Data Atlas

Knowledge work, operational and World Model data, captured where it happens.

Enterprise data

Every one of these businesses also runs the same machine underneath — scheduling, procurement, documentation, billing, compliance. That layer owes nothing to cement or cardiology; it is the work Enterprise-Bench measures, and it is here in eight industry variants. Each card lists it.

Textiles

Fabric Line

01

Cotton in one end, finished bedsheets out the other. Every step in house.

34,616 staff~US$835M/yr13 factories
Volume Per roll · per lot

Inspection, grading and lot disposition recur through every production run.

Counts on request
Data extracted
Inspection records carry the defect class, its severity and the disposition taken, with process parameters running alongside through spinning, weaving, processing and stitching. Achieved colour is measured against specification and paired with the recipe correction that followed. Order-to-machine allocation is captured with every re-plan made against it — and with it the reasoning a grader applies to a borderline lot, or a planner applies when a delivery date moves.
Enterprise operations
Order intake and costing, fibre and dyestuff procurement, production planning across thirteen sites, export documentation and buyer correspondence.

Automotive & Electronics

Auto Assembly

02

Kits in, cars out. Welded, painted, built and tested — plus the harnesses, seats and phones.

1,760 staff~US$702M/yr1 facility
Volume Per unit · per test

Every vehicle and device passes end-of-line test; every component lot is accepted or rejected.

Counts on request
Data extracted
Every end-of-line, continuity and functional test result is paired with the fault found and the repair made — across vehicles, wire harnesses, mouldings and mobile devices. Assembly defect records carry their disposition, component lots their acceptance result against specification. Behind those sit the diagnostic steps a technician takes from symptom to cause, and how acceptance limits get applied when the line is behind.
Enterprise operations
Knock-down kit ordering and clearance, supplier scheduling, dealer allocation and warranty administration.

Cement

Cement Kiln

03

Rock into cement at 1,450 °C. Two kilns that never stop.

6,987 staff~US$493M/yr2 plants
Volume Continuous · 24/7

Control moves at shift cadence, lab chemistry per sample, alarms as they come.

Counts on request
Data extracted
Kiln operating conditions are captured with the control move taken against them, raw-material and raw-meal chemistry with the feed change that followed, and clinker results with whatever upstream correction they triggered. Grinding conditions and additive ratios are recorded per cement specification, equipment alarms with the action chosen. The judgement layer is which drift warrants which response, and which alarms an experienced operator tolerates.
Enterprise operations
Fuel and raw-material contracting, dispatch and haulage scheduling, maintenance planning and spares procurement.

Chemicals & Pharmaceuticals

Pharma Batch

04

Soda ash to oncology tablets. Seven plants, and a regulator reading every batch record.

2,424 staff~US$407M/yr7 plantsGMP regulated
Volume Per batch · per QC result

Every batch carries a release decision; every out-of-spec result carries an investigation.

Counts on request
Data extracted
Batch manufacturing records carry in-process measurements, QC and laboratory results carry their disposition against specification, deviations carry severity and corrective action, and incoming raw-material and API lots carry their acceptance result. Written alongside, to a regulated standard, is the reasoning: why a batch was released or held, and whether a failed result was a testing error, a process deviation or a genuine product failure.
Enterprise operations
Regulatory documentation and change control, API sourcing, distributor and tender allocation, and the paperwork that follows every process change.

Energy

Power Block

05

660 megawatts, around the clock. Coal in, grid out, seawater carrying the heat away.

1,156 staff~US$313M/yr6 plants
Volume Continuous · 24/7

Control moves and alarms arrive constantly; chemistry and condition monitoring on a fixed cycle.

Counts on request
Data extracted
Combustion conditions come paired with the fuel and airflow trim applied, steam and turbine conditions with the control change made, and plant alarms with the prioritisation and response chosen. Condition-monitoring measurements carry the serviceability call; water and steam chemistry carries the treatment or dosing that followed. The expertise sits in how output, heat rate and wear get traded against each other, and which alarm in a cascade is the real one.
Enterprise operations
Coal contracting and fuel logistics, generation scheduling and settlement, outage planning and spares.

Healthcare

Hospital Floor

06

Cardiac and kidney hospitals. Triage to discharge, every episode documented.

