What we supply, and what makes each one worth having. Exact volumes and a reviewable
sample for any line below, on request.
Audio
Multichannel conversational audioRecorded with one speaker per channel rather
than separated afterwards, so speaker attribution is exact rather than inferred by a
diarizer. The front desk healthcare set is also transcribed, with timestamped speaker
role, topic and call outcome.
Conversational audio at scalePrimarily AAVE, multichannel throughout.
Single speaker read speechAcross a range of languages.
Video
EgocentricHead mounted capture of real factory and workshop work, including
stereo pairs with synchronised IMU.
Real world videoScored clip by clip for resolution, motion and content, so you
filter to what you want before paying for egress.
Agent trajectories and code
Complete agent trajectoriesIn ATIF format. The task, the model's reasoning,
every tool call and the result that came back, and the final answer. Coding agents,
gateway traffic and web research across Claude and GPT model families.
Source codeWhole repositories rather than loose files, selectable for authored
code rather than vendored dependencies.
Documents and books
BooksAs PDFs, most also with extracted markdown and their figures pulled out as
separate images.
Scientific and technical titlesCatalogued by subject, language and page count.
Trade paperworkScanned export and shipping documents.
Automotive technical documentationService manuals, wiring diagrams and factory
quality standards, across several manufacturers and languages.
Engineering and medical
CAD and 3DMechanical assemblies with drawings and STEP exports, Revit family
libraries, and jewellery as real NURBS solids rather than display meshes.
Medical imagingCT series as DICOM with the acquisition metadata intact.
Post-training
Reinforcement learning environmentsRunning multi-service systems with the
operational layer broken or removed, where the agent has to make the platform
genuinely work. Graded by restarting from the agent's own configuration, so nothing
hand-started survives. Frontier models do not solve them.
See the environments and their pass rates
Reviewing a sample
Every line above has a packaged sample, and samples are not zip files you have to take on
trust. They go into a viewer, one page per delivery, with the files rendered so they can be
judged: audio plays with its channels split, video plays, documents show every page, CAD
shows the geometry, and agent trajectories are browsable turn by turn.