Licensed datasets

Proprietary data, licensed at the source.

Clinical, operational, engineering and recorded-work data licensed directly from its owners, de-identified to an auditable standard, reviewed by practitioners, and delivered with a provenance record for every item.

record inspector · one record, five stages · illustrativeLive
// DO-MED-003 · emergency department note · source: partner health system Patient: Maria Keller MRN: HX-204-9917 DOB: 14/03/1969 Attending: Dr. J. Okafor Ward: Cardiology 4B Admitted: 02 Jun 2026 HPI: 57F presenting with 3 days of exertional chest tightness radiating to the left arm, worse on climbing stairs. Known hypertension on amlodipine. Denies syncope. ECG: sinus rhythm, 1 mm ST depression V4–V6. Troponin-I: 0.09 ng/mL → 0.31 ng/mL at 6 h. Plan: DAPT loaded, heparin infusion, cath lab booked for AM.

01 · Sources

Two ways a record enters the catalog

Either an institution licenses records it already holds, or a practitioner writes what does not yet exist. Nothing is scraped, and nothing enters the catalog without a signed licence.

licensed from institutions · signed with the ownerLicensed

Licensed from institutions

Health systems, enterprises, engineering organisations and contact centres license the records they already keep. We handle de-identification, normalisation and compliance; the owner keeps control of scope and term.

Hospital networksclinical notes · imaging · orders
Enterprise engineering orgsmonorepos · reviews · incidents
Operations deskstickets · CRM · procurement threads
Contact centres & studioscalls · egocentric video · speech

Hold data like this? License it →

authored by practitioners · to specificationCommissioned

Authored by practitioners

When the data you need has never been written, a practitioner writes it to your specification: worked differentials, pull-request verdicts with reference fixes, audit reasoning, grading rubrics. Each item records the reasoning, not only the answer.

MD
Consultant cardiologistdifferentials · guideline alignment
Medical
SWE
Staff engineer, paymentsverdicts · reference fixes · tests
Code
CA
Chartered accountantassertions · evidence chains
Finance
LLB
Commercial counselclause risk · redlines
Legal

02 · Catalog

Current and planned datasets

An illustrative slice of the catalog. Any entry can be sampled under NDA before licensing; sizes and coverage are pre-launch targets.

Code
Enterprise monorepo with review history
PRs · reviews · CI runs · 4.8k PRs · 38 contributors · Python / TS
pr, diff, review_thread, ci_result, merge_outcome
Sample →
Code
Incident post-mortems with linked commits
text + code · 620 incidents · 2019–2026
timeline, root_cause, remediation_diff, severity
Sample →
Company ops
Support ticket lifecycles
CRM exhaust · 190k tickets · 9 languages
ticket, routing, agent_actions, resolution, csat
Sample →
Company ops
Procurement email threads
text · 31k threads · 4 verticals
thread, parties(role), decision, attachments_meta
Sample →
Medical
Discharge summaries with expert differentials
clinical text · 42k notes · 11 specialties
note, differential, reasoning, labels, guideline_ref
Sample →
Medical
Radiology reads · MRI / CT
DICOM + report · 18k studies · de-identified pixels
series, findings, impression, expert_grade
Sample →
Audio
Multilingual support calls
speech · 12k hrs · English, Spanish, Hindi + 6
wav, diarisation, transcript, intent, outcome
Sample →
Audio
Earnings calls with analyst annotations
speech + text · 1.4k calls · 48 h avg
wav, transcript, extracted_kpis, reasoning
Sample →
Egocentric video
Warehouse picking, first-person
video · 2.1k hrs · 30 fps · gaze
clip, action_labels, object_tracks, task_outcome
Sample →
Egocentric video
Field maintenance walkthroughs
video + audio · 640 hrs · narrated
clip, step_labels, tool_use, narration
Sample →
Finance & Legal
Audit working papers
documents · 9k engagements · redacted
workpaper, assertion, evidence_chain, finding
Sample →
Finance & Legal
Commercial contract reviews
text · 27k contracts · 5 jurisdictions
clause, risk_flag, counsel_note, redline
Sample →

03 · Quality

Five stages, and what each one rejects

A record passes five stages before delivery. Duplicates, thin or malformed records, residual identifiers and anything a practitioner will not sign off are rejected along the way. On a typical corpus, 42% of ingested records do not ship.

01

License

Licensed from the owner, with the schema, consent basis and permitted use on file before any data is transferred.

02

Dedupe & normalise

Near-duplicates and malformed records removed; formats and units standardised.

03

De-identify & audit

Direct and quasi-identifiers removed, every redaction logged, residual risk reviewed by a person.

04

Practitioner review

Domain specialists reject anything incorrect, thin or inconsistent, and add the reasoning layer where commissioned.

05

Package & sign

Manifest, checksums and provenance attached; delivered as JSONL or Parquet.

Records remaining after each stage · illustrative corpus

Ingested
100%
After dedupe−14%
86%
After identifier audit−15%
71%
After practitioner review−8%
63%
Delivered−5%
58%
42%of ingested records are rejected. We deliver less and stand behind all of it.

04 · Licensing

Three licence structures

Scope, exclusivity and term are set by the institution that owns the data and enforced by the licence after delivery.

L-01

Exclusive

One licensee · fixed term

A dataset licensed or built for a single lab. No other party receives it during the term, and the source institution agrees to the exclusivity up front.

L-02

Shared

Limited licensees · scoped use

The same dataset licensed to a small number of teams, each with a defined model class and permitted use. Lower cost per record, with the same provenance record.

L-03

Commissioned

Built to your specification

You define the distribution, modality and expert layer; we source or author it. You hold the output under the terms agreed with the contributing institutions.