasks, drift scenarios and full trajectories.
Identity name, tier and fidelity, application, loop scope, version · Task goal, task space, fixed protocol, task rows, recall / generalisation split · Interface OpenEnv reset/step/state, tools, Docker, footprint, harness mapping (Scenario = task pack, Adapter = OpenEnv client, Executor = buyer sandbox or GD lab scheduler, Metrics = verifier report), self-hosted run (PBS or Slurm, no outbound network), full-pack GPU-hours · Verifier rubric, noise model, coverage statement, anti-hacking report, safety constraints · Baselines ≥3 models × 8 runs, human reference · Provenance data behind the twin, reproducibility file · Rights · Support.
Not sellable if empty
Coverage statement · anti-hacking report · baselines
Follows Datasheets for Datasets: unit of data, count and success/failure split, recall / generalisation tag per item, per-unit schema, raw files, how produced, instruments, human involvement, cleaning, labels, known issues, validation, intended and unsuitable uses, delivery · plus Rights (licence, provenance of rights, jurisdiction).
Not sellable if empty
Per-unit schema · validation · provenance of rights
The Corpus evidence-QA line descends from the generate-then-verify pipelines (Darwin-SIG, MatSci-Instruct, 2023–24): a generator model writes QA pairs from full-text papers and a separate verifier model checks accuracy and relevance. Cite the published versions, not internal decks.
| Sample | Kind / tier | Scope | What it shows |
|---|---|---|---|
| Echo-NanocrystalEmission | Physical · Echo · F4 | Realisation | cheapest entry; recipe in, predicted spectrum out; out-of-range queries end with zero reward; agent vs random, BO and warm-started BO at equal budget, success at k = 1, 3, 5, 12 |
| Pulse-ThinFilmCoating | Physical · Pulse · F4 | Realisation | step-by-step tools; per-step safety and feasibility; hidden drift episodes that reward agents who notice |
| T1-Live-ThinFilmCoating | Trajectory pack · F5 | Realisation | 50 real episodes, success and failure; full observation stream; 10 % re-read by a scientist |
| D-Expert-ThinFilmCoating | Expert data · F1–F2 | — | eval items, procedure rubrics, reasoning traces; two reviewers; filtered against frontier models; can start before the lab |
| Rule | Decision |
|---|---|
| Buyers | US and EU frontier labs and large model companies only. No mainland China buyers, no small labs. |
| Entity and data | AU or HK entity contracts; all data produced and held outside mainland China. |
| Infrastructure | Buyer runs training and inference; we supply environments, tooling and hardware access. |
| Exclusivity | Non-exclusive 1× by default · time-limited 2.5–3× (2–3 months) · perpetual only for commissioned sets, FrontierMath model: we keep a holdout and publication rights. |
| Pricing | Ranges held internally; every quote scoped per deal; no public price list. Market anchors: $200–2,000 per task · ~$10k per cross-site physical trajectory · $500–5,000 per expert eval item. |
| Acceptance | Agreed in writing before delivery; all five must pass. Label quality: two independent reviewers, adjudication, agreement ≥ [κ 0.8], critical-error rate ≤ [1 %] on a buyer-drawn audit sample · Schema completeness: every required field valid, raw evidence on every item, provenance_type on every record · Reproducibility: re-runs within the stated noise model; operational numbers met (latency distribution, concurrency, reset, failed-job rate, complete-run cost; for Live also prep, cleaning, calibration, repeats) · Leakage controls: no public or prior-delivery duplicates, search report attached, holdout never released · Usefulness: evaluation usefulness (separates models; holdout-variant scores track main set) or training uplift on a buyer-held benchmark, chosen per deal. A gap between two models or a mid-range reward is not acceptance evidence. |
| Transparency | Token use, tool-call log, per-episode trace, physical conditions, verifier version with every run. |
| Rights provenance | Our own production, or bought under a licence that is perpetual, worldwide, sublicensable, covers AI training and resale of derivatives. |
| # | Organisation | Team | Why first |
|---|---|---|---|
| 1 | Google DeepMind | science, RL environments | largest stated appetite for environments |
| 2 | Meta | post-training environments | buys at scale; backs OpenEnv |
| 3 | Microsoft Research AI for Science | materials, lab automation | needs physical verification it does not own |
| 4 | Mistral or another EU lab | post-training | EU buyer; less crowded |
Anthropic and OpenAI later: hardest to enter, and worth more once a public benchmark result exists.