Three products. One standard of proof.
Every row we ship carries the task it came from, the practitioner who authored it, the grader that checked it, and the verdict that grader returned. If any of those four is missing, it isn't finished.
Expert demonstrations
Not labels. Work product — the thing a professional would actually hand to a colleague, plus the reasoning that got them there.
- What it is
- Supervised trajectories authored by credentialed practitioners in the tools they use professionally. A financial model arrives as a workbook with live formulas. A patch arrives as a diff against a real repository at a real commit.
- Formats
- SFT demonstrations · preference and comparison sets with written justification for the ordering · rubric-scored reasoning traces capturing intermediate steps and discarded approaches · adversarial rewrites, where an expert is shown a model's output and asked to break it.
- What makes it different
- Most datasets record the destination. We record the route. The intermediate step a junior analyst gets wrong is worth more for training than the final number, and it is exactly what never gets published.
- How it's checked
- Programmatic verification wherever the domain allows it, plus a second independent expert on every submission. Disagreements escalate rather than average out.
Verifiable RL tasks
An environment is only useful if the reward is honest. We build the reward function before the task exists, and we test it against expert work before a model ever sees it.
- What it is
- Containerised, executable tasks with real tools — a shell, a repository, a spreadsheet engine, a filings corpus, an internal API surface. The model acts; the environment records the trajectory; the grader returns a score.
- Reward design
- Deterministic where possible: tests that pass, numbers that reconcile, proofs that check. Rubric-based with partial credit where determinism isn't available — and never a bare model-as-judge without a human calibration set behind it.
- Reliability
- Every grader is run across seed sweeps to surface flakiness before delivery. A task whose verdict changes between identical runs is a bug in our work, not noise in yours.
- Horizon
- Short single-tool tasks through multi-hour, multi-tool workflows with dependencies, dead ends, and recoverable failure states.
Private benchmarks
A public benchmark tells you how well a model memorised a public benchmark. We build the other kind.
- Contamination resistance
- Built from work product that has never been on the public internet, held in a registry that records exactly which items have been exposed to which parties and when. Exposure is tracked, not assumed away.
- Human baselines
- Every benchmark ships with scores from practitioners at defined seniority levels on the same items — so a model number means something in professional terms, not just relative to other models.
- Versioning
- Immutable, versioned suites. A score is reproducible against a specific suite version and a specific grader version, both recorded.
- Bespoke targets
- If you're chasing a specific capability, we build the eval for that capability rather than adapting one that nearly fits.
Your schema, your infrastructure
We adapt to your pipeline. You should not have to write an ingest shim for a vendor.
- Formats
- JSONL, Parquet, or a schema you define. Environments ship as container images with a task manifest and a grader entry point.
- Transport
- Delivered into your own S3 or GCS bucket, a private registry, or a Hugging Face private repo. We do not require you to pull from us.
- Provenance
- Every record carries a manifest: task spec ID, author credential class, grader version, verdict, review history, and timestamps.
- Rights
- Contributors assign rights in their work product before they are paid. You receive a clean chain of title, documented.
What we don't do
Saying this out loud saves everyone a discovery call.
We don't scrape. Every record is commissioned and rights-cleared. If a dataset can be scraped, you don't need us for it.
We don't ship synthetic data as expert data. Models are used inside our pipeline for drafting and triage. They are never the author of record, and anything model-generated is labelled as such in the manifest.
We don't use general crowd labour. The whole thesis is that the credential is the product. A task that a non-expert could complete is a task that isn't worth commissioning.
We don't resell. Work commissioned for you is yours. It does not turn up in a general corpus six months later.
Bring us a capability gap
Send the failure mode you're seeing. We'll tell you whether data fixes it — and say so if it doesn't.