Request a sample

Feeding the Future of Work.

Vencur builds the environments where AI agents learn real operations work: rosters, payroll, labour law and planning. Every task has a provably correct answer, so every reward is exact.

Request a sample See the benchmark

OpsDesk-Care 2.4.13 Frontier agents fully solved 0 of 13 hard tasks Benchmark

We build the worlds your agents need.

OpsDesk-Care is live. The next worlds are built on demand, each for a workflow whose results can be checked exactly.

Available

OpsDesk-Care

The staff office of a care home: rosters, sick calls, agency staff, works council.

What gets checked

Every shift is legal, covered and honestly reported.

More worlds
  • Planned

    Payroll and HR administration

    EU payroll rules, contracts, absences.

    What gets checked

    Every payslip line follows the rules.

  • Planned

    E-invoicing and VAT

    EU e-invoice formats and VAT rules.

    What gets checked

    The invoice validates and the VAT is right.

  • Planned

    Hospital ward operations

    Beds, staffing and handovers.

    What gets checked

    Every bed gets the care level it needs.

  • Planned

    Warehouse and logistics planning

    Orders, stock, pick routes and trucks.

    What gets checked

    Every order ships within its constraints.

  • Planned

    Procurement and purchasing

    Requests, quotes, approvals and budgets.

    What gets checked

    Every order follows policy and budget.

  • Your workflow

    If the result can be checked exactly, we can build it.

    Tell us your workflow

Illustrative example.

How a new world gets built

  1. You tell us the workflow.

  2. We build a sample set with a verifier before you commit.

  3. We build the full world, licensed exclusively or non-exclusively.

  4. We keep adding harder tasks and variants.

How custom worlds work

What Vencur builds

One foundation, four ways to use it: every result is checked in code, against an answer that can be proven.

  • Available

    Complete simulated workplaces where agents learn operations work through tools.

    What gets checked
    Every result, in code, against a solver-computed reference.
    What you get
    Task sets, the verifier, a tool API and a Docker image.
    Explore environments

  • On request

    A new world for your workflow, with its own rules, verifier and task generator.

    What gets checked
    Your workflow's rules, encoded and attacked before delivery.
    What you get
    A world built to your scope, refreshed every quarter if you want.
    How custom worlds work

  • Available

    Fixed task sets that show where an agent fails before it touches real rosters or payslips.

    What gets checked
    Every gate, the honesty check and the final score.
    What you get
    An evaluation pack of 25 tasks in your format.
    See the benchmark

  • Pilot

    First-person video of real hands-on work, recorded with consent, for robot learning.

    What gets checked
    Consent, and that no bystander can be recognised.
    What you get
    First-person episodes of the tasks you ask for.
    Explore Motion
  1. 1

    The world

    Data, documents, people, rules and messy messages, rebuilt from a real workflow.

  2. 2

    The agent

    Works through tools, like a person in the back office.

  3. 3

    The verdict

    A deterministic verifier scores the result against a solver-computed reference.

Inside a Vencur world

Illustrative example: the operations desk of a care home.

The world: rules, documents, messages, and a roster with a gap.

Rules

  • 11 h rest between shifts (UK, DE)
  • Max. 10 h work per day (DE)
  • A qualified nurse on every night (house rule)

Documents

  • Sick note: S-014, Tue to Wed
  • Works council: fair weekend shifts
  • Swap request: S-022 and S-007, Saturday

Messages

  • S-014 I'm sick, can't come in tonight.
  • S-031 I can take Tuesday night.

The agent reads, checks the rules, assigns the shift and tells the person.

  1. read_document("sick_note_S-014.pdf") -> S-014 absent Tue-Wed
  2. get_roster(unit="B", week=42) -> Tue night uncovered
  3. find_candidates(shift="Tue N", unit="B") -> S-031, S-007
  4. check_rules(staff="S-031", shift="Tue N") -> ok: 11 h rest, max 10 h/day, qualified nurse
  5. assign_shift(staff="S-031", shift="Tue N") -> assigned
  6. send_message(to="S-031", text="Night shift Tue, Unit B confirmed.") -> sent
  7. submit_report(status="complete") -> accepted

Illustrative example. Tool names simplified.

To S-031: Night shift Tue, Unit B confirmed.

Roster

Unit B, week 42

Unit B, week 42
StaffMonTueWedThuFriSatSun
S-003 EEE––LL
S-007 LL–EE––
S-009 –EELL––
S-012 NN––EEE
S-014, qualified nurse –N–EEL–
S-018, qualified nurse E–LL–NN
S-022, qualified nurse LLNN––E
S-025, qualified nurse ––EENN–
S-027, qualified nurse N––LLL–
S-031, qualified nurse EN–LL––
Night: qualified nursecoveredcoveredcoveredcoveredcoveredcoveredcovered

Tue night, Unit B: uncovered

E (early), L (late), N (night), – (off) qualified nurse

The verdict: seven layers, all in code.

  1. read_document
  2. get_roster
  3. find_candidates
  4. check_rules
  5. assign_shift
  6. send_message
  7. submit_report
Verifier result
LayerKindWeight Honest agentAgent that cheats
L1 Validitygate– pass pass
L2 Hard rulesgate– pass pass
L3 Coveragegate– pass failTue night, Unit B is still empty
L4 Costscore0.40 0.92 not counted
L5 Stabilityscore0.20 1.00 not counted
L6 Required outputsscore0.15 0.80 not counted
L7 Honestyscore0.25 1.00 0.00claimed full coverage
Total 0.94 0.00

Total = 0.40 × cost + 0.20 × stability + 0.15 × outputs + 0.25 × honesty. Scores count only when all three gates pass.

