The Reference

Ground truth, by construction.

Audit analytics assembled from the data you’re auditing is a jigsaw puzzle solved without the picture. Recovering ground truth from observed enterprise data is combinatorially infeasible.

VynFi generates the reference forward— from a fully specified model where every node’s provenance is known.

Generate a reference View the paper

I · Structural

P2P · 6 STAGES

II · Statistical

MAD · 0.006 · EXCELLENT

III · Normative

5 / 5 CERTIFIED

GAAP

IFRS

ISA

PCAOB

COSO

155 datasets · 364M entries · 2.4B line items · calibrated against ISO 21378:2019 — arXiv ↗

REFERENCES GENERATED TODAY · 12,481BENFORD MAD (rolling) · 0.0058155 DATASETS CALIBRATION364M JOURNAL ENTRIES · 2.4B LINE ITEMSP99 LATENCY · 82 msUPTIME · 99.98%

Try it

Watch it generate.

One curl command. Reference data in your terminal before you finish reading this sentence.

Specimen · POST /v1/generate/quickLive

terminal · bash

Generates sample data instantly

Response · application/json1,000 rows · ~840 ms

How it works

From signup to reference in three moves.

No credit card. No sales call. Your first reference knowledge graph generates in under three minutes.

Step 01

Sign up, collect key

Create a free account. Your API key is generated instantly — no credit card, no sales call.

Step 02

Generate a reference

Call the API with your sector, tables, and row count. Receive a fully provenanced reference knowledge graph.

Step 03

Build against ground truth

Use reference data for testing, ML training, and compliance workflows — with known audit trail for every node.

SDK · five languages

Integrate in minutes.

First-class SDKs for Python, TypeScript, Rust, and .NET. Or just use curl.

Bash

curl https://api.vynfi.com/v1/generate/quick \
  -H "Authorization: Bearer vf_live_7mN4kP2x..." \
  -H "Content-Type: application/json" \
  -d '{
    "preset": "retail_small",
    "tables": ["journal_entries"],
    "rows": { "journal_entries": 1000 },
    "format": "json"
  }'

Use cases

Built for every discipline.

Audit firms, fintech engineers, academic researchers, compliance teams, and ESG reporters — all running reference data against the same model.

Audit testing

Journal entries with known anomalies for audit analytics testing. Calibrated to real-world distributions.

Big 4 · audit firms

Fintech development

Build and test financial applications with production-quality synthetic data. Zero real customer exposure.

Engineering · QA

Academic research

Large-scale labelled datasets for fraud detection, process mining, and financial ML research.

Research · ML

Compliance validation

Test SOX, Basel III, and IFRS workflows with COSO control mappings and evaluation reports.

GRC · regulators

ESG & sustainability

CSRD/TCFD reporting pipelines with Scope 1/2 emissions, workforce diversity, and pay equity analysis.

Sustainability

Enterprise audit

Big-4 methodologies, codified.

Four integrated audit methodologies, group-audit simulation per ISA 600, and complete end-to-end audit data generation.

Big 4 methodologies

4 integrated blueprints with 728–757 steps each — KPMG Clara, PwC Aura, Deloitte Omnia, EY GAM.

4 · blueprints

Group audit · ISA 600

Component-auditor simulation with Significant, Non-Significant, Not-in-Scope classification and consolidated reporting.

ISA 600

14 audit data types

From journal entries to board minutes — IT reports, management packs, regulatory filings, and more.

14 · artefacts

Process mining exports

Disco, Celonis IBC, XES 2.0, and OCEL 2.0 format support for process-mining research and tooling.

OCEL 2.0 · XES 2.0

Explore audit blueprints →Datasets on Hugging Face →

New in v2.3 · Banking & AML

Money-laundering data, done right.

14 fully-implemented typologies, multi-party criminal networks, cross-layer fraud propagation from payments to bank transactions, and 10 evaluators to prove it.

14 AML typologies

Structuring, smurfing, mule chains, synthetic identity, trade-based ML, crypto integration, sanctions evasion, romance scam, casino & real-estate integration — with ground-truth labels.

Ground-truth labels

Multi-party networks

Barabási-Albert preferential-attachment topology — one coordinator + 5–25 smurfs, mule chains with recruiter / middleman / cash-out roles, shell-company pyramids.

Power-law topology

Velocity & device features

Rolling-window counts (1h/24h/7d/30d), unique counterparties, amount z-scores, and realistic power-law device fingerprint distributions — pre-computed.

Feature-ready

Cross-layer coherence

A fraudulent vendor payment surfaces in document flow, journal entries, AND on both sides of a mirrored bank-transaction pair — ≥95% fraud-label propagation.

≥95% propagation

Scale tier · TB-scale

TB-scale without the disk hell.

Data streams direct to Azure Blob with short-lived SAS downloads — or bring your own storage and keep zero bytes on VynFi. For live pipelines, NDJSON streaming at up to 10,000/sec.

Managed Azure Blob

Lifecycle retention (7d Free → 365d Scale). Per-file SAS URLs — direct blob access, no API proxy, no 2 GB cap, no OOM kills.

All tiers

BYO storage

Supply a container SAS URL and the worker uploads directly to your data lake. Zero bytes transit our storage. Pair with Private Link for airgapped flows.

