- Scope
- 50 to 100 tasks in one workflow family
- First vertical
- Finance and back-office operations: close, accounts payable, reconciliations, support, procurement
- Delivery
- Docker container in Harbor task layout
- Timeline
- Scoped per workflow family; agreed in writing before work starts
Environments
Mikra delivers RL environments seeded with a real company’s operations: a world of cloned tools, tasks written by practitioners, a verifier for every task, and a calibration report, shipped as a Docker container in Harbor layout.
- World
- Accounts payable, 2019–2025 · 3 tools: ERP, mail, document store
- Prompt
- Find the invoices paid twice in Q3 2024, post reversing entries, and write the controller a one-page memo.
- Reference
- Present 23 tool calls · 14 duplicates
RubricVerifier · Result
- All duplicate payments identifiedLedger state checkPass
- Reversing entries posted to the correct periodLedger state checkPass
- No ledger lines changed outside the duplicate setFile diffPass
- Memo lists each vendor, invoice, and amountRubricPass
- Memo states the control failure in one paragraphRubricFail
| Model | Harness 1 | Harness 2 |
|---|---|---|
| Model A | 19% | 31% |
| Model B | 6% | 13% |
| Model C | 0% | 0% |
Four parts, one container.
- 01
World
Cloned tools seeded with one company’s operating records: mail, chat, tickets, files, and ledgers, de-identified by Mikra before anything ships.
- 02
Tasks
A prompt and a rubric for each task, written by practitioners who have done the job, such as a controller or an accounts payable lead.
- 03
Verifiers
Deterministic checks on end state, step-level rubric lines, and a reference solution for every task.
- 04
Calibration report
Pass rates per model on two harnesses, with every setting disclosed, so you can reproduce our numbers before you train on them.
Real records in, verified tasks out.
A company licenses its operating data and uploads it to Mikra. Mikra de-identifies it. Raw records stop at that line. Everything built and shipped after it carries de-identified data only. De-identification removes or replaces identifiers; it does not make re-identification impossible. The risk is lower, not zero.
Checked before it ships.
The checks a post-training team runs on an environment before trusting its reward. We run them first and put the results in the calibration report.
- 01
Difficulty calibration
Every task fails on a single attempt by the strongest model we test and passes within eight attempts by at least one. Tasks outside that band are rewritten or dropped.
- 02
Trajectory review
Practitioners rank a sample of rollouts blind. We compare rubric scores against their ranking and rewrite any rubric line that disagrees.
- 03
Reward-hack probes
Each task is run by an agent that does nothing, an agent that edits tests or verifier files, and an agent told to cheat. All three must score zero.
- 04
Two harnesses
Calibration runs on two agent harnesses with fixed settings: model version, sampling temperature, tool budget, and turn limit. The report lists each one.
- 05
No public-corpus overlap
Worlds are built from licensed private records, so the tasks and their answers do not exist on the public web.
- 06
Held-out split
A share of every task family never ships for training. We keep it so a gain on the training tasks can be checked against tasks the model has not seen.
Start with one workflow.
What you get back
- The container: world, tasks, verifiers, and reference solutions
- The calibration report, with harness settings
- The held-out set, scored on your models at the end of the pilot
What we need from you
- The capability you want to train or measure
- The harness you run, so we calibrate on it
Every task has a source.
Your legal team gets the same record the data owner signed off on.
- 01
Data Passport
Every environment carries one: source, rights basis, de-identification method, review, and version. Your counsel reads it before the container.
- 02
Licensed data
Each world is built from data licensed by the company that owns it, with the rights documented. If the rights can’t be documented, the environment doesn’t ship.
- 03
No raw records in the container
Raw records never leave Mikra. Containers carry only de-identified data, and customers are contractually barred from re-identifying anyone. The risk is lower, not zero.
- 04
Exclusivity
Exclusive use of an environment is available for a fixed window, agreed per pilot.
Environments
Scope a pilot
with us.
Tell us the capability you want to train or measure and the harness you run. We scope a pilot for one workflow family and agree the timeline in writing before work starts.