1,578 staff~1.4M patients/yr3 facilities
Volume ~1.4M patients / yr

Each episode generates triage, orders, results to interpret and a disposition.

Data extracted
ECG studies come with the diagnostic classification recorded against them, laboratory panels with the escalation decision taken, imaging studies with their findings, medication orders with the pharmacist's review outcome, and triage inputs — symptoms, vitals, history — with the urgency and care pathway assigned. The clinical reasoning that led to each finding is the part that never reaches the discharge summary.
Enterprise operations
Admissions and scheduling, theatre and bed allocation, billing and insurance, pharmacy stock and consumables procurement.

Commercial Real Estate

Facility Ops

07

3.4 million square feet that can never go dark. 200+ tenants, every system running.

951 staff~US$18M/yr3.4M sq ft
Volume Per ticket · per alarm

Tickets and alarms arrive all day across 200+ tenancies, each triaged, routed and closed.

Counts on request
Data extracted
Tickets are captured end to end — symptom, system identified, priority, routing and resolution. BMS alarms carry the intervention that followed, planned inspection rounds the maintain-or-defer call, security events their severity and response, and each job the parts and contractor spend against it. The judgement is reading a tenant's symptom back to the system actually at fault, and ranking that tenant against a safety risk and a budget.
Enterprise operations
Lease administration and tenant billing, contractor procurement, service-level reporting and incident escalation.

Commercial driving · North America

Hard Miles

08

Five million hard moments on North American roads. 14–16 s a clip, every frame instrumented.

20,834 h video6 vehicle classesJan–Aug 2026
Volume 5,000,047 labelled
Data extracted
Each clip carries driver action straight off the vehicle bus — brake and turn-signal state, engine RPM and throttle on the OBD-II / J1939 subset — with per-frame velocity, UTC to the millisecond and a GNSS-or-bus source flag. The world-model layer sits alongside it: road-facing video at 10–20 fps, tracked objects with persistent IDs, 3-D boxes and range-rate, lane-boundary point sets, an ego-path polygon and a metric bird's-eye view 0–60 m by ±15 m. Every label carries the human adjudication behind it — which trigger fired, the road surface, whether the stack degraded.
Enterprise operations
Fleet administration, device provisioning across two hardware generations, and the review queue that turns a raw trigger into a labelled event.

Benchmarks

Coding benchmark

SWE-Epics

09

A thousand engineering epics, cut from the real history of large codebases. Not tickets — epics.

1,000 tasksGo · C++ · Python · TypeScriptVerifier-backed
Volume 1,000 tasks

Median: 6 atomic targets, 23.5 files changed, 679-file repository. 95% at least module-wide.

What a task is
Each task is multi-PR-derived: we curate a cohesive set of real changes out of a repository's longitudinal history and compose them into one statement of work. The codebase is mature, with existing conventions, consumers and constraints a change has to respect, and the work spans features, APIs, behaviour, bug fixes and components rather than a single localised patch. Domains run across databases and data systems, dev tooling, systems and runtimes, networking, security and web infra. Every atomic target is verifier-backed — a solution is executed against the targets, not scored for similarity to a reference diff.

Enterprise benchmark

Enterprise-Bench

10

Five hundred tasks against a synthetic enterprise. Tools, not an answer key.

500 tasks519,902 documentsMCP tool surfaceDeterministic + judge
Volume 500 tasks

A well-formed task lands on tens to low hundreds of targeted reads.

What a task is
The corpus is a standing synthetic enterprise — systems of record, documents and correspondence — reachable only through MCP servers. Nothing is handed over and nothing can be scraped: the model has to narrow 519,902 documents down to the tens or low hundreds that bear on the question, then exercise judgement on what it found. The query is the index into the corpus. Query narrows, documents resolve, judgement decides.

A task qualifies only if it survives four tests — it cannot be answered by query alone, by the documents alone, without judgement, or from a subset that was given rather than derived. That rules out the two easy failure modes: decorative documents, where retrieval already contains the answer and the corpus is set dressing, and brute force, which measures budget rather than competence. Grading is deterministic criteria plus judge.

Request more information.

Sources, record counts, schemas and sample data — on request.

Request access →