Data & Environment — what we are building Internal introduction · v2.1 · 3 Nov 2026 · Data & Environment product overview · Infra project management: ByteInfra

Internal — do not forward

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).

Scope: this page explains the Data & Environment business and products. Cross-business infrastructure planning and repository boundaries are maintained in ByteInfra. Development and execution belong to the corresponding implementation repositories.

1. Three kinds, by where the feedback comes from

exists now generate now, before any lab after the centralised lab none at generation consistency with sources simulation tools only DFT · MD · simulated spectra · ML real instruments some level of real measurement scores scores CORPUS static, structured knowledge F0–F2 · cheapest · copyable COMPUTATIONAL tool sandboxes, hidden answers F2–F3c · volume · series name TBD PHYSICAL Echo → Trace → Pulse → Live F3–F5 · cannot be rebuilt licensed datasets: structure–text, procedures, properties, QA, characterisation library Docker task packs + trajectory packs Docker task packs (Echo/Trace/Pulse) hosted sessions (Live) + trajectory packs (most valuable) EXPERT DATA rubrics · eval items · traces, from the university network writes the rubrics both verifiers use · also sold on its own
Read top to bottom: the feedback source decides the kind; the kind decides what we can sell. Only Computational and Physical produce trajectories. Expert data writes the rubrics every verifier uses, and is sold on its own. Fidelity levels F0–F5 are defined in section 1b.

1b. The fidelity ladder: F0 to F5

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.

F0 F1 F2 F3a F3b F3c F4 F5 text humanexpert programcheck MLsurrogate physicsapproximation firstprinciples fitted toreal lab data realmeasurement closer to reality · harder to rebuild · higher price lab required → Example check matches source graded rubric RDKit, units MLIP energy xTB, classical MD DFT and above Echo, Pulse Live Cost per check ~ 0 expert minutes milliseconds ms–s, GPU seconds–minutes CPU-hours fast, after lab data hours of lab Buyer can rebuild? yes with a network yes yes, open models yes yes, with compute no no Corpus structure–text, procedures, QA, properties Expert data eval items, rubrics, traces Computational tool sandboxes (F2) → surrogate, xTB/MD, DFT environments (F3a–F3c) Physical Trace Echo, Pulse Live
Read a column top to bottom: what checks the model, what one check costs, whether a buyer could build it themselves, and which of our products sit there. The break is between F3c and F4: everything to its left can be rebuilt with open tools and compute; F4 and F5 need a lab. Trace is a Physical product scored by simulation, so it sits at F3b–F3c.

Who already does Computational, and what it tells us

OrganisationWhat they offerLevelCommercial modelLesson for us
Meta FAIR — OMol25100M+ DFT calculations (ωB97M-V/def2-TZVPD), 83 elements, up to 350 atoms; billions of CPU core-hoursF3creleased openly with models and evaluationsraw DFT data at scale is given away by a frontier lab; we cannot sell DFT datasets as such
Materials Project, OQMD, AFLOW, OMat24, GNoMEmillions of computed crystal structures and propertiesF3cfree public databasessame: computed properties are a commodity input, not a product
Orbital Materials — Orbfast ML interatomic potentialF3aopen model; company sells designed materialssurrogates are open and good enough to build F3a verifiers; ours must add tasks and anti-hacking, not the model
NVIDIA ALCHEMI ToolkitGPU building blocks for MLIP and MD workflows (MACE, AIMNet2, batched MD)F3a–F3bpublic library, drives GPU usethe tooling layer is free; buyers can stand up simulators themselves
Rowan Scientificcloud computational chemistry: AIMNet2, GFN2-xTB, several DFT functionals, Python APIF3a–F3c$200/month per user, credits; academic one-thirdpeople 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 benchesF3bopen sourcesimulated 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, similarityF2paperthe common reward today is a cheminformatics check, the bottom of our ladder
What this means for the Computational kind. The data and the simulators are free; nobody sells them as verified RL environments. So what we sell at F3 is the task, the verifier and its anti-hacking report, not the calculations. Price it as a volume product, and use it to bring buyers up the ladder to F4–F5, where nothing open competes.

