RL environments built from real expert work.
Train and evaluate agents on long-horizon workflows authored by verified practitioners, using resettable software, deterministic verifiers, and human performance baselines.
The most valuable work never made it to the internet. It lives in how a payroll specialist closes a period, how a claims adjuster reads a file, how a dispatcher talks a rate down on a live call.
Environments compound. Datasets deplete. We turn expert workflows into environments you can rerun, grade, and improve against, not one-time datasets.
Every expert workflow becomes four reusable assets
One authoring flow produces a task set, a benchmark, an RLVR environment, and a correction corpus. Use them individually or as one training and evaluation stack.
RL environments
Real business software, self-hosted in isolated microVM sandboxes. Every environment ships with deterministic seeds, per-episode state resets, snapshot restore, and full action telemetry, built to survive thousands of eval runs a day without polluting state.
Long-horizon task datasets
Multi-hour workflows elicited from working practitioners and confirmed by them, not written by contractors. Ground truth is a set of acceptable outcomes with tolerance bands, mandatory invariants, and disqualifying conditions.
Rubrics & verifiers
Expert-weighted rubrics compiled into code. Database assertions, numeric reconciliation, and audit-log checks grade the work. Model-graded criteria are capped, labeled, and expert-audited, never the default.
Benchmarks
Frozen, expert-baselined benchmark cuts with public leaderboards and reproducible harnesses. The human baseline is the authoring expert's own recorded cold run. Private held-out sets are available to license.
Reward integrity is the product.
Many agent benchmarks reward outputs that look correct without verifying the resulting system state. Ours grade with code: state assertions, reconciled ledgers, and audit-trail checks. Model-graded criteria are capped at a fifth of any corpus, labeled, and expert-audited.
Every verifier must reproduce the authoring expert's own recorded run before it ships.
db_assertState assertions against the environment's own database. The record either exists in the right state or it does not.
numeric_reconcileTotals, ledgers, and registers reconciled within expert-set tolerance bands. The ledger ties or the episode fails.
file_diffProduced artifacts diffed against acceptable outcome sets, not a single golden file.
audit_log_assertThe path matters. Required intermediate actions are checked in the environment's audit trail.
invariant_assertMandatory invariants that must hold at every step. Breaking one is a disqualifying condition with a terminal penalty.
From working expert to running environment
Every step hardens real expert work into something a lab can train against.
- 01
Source the practitioner
Our private expert network reaches verified specialists in occupations labs cannot staff. Identity and skill verification gate every cohort.
- 02
Elicit the work
We capture how the work actually gets done. The decisions, the exceptions, the tricks that never made it to the internet.
- 03
Author in the sandbox
The expert walks their own tasks inside a recorded environment on real software. That cold run becomes the human baseline and the ground truth.
- 04
Compile the reward
Rubric weights compile into shaped reward functions. Checkpoints become intermediate signals, disqualifiers become terminal penalties.
- 05
Verify deterministically
Every verifier must pass the authoring expert's own recorded run before it ships. If the check cannot reproduce the expert, it does not ship.
- 06
Refresh on the model clock
Agent failures reopen the task list at higher difficulty. Every frontier release gets fresh variants, harder bands, and a re-run leaderboard.
Operational expertise that is difficult to recruit inside a lab
Payroll administrators, claims adjusters, dispatchers, and other working specialists are rarely represented in technical hiring networks. We built a sourcing and verification process to reach them directly.
Public benchmarks in progress. Private holdouts available.
Public releases include a leaderboard and reproducible harness. Private held-out sets stay outside the public corpus and are available to license.
Recruiter-Bench
Long-horizon, multi-stakeholder recruiting workflows: sourcing, screening, coordination, and recovery when humans behave like humans. Most agents fail the hard band.
ATS-Bench
Can agents build, fix, harden, and migrate integrations against a live applicant tracking system? 50 public tasks, 200 private, graded by webhook-verified system state.
Tell us where your model breaks. We build the environment.
Sample task packets, environment access, and benchmark holdout licensing for AI researchers, data teams, and labs. Exclusive cuts available.