Expert workflows
Local accountants, tax and legal professionals solving real tasks step by step, turned into thousands of checkable cases.
LIVE · ENGINE AND EXPERT PORTALNew How our own checks caught a loophole in our test Teach AI how the rest
Teach AI how the rest
of the world works.
Consented expert workflows, computer use and local languages from the Global South, and the environments to train and test AI agents on them. Every contributor paid.
00 / START HEREThree ways to work with Proveron
Choose your path
01 / TRAIN AND EVALUATEAI labs · agent companies
Data and environments for the markets your models miss.
License consented expert workflows, computer-use recordings and local-language data, and run your agent through fresh business worlds with replayable, record-checked scores.
See the benchmark 02 / TRY BEFORE YOU DEPLOYOperations, finance and risk teamsKnow what an agent would do with your work.
Compare AI tools side by side on a simulated copy of your workflow. Shadow mode, in development, runs them on records you export and compares their proposals with what your staff actually did.
How it works 03 / CONTRIBUTEProfessionals · speakers · practitionersYour knowledge teaches AI. You get paid.
Write realistic tasks from your own work, record how you use the systems you know, or contribute in your language. Paid per approved task, with every decision explained.
Work with us00 / WHAT WE SUPPLYMost of the world is missing from AI's training data
Three kinds of consented data, and the environments to use them.
Computer use
People operating the software these economies run on: local accounting packages, bank portals, government sites.
IN DESIGNLanguages
Text, speech and translation in under-served languages, from Hausa, Yoruba and Swahili to Arabic dialects.
IN DESIGNEnvironments
Simulated companies built from that knowledge, to train and test AI agents with record-checked scores.
BUILT · FINANCE FIRSTTHE ENGINE AT WORKNo people in the loop: every step below is automated
One case, from rule to grade, in seconds.
An invoice lost its purchase order, and its note points to the wrong one. The engine planted that trap; the agent has to read the document, notice the note is wrong, and prove the right order from the records.
- 01PackExpert rules, written once
- rule invoice_po_exists
- problem invoice_missing_po_link
- clue PO reference · document only
- credit restore +1.0 · guess −1.0
- 02WorldA new company, generated
- seed 42 · 350 orders · 250 invoices
- INV-00029 Lopez Ltd $6,167.70
- → purchase order missing
- invoice PDF · note: PO-00049
- 03AgentOrdinary tools, every step logged
- read_artifact INV-00029.pdf
- PO-00049 → other supplier ✗
- match Lopez Ltd · $6,167.70 → PO-00316
- repair_invoice_po_link → PO-00316
- 04GradeChecked against the records
- outcome restore_original +1.0
- misleading note caught
- safety failures 0
- replay same final state ✓
Real values from development world seed 42 (engine 0.4.11, finance pack 0.2.13). The agent steps show one correct way to solve the case; it is not a published model score.
01 / WHY PROVERONWhat a score is worth depends on the test
A benchmark is only as good as its world.
Agents look capable on clean demos and fixed task lists. Proveron tests them where real work happens: linked records, conflicting documents, rules that differ by country, and changes that need approval.
Read the methodFresh worlds, not fixed tasks
Every test is a newly generated company. Fixed task sets can leak into training; ours cannot be memorised, and held-out worlds are retired after use.
Graded by the records, not an AI judge
The score comes from what the agent actually changed, checked against a proven answer key. Expert judgements are reported separately.
Validity you can inspect
Every release must defeat shortcut scripts, including one that never opens a document. When our own check finds a loophole, we fix the test and say so.
Built for each jurisdiction
Tax, invoicing, payments and approvals differ by country. Packs capture how the work is done in each one, from the practitioners who do it.
From test to deployment
The same tools, rules and audit trail that grade an agent can later govern it on real systems. Proof first, then production.
THE SYSTEM
From a rule on paper to a result you can replay.
Each stage has a job. The agent only sees the world and its allowed tools. The hidden answer stays with the evaluator.
PROVERON / PROCESS VIEW01 / 04
DOMAIN PACK
Describe the work.
A versioned pack defines the records, relationships, rules, actions, and outcomes of one workflow.
02 / ENVIRONMENTSOne engine. Jurisdiction by jurisdiction.
A world for the work you need to test.
Our first pack covers finance operations. Next come invoicing and tax workflows across several countries, each built and reviewed with local practitioners.
