Source-linked AI summary
Generating a Consistent Enterprise: Synthesis and Reference-Free Evaluation of Multi-System Business Data
Benjamin Gruenbaum, Doron Porat, Assaf Natanzon, Roy Zavida, Chen Dinachi, Or Itzahary, Omer Niv
TL;DR
Synthetic enterprise data lacks both the confidential real dataset needed for training and the paired reference needed for conventional evaluation. The paper generates consistent fictional companies and question-conditioned databases from specifications, then evaluates them with reference targets and reference-free checks. Across 23 companies, mean realism rose from 60.3 to 99.1, while detector flags fell from 55.2% to zero.
Problem
Confidentiality prevents using production records, while conventional synthetic-data methods require real training data and reference-based evaluation.
Method
A deterministic entity graph projects one fictional enterprise across 66 product simulators, while reference targets and reference-free instruments guide and assess generation.
Results
60.3 to 99.1 mean realism and 55.2% to zero detector flags were measured across 23 generated companies under fixed evaluation instruments.
Takeaways & Limitations
The system provides reproducible multi-system enterprise data without customer records and makes its statistical assumptions explicit.
Takeaways & Limitations
The generator is only as accurate as its reference targets, and the evaluation does not compare Era by Eon with other specification-driven generators.
Abstract
from arXiv · showhide
Synthetic relational data is normally produced by a model trained on a real dataset, and its quality is measured as the distance to that dataset. This paper describes a generator that has no real dataset at either end. Given an industry, a company size, a business model, a set of business applications, and a random seed, it produces a complete fictional enterprise: a workforce, a customer base, sales deals, support tickets, recorded calls, chat messages, and documents, all consistent with one another. One entity graph is projected into the native formats of 66 business products, so the same customer appears in the CRM, the support desk, and the call system under one identity. Because no real counterpart exists, realism is built in from cited reference statistics and verified by reference-free measurement: a five-axis scorecard of 28 statistical checks, an adversarial detector that hunts for the marks of synthetic generation, and a set of soundness checks that include a classifier test against an independently shuffled copy of the data. Because these instruments existed before the generator was tuned, progress is measured under a fixed yardstick: over 23 generated companies, mean realism climbed from 60.3 to 99.1, the weakest company from 41.1 to 94.9, and the detector, which initially flagged 55.2% of all records, now flags none. The scores hold on a seed never used during development. A second generator builds relational databases from a list of business questions. It forces qualifying rows for each answerable question, adds controlled near misses, and computes exact labels from the finished tables. The generator runs as a hosted service at https://console.era.eon.io. A company built there to a specification is served through its simulators over MCP and REST, and the simulators are also published as container images for offline use
1 Introduction
The paper presents a reference-free system for generating consistent fictional enterprises across many business applications and evaluating their realism without confidential real data. Fixed pre-existing instruments measured improvement from 60.3 to 99.1 mean realism while detector flags fell from 55.2% to zero.
- Motivation: The system addresses confidential enterprise data needs because production records cannot be used and existing public substitutes are small, narrow, or stale.This need is especially acute for agents that read multiple business systems.
- Enterprise generation: A deterministic entity graph is projected into 66 product-native formats, making cross-system consistency a consequence of rebuilding the same graph from one seed.Each simulator stores only the portion its product would keep and uses vendor-style identifiers.
- Reference-free realism: Reference targets from operational experience and published sources shape volumes, values, timestamps, text, and deliberate real-world correlations.Each target records provenance, while unsourced targets are marked as estimates.
- Evaluation: A five-axis scorecard with 28 checks, an adversarial detector, and soundness tests evaluate generated data without paired real data.The soundness checks include a classifier test against an independently shuffled copy to detect missing dependencies.
- Evaluation: 60.3 to 99.1 mean realism, 41.1 to 94.9 for the weakest company, and 55.2% to zero detector flags across 23 generated companies.The instruments predated tuning, and an unseen development seed scored 91 with no flagged records.
- Additional generator: The paper contributes a question-conditioned generator that forces qualifying rows for business questions and recomputes exact labels from finished tables.The broader system is available through a hosted console and published simulator images.
2 Related work
The paper situates its enterprise generator between learned synthesis and specification-driven systems. Its novelty is extending cited-statistic, repeatable generation across many applications with correlated enterprise data, documents, and free text.
- Learned tabular synthesis: Learned tabular generators train on real data and evaluate similarity to that training data, whereas this paper targets a setting without a real reference dataset.The related work includes Synthetic Data Vault, CTGAN, and SDMetrics.
