Synthetic Business Worlds
Relational ERP ledgers, messy scanned invoice PDFs, and bank feeds with mathematically proven ground truth.
BUILT · FINANCE FIRSTNew 12-Model Empirical Agent Benchmark Live The Flight Simulator
The Flight Simulator
for Enterprise AI Agents.
PwC found that 89% of digital operations investments fail to deliver ROI because off-the-shelf AI breaks on messy back-office records. We stress-test agents across synthetic business worlds, verify their safety, and help companies safely deploy the winners.
00 / START HEREThree ways to work with Proveron
Choose your path
01 / FOR ENTERPRISES & AGENCIESOperations, finance and agency leaders
Deploy AI agents with proof, not promises.
PwC found that 89% of digital operations investments fail to deliver ROI. We benchmark candidate agents on your workflow, run a read-only shadow trial on your past records, and oversee a safe, governed implementation.
Book deployment consultation 02 / FOR AGENT BUILDERSAI agent startups · applied AI teamsThe flight simulator before you touch real money.
Stress-test your agent against 50 adversarial business worlds for free. Catch silent failures, unapproved balance write-offs, and document bypasses before your enterprise clients do.
See 12-model leaderboard 03 / FOR GLOBAL CONTRIBUTORSAccountants · auditors · practitioners worldwideYour knowledge keeps AI honest. You get paid.
Earn money planting real operational traps or auditing records in our Adversarial Bounty Game. Powered by MIT's Bayesian Truth Serum, rewarding genuine local expertise.
Join the bounty network00 / WHAT WE SUPPLYClosing the 89% enterprise AI failure gap
The simulation engine, the truth protocol, and the deployment bridge.
Shadow Mode & Implementation
Non-destructive testing on historical records, followed by governed deployment with dual-approval limits.
READY · ENTERPRISE PILOTSGlobal Truth & Bounty Protocol
Real people in emerging markets earning income authoring and auditing operational traps under Bayesian Truth Serum.
LIVE · BOUNTY NETWORKVerifiable RLVR Environments
High-volume, seed-isolated reasoning gyms for post-training models via Prime Intellect Hub and Gymnasium.
READY · LABS & HUBSTHE 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 prompts
Every test is a newly synthesized company. Fixed task sets can leak into training; ours cannot be memorised, and held-out worlds are retired after evaluation.
Graded by the records, not an AI judge
The score comes from immutable state changes left on the relational ledger, checked against a proven answer key. Zero flaky LLM judges.
Anti-shortcut document ground truth
Every release must defeat shortcut scripts that bypass documents. Agents are strictly forced to inspect OCR artifacts to solve cases.
Money-movement safety guards
Actions exceeding statutory or organizational spending thresholds are penalized if executed autonomously without managerial escalation.
From shadow test to governed rollout
The same tools, rules, and audit trails that grade an agent in simulation govern its execution in live production. Proof first, then deployment.
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 worldwide sign in to plant traps, audit records blind under Bayesian Truth Serum, and earn cash per verified scenario. Active onboarding across emerging markets.
Join the Bounty Network Open contributor network.Read the evidence behind an agent run.
Enterprises and AI labs sign in to inspect their private evaluation batches, verify trajectory hashes, and download answer-safe diagnostic reports.
Request Batch Access Design partner access.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.
Deploying an AI agent? Deciding which AI tool to trust with your operations? Submitting your agent for a free 50-seed stress test?
Start a conversation Direct partner channel · [email protected]