01 - DATA

Training data from behind closed doors

Data licensed from the institutions that made it, plus datasets written from scratch by experts in the top 1% of their field.

Request sample data
sample_inspector - raw → training-ready
RAW · PII DETECTED licensed at source · provenance + terms attached
SUPPLY
Two supply chains

Where our data comes from

We license data at the source, or our experts author it for SFT and RLFT.

lh2 · supply chains live
01 · Institutional datasets
Licensed at the origin. Hospitals, enterprises, and factories license the data they already generate - de-identified inside their perimeter, with provenance and terms attached to every record.
Enterprise codebasesmonorepos · PR reviews
+ def merge(pr):   if pr.approved:     return ci.ship(pr) - # TODO: review
✓PR #482 · approved
Real enterprise repos with full review history
Clinical archivesnotes · imaging · labs
DICOM · de-identified1,204 studies
De-identified inside the source perimeter
Operational exhaustworkflows · tickets
ticket opened #SR-4471
routed tier 2
resolved 4h 12m
How real companies actually run, per record
Recorded workcalls · meetings · floor video
REC00:41:12 · 9 langs
Multimodal capture from live operations
02 · Expert-built datasets
Authored and evaluated by practicing domain specialists across engineering, legal, medical, and financial sectors. These professionals curate the ground-truth data that supervised fine-tuning (SFT) imitates, and provide the precise preference signals that optimise reinforcement learning (RLFT).
Reasoning tracesstep-by-step derivations
given · hedged 1.09, settled 1.12
rule out · rate error, hedge held
weigh · gap sits in settlement leg
conclude · FX variance, not revenue
4 of 6 steps · derivationauthored
Authored from scratch when the data does not exist
Golden Datasetsrubric-graded answers
goldRevenue cut-off at year end
correctness4/4
citation3/4
reasoning2/4
✓graded vs gold · 9 / 12
Vetted specialists grade quality against a rubric
Preference datarankings · critiques
A
vs
B
ranked by 3 specialistsκ 0.86
How specialists actually think, captured as preferences
Adversarial setsedge cases · probes
probe setRelated-party disclosures
✗10consolidation · minority stake
!1 of 18 probes breaks it
The failure modes only a practitioner knows to look for
CATALOG
The catalog

Datasets, as they ship

A live view of the catalog. Pick a vertical to see the fields that matter for it. You sample any entry before you license it.

QUALITY
Quality control

Most records don't get through

Four gates sit between raw source and delivery. Here is what each one keeps, by volume.

Ingestedraw records from source
0%
Provenance verifiedconsent + license chain intact
0%
De-identification auditautomated strip + human spot audit
0%
Expert reviewdomain specialists grade quality
0%
Shippedcontamination-screened, training-ready
0%
LICENSE
Ways to work with us

Three ways to get the data

License what already exists, commission something new, or run a standing pipeline. Pick the one that matches your timeline.

001

Off-the-shelf datasets

Datasets from the catalog. You sample and evaluate them, and we deliver in weeks.

002

Custom collection

Tell us the capability gap. We find the institutions, recruit the experts, and build to your spec.

003

Ongoing data programs

A standing pipeline: continuous collection, refresh cycles, and a dedicated pod working to your training calendar.

Sample it first

Request a sample from any vertical and run it through your own pipeline.

Request sample data