RL environments as a service

Pebble builds and operates RL environments as a service. You bring a business process. We map it into steps, define a verification check for each step, and hand back software that runs the process on AI and checks every step of its own work. The environment is yours: host it on your infrastructure or ours, and keep every run record as a dataset you can train on.

Most companies selling RL environments sell to labs, organized by modality: coding environments, computer-use environments, browser environments. Pebble sells to the companies that own the process. The unit is not a modality, it is a process: reconciling invoices, triaging safety reports, adjudicating claims, onboarding a customer, running order to cash. Where the other vendors sit is mapped in the RL environments market map.

What is an RL environment as a service?

An RL environment as a service is a vendor-built, vendor-maintained environment in which an AI agent does real work and every step of that work is checked against a defined standard. The environment provides three things a bare LLM call does not: a step map of the process, a verifier at each step, and a run record that shows exactly what happened. The "service" part means you do not staff an environment-engineering team. Pebble scopes the process, builds the environment, exercises it against real runs, and maintains it as the process changes. The method is written up in our white paper, Scaling Experimentation, and you can see a public environment we run at Rockie.

How does Pebble verify each step?

Every step in the environment gets one of three check types, in strict order of preference:

1. Computed - verified by code, no model involved in grading. Do the totals reconcile. Does the reported count match the sensor count. This is the gold standard, and we push as many checks as possible into this category.
2. Classified - a small model trained for that one check, like a classifier that catches an illegible field. Trained on your data or open data, and scored against held-out examples before it is trusted.
3. Judged - for outcomes that need human taste. Your operators annotate a few hundred pass/fail examples, we train a judge on them, and we report judge-human agreement on held-out examples. Where agreement is too low, a human stays in the check.

When a step fails its check, the run stops there, gets fixed, and resumes at that step, not from scratch. When something goes wrong you get a step number, not a transcript to scroll through. This is agent verification in the sense that matters: not verifying who the agent is, but verifying the work it did. The same checks that catch failures in production are what make the environment usable for reinforcement learning with verifiable rewards later.

What do you own at the end?

Everything that matters. The environment runs as software you host or we host. Every run's record is appended to a training set that belongs to you, in plain data formats with no vendor attached. Engagements climb a three-stage ladder:

• 01 RUN - working software. One process scoped, the environment built, running on a frontier model to start.
• 02 PROVE - a full audit trail. Every run logged against the full set of checks. Compliance gets an audit trail, postmortems start from a step number, and reliability becomes a percentage read off run history.
• 03 OWN - a model and dataset you own. An open-source model is run through the environment at volume, checks score every run, and passing runs weigh more in training. The result is a model tuned to the work at a fraction of frontier API cost per run, that no one can reprice or deprecate out from under you. When a stronger open-source base ships, the same run records train its replacement.

How does an engagement work?

Four stages, starting with a call with the person who builds the environments:

1. Scope - a call. If an environment is the wrong tool for the job, we say so on the call.
2. Discover - about a week inside your systems, under your access model. We keep the step map and none of your data. It ends in a written proposal: end goal, work packages, price.
3. Build - the environment is exercised against real runs before it is called done.
4. Run - handoff on your infrastructure or ours, plus a retainer for working sessions, maintenance, and the next stage of the ladder.

There is no public price list because no two processes have the same step map. The proposal after discovery states a fixed price per work package. For how the market prices this class of work, see what RL environments cost.

Who is this for?

If AI does work inside your product and the results are hard to trust or expensive to produce, this is for you. The pattern fits processes with real stakes: multi-source inputs that must become clean structured output, outcomes with safety or compliance weight, and volumes where frontier API pricing hurts. It also fits AI product teams that need environments and verifiable rewards for their own agents but do not want to build environment engineering in-house.

Frequently asked questions

Is this the same as what Scale AI or Prime Intellect sells?

No. Frontier labs buy environments in bulk to make general models better at broad skills. Pebble builds one environment around one company's process so that company can run that process on AI, prove it works, and own the resulting model and data. Same underlying technology, different customer and different deliverable.

Do we need to be doing reinforcement learning to use this?

No. Stage 01 runs your process on a frontier model with checks and resumable runs, which is valuable on its own. The RL part becomes relevant at stage 03, when your accumulated run records post-train an open-source model. The environment is built RL-ready from day one so nothing is thrown away.

What does "verified" actually mean here?

It means each step's output is checked by code where possible, by a tested classifier where code cannot decide, and by a judge trained on your operators' annotations where taste is required, with agreement rates reported. See the glossary for the full vocabulary.

How do we start?

Write to [email protected] for a scoping call. You will be talking to the founder who builds the environments.