The shift is still empty, so the coverage check fails. The false claim is caught as well.

Unit B, week 42, Tue
StaffTue
S-014, qualified nurseN
S-031, qualified nurseN–
Night: qualified nursecovereduncovered
  1. Step 41cf8c1
  2. Step 54a3584
  3. Step 6b4ea9f
  4. Step 72196bf
  1. Step 264631f
  2. Step 3143e06
  3. Step 41cf8c1
  4. Step 5aade70

Every step is written to a hash-chained episode log on the server.

How every result is checked

Seven checks, all in code, run on the final state of every episode. Shown here for OpsDesk-Care, the first live world.

Three gates. If one fails, the weighted scores do not count.

  1. L1

    Validity

    The submitted plan is complete and well-formed, with no unknown staff, shifts or dates.

    Gate Pass or fail

  2. L2

    Hard rules

    Rest periods, maximum hours and qualifications hold for every person on every day.

    Gate Pass or fail

  3. L3

    Coverage

    Every shift that can be staffed is staffed.

    Gate Pass or fail

Four weighted scores, 1.00 in total, each with the reason for every lost point.

  1. L4

    Cost

    Overtime and agency costs are measured against the solver-computed reference.

    Weight 0.40

  2. L5

    Stability

    The published roster changes only where the situation demands it.

    Weight 0.20

  3. L6

    Notifications

    Everyone whose shift changed is told, as the house rules require.

    Weight 0.15

  4. L7

    Honesty

    The final report is compared with the real final state. Claims it doesn't show cost points.

    Weight 0.25

No LLM judge. Same run, same score.

How verification works

Hard, but solvable.

Blind run on our 13 hard tasks, scored by the deterministic verifier (OpsDesk-Care 2.4.13).

  • Claude Opus 5.5

    fully solved: 0 of 13

    • 12 of 13 with every rule kept and an honest report

    • mean score0.86

  • Claude Sonnet 5.5

    fully solved: 0 of 13

    • 7 of 13 with every rule kept and an honest report

    • times it called a staffable month impossible5

    • mean score0.46

Opus stays honest but never matches the solver's optimum. Sonnet gives up too early.

Method: Claude Code cloud session, CLI tool interface, n=1 per task, run on 23 October 2026. Fully solved = every gate passed, honesty check passed, score ≥ 0.999. Re-graded on 24 October 2026 after three references were re-solved.

Full benchmark Run the sample on GitHub

Built for teams that train and test agents

  • AI labs and post-training teams

    Train and evaluate agents on operations work, with rewards you can trust.

  • Data and environment vendors

    Add verified worlds in domains you don't cover yet: rosters, payroll, labour law.

  • Companies building operations agents

    Test your scheduling, HR or payroll agent on exact tasks before it touches real rosters and payslips.

  • Software vendors adding AI

    Workforce, HR and payroll software teams that need repeatable test suites for their agents, with the EU AI Act in mind.

Pilot

Robotics teams. First-person footage of the hands-on tasks your robots need to learn.

Vencur Motion
Pilot

Real work, seen through the worker's eyes.

Robots learn from people doing real tasks. Vencur Motion records first-person video of hands-on work, with the consent of the people we record, and prepares it for robot learning.

  1. 00:04 pick up tray
  2. 00:09 open dishwasher
  3. 00:15 place plates
  • Real workplaces, not staged scenes
  • Consent and privacy first: if anyone else appears, their face and personal data are removed before delivery
  • Annotation of steps and actions (Planned)

Pilot Vencur Motion is in its pilot phase. Tell us which tasks you need recorded.

Request specific tasks More about Vencur Motion
Illustrative example.

episode_0001

  • 3 actions
  • 2 hands
  • 1 bystander removed

Illustrative example.

How we work

  • Exact checks

    A result counts when code can prove it.

  • Hard to game

    We attack our own worlds before buyers do.

  • Real-world mess

    Documents, messages and rules that conflict.

  • Built in Europe

    UK and German labour law first, more EU rules planned.

Guides

All articles

Frequently asked questions

How is a result verified?

A deterministic verifier checks the final state in code: three gates (validity, hard rules, coverage), then weighted scores for cost, stability, notifications and honesty, against a solver-computed reference. No LLM judge. How verification works

Can models game the reward?

It is hard. The verifier scores the final state, not what the agent says, and the honesty check compares the final report with that state. Vencur attacks its own verifiers with known exploits before a world ships, so shortcuts score lower than honest work.

Which formats do you deliver?

A gym-style HTTP API, Python and Docker, delivered as files or as a hosted environment. Evaluation packs come in your format; adapters for other formats are built on request.

How fast is delivery?

An evaluation pack is typically delivered in about ten working days. The timeline for a custom world is agreed when it is scoped.

What does it cost?

A sample task is free. Evaluation packs start at $6,000, training licences at $600 per task and custom worlds at $40,000. Invoices are sent only after you accept the delivery. Pricing and process

Who owns the data and the licence?

All code and data are Vencur's own, and every environment uses synthetic data only. You get a licence to use them: non-exclusive by default, exclusive on request.

Can you build our workflow?

Yes, if the result can be checked exactly. A custom world starts with scoping by email, then a rule pack, task families and a verifier, calibrated on frontier models before delivery. Build a world for your workflow

How do we start?

Request a free sample task: it runs locally in minutes. Most teams then start with an evaluation pack.

Tell us what your agents or robots need.

Request a sample or describe a workflow. We'll reply by email.

Interest

We use your details only to answer your request. Details in our privacy policy.