For AI labs and data vendors

RL environments built from production code

Software engineering tasks with hidden verifiers, mined from a real codebase, calibrated against real models and reviewed for reward hacking. Delivered in Harbor and HUD formats by a senior pod under one contract.

Start a pilot How we grade

EltexSoft builds reinforcement learning environments for AI labs and data vendors. Each environment is a Dockerized production codebase with hidden test verifiers, delivered in Harbor and HUD formats, with every task reviewed for reward hacking before it ships. A senior pod of four engineers does the work under one EltexSoft contract. Pilots are fixed price, 20 to 50 tasks, and the buyer grades them.

What we build

Environments a trainer can trust

How we grade

The grader is the product

A task that grades wrong teaches the model the wrong thing, at scale. Every task passes the same validator before a model ever sees it.

  1. Step 1

    Base run

    The untouched snapshot must fail every fail-to-pass test and pass every pass-to-pass test. Three times, to catch flakes.

  2. Step 2

    Gold run

    The reference solution must pass everything. Three times. A task whose own solution is flaky is not a task.

  3. Step 3

    Clock shift and scrub

    A run under a shifted system clock catches time-dependent tests. Two independent secret scanners check the image and the git history, and the data is synthetic seed, never production records.

  4. Step 4

    Reward

    Zero if any pass-to-pass test fails. Otherwise the fraction of fail-to-pass tests passing, or a binary score if your trainer needs one. The rule is written into each task, not left to the grader.

  5. Step 5

    Second-engineer review

    Instruction clarity, test leakage between tasks, shortcut attempts and a difficulty estimate, recorded per task before it reaches calibration.

Pilot terms

Fixed price. You grade it.

A pilot is the cheapest way for both sides to find out whether the tasks fit your trainer. The terms are short on purpose.

Price
Fixed, agreed before the start
Scope
20 to 50 tasks on one environment
Grading
You grade them, with your rubric and your models
Formats
Harbor, HUD or both
Rates
On request
Ask for pilot terms

Who does the work

A senior pod under one contract

Four engineers who have shipped production software together, not a crowd of anonymous task writers. One agreement, one point of contact, one quality bar.

  • Task authors, backend

    Two senior engineers who mine the history, write the instructions and the gold solutions, and select the tests.

  • Task author, frontend and full stack

    One senior engineer for the tasks that cross the API boundary into the React client.

  • QA and calibration

    One engineer who authors nothing. Runs the validator, attempts the shortcuts, runs the model rollouts and signs off per task.

Need the same roles on your own environments? Hire RL environment engineers as a pod or one at a time.

Trust

What is in the contract

FAQ

Common questions

What is an RL environment for coding agents?
A reinforcement learning environment for coding agents is a runnable codebase plus a set of tasks, each with an instruction and a hidden verifier. The agent edits the code, the verifier runs the tests the agent could not see, and the result becomes the reward signal used to train or evaluate the model. The environment has to be deterministic and self-contained, which is why ours ship as a single Docker image with the database, cache and workers inside and the network turned off.
What is a hidden verifier?
A hidden verifier is the set of tests used to grade a task that the agent never sees during the episode. It usually has two parts: fail-to-pass tests that fail on the starting code and must pass after the change, and pass-to-pass tests that already pass and must keep passing. Hiding them prevents the agent from editing the tests or special-casing their inputs instead of fixing the code.
How do you prevent reward hacking?
Reward hacking is when an agent gets the reward without doing the intended work. We attack it before the task ships: an engineer who did not author the task tries to pass the verifier by editing tests, hard-coding outputs or reading the solution from git history. Pass-to-pass tests make partial fixes score zero. Tasks with any shortcut go back to the author, and the review note stays attached to the task.
Which formats do you deliver?
Harbor task packages (a manifest, an instruction, a test selection, a test patch and a gold solution per task) and HUD environments with the same grader. Both run against the same Docker image, so a task graded one way grades the same the other way. If your trainer needs a different manifest, that is a formatting step, not a rebuild.
How does a pilot work and what does it cost?
A pilot is a fixed-price engagement of 20 to 50 tasks on one environment. You tell us the task profile you want (difficulty mix, backend or full stack, format), we deliver the tasks with their calibration data, and you grade them with your own rubric and your own models. The price is agreed before the start and rates are available on request.
Do you work with data vendors as well as labs?
Yes. Data vendors and environment platforms that need to fill an order for software engineering environments can contract the pod under a single EltexSoft agreement, with IP assignment and the same NDA, verification and insurance terms. We do not name labs or vendors on this site.
Who owns the code and the tasks?
The contract assigns the delivered tasks and environment work to the buyer. The pod engineers have already assigned their work to EltexSoft, so there is no gap in the chain. Where an environment is built on a codebase we own, the licence terms for that codebase are set out in the same contract.

Tell us the task profile you need.

Difficulty mix, backend or full stack, Harbor or HUD. An engineer replies within one business day with pilot terms.

Start a pilot