Fig. 01 Environment ap-ledger-01 · three rollouts

The environment factory for high-end science: simulation-based RL

DataOrigin builds simulation-based environments for frontier labs in hard science: Nuclear physics, CAD, metallurgy, civil, mechanical engineering and software development. Others sell RL tasks; we sell the lab. We train the next skill, multi-step research, where agents design experiments, run simulations and debug results. The simulator grades every step, so there is almost no room for reward hacking. Code models had GitHub to learn from. Physics up to nuclear and civil have nothing like it, and that's where models are weakest, so models cannot pick it up from the internet in pre-training. 1,000+ India's PhDs audit the tasks at a fraction of the cost, and every model failure we find makes the next environment harder.

01 · Environment. A replica of the production system, seeded with de-identified state. Every action the agent takes changes it.

environmentrolloutgradedataset
1,208 invoices · 1,214 bank lines · identities pseudonymised
Productsenvironments · datasets · graders Monetise your datafor institutions that hold it

6

production systems replicated as environments

820

tasks written from real tickets and incidents

4,100

graded checkpoints across the task set

160

practising specialists writing and grading tasks

42%

of ingested records rejected before delivery

0

scraped or unlicensed records in the catalog

Live · OriginBench 1.0 Repository-scale coding, graded by hidden verifiers. 30 tasks in 12 languages, 7 frontier models, 630 graded runs. Claude Opus 5.5 leads at 60.0% pass@3. View the leaderboard →

Fig. 02 Why this exists

Benchmarks measure the test. Production measures the job.

Sample tasks from the environments we are building. Pick a domain, open a task and see how frontier models responded.

Every lab reads the same web and buys the same chips. What still separates models is the distribution they were trained on, and the distributions that matter now, how a claim is settled, a build fixed, a chart reconciled, were never published. They sit inside institutions. Reaching them takes a licence, not a crawler.

01
Coverage

Institutional records cover the tasks models fail at today: diagnosis, code review, claims handling, reconciliation.

02
Process, not only outcomes

Public text keeps the conclusion and drops the reasoning. Trajectories inside a real system keep both.

03
Measured, not assumed

A benchmark score is a claim. A graded run inside a replica of the production system is a measurement.

See the published benchmark →OriginBench 1.0 · 7 frontier models · 630 graded runs
Same models, two scoresmean of five seeded runs · seven models
REPO-CI-04240 tasks · repository & CI
86–9226–50
−51
AP-LEDGER-01150 tasks · accounting close
84–9036–58
−40
CLINIC-EHR-03140 tasks · clinical
88–9410–27
−73
0%25%50%75%100%
public benchmarkinside the environmentgap in points

Fig. 03 How it works

One loop. Four things you can hold.

Every product is a stage of the same loop, and every stage leaves an artifact a lab can inspect. Shown for one episode of ap-ledger-01.

1Environment

A replica, seeded with licensed state

A snapshot of a production system its owner licensed to us, identities removed, surrounding services stubbed to production behaviour.

snapshot
INV-4471 · V-093112,400.00
INV-4472 · V-02083,180.50
INV-4476 · V-1140paid
INV-4480 · V-11408,900.00
licenceDO-L-2026-018

2Rollout

The agent works a real task

Tasks are written from real tickets, with the edge cases, approval steps and failure modes left in. Every action changes state.

trace
match_payment✓ 4471
match_payment✓ 4472
flag_duplicate4480 held
post_journal✓ JE-0912
write_note✓

3Grade

A practitioner's rubric scores it

A gate, then weighted criteria, written and applied by someone who does the job. Deterministic where possible, judgement where not.

rubric v3
gatebalances ✓
0.30 no duplicate posted✓
0.25 approval path✓
0.25 audit note0.4
score0.65

4Dataset

The episode ships as one row

Graded trajectories become datasets with licence and reviewer on every row, or the environment runs live as a reward source.

jsonl
{ "episode": "ep-0412",
  "env": "ap-ledger-01",
  "score": 0.65,
  "licence": "DO-L-2026-018",
  "reviewed_by": "DO-EXP-0688",
  "pii_residual": 0 }

Fig. 04 Environments

Six production systems, replicated.

