Skip to content
ARCHYTASAI

Hyper-realistic · Deployment-faithful · RL environments

Most RL environments are clean rooms.
Real work isn’t.

Archytas builds hyper-realistic, deployment-faithful RL environments. Tools refuse, records disagree, people want things they don’t say, and the agent’s authority runs out. Reward goes only to verified work and honest reports, whether or not the agent thinks it’s being watched.

Discuss an environmentSee the example episode

$ env run ledgerops --verify

env/ledgeropsExample episode
period.closed1.00
bank.matched0.98
duplicate_payment.resolved0.00
cash.bank_equals_ledger0.00
approvals.policy_FIN70.00
Agent: “October is closed.”Reward 0.00

Built for

Frontier AI labs

AI-native companies

Training infrastructure teams

Part of Deep Insight Labs

The problem. Agents trained in clean rooms learn to push through.

In a clean room, every tool call succeeds, sources agree and the agent’s report is taken at its word. Agents trained there claim success after a tool refused, pick one of two conflicting records without saying so, and act without the authority to. They need to practice where those things happen, and where the result is checked on system state.

  • P.01

    01Verified end state

    The agent has to finish the workflow: change the right records, respect constraints and handle exceptions. We score the system state it leaves behind.

    1. 1.1Final database state is the ground truth
    2. 1.2Well-written answers earn nothing alone
    3. 1.3Every check is an executable assertion
    Scored
    Final database state
    Not
    How the answer reads
  • P.02

    02Rewards that hold up

    Expert-approved outcomes and executable assertions. Checks accept alternative correct solutions, and we test for reward hacking before training.

    1. 2.1Alternative correct paths accepted
    2. 2.2Exploit tests for every grader
    3. 2.3Expert review where judgment is needed
    Scored
    Assertions on outcomes
    Not
    A single grader’s opinion
  • P.03

    03Built for training

    Environments generate reusable episodes, vary difficulty and keep held-out evaluations separate, so you can measure whether a model got better.

    1. 3.1Seeded, resettable episodes
    2. 3.2Graded scenario difficulty
    3. 3.3Held-out evaluation splits
    Scored
    Held-out improvement
    Not
    Difficulty for its own sake

Two terms. Hyper-realistic describes what goes in. Deployment-faithful is what we measure coming out.

How we build and measure them
  • DEF.01The default

    Clean room

    An RL environment where every tool call succeeds, sources agree, the agent works alone, the objective and the grader are visible, and the agent's account of its own work is taken at its word.

    What agents learn there

    1. 1.1Claim success after a tool refused the action
    2. 1.2Pick one of two conflicting records without saying so
    3. 1.3Keep going when the job was to ask, escalate or stop
  • DEF.02How we build

    Hyper-realistic

    An environment that keeps what clean rooms remove: tools that refuse and fail, records that conflict or go stale, intent nobody fully states, other parties with goals and authority of their own, and limits on what the agent may do. Built from real system prompts, tool schemas, documents and workflows, not placeholders.

    In practice

    1. 2.1Refusals, failures and conflicts planted with a known answer
    2. 2.2Simulated parties grounded in real interaction data
    3. 2.3Real artifacts in place of synthetic stand-ins

    Tested by

    conditions_covered · simulator_fidelity

  • DEF.03What we measure

    Deployment-faithful

    An environment whose episodes an agent can't reliably tell apart from production, and whose scores predict how the agent behaves once it's deployed. Hyper-realism describes what goes into an environment. Deployment faithfulness is what we measure coming out.

    In practice

    1. 3.1The grader stays out of the agent's view
    2. 3.2Watched and unwatched versions of a task score the same
    3. 3.3Evaluation on held-out people, workflows and repositories

    Tested by

    realism_win_rate · oversight_gap · transfer

Explore. The research, the environments, the method and the team behind them.

Who we work with. Built for the teams training AI agents.

For frontier AI labs

Hard, expert-grounded tasks and environments for post-training.

  • 1.1Agentic reasoning and tool use
  • 1.2Long-horizon execution against real state
  • 1.3Domain depth in finance and enterprise operations
  • 1.4Held-out evaluations kept separate from training data
Discuss your training needs

For AI-native companies

Turn your workflows and production failures into training and evaluation assets.

  • 2.1Environments modeled on the workflow you need automated
  • 2.2Evaluations built from the failures you see in production
  • 2.3Rewards that encode your definition of done
  • 2.4A confidential scoping call before any data changes hands
Discuss a custom environment

Work with Archytas. Bring us a workflow your model struggles with.

We build custom datasets, evaluations and RL environments for AI labs, AI-native companies and training infrastructure partners.

What happens next

  1. 01You tell us the workflow or capability
  2. 02We reply to set up a confidential scoping call
  3. 03We propose data, environments or evaluations to fit

Or email

info@deepinsightlabs.ai
I’m reaching out from

For example: financial reasoning, long-horizon tool use

Please don’t include confidential data. We’ll set up a confidential scoping call before anything sensitive is shared. See our privacy notice.