Product boundary note · Mission Delivery Notes
Keep source data inside the boundary.
Decoy is being developed to create a realistic synthetic substitute inside the customer boundary. The customer reviews what may leave before development begins.
Build with a Realistic Substitute
Modernization teams need data with the right schemas, volumes, distributions, and cross-table relationships. Sensitive financial, healthcare, acquisition, and personnel records may have to remain inside the approved environment. Decoy is intended to create the substitute there, then let the customer approve what leaves.
That boundary matters more as AI expands implementation, testing, analysis, and documentation capacity. Every dataset introduced to the delivery workflow is one careless prompt away from a model provider, one test fixture away from a repository, or one log statement away from disk. Accountable people and approved controls still matter, but a technical boundary should not depend on perfect attention alone.
So we are building Decoy.
How It Works
Decoy is being built as a single compiled binary that runs inside the customer's approved environment, including an ATO boundary, VPC, or air-gapped network. It uses read-only connections to approved S3 buckets, databases, data lakes, Dataverse, and flat files, and the design does not require it to phone home. Its current generator library supports over 310 data types, including names, addresses, dates, currency amounts, UUIDs, IBAN numbers, latitude/longitude pairs, MAC addresses, and NAICS codes.
| Boundary | Contents |
|---|---|
| Proposed customer-side work | Approved data sources, Decoy binary, approved AI service, validation loop |
| Outputs reviewed before transfer | Synthetic dataset, documentation, provider code |
| Proposed development-side work | Application development, testing, and AI-assisted work using customer-approved synthetic outputs |
How generation worksFrom schema analysis to customer review
In the intended workflow, Decoy pairs with an AI agent authorized to operate inside the boundary, such as a model running in the customer's approved environment. The agent analyzes schemas, field semantics, explicit and implied relationships, embedded structures, and data quirks that developers would need to know. The actual provider, permissions, retention rules, and human review remain customer-approved constraints.
Then the AI drafts code for custom data providers. The generation rules cover every field, relationship, and approved distribution. If your vendor table has 200 entries and your invoice table references them with realistic frequency distributions, the synthetic data can preserve that behavior. If even the distribution shape is sensitive, it can be masked instead.
The generated provider code runs, produces synthetic data, and if validation finds wrong types, broken constraints, or missing relationships, the AI agent revises it. Named people remain responsible for reviewing the configuration and evidence; the customer decides what is authorized to leave the boundary.
What the Contract Can Require
Require read-only access, keep source data inside the approved environment, define the permitted outputs, retain validation evidence, and let a designated customer official decide what may cross the boundary.
What Crosses the Trust Boundary
In the intended operating model, only three reviewed artifacts leave your environment:
The intended output matches the approved schema, row volume, selected statistical properties, field lengths, nested structures, and referential relationships while excluding source values.
Schema maps, field profiles, relationship graphs, and data quirks give developers what they need to understand the data without seeing it. The customer keeps a copy for onboarding and transitions.
The generation rules contain no real data. They provide the logic for producing synthetic data that matches the original's shape and can be reused when schemas evolve.
Decoy writes its output to a customer-controlled location, such as an S3 bucket or shared drive. You review everything before we see it. By default, names, identifiers, text content, and specific source values are excluded; approved numeric distributions may remain unless the customer enables distribution masking.
Documentation packageWhat Decoy is designed to leave behind
Generating a useful substitute requires a documented model of the schema and relationships. Decoy is designed to leave that non-sensitive guide with the customer:
Schema Map
Every table, collection, and file, including fields, types, nullability, and constraints, generated from the current run.
Field Profiles
Per field: cardinality, min/max/average for numerics, max string lengths, enum values for categoricals, embedded structure notes (pipe-delimited, nested JSON, arrays).
Relationship Graph
Explicit foreign keys plus discovered soft-references, including matching GUIDs, identical long numbers across tables, and implied parent-child patterns the original developers never documented.
Data Quirks Log
Things a human analyst would discover after a week of exploration: "field labeled 'date' contains epoch milliseconds," "vendor_code uses 3 different formats across time periods," "free text field maxes at 4,200 characters but averages 80."
Provider Reference
What each custom generator does and why it exists, so developers on our side understand exactly what the synthetic data represents.
The agreement and applicable clauses determine rights in the documentation package. The package can support onboarding, transitions, and institutional knowledge.
Preserve the Relationships Applications Use
Random values can test field types. Real applications need the relationships between invoices, vendors, rates, categories, and time to remain coherent. That is the difference Decoy is designed to preserve.
- Random values by type
- No cross-table relationships
- Uniform distributions
- Fixed row counts
- Breaks application logic on first real query
- AI-analyzed field semantics
- Configured referential relationships preserved
- Approved distribution shapes
- Source-volume matching when required
- Application logic exercised against the substitute
What random fixtures missThe relationships application logic depends on
Real data has relationships. Invoices reference vendors. Vendors have rate histories. Rate histories cluster by job category. Job categories map to contract vehicles. None of these relationships exist in randomly generated data, and without them your application logic has nothing meaningful to exercise.
Decoy is designed to preserve the relationships the application needs to exercise: explicit foreign keys, discovered soft-references, approved frequency distributions, cardinality ratios, and temporal patterns. The AI agent can surface relationships the original developers may not have documented because it examines both the data and the schema inside the authorized environment.
Why This Can Unblock Delivery
Access to sensitive data often delays development. Decoy is designed to provide realistic synthetic records and a reviewable boundary artifact without sending the original financial, healthcare, or acquisition records to Intelligrit.
The benefit is operational as well as protective: when every local test record is synthetic, teams can use approved development and AI workflows without turning each iteration into a new data-handling exception.
Running DecoyConfiguration, permissions, and reruns
Decoy uses a configuration file for approved data-source connections and optional instructions for corner cases. The customer or Intelligrit can prepare it as the engagement permits. The intended deployment grants read-only source permissions because Decoy does not need to write to the original systems.
The AI agent running in your environment handles the rest: schema discovery, relationship mapping, provider generation, synthetic data output, and documentation. When your data structures change through new columns, tables, or relationships, you rerun Decoy. Most government data structures are fairly stable, so the intended use is occasional rather than a continuous pipeline.
A note on availability: Decoy is in development and currently offered only through an Intelligrit engagement. Scope and controls depend on the customer environment, data sensitivity, and approved AI operating boundary.