- Specification-driven generators: The generator extends specification-driven synthesis from one schema to a whole enterprise with many applications sharing one identity.It adds correlated volumes, a business calendar, real documents, and free text.
- Benchmark environments: Enterprise agent benchmarks are hand-built, while this system generates controlled dirty values and labeled unanswerable questions during database generation.The approach connects synthetic enterprise environments with text-to-SQL benchmark requirements.
3 One specification, one graph
The generator accepts a five-value company specification and builds one deterministic entity graph whose projections remain consistent across systems. Separate random streams and dependency-aware generation contain changes while preserving shared identities, relationships, and reproducibility.
- A five-value specification—industry, company size, business model, application portfolio, and seed—produces one entity graph containing the fictional company’s people, customers, and business activity.
- Dependency-ordered generation builds later records from earlier ones, while the graph retains operational texture such as duplicates, prospects, departed owners, and service accounts.
- 66 simulators can serve one consistent company because the same specification and seed reproduce the company byte for byte on any machine.
- Independent subsystem streams prevent unrelated random draws from shifting when code changes, so generated diffs remain localized rather than changing the entire corpus.
- Dependency-aware streams preserve one-way consistency: downstream calls read existing customer records, so customer changes propagate into calls but call-generation changes cannot alter customers.
- The generator uses reference targets for both synthesis and evaluation, and a 20,000-employee company with about 250,000 records takes about nine minutes and occupies about 25 MB.
4 Projection into 66 product formats
The system projects one generated enterprise into 66 product-shaped interfaces while preserving cross-product identity and shared facts. It also renders entity-linked business documents as deterministic native files.
- 4.1 Simulators as projections: Each of the 66 simulators independently rebuilds or loads the entity graph and exposes the subset of records its business product would store through REST and MCP interfaces.
- 4.1 Simulators as projections: Product-specific projections can represent the same call differently while preserving participants, time, and customer across systems such as Gong and Salesforce.
- 4.1 Simulators as projections: The application portfolio controls which simulators expose records, without changing the underlying graph or the people, customers, deals, calls, and tickets generated from the same seed.
- 4.2 Shared and vendor-shaped identifiers: Shared graph identifiers and stable opaque external identity keys support cross-product reconciliation without relying on names or email addresses.
- 4.2 Shared and vendor-shaped identifiers: Optional vendor-shaped identifiers satisfy product-specific syntax through deterministic mapping, collision checks, re-probing, and reverse links to graph identifiers.
- 4.3 Documents as real files: Documents use purpose-specific templates, such as security reviews and pricing worksheets, whose structures encode expected sections and whose values can satisfy associated deal totals.
- 4.3 Documents as real files: Document facts come from associated graph records, while file identifiers generate local variation so proposals, worksheets, and records remain mutually consistent but nonidentical.
- 4.3 Documents as real files: The renderer produces deterministic DOCX, XLSX, and PDF files with preserved structure and fixed metadata, allowing identical files across storage products.
5 Question-conditioned database generation
A second generator designs relational databases from business questions, validates the resulting schema, and populates tables so answerable questions have controlled, exactly labeled data. It also supports realistic imperfections and deterministic successive-day snapshots.
- The question-conditioned generator creates company-specific application and analytics tables separately from the 66 product projections.
- Users provide questions covering filters, aggregates, temporal comparisons, and missing relationships, from which the system infers required entities, relationships, schemas, and exact answers.
- The language model proposes designs during a bounded design phase, while deterministic validators check and repair keys, references, domains, dependencies, and question instructions.
- Once accepted, the stored schema remains fixed across regenerations: the same stored model and seed reproduce rows and labels, whereas rerunning design can produce a different schema.
- Seeded population fills tables in dependency order and plants qualifying rows so each answerable question has a nonempty result rather than relying on chance.
- Near-miss rows make individual question conditions meaningful, while controlled capitalization, whitespace, and placeholder flaws add imperfections without breaking relationships or labels.
- Labels are computed by executing every question on completed tables, with comparison rules for sets, keyed mappings, and numeric tolerances.
- Successive database days apply deterministic inserts, updates, and deletes while preserving keys and references where possible.
6 Evaluation without a reference
The paper evaluates generated enterprises without paired real data by combining plan-based checks, a five-axis scorecard, adversarial pattern detection, and an unseen-seed regression gate. These instruments measure distributions, relationships, time, customer concentration, text, and structural plausibility while reporting coverage and limitations.
- Reference-free checks use the generation plan, transformations of generated rows, and reference targets; question-conditioned databases and entity graphs receive distinct evaluation instruments.
