ENVIRONMENTS

Where models learn to act

We build RL environments, evals and benchmarks out of real enterprise workflows. The people who do that work for a living grade the runs.

lh2-env - episode runner EPISODE LIVE
MEASURE
What we build

Three instruments, one question: can it do the job

Think of it likeA flight simulator
Graded byA licensed examiner
Scored againstAn unseen test flight
01 / ENV
RL environments

The agent practises the task in a working copy of your software, failing and improving well before production.

multi-steptool useexpert reward
02 / EVAL
Evaluations

A practitioner grades that same work against criteria drawn from how your field judges it.

rubric-gradedcontamination-freeprivate
03 / BENCH
Benchmarks

A standard test built from tasks nobody published, so a model cannot memorise its way to a score.

domain-deepversionedheld-out
The goal
CAPABILITY

A model that can do the job.

THE GAP
Why it matters

Where public scores stop being useful

A score only means something once you know what got tested and who graded it. Here are three of ours, next to the public suites the same models sit on.

LH2-SWE-REPO

The model fixes failing tests, reviews PRs and closes incidents inside a licensed enterprise monorepo, under live CI, with the original issue thread for context.

240 tasksRubricManifestTracesSample task
Model AFRONTIER · 202664% ± 4.1%91% public
Model BFRONTIER · 202657% ± 4.3%88% public
Model COPEN · 202561% ± 4.6%84% public
0%20%40%60%80%100%

LH2-OPS-AGENT

The agent carries a workflow across CRM, ticketing and docs: six connected applications, with the escalations and hand-offs an ops team runs.

180 workflowsRubricManifestTracesSample task
Model AFRONTIER · 202658% ± 4.4%89% public
Model BFRONTIER · 202654% ± 4.5%86% public
Model COPEN · 202549% ± 4.9%82% public
0%20%40%60%80%100%

LH2-CLIN-REASON

Longitudinal patient cases built from de-identified records. Work the chart, order the tests, commit to a diagnosis, write it up.

150 casesRubricManifestTracesSample task
Model AFRONTIER · 202666% ± 3.9%93% public
Model BFRONTIER · 202660% ± 4.2%90% public
Model COPEN · 202555% ± 4.7%87% public
0%20%40%60%80%100%
CATALOG
The catalog

Environments built from real work

Six domains, each a working simulation of a job someone does for a living. A model can practise, fail safely, and improve before it touches production.

OBS — what the agent can seeThe information available to it at each step, like a dashboard it reads.
ACTIONS — what it can doThe moves it is allowed to make, like the buttons it can press.
REWARD — how it is scoredThe expert rubric that decides whether it actually did the job.

Software engineering

A live codebase: broken tests, open PRs, failing CI. Can it fix things the way a senior engineer would?

OBSrepo·CI logs·issue threadACTIONSread·patch·run·commitREWARDtests + senior-engineer rubric

Agentic SaaS operations

CRM, helpdesk and docs, which is an ops analyst's morning. Fill the forms, route the tickets, escalate the right ones.

OBSapp state·tickets·emailACTIONSnavigate·fill·route·escalateREWARDtask completion + ops rubric

Clinical reasoning

A de-identified patient file. Decide what is wrong, order the right tests, write it up. Physicians grade it.

OBSchart·labs·imaging reportsACTIONSorder·diagnose·documentREWARDphysician panel rubric

Finance & legal ops

Real filings and contracts. Reconcile the numbers, redline the clauses, file. CAs and counsel grade it.

OBSworkpapers·contracts·filingsACTIONSreconcile·redline·fileREWARDCA / counsel rubric

Embodied, from factory-floor capture

A robot in a simulation built from real production footage. Grasp, place, move, and stay inside the safety envelope.

OBSvideo frames·stateACTIONSgrasp·place·navigateREWARDtask success + safety constraints

Red-team arenas

Security practitioners try to break the agent. It is scored on whether it spots the attack and holds its guardrails.

OBSagent surface·tools·memoryACTIONSinject·escalate·exfiltrateREWARDbreach detection + severity
SIGNAL
The payoff

How the training signal gets built

01

Capture the workflow

We instrument the real task, whether that is a repo, a chart or a workflow, from a licensed source.

02

Define the action space

We model the tools, states and transitions the job allows.

03

Write the reward

Domain experts write the rubrics and grade the trajectories, which makes the reward hard to game.

04

Calibrate & version

We run inter-rater checks, hold out splits, and version each release.

Task success across RL training, same base model
90% 60% 30% 0 RL training steps → trained in LH2 environments static dataset only
DEPLOY
Delivery

Run it your way

An environment has to sit where the work already lives. However closely you hold your data, one of these three lets your models train against it without moving it.

001

LH2-hosted

We run the environments and you hit an API. It is the quickest route from contract to first episode, and the task manifest, grader rubrics and per-episode traces all come through the same endpoint.

002

Inside your perimeter

We ship containerised environments into your own infrastructure. You keep the network boundary, the logs and the weights. We send the environment, the rubrics and the version history.

003

Built with your team

A joint programme on shared IP terms. Your practitioners write the reward alongside our calibration process, and you decide whether to publish the benchmark or keep it.

Benchmark against reality

Run your model through one environment and look at the gap.

Request an eval run