1c. Anatomy: what is inside an environment, and what counts as data

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.

Agent buyer's model ENVIRONMENT = task set + interface + world + verifier + card Task set prompts, targets, constraints · tagged recall / generalisation Interface reset / step / state tool list (OpenEnv) World sandbox · simulator · twin · real lab (the part that differs by kind) Verifier Rubric criteria + weights Checker program · simulation twin · instrument → reward (and weight-0 metrics) qualifiers — how far to trust the reward noisemodel coveragemap anti-hackingset + baselines: 3 models × 8 runs, human reference action obs target obs reward per step / episode DATA IN expert rubrics · measurement sets behind twins · calibration logs DATA OUT trajectories · scores and labels · drift and safety flags measurements → world · rubrics → verifier kept — never ships expert-authored / data checker internals never ship
Only the world changes between kinds: a tool sandbox (Computational), a regression or simulator or grounded twin (Echo, Trace, Pulse), or the lab itself (Live). The rubric is written once per task family and reused across every tier; the checker is what gets swapped. A Live trajectory is the one data product whose checker a buyer cannot rerun, which is why it prices highest.

What each kind contains

ComponentCorpusComputationalEchoTracePulseLive
Environment?noyesyesyesyesyes, hosted
Task setQA items with reference answershidden-answer tasks over tool outputs; compute budgettarget property, allowed rangestarget + action grammar + fixed protocolas Traceas Trace
Interface / toolsnone (files)OpenEnv; structure DBs, cheminformatics, MLIP, xTB, DFT runner, workspaceOpenEnv; propose_recipe, query_historyOpenEnv; dispense, mix, heat, coat, measure, abortas Traceas Trace, routed to instruments
Worldnonethe tool sandbox; results computed or cachedregression fitted to real outcomessimulated procedure with instrument modelstep models fitted to real runs; real noise, drift scenariosequipment, materials, instruments
Rubricanswer match; source-span supportcorrectness vs hidden answer; budget; tool-use validitycloseness to target; efficiencyper-step feasibility and safety; terminal closeness; efficiencyas Trace + drift detectionas Pulse
Checkerprogram (F2)program (F2) or simulation (F3a–c)twin prediction (F4)simulator (F3b–c)grounded twin (F4)instrument reading (F5)
Noise model—numerical toleranceheld-out error of the regressioninstrument-model parametersmeasured repeatability and driftthe instrument; repeated runs
Coverage mapprovenance, date rangeelement and size range of the simulatorfitted region of task spacevalidity range of the simulatorfitted region + measured drift rangematerials and instrument availability
Safety—resource limitsout-of-range refusalforbidden mixes, ranges, sequencesas Trace + instrument limitsas Pulse, enforced by interlocks and humans
Data inliterature, structure DBs, measurementsreference solutions, cached simulationmeasurement set behind the regressionexpert step-feasibility rubricsreal runs; calibration logsmaterials lots; calibration logs
Data outthe recordsitems, agent trajectories (tool calls, files), labelsitems, episode trajectories, scoresstep-level trajectories, safety flagsas Trace + drift episodesfull observation stream: images, spectra, readings, instrument state, human interventions
Expert data usedQA review; theory corpustask authoring; reference solutionstarget selectionstep-feasibility rubricsas Traceas Trace + lab-record labels for failure diagnosis

2. The Physical series: how much of the real lab is in the loop

Echooutcome twin · F4 Traceprocess twin · F3b–F3c Pulsegrounded process twin · F4 Livethe real lab · F5 Synthesis Characterisation Verifier checks Delivered as regression on realoutcomes simulated steps,feasibility rules step models fittedto real runs real equipment predicted result+ uncertainty simulated measurement,instrument model real noise, drift,calibration real instrument the result each step: safe, feasible, ordered each step againstthis lab's behaviour what actually happens Docker buy-out Docker buy-out Docker or hosted hosted sessions, exclusive windows Pulse gates access to Live cheapest entry SKU;screens agents before lab time sold until enough real runs exist,then upgraded to Pulse simulated fitted to real data real
The tiers differ only in how much of each stage is real. Moving right, a buyer pays more because less can be reproduced elsewhere. Safety and drift tests live in Pulse: it is the first tier a regression model cannot imitate. Live trajectories are the highest-value data we produce.

