RL environments built from real enterprise work

The General Data Company builds reinforcement-learning environments for AI labs from real, consented enterprise workflows: task specifications grounded in genuine operational records, realistic working state, and programmatic verifiers that score every attempt. The rights are cleared at the source. The material is held out from the open web and has never been published.

Ships as

Task specification, working state, programmatic verifier, held-out split

Rights

Written chain of title from the originating estate to your license

Exclusivity

Available at task-collection level, withheld from every other buyer

Why labs buy environments, not just corpora

Frontier models are increasingly trained with reinforcement learning on verifiable tasks. A static document corpus tells a model what work looks like; an RL environment lets a model attempt the work and be graded on the outcome. That grading signal (a programmatic verifier that decides whether the attempt succeeded) is what makes an environment usable for training, not just evaluation.

The scarce input is no longer text. It is realistic, verifiable tasks drawn from work that actually happened, and that never appeared on the open web, so nothing about the task leaks into pre-training data.

What a TGDC environment contains

• A task specification derived from a real workflow: the goal, the starting state, and the materials an operator actually had.
• Realistic working state assembled from rights-cleared operational records (correspondence, contracts, ledgers, forms), anonymized and verified per batch.
• A programmatic verifier that scores each attempt against the ground truth of how the work resolves, so the environment yields a usable reward signal.
• A held-out evaluation split so labs can measure progress on tasks the model has never seen in any form.

Where the tasks come from

Every environment is grounded in TGDC's three sourcing channels: insolvency estates, which administrators release only under a written mandate, workflow capture recorded with consent while the work is happening, and personal documents contributed through our app with on-device anonymization. Nothing is scraped. Because the underlying records were never published, the resulting tasks are genuinely unseen by every model in existence.

The gap we focus on is expert work inside real operations. These are the reasoning-heavy back-office, legal, financial and administrative workflows that public web data cannot teach, because that work was never on the web.

Licensing and exclusivity

Environments are licensed per task collection with written terms. Exclusive licensing is available at the task-collection level: an exclusive collection is delivered to one customer and withheld from every other. Each delivery ships with provenance documentation and per-batch anonymization reports, so procurement and legal teams can trace exactly where every artifact came from and what rights attach to it.