- Plan-adherence, question-specificity, and label-integrity checks test declared domains and distributions, whether every question condition matters, and whether stored answers are uniquely and fairly comparable.
- Dependency structure is tested by independently shuffling columns, exposing rows whose individual distributions look plausible but whose within-row relationships are unrealistic.
- Cardinality-estimation error compares exact query counts with estimates based on uniform values and independent conditions using q-error.
- The five-axis scorecard contains 28 checks covering individual fields, field relationships, time, customer-level distributions, and text.
- Checks with too little data are marked unmeasured and excluded, so reports state how many checks actually ran.
- The adversarial detector searches for synthetic signatures including numbered duplicates, cyclic category proportions, regular timestamps, repeated text, constant fields, and narrow value ranges.
- A zero detector flag means none of its covered patterns was found, not that the data is proven indistinguishable from real enterprise data.
7 Results
A paired evaluation of 23 specifications found that the realism profile substantially improved statistical scores and removed detected synthetic artifacts under a fixed evaluation yardstick.
- 7.1 Measured improvement under fixed instruments: The paired design held company specification and seed constant, isolating the realism profile from industry, size, and random-input differences.The legacy profile used uniform counts, independent fields, simple timestamps, and repeated templates; the revised profile added reference distributions, relationships, varied text, and causal sequences.
- 7.1 Measured improvement under fixed instruments: 99.1 versus 60.3 mean realism, and 94.9 versus 41.1 for the lowest-scoring company, after the generator revisions.Table 2 reports scores for 23 companies under one fixed 0–100 scorecard.
- 7.1 Measured improvement under fixed instruments: 55.2% of legacy records contained at least one detected artifact, while revised companies contained none.On a separate continuous-integration seed, the lowest company score was 91 and no records were flagged.
- 7.2 What the instruments caught: The revised generator increased transcript diversity from 42 fixed sentences to 19,794 distinct utterances in one hyperscale company.The most common sentence fell from 2,936 occurrences to 631, and calls were assigned one subject rather than combining unrelated subjects.
- 7.2 What the instruments caught: The scorecard and detector are complementary: one measures statistical prevalence, while the other identifies records bearing known generation artifacts.Strong statistical scores can coexist with obvious artifacts, and an empty detector report can coexist with unrealistic statistical structure.
8 Availability
Era by Eon provides hosted, read-only synthetic companies through MCP and REST, with per-system container images enabling consistent offline use.
- 8 Availability: The hosted service generates a specified company and starts one simulator per selected business system, exposing vendor REST APIs and equivalent MCP operations.A company receives one token, or a token scoped to a single simulator.
- 8 Availability: Hosted companies expose inception, current, and one-quarter-later day states, while remaining read-only and synthetic.An agent can read the company but cannot write into it or change it.
- 8 Availability: Each simulator is also published as a Docker container that runs offline and rebuilds the same company from its seed and anchor date.Composing several images serves one consistent company without exchanging data; pinning the seed, anchor date, and image tag preserves the served company.
- 8 Availability: The service is intended for building, testing, and demonstrating software before it encounters customer records, not for replacing real-data staging.Every environment is synthetic.
9 Limitations
The evaluation is bounded by the generator’s reference targets, seeded templates, and comparison design, leaving important gaps relative to real enterprise data.
- 9 Limitations: Incorrect reference targets can produce incorrect distributions that the scorecard will not reveal when it evaluates against those same targets.Ten of thirteen targets have recorded provenance, while three are marked as estimates.
- 9 Limitations: Seeded templates provide varied, internally consistent text but cannot reproduce the full range of human writing, so stronger detectors may still distinguish it.The missing range includes style, ambiguity, error, and context.
- 9 Limitations: The reported evaluation compares two versions of the same generator rather than Era by Eon with other specification-driven generators.The detector has not been calibrated against a large corpus of real enterprise records because those records are confidential.
10 Conclusion
The paper presents a deterministic, reference-free approach to generating consistent synthetic enterprises and question-conditioned databases, with reproducible evaluation and public deployment.
- 10 Conclusion: A shared entity graph keeps people, customers, and business events consistent across 66 product simulators while reference targets shape data without confidential customer records.Era by Eon generates a complete fictional company from a compact specification and seed.
- 10 Conclusion: The question-conditioned generator creates validated schemas, deterministic plans, populated tables, and labels computed from final rows.Successive day states add controlled change while preserving keys and relationships.
- 10 Conclusion: Reference-free checks using generation plans, transformed rows, and reference targets exposed structural, timing, and text problems that simple field-level statistics missed.The resulting companies are available through the hosted service and per-system offline container images.