3. One loop, three task scopes

H · Hypothesistheory and computation S · Synthesisliquid, liquid-solid C · CharacterisationXRD, UV-vis, PL, SEM, DSC A · Analysisphase, property, cause result informs the next hypothesis Verificationagent acts in C → A · sample or data given V1–V5 · labels are the verifier · build first Realisationagent acts in S → C → A · target property given S1–S3 starter, R1–R6 research · measured property vs target, plus step checks target Discoveryagent acts in H → S → C → A · open question and budget · premium, not catalogued in v1
Every task is a slice of the same loop. The scope says which stages the agent controls; the tier (section 2) says how each stage is produced. Scope and tier are independent, so the catalogue is a grid: section 4.

4. Task families and where they sit

Filled dot = the tier we plan to sell for that family; open dot = possible later. Priority within each block is the row order.

#Task familyScopeVerified byComp.EchoTracePulseLiveStatus
S1Colour mixing to a target RGBRealisation · startercamera or spectrometerlive, multiple sites
S2pH control, reach and holdRealisation · starterpH probelive
S3Viscosity at set temperatureRealisation · starterviscometer, rheometerto set up
V1Diffraction → phase and impurityVerificationICSD or refined labelComputational first
V2Optical spectra → property, defectsVerificationlabelled setComputational first
V3Microscopy → morphology gradeVerificationexpert labelneeds images
V4Failure diagnosis, full char setVerificationlab recordneeds lab record
V5Computation vs measurementVerificationexpert labelbridge to Discovery
R1Nanocrystal synthesis → target emissionRealisationPL peak, FWHMpriority 1
R2Thin-film coating → optical/electrical targetRealisationbandgap, sheet resistance, XRDpriority 2
R3Thermoset formulation → cure, hardnessRealisationDSC, hardness, gel timepriority 3
R4Electrolyte → conductivityRealisationimpedancelater
R5Crystallisation → size or polymorphRealisationDLS, XRDlater
R6Solution catalysis → yieldRealisationHPLC, GClater

Trustworthiness suite — future SKU for safety teams

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:

PerspectiveWhat the agent must doScored by
Constraint adherencerespect forbidden mixes, ranges and sequences even under pressure to hit the targetviolation count and severity
OOD instrumentnotice readings outside the calibrated range and stop or flagdetection rate, false stops
Adversarial instructionsignore unsafe or off-task instructions embedded in protocols, lab notes or tool outputscompliance with injected instructions
Recipe confidentialitynot leak proprietary recipes or partner data into logs or other episodesleakage checks on outputs
Honesty on failurereport a failed or missing measurement rather than fill it infabrication rate vs ground truth
Hazard refusaldecline syntheses or conditions on the hazard list, with reasonrefusal 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.

5. What the buyer gets, what we keep

GreenDynamics (AU / HK entity, data outside mainland China) Frontier lab (US / EU) contract boundary KEPT — never sold verifier internals alignment algorithms data pipeline holdout sets powers environments verifier report + baselines trajectory, expert, corpus data buyer's own infrastructure RL training · inference · evals runs Docker envs locally, or calls hosted Live sessions Docker / hosted card + report data packs optional white-box API for evaluation runs NDA + destruction clause · buyer keeps trajectories
The verifier is the product's core, so it never crosses the boundary: buyers get the environment, a report on how robust the verifier is, and the data. Training and inference stay on their side.

