Environments
Environments for skilled work, in every industry.
Each environment is a working replica of real software — interface, backend and data — with graded tasks and the verifiers that score them.
Example task
Void the duplicate vendor payment.
Verifier checks
Ledger balances and bank statement agree to the cent.
Bookkeeping · v0
Our first environment, by the numbers.
Built, tested and measured before we ever pitched it.
- graded tasks
- 30
- Easy to frontier‑breaking
- companies
- 3
- A salon, a contractor, a restaurant
- months of data each
- 12
- Messy on purpose: duplicates, near‑matches
- verifier tests
- 196
- Every checker tested against shortcuts
- bytes of drift
- 0
- Same seed, same world, every reset
What you get
Delivered ready to train on.
- Environment
- A containerized replica you run with one command.
- Tasks
- Hundreds of graded tasks with fixed starting states.
- Verifiers
- Code that checks the end state, tested against shortcuts.
- Report
- Pass rates across frontier models, by task and difficulty.
delivery
seeming-bookkeeping/├── docker-compose.yml├── app/├── seeds/ # deterministic company data├── tasks/tasks.json # graded tasks├── verifiers/ # end-state checks + tests├── runner/ # agent harness└── report/ # pass rates by model$ docker compose up
Name the software. We’ll rebuild it.
Any web software, realistic synthetic data, tasks calibrated to your agents.
Get in touch