Pick an environment and the console loads a task from it: the trace, the rubric, the row it became, and the full task card with instruction and method. It advances on its own until you touch it.

environments · one task loaded from eachlive

—

—

0.00
graded by —
lead run shown · five seeded runs per model

Fig. 05 Data licensing

Six kinds of records, licensed at the source.

The records behind the environments are also licensed on their own. Each comes directly from the institution that produced it, is de-identified inside its walls, and ships with a licence ID and a named reviewer on every row. One sample of each is below.

mp4 + jsonl · 2 of 6 clips in the packDO-VID-006 / DO-VID-009 · first-person · step labels, object tracks
cam-01 · ego · 720p3003:05 → 03:21
step 4/7slice chicken · into the panstir
sample clip · in the sample pack
DO-VID-006 · home cooking · 16 s of clip 0412step labels · object tracks
cam-02 · ego · 720p3011:35 → 11:53
step 5/9tin pads · re-solder boardinspect
sample clip · in the sample pack
DO-VID-009 · electronics repair · 18 s of clip 0207tool use · step labels
faces never in frame · excerpt chosen without captions or identifying marksstep labels and object tracks as jsonl per framesound muted on the page, kept in the pack

Fig. 06 Graders

Rubrics written by people who do the job.

Practising physicians, staff engineers, chartered accountants and counsel write the tasks and grade the runs. Credentials are confirmed with the issuer, the exam is timed, the trial is paid, and every grader is re-calibrated weekly on seeded gold sets.

applied
8,400
verified with issuer
−87%1,100
passed the exam
−63%410
passed the paid trial
−54%190
admitted
−16%160
admitted · bars on a square-root scale1.9%
MD
Consultant physicianscardiology · oncology · radiology · emergency
clinical
SWE
Staff and principal engineerspayments · infra · ML systems
code
OPS
Operations leadssupport · CRM · procurement · warehouse
operations
CA
Chartered accountants and auditorsassurance · forensic · tax
finance
LLB
Commercial counselcontracts · compliance · IP
legal
PhD
Research scientistsML · statistics · rubric design
evals

Fig. 07 Provenance

Sourced worldwide. Graded by practitioners.

Every record traces to a licence signed by its owner and a reviewer who practises in the field. Built so a hospital or a bank can sign without lowering its standards, and a lab can audit what it received.

001De-identification before transfer

Identifiers are removed inside the owner's environment. Each redaction is logged and residual risk is reviewed by a person.

002A licence behind every record

Each record traces to a licence ID, a consent basis and a permitted use, written into the manifest.

003Owner-defined terms, enforced

Scope, exclusivity, term and revocation are set by the owner and enforced contractually after delivery.

004Auditable end to end

Encrypted pipelines, per-access logs and audit rights for both the owner and the licensee.

src · where records are licensed from  ·  rev · where graders sit

Fig. 08 For data owners

License the records you already keep.

Hospitals, enterprises, engineering organisations and contact centres license what they already keep. We de-identify inside your walls, build and deliver, and report every licensee.

clinicalcode & CItickets & CRMcallsfirst-person videodocuments
  1. 01
    What you hold

    Clinical notes and imaging. Code, reviews and CI. Tickets, CRM, email and procurement. Calls, meetings and first-person video. The by-product of the work, not the work itself.

  2. 02
    How it is licensed

    A one-time licence or a revenue share. You set scope, term, exclusivity and revocation, and the contract enforces them after delivery. No resale outside the agreed scope.

  3. 03
    De-identification inside your walls

    Direct and quasi-identifiers removed in your environment before anything moves. Every redaction logged; residual risk reviewed by a person; spoken identifiers cut from audio.

  4. 04
    Payment and reporting

    Per-delivery statements showing which licensee received what, under which terms, and what was paid. Revocation is honoured downstream. Nothing moves until the licence is signed.

For AI labs and enterprise teams

Training or evaluating a model?

Tell us the capability gap. If an environment or a dataset covers it, you will have samples under NDA within days.

No newsletters. We reply by email.

For data owners

Holding data that could train models?

License it on your terms. We handle de-identification, compliance and delivery; you keep control of scope, term and exclusivity.

Scope, term and exclusivity stay yours to set.

partnerships@dataorigin.ai · we reply within two business days