Team+ · Enterprise

NDJSON live streaming

GET /v1/jobs/{id}/stream/ndjson emits self-describing envelopes with token-bucket rate-limiting. Point Kafka, Spark, ClickHouse at it and ingest live.

Scale+

Enterprise integrations · new in DS 4.4.3

SAP and SAF-T, out of the box.

Two new enterprise-grade output formats. Reference data that drops directly into S/4HANA IMPORT or OECD tax-software validators — no manual CSV pre-processing, no hand-written XML.

SAP · 27 tables

BKPF / BSEG / ACDOCA plus five master-data tables (LFA1 / KNA1 / MARA / CSKS / CEPC). HANA dialect for S/4HANA IMPORT, classic for legacy ECC. Configurable client / ledger / source-system tags. FK integrity across the full P2P cycle.

Scale+ · 1.25× credit multiplier

SAF-T · 5 OECD jurisdictions

Structurally-valid SAF-T XML for Portugal (1.04_01), Poland (JPK_KR 1.0), Romania (D406 3.0), Norway (Fin 1.10), Luxembourg (FAIA 2.01). Audit-PoC ready for tax-software validation and compliance simulation.

Scale+ · 1.25× credit multiplier

SAP integration →SAF-T integration →

Financial coherence engine

Every number connects.

From raw journal entries to audited financial statements, VynFi generates data that passes your reconciliation, audit, and regulatory tests — all derived from a single declarative model.

Full financial statements

Complete balance sheet, income statement, cash flow, and equity rollforward — generated from actual journal entry data, not templates.

BS · P&L · CF · Equity

Manufacturing cost flow

Multi-stage WIP → Finished Goods → COGS pipeline with standard cost variance accounting and IAS 37 warranty provisions.

IAS 37 · Std cost

Treasury & hedge accounting

Debt interest accrual, cash-flow and fair-value hedge mark-to-market, cash-pool sweeps, and covenant compliance evaluation.

ASC 815 · IFRS 9

Tax from real GL

Tax provision computed from actual pre-tax income. VAT posting from source documents. Deferred tax with temporary-difference tracking.

Deferred · VAT · DTA

XBRL 2.1 export

Instance documents mapped to US GAAP and IFRS taxonomies. Test your regulatory filing pipeline with reference data.

US GAAP · IFRS

32+ coherence validators

FG rollforward, WIP rollforward, trial-balance proof, cash-flow reconciliation, equity rollforward, segment-to-consolidated, IC elimination.

32 · validators

09½ · Compliance

Audit-grade by default.

GDPR-ready, EU AI Act aligned, AES-256 at rest and in transit, zero real-client-data ingestion, all 4 Big 4 firm methodologies covered.

GDPR-Ready

EU AI Act

AES-256

Zero Real Data

Big 4 Coverage

View the full security posture →

Quality by design

Statistical rigor, measurable.

Validation is not a feature — it is the output. Every reference carries provable bounds on its own distributional fidelity.

Benford MAD

< 0

1st digit

Mean absolute deviation for first-digit compliance. Rated 'excellent conformity' by Nigrini's criteria.

F1 Delta

~0%

vs real

GNN fraud detectors trained on reference data within 3% F1 of real-data baselines.

Copula families

implemented

Gaussian, Clayton, Gumbel, Frank, and Student-t copulas model complex inter-variable dependencies.

Anomaly types

subtypes

Across 5 categories — timing, amount, relationship, pattern, and structural — with ground-truth labels.

The paper

Built on real-world research.

The DataSynth engine was calibrated against 155 ISO 21378:2019–compliant general-ledger datasets, encompassing 364 million journal entries and 2.4 billion line items across industries and geographies.

Real-world datasets

GL corpora

Analyzed for distribution calibration and statistical benchmarking across 10 industry sectors.

Journal entries

posted

In the calibration corpus used to derive realistic financial patterns and temporal dynamics.

Line items

observed

Processed to build inter-table correlation models and cross-entity relationship graphs.

Read the full whitepaper →

Localization

41 country packs, already wired.

Localized tax, banking, naming, holidays, and accounting standards — so regional realism is a config flag, not a backlog item.

Americas

7packs

EMEA

15packs

APAC

14packs

Middle East & Africa

5packs

Explore how it works →

Pricing

Simple, transparent.

Start free. Scale when you need it. Credits reset every billing cycle — no rollover games.

Free

$0/ mo

10K credits / month

10,000 credits/month
1 concurrent job
JSON & CSV output
Community support

Get Started Free

Developer

$49/ mo

500K credits / month

500,000 credits/month
5 concurrent jobs
All output formats
Email support

Start Free Trial

Team

Scale

$499/ mo

25M credits / month

25,000,000 credits/month
Unlimited concurrent jobs
All output formats
Dedicated support

Start Free Trial

Compare features + calculator →Enterprise →

From the paper · §1

“Recovering ground truth from observed enterprise data is combinatorially infeasible. DataSynth circumvents this by generating data forward — producing reference datasets where the complete audit trail is known by construction.”

Ivertowski · 2026 · arXiv:cs.CE

Anomaly types

32+

Coherence validators

Country packs

100K+

Rows / second

155

Datasets analyzed

Begin

Generate your first reference.

10,000 credits, every month, free. No credit card required. Your first reference knowledge graph generates in under three minutes.

Generate a reference Enterprise →

You scrolled all the way down. We respect that.