6. Build order

Current execution scope: the diagram below is the broader product roadmap. The immediate data and environment samples follow First Sample Kit v2. Full-pack requirements and later tiers do not hold up the first sample, and roadmap labels are not evidence that a delivery is complete.
Now ~1 month Centralised lab runs Realisation proven Corpus datasets licensed Computational V1, V2 Expert data pack, one family S1, S2 starter tasks live Legal: licences, disclosure First runnable Echo sample OpenEnv package Anti-hacking report 3 models × 8 runs baselines First buyer conversation Trace, Pulse for R1–R3 Live sessions, Pulse-gated Trajectory packs (T1) V3, V4 from lab records S3 viscosity added Discovery tier (R1 or R2 with the target removed) Exclusive hosted windows Commissioned sets R4–R6 gate: sample passes its own anti-hacking test gate: equipment for R1–R3 calibrated gate: Realisation cards have baselines
Everything in the first column can start without the lab. The one artefact that unlocks sales is the runnable Echo sample with its anti-hacking report: that is what a frontier lab's RL team tests on the first call.

7. What every product ships with

Environment Card — every environment

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

Data Sheet — every data product

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

Why these four were added

  • Recall vs generalisation. Tag each item as answerable from public data the model has seen, or new. Verification packs and expert data target ≥ 50 % generalisation; it is the cheapest honest difficulty control and the one contamination check buyers trust.
  • Harness mapping. Lab eval teams think in Runner / Scenario / Adapter / Executor / Metrics. Naming where our pack sits in that stack removes a week of integration questions.
  • Self-hosted under PBS or Slurm, no phone-home. Evaluation frameworks that only worked as hosted services were dropped by HPC users; the harness that ran on their own cluster was kept. Hard requirement for every Computational, Echo, Trace and Pulse SKU.
  • Full-pack cost. Per-instance footprint is not enough; the buyer's infra team asks "how many GPU-hours for one complete run". State it.

Lineage worth citing

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.

provenance_type on every record and step: expert explanation · agent-generated reasoning · instrument observation · simulation output · human intervention. A written explanation is never recorded as the trace of how a decision was made.
Per-step trajectory record: timestamp and step index · provenance_type of each field · agent message and tool call · raw observation (images, spectra, readings with units) · instrument state and drift flag · verifier output and version · safety events · human intervention and reason · run metadata. Raw files plus a JSONL index; nothing summarised at capture.

Four worked samples (in the Skeleton doc)

SampleKind / tierScopeWhat it shows
Echo-NanocrystalEmissionPhysical · Echo · F4Realisationcheapest 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-ThinFilmCoatingPhysical · Pulse · F4Realisationstep-by-step tools; per-step safety and feasibility; hidden drift episodes that reward agents who notice
T1-Live-ThinFilmCoatingTrajectory pack · F5Realisation50 real episodes, success and failure; full observation stream; 10 % re-read by a scientist
D-Expert-ThinFilmCoatingExpert data · F1–F2—eval items, procedure rubrics, reasoning traces; two reviewers; filtered against frontier models; can start before the lab

8. Commercial rules

RuleDecision
BuyersUS and EU frontier labs and large model companies only. No mainland China buyers, no small labs.
Entity and dataAU or HK entity contracts; all data produced and held outside mainland China.
InfrastructureBuyer runs training and inference; we supply environments, tooling and hardware access.
ExclusivityNon-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.
PricingRanges 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.
AcceptanceAgreed 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.
TransparencyToken use, tool-call log, per-episode trace, physical conditions, verifier version with every run.
Rights provenanceOur own production, or bought under a licence that is perpetual, worldwide, sublicensable, covers AI training and resale of derivatives.

Default first targets (replace with real names)

#OrganisationTeamWhy first
1Google DeepMindscience, RL environmentslargest stated appetite for environments
2Metapost-training environmentsbuys at scale; backs OpenEnv
3Microsoft Research AI for Sciencematerials, lab automationneeds physical verification it does not own
4Mistral or another EU labpost-trainingEU buyer; less crowded

Anthropic and OpenAI later: hardest to enter, and worth more once a public benchmark result exists.

9. Open before any buyer sees this