DEVELOPMENT PACK01 / FINANCE OPERATIONS
Finance, in a world of linked records.
Purchase orders, invoices, payments and bank entries, with the invoice PDFs, remittance emails and statements an agent must read to decide.
WHAT COMES NEXT
Each new pack needs its own rules, evidence and expert review.
E-invoicing and VAT, several jurisdictionsPLANNED
Payroll and withholdingPLANNED
Trade and customs documentsPLANNED
Healthcare administrationPLANNED
The finance environment is synthetic and under development. Professional review of its real-world fidelity is still pending.
03 / THE EVIDENCEInspectable by design
THE QUESTION ISN'T “DID IT SOUND RIGHT?”
Show the decision. Show the consequence.
A run records the tools an agent called, the documents it read, the records it changed and the final state. The grader compares that work with declared outcomes and reports where it succeeded or failed.
- 01 Repeat the world from its seed
- 02 Replay the agent's actions
- 03 Read an answer-safe failure report
ILLUSTRATIVE REPORT / DEVELOPMENT WORLDPROVERON / EVAL
RUN REVIEWWhat happened
VERIFIED TRACEWhat happened
when the agent acted?
One case, a clear trail.
View the evidence the agent saw, its action, the resulting state, and why the outcome was scored that way.
01 / OBSERVEPublic records and documentsvisible
02 / ACTPermitted tool calllogged
03 / VERIFYRule and outcome checksprotected
This is a product illustration, not a published model score.↗
Reproducible
Same pack and seed, same generated state. Versioned inputs make changes visible.
Answer-safe
Agents use public tools; hidden keys stay outside the agent's view.
Honest about limits
Report scope, missing reviews and evaluation status are stated beside results.
04 / PROVERONBENCHEmpirical Agent Evaluations
A MEASURE THAT CAN BE QUESTIONED
Leaderboard grounded in hard proof.
12 frontier and open-weight AI agents evaluated on reproducible synthetic business environments with complete trajectory audits, tool-use tracking, and safety policy enforcement.
PROVERON / BENCHMARK PROGRAMME 12 MODELS EVALUATED
12
Agents tested
DeepSeek, Qwen, Kimi, GLM, GPT & Gemini
04
Evaluation lenses
Execution · reliability · safety · validity
1.000
Top score
DeepSeek V4 Pro (document bypass noted)
Full benchmark report & trajectory audits available for public inspection.Explore live leaderboard ↗
05 / WHO IT'S FOROne foundation, different questions
Built for teams that need more than a demo.
Find the failures before production.
Test an agent against changing synthetic companies. Study individual tool calls, regressions and repeat attempts. Compare models side by side on the same worlds.
Discuss agent testingSee what an AI tool would do with your work.
Compare candidate tools on a declared copy of your workflow. In shadow mode (in development), they work records you export, propose changes without writing anything, and we compare their proposals with your staff's decisions.
Discuss an assessment06 / PROFESSIONAL KNOWLEDGE
The people who do the work should shape the test.
Practitioners write realistic tasks from their own work, solve other experts' tasks blind, and review how rules are translated. Every contribution is reviewed, explained and paid per approved task.
Work with usHOW EXPERTS CONTRIBUTEDESIGNED / PILOT PENDING
01
Author a taskA realistic assignment with evidence and criteria.
02
Solve blindWork another expert's task without the answer.
03
Review a ruleCheck the plain-language translation against practice.
04
Release with evidenceKeep scope, version and disagreements visible.
No expert-verified release is claimed for this first environment.
07 / WORKSPACESSeparate access for separate roles
FROM THE PUBLIC STORY TO THE WORK
Your workspace begins here.
The two portals serve different people. Neither exposes the operator Studio or a hidden answer key.
Help define what good work looks like.
Practitioners sign in to take paid tasks, author and solve scenarios and follow every review decision. The portal is in local testing; public onboarding awaits approved terms.
Public access is not configured.Read the evidence behind an agent run.
Customers sign in to inspect their own batches and answer-safe reports. The local portal uses issued credentials; public hosting is pending.
Public access is not configured.Operator Studio is private and intentionally has no link from this public site.
08 / START A CONVERSATION
Put your agent's decisions under a brighter light.
Building an agent? Deciding which AI tool to trust with your operations? Bringing professional knowledge into AI testing?
Preview only: a verified contact channel is added before public launch.