Environments With a Past.

Digital twins of the industry-specific software a business ran, seeded with its de-identified history. Tasks come from what was actually queued; graders check the result against what the outside world said.

Get data

rl environment · anatomy

state · actions · task · grader

State

practice management twin · 4.2 yrs · 214,806 claims · 31,119 patients · 46 payers · all de-identified, all consistent

Action space

typed API / MCP: 38 tools · rendered UI: keyboard and pointer · every action logged

Task

day 2023-08-08: 12 claims queued, 3 denials to work, 1 pre-auth expiring · seed from the record

Grader

claim state vs payer adjudication · ledger balance vs actual · rubric where judgement · answer key from what happened

Most environments are written. Ours are recovered: the state, the tasks and the answer keys come from a real business, and only the action surface is built.

The twin

The practice management system, the claims or estimating platform, the tax suite, the case manager, the PSA. Rebuilt from the data model and the workflows the records exhibit, not from the vendor's binaries, and seeded with the corpus so every task refers to something with a history.

Faithful where it matters

The original staff can complete the original tasks in it, which is one of our acceptance tests. Quirks the work depended on are reproduced and documented.

No vendor assets

No binaries, no proprietary UI, no redistributed code. The data model and the workflow are what is copied.

Two action surfaces

A typed API or MCP server for tool-using agents, and a rendered interface with keyboard and pointer for computer-use agents. The same task suite runs against both, and every action is logged.

Containers

Delivered as containers with the twin, the seed data, the task suite and the graders. Reset to any day in the record.

Difficulty from history

How much of the past the agent is shown is a dial. The same task is harder with less context.

Graders with an answer key

Programmatic where the outcome is programmatic: the claim state, the ledger balance. Rubric-based where judgement is involved. Reference-trajectory where process matters. Every grader has an answer key from what really happened.

Outside-world verifiers

The payer's adjudication, the agency's acceptance, the court's docket entry. Not a model's opinion of the result.

Bring your own

The twin exposes its full state, so a lab can author tasks and graders against it, or have us build task families to a specification.

Coverage

The software no benchmark covers, and the outcomes it records.

Categories rather than vendors here. Ask for the current inventory by vertical.

Vertical	Software in the twin	What the record says happened
Revenue cycle and medical billing	Practice management and billing platforms, clearinghouse and payer portals	Claim adjudication, denial codes, days in A/R
Insurance, adjusting and TPA	Claims management, estimating, policy and agency management	Settlement, subrogation, reserve accuracy
Title, escrow and closing	Title production, escrow accounting, e-recording	Closed, curative required, funding timing
Accounting and tax	Tax suites, bookkeeping, payroll, practice management	Return accepted or amended, notices, client retained
Law	Case and matter management, docketing, trust accounting, e-filing	Docketed, granted, settled, deadlines met
Clinics and practices	Practice management, scheduling, charting, imaging, recall	Treatment completed, recall kept, claim paid
Collision repair and dealers	Estimating, shop management, dealer management systems	Estimate approved, supplement accepted, cycle time
Managed IT	PSA and ticketing, RMM, documentation, quoting	Resolved within SLA, reopened, escalated, renewed

Also in for labs